# Crowdee > Crowdee is a hybrid content-verification platform that combines AI-powered analysis with human-in-the-loop crowd review to produce auditable, defensible authenticity verdicts for newsrooms, compliance teams, and AI researchers. ## Company Information - Company Name: Crowdee GmbH - Company Seat: Berlin, Germany - Headquarters Address: Zehdenicker Str. 5, 10119 Berlin, GERMANY - Managing Directors: Dr.-Ing. Tim Polzehl and André Beyer - VAT ID: DE304624889 - Registration No.: HRB 173232 B - Register Court: Amtsgericht Berlin-Charlottenburg - Foundation Date: 06.01.2016 - Timezone: Europe/Berlin - Opening Hours: Monday to Friday from 09:00 to 18:00 - Main Currency: Euro (€) - Spoken Languages: English, German - Industries: Artificial Intelligence (AI), Information Technology (IT), Media Technology, Content Verification, Fact-Checking, Crowdsourcing, Research and Development (R&D) - Company Size: SME with <10 Employees - VAT Rate: 19% on all offered products and services - Operating Status: active - Company Type: German GmbH (limited liability company, for profit) - Business Model: B2B SaaS - Domain: crowdee.ai - [Homepage](https://crowdee.ai) - Phone Number: +49 30 330 210 64 - Fax Number: +49 30 330 210 65 - Mail Address: Crowdee GmbH, Zehdenicker Str. 5, 10119 Berlin, GERMANY - [Contact Form](https://crowdee.ai/contact) - E-Mail Address: hello@crowdee.ai - Press E-Mail Address: press@crowdee.ai - Security E-Mail Address: security@crowdee.ai - Sales E-Mail Address: sales@crowdee.ai - Support E-Mail Address: support@crowdee.ai - Careers E-Mail Address: careers@crowdee.ai ## Brand Information - Mission: To verify content at scale by combining AI-powered triage with human-in-the-loop expertise — producing auditable, defensible verdicts for media agencies, newsrooms, and compliance teams. - Vision: To become the intelligence layer for media agencies, enabling faster and more accurate content verification workflows that scale with content volume without compromising quality. - Tagline: Verify content at scale. With humans in the loop. - [Registered Trademark](https://euipo.europa.eu/eSearch/#details/trademarks/017883439): Crowdee - Key Focus: Hybrid AI + human content verification, auditable verdicts, DSA and EU AI Act compliance, research-backed media intelligence. - Origin: Born from the BMBF-funded news-polygraph research project with Deutsche Welle, rbb, and Fraunhofer IDMT. ## Platform Overview Crowdee is a hybrid content verification platform for newsrooms, compliance teams, and AI researchers. It combines AI-powered triage with over 1 million screened crowd workers to produce auditable, defensible verdicts — not just confidence scores. The platform supports 14 built-in content verification pipelines and 7 Language Technology (LT) pipelines across text, image, audio, video, and multimodal content, orchestrated through a single unified workflow. Verification pipelines produce auditable authenticity verdicts backed by AI analysis and optional human crowd review. LT pipelines extract structured linguistic data — transcriptions, named entities, translations, detected languages, and OCR text — from any file, independently of verification. Every verification result includes inter-rater agreement statistics, timestamp trails, and exportable compliance documentation — ready for DSA Article 34/35 and EU AI Act reporting. **MCP Server:** Crowdee exposes a Model Context Protocol (MCP) server at `POST https://api.crowdee.ai/v2/mcp` (Streamable HTTP transport). Authenticate with X-API-Key: crw_... — same as the REST API. The server registers 46 tools covering projects, datasets, the unified Data Platform pipeline catalog (running and monitoring cleaning/dedup/LT pipeline runs), AI Platform finetuning (base model catalog, training-dataset assembly, finetune runs, model registry), verification pipeline runs, Language Technology pipeline runs, crowdsourcing jobs and answers, content gathering (jobs, gathered items, dataset export), AI output evaluation, and continuous monitoring schedules, plus 4 resources for the verification, Language Technology, unified Data Platform, and AI Platform base-model catalogs. Compatible with any MCP client including AI coding assistants and agent frameworks. **Key Platform Screens:** - Dashboard: Command center showing all content verification and crowd-sourcing projects at a glance - Chat Agent: Conversational AI assistant inside the platform for guided verification workflows and pipeline orchestration - Content Verification: Upload content, choose from AI verification and Language Technology pipelines, review results with confidence scores and visualizations; pipeline runs update in real time via WebSocket notifications - Crowd Sourcing: Design survey templates, publish crowdsourcing jobs to the worker pool, collect structured human judgments; review and accept/reject answers, export results as JSON/CSV or convert into datasets for further pipeline runs - Datasets: Manage named multi-modal file collections (image, audio, video, text, document) with full version history; trigger cleaning, deduplication, format conversion, PII redaction, train/val/test splitting, crowd-assisted labeling, per-file enrichment, and Language Technology pipeline runs — all from the unified Data Platform pipeline catalog — to produce processed derived versions ready for verification or crowdsourcing - AI Monitoring: Create recurring schedules that sample verification/LT runs and dispatch crowd evaluations, tracking drift against each schedule's own trailing baseline - Reports: Review verification analytics, drill into pipeline run history, and export dossiers for compliance or editorial review ## Products Crowdee ships five crowd-powered products, all built on the same underlying job engine, input-data model, and API: - [Content Gathering](https://crowdee.ai/content-gathering): a trained crowd searches social platforms, messengers, and the open web against your criteria, returning matching posts, articles, and multimedia with engagement metrics and deduplication. - [Multimedia Verification](https://crowdee.ai/content-verification): the core product — 14 AI + crowd pipelines detecting manipulation, deepfakes, and synthetic content across text, image, audio, and video. - [Source Research](https://crowdee.ai/source-research): live web research plus crowd and AI analysis trace a piece of content back to its source, the accounts behind it, and a credibility assessment. - [AI Result Evaluation](https://crowdee.ai/ai-output-evaluation): a crowd panel rates the clarity, evidence, and actionability of any AI-generated verification result, feeding a durable feedback log. - [Continuous Monitoring](https://crowdee.ai/continuous-monitoring): a scheduled, always-on version of AI Result Evaluation that flags quality drift automatically against each schedule's own trailing baseline. See the [Products hub](https://crowdee.ai/products) for an overview of all five. ## Crowd Platform Alongside the five products, Crowdee offers the [Crowd Platform](https://crowdee.ai/crowd-platform): a self-serve tenant on Crowdee's own platform stack for organizations that want to run their own crowd program rather than buy a delivered result. It wraps the same client app, worker app, and backend API Crowdee's own agency team uses, scoped to an isolated organization with its own panel, projects, and audit trail. Five modules cover the full lifecycle: - Panels & Workers: organization-scoped worker panels, worker administration and screening, and a dedicated worker app for task intake. - Tasks & Templates: crowdsourcing task templates, input data upload and management, and survey-based task authoring. - Pipelines & Datasets: pipeline catalog and queue orchestration, per-run datasets you can export, and feedback reports on completed work. - Payouts & Compliance: payout request approval and payment issuance, credit ledger and transaction history, and organization-level compliance records. - Monitoring & Audit: scheduled quality monitors, run-level audit trail, and reports across projects. The Crowd Platform is infrastructure you operate, distinct from the five products above (which are delivered results) and from the [Verification Platform](https://crowdee.ai/platform) (a managed verification workflow). No separate pricing tier exists yet; access is arranged via [a demo call](https://crowdee.ai/demo). ## Data Platform Alongside the Verification Platform and Crowd Platform, Crowdee offers the [Data Platform](https://crowdee.ai/data-platform): the layer where the files behind those other offerings actually live. It centres on two primitives — versioned **datasets** and a unified **pipeline catalog** — and is the third of Crowdee's four solutions (Verification Platform, Crowd Platform, Data Platform, AI Platform). - Datasets: named, versioned collections of files (image, audio, video, text, document, or multimodal). Every cleaning, conversion, redaction, or split creates a new derived version; the raw original is never overwritten, giving a full audit trail back to the source files. - Unified pipeline catalog (13 pipelines): merges dataset-cleaning and Language Technology pipelines into one catalog, callable via `GET /v2/data-platform/pipelines/catalog` and `POST /v2/data-platform/pipelines/:slug/run`: - `clean-audio-silence` — trims leading/trailing silence from audio files. - `convert-format` — converts audio files between mp3, wav, ogg, flac, and m4a via ffmpeg (image/video/document conversion planned). - `pii-redact-text` — detects and redacts PII (names, emails, phone numbers, addresses, ID numbers) from text files using an LLM, producing a new redacted file. - `dedup-content-hash` — computes a SHA-256 content hash per file and flags exact-duplicate groups for manual review; non-destructive. - `split-dataset-version` — partitions a dataset version into train/validation/test versions by random assignment, reusing files by reference. - `crowd-label-classify` — has a crowd assign a label from a taxonomy to every file, with independent per-file majority-vote consensus once enough workers have responded to that file. - Plus all 7 Language Technology pipelines (see below). - The Data Platform is the seam between the other two offerings: a dataset version's files can seed a Crowd Platform job's input data, and a Verification Platform pipeline run can target a dataset version directly. Content-gathering results (a Crowd Platform feature) can flow into a new dataset without a manual re-export step. The Data Platform has no separate pricing tier — most pipelines run on the same credit system as the rest of the platform (see Pricing below); access is arranged via [a demo call](https://crowdee.ai/demo). ## AI Platform Alongside the Verification Platform, Crowd Platform, and Data Platform, Crowdee offers the [AI Platform](https://crowdee.ai/ai-platform): a self-serve environment for finetuning small, open-source language models (SLMs) on your own data. Training sets are assembled directly from Data Platform dataset versions (specifically ones already run through the Crowd-Assisted Labeling pipeline) or from completed Crowd Platform crowdsourcing job answers, so there is no separate data-labeling step required. - Base model catalog: a curated set of small, LoRA-friendly open-source models — Qwen3.5 (0.8B/2B/4B), Ministral 3 3B, Granite 4.1 3B, and Gemma 4 E2B — chosen to finetune quickly and affordably without GPU infrastructure. `GET /v2/ai-platform/base-models` lists the catalog. - Training data: `POST /v2/ai-platform/finetune-datasets` builds a training manifest from either a `dataset_version` (crowd-labeled files) or a `crowd_job` (accepted answers), stored as JSONL in S3. - Finetuning: `POST /v2/ai-platform/finetune-runs` starts a LoRA finetuning run on a chosen base model at a flat, upfront credit cost. Runs execute on a separate Python worker that polls the job table directly (no BullMQ) — the only job type in Crowdee not dispatched through the usual queue. - Model registry: every finetuned model is versioned, converted to GGUF, and registered under an organization-scoped tag in Crowdee's shared Ollama runtime (`GET /v2/ai-platform/models`) — never shared across tenants. - Benchmarking: reuses the existing AI Result Evaluation flow (`POST /v2/projects/:id/ai-output-evaluations/external`) to get an independent, crowd-rated transparency score on a finetuned model's outputs before relying on it in production. Finetuning runs are triggered manually in this release; automatic retraining off AI Result Evaluation feedback or Continuous Monitoring drift flags, and wiring a finetuned model in as a drop-in replacement inside Verification or Language Technology pipelines, are both on the roadmap rather than available today. The AI Platform has no separate pricing tier — finetuning runs cost 100-200 credits depending on the base model; access is arranged via [a demo call](https://crowdee.ai/demo). ## Verification Pipelines (14 built-in) These pipelines produce an authenticity verdict (authentic / manipulated / synthetic / inconclusive / unverified) with a confidence score and supporting indicators. The `synthetic` verdict flags content generated from scratch by AI, distinct from an edited-but-real `manipulated` original. Pipelines are grouped by content modality: **Text:** - Political Bias & Sentiment Analysis - Veracity Estimation with confidence scores - Knowledge Base Retrieval (231K+ fact-check articles across 20 languages) - Entity & Quote Extraction - Key Message Matching - Contextual Evidence Retrieval **Image:** - Manipulation & Splicing Detection - Synthetic / AI-Generated Image Detection - Visual Geolocation Verification - Object & Logo Recognition - Copy-Move Forensics **Audio:** - Audio Tampering & Splice Detection - Synthetic Speech / Deepfake Detection - Speaker Matching & Voice Comparison - Audio Reuse & Fingerprinting - Music–Speech Segmentation - Automatic Transcription (ASR) **Video:** - Video Synthesis & Deepfake Detection - Keyframe Extraction & Scene Analysis - Temporal Tampering Detection - Face-Swap & Reenactment Spotting **Multimodal:** - All-in-One Orchestration Pipeline - FAST Lightweight Screening - Full Manipulation Detection Suite - Cross-Modality Consistency Checks **Crowd Layer:** - Crowdsourced Text Verification - Crowdsourced Image Verification - Crowdsourced Audio Verification - Crowdsourced Video Verification - Quality Assurance & Inter-Rater Agreement **Source & Origin:** - Source Credibility & Provenance Tracing (`verify-source-credibility`): live web research to trace a claim or piece of content to its origin, AI synthesis of a source pointer and category (government/private/company), and crowd fact-checker confirmation before a final credibility score ## Language Technology (LT) Pipelines (7 built-in) LT pipelines extract structured linguistic data from files. Unlike verification pipelines, they produce text and data output — not verdicts. They can run on individual project files or on all files in a dataset version in a single API call. | Pipeline slug | Input modality | Output | |---|---|---| | `lt-language-id-text` | text / document | Detected languages with BCP-47 codes and confidence scores | | `lt-language-id-audio` | audio | Spoken language detected by Whisper | | `lt-entity-detection-text` | text / document | Named entities (PERSON, ORG, LOC, DATE, EVENT, MISC) | | `lt-entity-detection-audio` | audio | Transcription → named entity extraction | | `lt-transcription` | audio / video | Full transcript with segment-level timestamps (Whisper) | | `lt-translation` | text | Translation to any target language; auto-detects source language | | `lt-ocr` | image / PDF | Extracted text; pdftotext for digital PDFs, Ollama vision for scanned documents | **API surface (unified, current):** - `GET /v2/data-platform/pipelines/catalog` — list all 13 Data Platform pipelines (LT + cleaning-style), with `category`/`executionMode`/`appliesTo`/`targetTypes` - `POST /v2/data-platform/pipelines/:slug/run` — trigger any pipeline against a project's files or a dataset version - `GET /v2/data-platform/pipelines/runs` / `GET /v2/data-platform/pipelines/runs/:runId` — list/retrieve runs, each with its per-file/per-stage breakdown **API surface (legacy, still functional as a deprecated alias):** - `GET /v2/lt-pipelines/catalog` — list available LT pipelines - `POST /v2/lt-pipelines/project/:projectId/run` — trigger LT pipeline on selected project files (one run per file) - `GET /v2/lt-pipelines/project/:projectId/runs` — list LT runs for a project - `POST /v2/lt-pipelines/dataset/:datasetId/versions/:versionId/run` — trigger LT pipeline on all files in a dataset version - `GET /v2/lt-pipelines/dataset/:datasetId/versions/:versionId/runs` — list LT runs for a dataset version - `GET /v2/lt-pipelines/runs/:runId` — retrieve a single LT run with full result **Result storage:** one `lt_pipeline_runs` record per file via the legacy routes above; the unified route stores one `data_platform_pipeline_runs` record per invocation plus one `data_platform_pipeline_stage_runs` record per (file × stage). The `result` JSON field contains `{ pipelineSlug, stages: [{slug, name, output, rawText}], output, rawText }`. The top-level `rawText` is the primary plain-text output (e.g. transcription text, extracted OCR text, translated text). The `output` object is the structured, pipeline-specific result (e.g. entity list, language codes, translation metadata). ## Datasets Datasets are managed, versioned collections of media files that exist independently of projects. A dataset has a single modality (image, audio, video, text, document, or multimodal) and progresses through a lifecycle of versions as files are cleaned and enriched. **Supported modalities:** image (JPEG/PNG/WebP/TIFF/GIF), audio (MP3/WAV/FLAC/OGG/M4A), video (MP4/MOV/WebM/AVI), text (TXT/MD/HTML), document (PDF), multimodal (mixed types). **Version lifecycle:** - `raw` → files uploaded directly; immutable once created - `cleaning` → a cleaning pipeline is running - `cleaned` → cleaning completed; files are processed - `enriching` → per-file enrichment jobs are running - `cleaned_and_enriched` → enrichment completed on a cleaned version - `failed` → processing failed (see error field) **Cleaning-style pipelines (Data Platform):** `clean-audio-silence` (trims leading/trailing silence), `convert-format` (audio format conversion), `pii-redact-text` (LLM-based PII redaction) each produce a new derived version without mutating the source. `dedup-content-hash` (SHA-256 duplicate detection) and `split-dataset-version` (train/val/test partitioning) work directly against the existing version instead. All are listed at `GET /v2/data-platform/pipelines/catalog` (the legacy `GET /v2/datasets/catalog/cleaning-pipelines` still lists just the original cleaning pipeline). **Deduplication review:** flagged duplicate files can be soft-excluded from a version (non-destructive, reversible) via `PATCH /v2/datasets/:id/versions/:versionId/files/:fileId` with `{ "excluded": true }`. **Enrichment:** triggered per file to extract metadata, technical properties, and AI-derived features. Enriched files are required before a dataset version can be used as input for a verification pipeline run. **Language Technology and crowd-assisted labeling on datasets:** any of the 7 LT pipelines, or the `crowd-label-classify` pipeline, can be triggered on an entire dataset version in a single API call via `POST /v2/data-platform/pipelines/:slug/run` (crowd-label-classify additionally requires a `projectId`, since the crowd job it creates is project-scoped). **File provenance:** every cleaned file records a `sourceFileId` linking it to its raw original, giving a complete audit trail from upload to processed output. **API surface:** `GET/POST /v2/datasets`, `GET/PATCH/DELETE /v2/datasets/:id`, `POST /v2/datasets/:id/files` (upload), `DELETE /v2/datasets/:id/files/:fileId`, `GET /v2/datasets/:id/versions`, `GET /v2/datasets/:id/versions/:versionId`, `PATCH /v2/datasets/:id/versions/:versionId/files/:fileId` (soft-exclude), `POST /v2/datasets/:id/versions/:versionId/clean`, `POST /v2/datasets/:id/versions/:versionId/enrich`, `GET /v2/datasets/:id/versions/:versionId/export` (returns 24 h presigned S3 download URLs), `POST /v2/data-platform/pipelines/:slug/run` (trigger any Data Platform pipeline against a dataset version). The same datasets router is also mounted at `/v2/data-platform/datasets/*`. ## Key Statistics - 1M+ Global Crowd Workers: Multilingual, multi-skilled workforce for human-in-the-loop verification and labeling - 231K+ Fact-Check Articles: Knowledge base aggregating major German and international fact-checkers across 20 languages - 14 Verification Pipelines: Built-in AI pipelines producing authenticity verdicts across text, image, audio, video, and multimodal content - 13 Data Platform Pipelines: Dataset cleaning, deduplication, format conversion, PII redaction, train/val/test splitting, crowd-assisted labeling, and all 7 Language Technology pipelines (transcription, language identification, NER, translation, and OCR) - 10+ Years of Production Experience: Crowd-based verification solutions in production deployments - 3+ Years of Research: BMBF-funded research developed with DW, rbb, and Fraunhofer - 24/7 Pipeline Availability: Automated verifications run around the clock with progress tracking and alerting ## Compliance & Security - Standards: GDPR - Regulatory Readiness: DSA Article 34/35 compliant; EU AI Act ready with human-in-the-loop documentation - Every content decision leaves a full audit trail including timestamps, inter-rater agreement scores, and reviewer reasoning - Data remains private and fully under customer control ## Use Cases - Newsroom Fact-Checking: Triage and verify claims from text, social media, images, and video at high volume with full audit trails - DSA Compliance: Document human oversight for every content decision; generate evidence packages for regulators - AI Research & Ground Truth: Collect structured human judgments with inter-rater agreement statistics for ML dataset creation - Deepfake & Manipulation Detection: Forensic-grade analysis across audio, video, and image content - Media Monitoring: Track claims, sentiment, and reputation across multiple content sources and languages - Trust & Safety: Content moderation and policy enforcement for digital platforms with documented human oversight per decision - Government & Regulatory Bodies: Generate DSA Article 34/35 evidence packages with full audit trails for regulatory submission ## Key Differentiators - Produces auditable verdicts, not just confidence scores — every result includes inter-rater agreement statistics, timestamps, and reviewer reasoning - Human-in-the-loop workflow catches edge cases that AI-only tools miss - Dossiers are exportable as PDF and JSON, ready for editorial or regulatory review - No per-seat pricing — costs scale with verification volume and crowd hours, not headcount - API-first design — integrates into existing content workflows via REST API - Full audit trail on every content decision: timestamps, inter-rater agreement scores, and reviewer rationale ## Crowdworkers Crowdee maintains a waitlist for independent contractors who perform verification tasks for newsroom and compliance customers. Crowd workers provide the human-in-the-loop layer that turns AI pipeline outputs into auditable verdicts. - [Waitlist](https://crowdee.ai/crowdworkers) - Pay model: paid per task, independent contractor status, weekly payouts via SEPA bank transfer or PayPal - Languages: English, German, French, Spanish, Italian, Portuguese, Polish, Arabic, Hebrew, Turkish, Chinese, Japanese, Korean, Hindi, Russian, Ukrainian - Task types: fact-check claims against trusted sources, transcribe and review audio, annotate images and video, review AI pipeline outputs - Requirements: native or near-native language ability, desktop/laptop browser, identity verification before first paid task, EU bank account or PayPal - Onboarding: small cohorts with qualification tasks and identity check before going live ## Security & Infrastructure Crowdee processes content that newsrooms, broadcasters, and compliance teams cannot afford to leak. Security and data protection are treated as primary product features. - Hosting: website, dashboards, databases, and self-hosted analytics all run on dedicated servers at netcup GmbH data centres in Germany - Workflow engine: the backend workflow engine powering the chat assistant and Content Verification pipeline runs on Crowdee-owned hardware in Berlin - No third-party CDN, serverless platforms, or hyperscaler-managed infrastructure in the request path - Analytics: self-hosted Matomo — no Google Analytics, no third-party tracking of visitors - Customer content is never used for AI training - Application security: TLS everywhere, signed tokens, origin enforcement (CSRF), rate limiting, scoped API credentials, proof-of-work CAPTCHA (Altcha — self-hosted, no third-party CAPTCHA service) - GDPR compliant; DSA Article 34/35 ready; EU AI Act ready with human-in-the-loop documentation - Responsible disclosure: security@crowdee.ai ## Support - [FAQ / Knowledge Base](https://crowdee.ai/support) - [Submit a support ticket](https://crowdee.ai/ticket) - Live chat assistant available on the website - Support email: support@crowdee.ai - Response time: within one business day - Office hours: Monday to Friday, 09:00–18:00 Europe/Berlin ## Pricing - Quotes: Crowdee GmbH creates custom quotes for verification projects. Default quote validity period is 30 days. - No per-user seat fees — costs scale with verification volume and crowd hours actually used. - Plans: - Free: €0, no credit card required. 1 image verification per IP per 24 hours. Supports JPEG, PNG, GIF, WebP, BMP, HEIC, AVIF up to 10 MB. Returns verdict, confidence score, and supporting indicators. Optional email delivery of the report. [Try the free demo](https://crowdee.ai/content-verification) - Pay-as-you-go: From €0.05 per verification. Self-serve credits, no commitment. Includes all AI pipelines (text, image, audio, video, multimodal), programmatic REST API access, bulk uploads and batch processing, dossier export in PDF and JSON, and standard email support. - Managed Campaigns: Custom pricing, scoped per project. Crowdee's team designs, runs, and delivers the full verification campaign. Includes crowd task design, quality assurance, and final reporting. - Enterprise: Custom pricing. API quotas, SLA, SSO, and on-premise deployment option available. Contact sales@crowdee.ai. - Crowd costs: Typically €15 per hour worked, exclusive of VAT. Rates may be higher for specialized or complex tasks. A 40% crowd panel provider fee applies on top of crowd task rewards. - Data Platform pipeline costs (credits): dataset cleaning, format conversion, deduplication, and train/val/test splitting are free; PII redaction is 50 credits per file; Language Technology pipelines (transcription, translation, OCR, language ID, entity detection) are 100-350 credits per file; crowd-assisted labeling is 40 credits per worker response (typically 3 responses per file). - AI Platform finetuning costs (credits): a flat, upfront 100-200 credits per finetuning run depending on the base model; benchmarking a finetuned model with the crowd costs the same as one AI Result Evaluation. - Hourly rate: €200 per hour excluding VAT for managed services and consulting - Payment term: Net 30 days from invoice date - Payment Options: PayPal and Bank Transfer - VAT Rate: 19% on all offered products and services - [Book a Demo](https://crowdee.ai/demo) ## Links - Homepage - [Homepage](https://crowdee.ai) - [Verification Platform](https://crowdee.ai/platform) - [Crowd Platform](https://crowdee.ai/crowd-platform) - [Data Platform](https://crowdee.ai/data-platform) - [AI Platform](https://crowdee.ai/ai-platform) - [Content Verification (free demo)](https://crowdee.ai/content-verification) - [Products](https://crowdee.ai/products) - [Content Gathering](https://crowdee.ai/content-gathering) - [Source Research](https://crowdee.ai/source-research) - [AI Result Evaluation](https://crowdee.ai/ai-output-evaluation) - [Continuous Monitoring](https://crowdee.ai/continuous-monitoring) - [About us](https://crowdee.ai/about) - [Contact](https://crowdee.ai/contact) - [Book a Demo](https://crowdee.ai/demo) - [Careers](https://crowdee.ai/careers) - [Pricing](https://crowdee.ai/pricing) - [Become a Crowdworker](https://crowdee.ai/crowdworkers) - [FAQ](https://crowdee.ai/faq) - [Support](https://crowdee.ai/support) - [Submit a Ticket](https://crowdee.ai/ticket) - [Security](https://crowdee.ai/security) - [Blog](https://crowdee.ai/blog) - [Research Projects](https://crowdee.ai/projects) - [Project: news-polygraph](https://crowdee.ai/projects/news-polygraph) - [Project: ALADAN](https://crowdee.ai/projects/aladan) - [Data Subprocessors](https://crowdee.ai/subprocessors) - [Status](https://status.crowdee.ai/) - [Terms and Conditions](https://crowdee.ai/terms) - [Impressum](https://crowdee.ai/impressum) - [Privacy Policy](https://crowdee.ai/privacy) - Subdomains - [Platform (app)](https://platform.crowdee.ai) - [Documentation](https://docs.crowdee.ai) - [System Status](https://status.crowdee.ai) - Social Media - [LinkedIn](https://www.linkedin.com/company/crowdee) - [Instagram](https://www.instagram.com/crowdee.datalab) - [X](https://x.com/CrowdeeDataLab) - [Bluesky](https://bsky.app/profile/crowdee-datalab.bsky.social) - [Facebook](https://www.facebook.com/crowdee.datalab) - External - [Crunchbase](https://crunchbase.com/organization/crowdee) - [Google Maps](https://maps.app.goo.gl/q3PQ5U1Xqq1ZnVXo8) ## Trusted By Deutsche Telekom, T-Systems, DFKI, TU Berlin, Deutsche Welle (DW), rbb, Fraunhofer IDMT, Holtzbrinck, Expert AI, Ubermetrics, Vocapia, Lingea, IANUS Technologies, Santer Reply, Intapp ## Research Projects - news-polygraph: BMBF-funded project to develop an AI technology platform identifying targeted disinformation by integrating software tools for claim detection, evidence retrieval, stance classification, and credibility scoring — validated by real journalists and fact-checkers. - Start: 2023 / End: 2026 - Consortium: DFKI, delphai / Intapp, Deutsche Welle, Fraunhofer IDMT, rbb, transfermedia GmbH, TU Berlin, Ubermetrics Technologies, Crowdee GmbH - [Website](https://crowdee.ai/projects/news-polygraph) - ALADAN: AI-based language framework for defense, addressing challenges around classified data, developed with four EU SME partners. - Start: 2022 / End: 2026 - Consortium: Vocapia Research, Lingea, IANUS Technologies Ltd, Crowdee GmbH - [Website](https://crowdee.ai/projects/aladan) - DoNotFear: Crowdsourcing for data collection around perceived safety on public transport (EIT-Digital). Data collection from people on the move. - Start: 2019 / End: 2019 - Consortium: ARTE TU Berlin, QUL TU Berlin, Crowdee GmbH - ICU: Internal crowdsourcing platform for GASAG (EIT-Digital). - Start: 2018 / End: 2019 - Consortium: TU Berlin, GASAG AG, Crowdee GmbH, InnoZ GmbH, IZT gGmbH - BRIDGE: Skills-based inclusion of migrants — crowdsourced skills tests for practical, vocational, and academic assessment in Italian welcome offices (EIT-Digital). - Start: 2018 / End: 2018 - Consortium: Engineering, Expert System, FBK, TU Berlin, Crowdee GmbH - ERICS: Crowd-supported automatic translation system for underrepresented languages (Farsi, Arabic) for Germany's refugee welcome portal HandbookGermany.de (EIT-Digital). - Start: 2017 / End: 2018 - Consortium: T-Systems Multimedia Solutions, Deutsche Telekom AG, Aalto University, DFKI, TU Berlin, Crowdee GmbH - ALMeS: Crowd-monitored predictive maintenance pipeline for IoT sensor data from production machinery (EIT-Digital). - Start: 2017 / End: 2017 - Consortium: Crowdee GmbH, Santer Reply, STMicroelectronics, Cohaerentia, FBK, TU Berlin, KONUX - PST: Data anonymization and pseudonymization pipeline enabling crowdsourcing with enterprise NLP data at scale (EIT-Digital). - Start: 2016 / End: 2016 - Consortium: Starbytes Reply, ELTE, TU Berlin, Crowdee GmbH ## Blog - Evaluate your own AI's output: external submissions land in AI Result Evaluation - Date: 2026-07-03 - Author: André Beyer - Category: Product - Excerpt: AI Result Evaluation used to only work on Crowdee's own pipeline runs. Now it works on anything with a verdict and an explanation — including your own model — and we didn't need a new table to do it. - [URL (EN)](https://crowdee.ai/blog/evaluate-your-own-ai-external-submissions) - [URL (DE)](https://crowdee.ai/de/blog/eigene-ki-ergebnisse-bewerten-externe-einreichungen) - Continuous Monitoring: self-baselining drift detection for AI pipelines - Date: 2026-07-03 - Author: André Beyer - Category: Product - Excerpt: Our first recurring job scheduler, why drift detection compares a schedule against its own history instead of a fixed threshold, and the one honest limitation we shipped with rather than wait to solve. - [URL (EN)](https://crowdee.ai/blog/continuous-monitoring-self-baselining-drift-detection) - [URL (DE)](https://crowdee.ai/de/blog/kontinuierliches-monitoring-selbstkalibrierende-drift-erkennung) - AI Result Evaluation: crowd-rated transparency for every AI verdict - Date: 2026-06-29 - Author: André Beyer - Category: Product - Excerpt: A confidence score tells you how sure the model is, not whether its explanation actually makes sense to a person. Here's how we built a crowd panel to rate that, and where the ratings go afterward. - [URL (EN)](https://crowdee.ai/blog/ai-result-evaluation-crowd-rated-transparency) - [URL (DE)](https://crowdee.ai/de/blog/ki-ergebnisbewertung-crowd-bewertete-transparenz) - Source Research: tracing content back to where it actually came from - Date: 2026-06-22 - Author: André Beyer - Category: Product - Excerpt: Why authenticity and provenance are separate questions, how a new web-research stage type lets pipelines search the live web instead of reasoning from training data, and what verify-source-credibility actually returns. - [URL (EN)](https://crowdee.ai/blog/source-research-tracing-content-to-its-origin) - [URL (DE)](https://crowdee.ai/de/blog/quellenrecherche-inhalte-zum-ursprung-zurueckverfolgen) - Synthetic content detection: telling AI-generated media apart from manipulated media - Date: 2026-06-15 - Author: André Beyer - Category: Product - Excerpt: Why 'manipulated' and 'synthetic' are different questions, which pipelines now score synthetic likelihood, and how we fixed a prompt bug that was silently capping what our own models could report. - [URL (EN)](https://crowdee.ai/blog/synthetic-content-detection-ai-generated-media) - [URL (DE)](https://crowdee.ai/de/blog/synthetische-inhalte-erkennung-ki-generierte-medien) - Content Gathering: crowd-powered discovery across the open web - Date: 2026-06-08 - Author: André Beyer - Category: Product - Excerpt: Why open-ended discovery needed a different data model than our verification pipelines, how deduplication works, and how gathered content flows straight into verification. - [URL (EN)](https://crowdee.ai/blog/content-gathering-crowd-powered-web-discovery) - [URL (DE)](https://crowdee.ai/de/blog/content-gathering-crowd-basierte-web-recherche) - The Future of Crowdsourcing in AI Data Annotation - Date: 2026-06-01 - Author: Alea Annacker - Category: Innovation - Excerpt: How distributed human intelligence is shaping the next generation of machine learning datasets. - [URL (EN)](https://crowdee.ai/blog/the-future-of-crowdsourcing) - [URL (DE)](https://crowdee.ai/de/blog/die-zukunft-des-crowdsourcing) - Cleaning pipelines catalog: from audio silence trimming to per-file enrichment - Date: 2026-05-25 - Author: Alea Annacker - Category: Product - Excerpt: How the cleaning pipelines catalog works, what audio silence trimming does to your files, how file provenance is preserved through the clean → enrich → verify workflow, and how new pipelines get added. - [URL (EN)](https://crowdee.ai/blog/cleaning-pipelines-catalog-audio-silence-trimming) - [URL (DE)](https://crowdee.ai/de/blog/cleaning-pipelines-catalog-audio-stille-trimming) - Verification contexts: how projects set the structured inputs pipelines read - Date: 2026-05-11 - Author: Markus Hadick - Category: Product - Excerpt: How verification contexts work as named, reusable sets of claims that pipelines check content against, why locking in context at run start keeps results consistent, and how to manage contexts day to day. - [URL (EN)](https://crowdee.ai/blog/verification-contexts-structured-inputs-pipelines-read) - [URL (DE)](https://crowdee.ai/de/blog/verification-contexts-strukturierte-inputs-pipelines-lesen) - File upload and the enrichment pipeline: the missing pre-requisite for verification - Date: 2026-04-27 - Author: Markus Hadick - Category: Product - Excerpt: Why enrichment must run before most verification pipelines, what it extracts for each modality, and how OCR fits into the picture. - [URL (EN)](https://crowdee.ai/blog/file-upload-enrichment-pipeline-prerequisite) - [URL (DE)](https://crowdee.ai/de/blog/file-upload-enrichment-pipeline-voraussetzung) - Dataset version lifecycle: raw → cleaning → cleaned → enriching → cleaned_and_enriched - Date: 2026-04-13 - Author: Markus Hadick - Category: Product - Excerpt: How Crowdee models multi-modal file collections as versioned datasets, tracks derived versions through cleaning and enrichment states, and preserves full file provenance back to the source. - [URL (EN)](https://crowdee.ai/blog/dataset-version-lifecycle-raw-cleaning-cleaned-enriching) - [URL (DE)](https://crowdee.ai/de/blog/dataset-version-lifecycle-raw-cleaning-cleaned-enriching) - The crowd job lifecycle: publish → workers → answers → accept/reject - Date: 2026-03-30 - Author: Alea Annacker - Category: Product - Excerpt: A walkthrough of every stage a crowd job passes through, from initial creation and dry-run testing to worker assignment, answer review, and downstream export. - [URL (EN)](https://crowdee.ai/blog/crowd-job-lifecycle-publish-workers-answers) - [URL (DE)](https://crowdee.ai/de/blog/crowd-job-lifecycle-publish-workers-answers) - Input data: per-task variation in crowd jobs - Date: 2026-03-16 - Author: Alea Annacker - Category: Product - Excerpt: How Crowdee's input data model — sets, lists, groups, and variants — powers per-task variation in crowd jobs, and how accepted answers can feed back as the next job's input. - [URL (EN)](https://crowdee.ai/blog/input-data-per-task-variation-crowd-jobs) - [URL (DE)](https://crowdee.ai/de/blog/input-data-pro-task-variation-crowd-jobs) - SurveyJS template design: how to design a survey that produces verifiable crowd judgments - Date: 2026-03-02 - Author: Alea Annacker - Category: Product - Excerpt: A practical guide to designing SurveyJS-based task templates: version history, per-task variation, and the design principles that produce crowd judgments you can actually act on. - [URL (EN)](https://crowdee.ai/blog/surveyjs-template-design-verifiable-judgments) - [URL (DE)](https://crowdee.ai/de/blog/surveyjs-template-design-pruefbare-urteile) - The public support chat: consent-gated chat widget for visitors - Date: 2026-02-16 - Author: Karina - Category: Product - Excerpt: How the Crowdee homepage loads the support chat widget only after explicit cookie consent, and how the async ticket form complements it for tracked support requests. - [URL (EN)](https://crowdee.ai/blog/public-support-chat-consent-gated-widget) - [URL (DE)](https://crowdee.ai/de/blog/oeffentlicher-support-chat-consent-gated-widget) - Getting started and authentication: how to issue an API key and make your first call - Date: 2026-01-19 - Author: Karina - Category: Developer - Excerpt: A practical overview of the Crowdee auth options, how to issue an API key, and where to find the step-by-step integration guide. - [URL (EN)](https://crowdee.ai/blog/getting-started-authentication-api-key) - [URL (DE)](https://crowdee.ai/de/blog/getting-started-authentifizierung-api-key) - The llms.txt endpoint: LLM-optimised API documentation - Date: 2025-12-08 - Author: André Beyer - Category: Innovation - Excerpt: Why we added a plain-text API summary alongside our full API reference and interactive docs, and how AI tools use it to generate correct integration code without parsing complex schemas. - [URL (EN)](https://crowdee.ai/blog/llms-txt-endpoint-llm-optimised-documentation) - [URL (DE)](https://crowdee.ai/de/blog/llms-txt-endpoint-llm-optimierte-dokumentation) - API keys and organization switching: how Crowdee keeps organizations isolated - Date: 2025-11-10 - Author: Markus Hadick - Category: Developer - Excerpt: A look at how Crowdee issues and scopes API keys, keeps organizations cleanly separated, and lets users switch between them without friction. - [URL (EN)](https://crowdee.ai/blog/api-keys-and-organization-switching) - [URL (DE)](https://crowdee.ai/de/blog/api-keys-und-organisation-switching) - OpenAPI, an interactive explorer, and llms.txt: the three layers of the developer experience - Date: 2025-10-13 - Author: André Beyer - Category: Developer - Excerpt: How we serve the same API surface three different ways — as a machine-readable spec, an interactive explorer, and a plain-text summary for LLMs — and why each layer exists. - [URL (EN)](https://crowdee.ai/blog/openapi-scalar-llms-text-developer-experience) - [URL (DE)](https://crowdee.ai/de/blog/openapi-scalar-llms-text-entwicklererfahrung) - The MCP server in depth: connecting AI agents to Crowdee - Date: 2025-09-15 - Author: André Beyer - Category: Innovation - Excerpt: A complete walkthrough of the Crowdee MCP server: what it exposes, how authentication works, and how AI agents can drive the full platform programmatically. - [URL (EN)](https://crowdee.ai/blog/mcp-server-deep-dive) - [URL (DE)](https://crowdee.ai/de/blog/mcp-server-im-detail) ## Team - Dr.-Ing. Tim Polzehl — CEO - André Beyer — CTO - Alea Annacker — Tech - Karina — Operations ## Careers - Job Openings: No open positions at the moment. Unsolicited applications welcome via email. - Careers E-Mail Address: careers@crowdee.ai - [Careers Webpage](https://crowdee.ai/careers) - 100% remote work possible ## FAQ ### Platform & Getting Started Q: What is Crowdee and what does the platform do? A: Crowdee is a hybrid content-verification platform that combines AI-assisted analysis pipelines with human-in-the-loop crowd review. It helps newsrooms, compliance teams, fact-checkers, and AI researchers detect manipulated or synthetic media, retrieve evidence, and produce auditable verdicts — not just confidence scores — across text, image, audio, video, and multimodal content. Q: Who is Crowdee built for? A: Crowdee is designed for professional and organisational users: newsrooms and editorial teams, compliance and trust-and-safety functions, regulatory and policy teams (e.g. for DSA Art. 34/35 reporting), fact-checking organisations, and AI researchers building or evaluating verification datasets. The platform is not intended for consumers or users under the age of 16. Q: Is there a free way to try Crowdee? A: Yes. [The free Content Verification demo](https://crowdee.ai/content-verification) lets you upload a single image and receive an AI-assisted authenticity verdict with a confidence score and supporting indicators. No account or credit card is required. One verification per IP address per 24 hours applies. Q: How do I get started with a paid plan? A: [Book a 30-minute demo](https://crowdee.ai/demo) with our team. We will discuss your use case, walk you through the relevant pipelines, and scope the right commercial arrangement — whether that's pay-as-you-go credits, a managed campaign, or an enterprise integration. Q: What types of content can Crowdee analyse? A: Crowdee covers five modalities: text (political bias, veracity estimation, entity and quote extraction, fact-check knowledge-base retrieval in 20 languages, translation), image (manipulation and AI-generation detection, visual geolocation, OCR text extraction), audio (tampering detection, synthetic-speech detection, transcription, language identification, named entity recognition), video (deepfake and temporal tampering detection, transcription), and multimodal orchestration with cross-modality consistency checks. Language Technology (LT) pipelines are available as a separate layer for extracting structured linguistic data from any file — independently of verification. Q: How many verification pipelines does Crowdee offer? A: Crowdee ships 14 built-in content verification pipelines and 7 Language Technology (LT) pipelines — 21 in total across the five supported modalities. Verification pipelines produce authenticity verdicts (authentic/manipulated/synthetic/inconclusive/unverified). LT pipelines extract structured linguistic data such as transcripts, named entities, translations, detected languages, and OCR text. The exact set available to your organisation depends on your plan. Q: What is a verification dossier? A: A dossier is an exportable report (PDF + JSON) bundling all pipeline outputs, confidence scores, retrieved evidence, crowd verdicts, and a full audit trail for a single verification case. It is designed to be submitted directly to editors, legal teams, or regulators without further formatting. Q: What is a confidence score? A: A confidence score is the model's self-reported certainty for a given verdict. A high score means the model is internally consistent — it does not guarantee correctness. Identical inputs may yield different verdicts as models are updated. For consequential decisions we recommend combining AI verdicts with crowd review. Q: In how many languages does Crowdee's fact-check knowledge base operate? A: The knowledge-base retrieval pipeline covers more than 231,000 fact-checks in 20 languages, sourcing from established international and regional fact-checking organisations. Q: Can I use Crowdee without creating an account? A: The free Content Verification demo is fully anonymous — no account required. For API access, managed campaigns, repeat use beyond the daily free limit, or any paid feature, an Individual Agreement and account are needed. Q: Is Crowdee suitable for both one-off investigations and ongoing monitoring? A: Yes. Pay-as-you-go credits suit one-off investigations or evaluation projects. Managed campaigns are ideal for time-bound initiatives such as election monitoring. Enterprise agreements support continuous, high-volume monitoring with dedicated capacity and negotiated SLAs. Q: What is the difference between AI pipelines and crowd-assisted review? A: AI pipelines process content automatically using machine-learning models and return verdicts within seconds. Crowd-assisted review adds one or more independent human raters per item for corroboration, quality control, and auditability — it takes longer but produces verdicts that are harder to challenge and more defensible in editorial or regulatory contexts. Q: Is Crowdee available in German? A: The Crowdee website and customer-facing communications are available in English and German. Pipeline outputs are language-agnostic where technically possible; crowd task interfaces can be configured in the languages available in the active worker pool. Q: Is Crowdee a German company? A: Yes. Crowdee GmbH is incorporated in Germany and headquartered in Berlin at Zehdenicker Str. 5, 10119 Berlin. All invoicing and data hosting is conducted under European terms. Q: Where can I find the platform documentation? A: [The interactive API explorer](https://api.crowdee.ai/v2) (Scalar UI) is accessible. For further questions, reach us via [our support page](https://crowdee.ai/support) or [our contact page](https://crowdee.ai/contact). Q: Does the platform include an AI chat assistant? A: Yes. The platform includes a built-in Chat Agent — a conversational AI assistant that helps you plan and execute verification workflows, query pipeline results, and get guided recommendations without leaving the platform. It is available to all authenticated users from the main navigation. Q: What is the Datasets feature? A: Datasets are managed, versioned collections of media files that exist independently of projects. You upload a corpus of files once — images, audio, video, text, or documents — and then apply cleaning pipelines, per-file enrichment, and Language Technology pipelines without duplicating the underlying content. Each processing step creates a new derived version, so you always have a full audit trail from raw upload to processed output. Enriched dataset versions can be used directly as input for verification pipeline runs, and LT pipelines (transcription, NER, translation, OCR, language identification) can be triggered on all files in a version in a single API call. ### Content Verification & AI Pipelines Q: Which content modalities does Crowdee support? A: Crowdee supports text, image, audio, video, and multimodal content. Each modality has dedicated pipelines tailored to its specific manipulation and authenticity signals. Multimodal verification orchestrates cross-modality consistency checks across two or more content types simultaneously. Q: What can Crowdee detect in images? A: Image pipelines cover manipulation and splicing detection, copy-move forensics, AI-generated image detection, and visual geolocation estimation. Each pipeline returns a verdict, a confidence score, and supporting forensic indicators that are bundled into the dossier. Q: What text analysis pipelines are available? A: Verification text pipelines include political bias and sentiment analysis, veracity estimation, named entity and quote extraction, and knowledge-base retrieval against 231,000+ international fact-checks in 20 languages. Results include source links and reliability annotations. In addition, Language Technology (LT) pipelines offer standalone text capabilities: named entity recognition across PERSON/ORG/LOC/DATE/EVENT types, automated translation into any major language (with auto-detected source language), and language identification — all returning structured JSON output rather than a verdict. Q: How does audio verification work? A: Audio verification pipelines detect tampering and cut-and-splice artefacts, identify synthetic or AI-generated speech, perform speaker matching against reference audio, and produce ASR transcriptions. Outputs include a manipulation verdict plus a timestamped analysis where applicable. For pure language processing without a verification verdict, the dedicated LT pipelines provide standalone transcription (Whisper, with segment timestamps), spoken language identification, and named entity recognition from the transcript. Q: What does video verification cover? A: Video pipelines detect deepfakes (face-swap and re-enactment), temporal tampering and frame-level inconsistencies, and AI-generated video artefacts. Results include a per-clip verdict, a confidence score, and flagged frames or segments for human review. Q: What is multimodal verification? A: Multimodal verification orchestrates two or more modality pipelines on a single case and performs cross-modality consistency checks — for example, comparing a video's audio track to its visual content, or matching quoted text to the speaker visible on screen. This is particularly useful for detecting sophisticated composite fakes. Q: How accurate are AI verdicts? A: Verdicts are probabilistic estimates produced by machine-learning models and may be incomplete, biased, or incorrect even when confidence is high. Identical inputs may yield different verdicts as models are updated. Accuracy varies by modality, content type, and manipulation method. For consequential use cases, we strongly recommend adding crowd review for human corroboration. Q: Can I rely on Crowdee's verdicts for editorial or legal decisions? A: Outputs are informational and are not, by themselves, legal, journalistic, regulatory, or other professional advice. They do not constitute a definitive determination of authenticity, manipulation, defamation, or copyright infringement. You remain responsible for reviewing each output and applying your own editorial and compliance controls before publication or regulatory submission. Q: How often are AI models updated? A: Models are updated regularly as new research and training data become available. Because of this, the same input may produce slightly different results over time. Model versioning is tracked in the audit trail included in each dossier, so historical verdicts remain reproducible and auditable. Q: What is the fact-check knowledge base? A: The knowledge-base retrieval pipeline queries an index of 231,000+ professional fact-checks sourced from established fact-checking organisations across 20 languages. It returns the most semantically relevant fact-checks for a given claim, together with their verdict labels and source URLs, so journalists can cross-reference existing coverage instantly. Q: Can I compose custom pipelines? A: Yes. Custom pipeline composition — selecting which modality pipelines to run, in what order, and with which quality gates — is available as part of managed campaign and enterprise agreements. Our team works with you during scoping to design a pipeline that matches your editorial or compliance workflow. Q: How quickly do AI pipelines return results? A: Fully automated AI pipelines typically return results within seconds for image and text, and within minutes for longer audio or video content depending on duration and the number of pipelines run. Crowd-assisted review takes additional time proportional to the task scope agreed during scoping. Q: What is a verdict? A: A verdict is the platform's determination for a given piece of content — for example, 'likely manipulated', 'consistent with authentic', or 'AI-generated'. It is accompanied by a confidence score, a set of supporting indicators, and a full audit trail. Verdicts from both AI pipelines and crowd review are included in the exported dossier. Q: What output formats are available? A: Dossiers are exportable as PDF and JSON. The JSON format is structured for programmatic ingestion into newsroom CMSs, compliance management systems, or internal tooling. The PDF format is designed for direct submission to editors, legal teams, or regulatory bodies. Q: Does Crowdee detect AI-generated images specifically? A: Yes. A dedicated AI-generated image detection pipeline analyses content for artefacts characteristic of generative models (GANs, diffusion models, and others). It returns a verdict indicating the likelihood that the image was synthetically generated, along with a confidence score and supporting features. Q: Can I monitor pipeline run progress in real time? A: Yes. The platform uses WebSocket notifications to push live stage-by-stage progress updates to your browser as a verification pipeline run executes. You can see each stage transition — from pending through running to completed or failed — without refreshing the page. This is especially useful for longer multi-stage pipelines involving crowd review steps. ### Crowd & Human Review Q: What is crowd-assisted human review? A: Crowd-assisted review distributes content items to multiple independent crowd workers who complete structured verification tasks — rating authenticity, flagging manipulation artefacts, or cross-checking contextual claims. Their individual judgements are aggregated, quality-controlled, and combined with AI pipeline output to produce a corroborated, auditable verdict. Q: How are crowd workers selected for a task? A: Workers are matched to tasks based on their language skills, domain knowledge, past performance, and quality scores on similar tasks. For specialist tasks — such as legal document review, regional language fact-checking, or domain-specific media analysis — dedicated worker pools with relevant qualifications are assembled. Q: How many workers review each item? A: The number of independent raters per item is agreed during campaign scoping and depends on the required confidence level, task complexity, and regulatory or editorial standards you need to meet. Typical configurations use three to five raters per item; higher counts are available for sensitive or high-stakes cases. Q: What is inter-rater agreement (IRA)? A: Inter-rater agreement is a statistical measure of how consistently independent workers rate the same item. High IRA indicates that the verdict is robust and reproducible; low IRA flags the item for further review or escalation. IRA metrics are included in the exported dossier. Q: What are gold questions? A: Gold questions are items with known correct answers that are embedded invisibly into the task flow. Workers are not told which questions are gold. Their performance on gold questions is tracked and used to identify inattentive or low-quality contributors, whose responses are down-weighted or excluded from aggregation. Q: How is overall quality controlled? A: Quality control combines inter-rater agreement analysis, gold question pass rates, attention checks, task completion time monitoring, and longitudinal performance tracking. Workers who consistently underperform are removed from active pools. These controls are configured per campaign and their parameters are disclosed in the dossier. Q: How long does crowd review take? A: Turnaround depends on task complexity, the number of raters per item, the volume of items, and worker availability. Simple binary tasks typically turn around within hours; multi-step, expert-level review of complex media may take one to two days. We scope expected turnaround as part of every managed campaign. Q: Can I customise the review task? A: Yes. Review tasks are built using survey templates that can be tailored to your workflow — including the rating scale, decision categories, evidence-capture fields, and the instructions shown to workers. Custom templates are designed collaboratively with our team during campaign scoping. Q: Do I interact with crowd workers directly? A: No. Crowdee manages the entire worker pool — recruitment, qualification, task assignment, quality control, payment, and dispute resolution. You interact only with Crowdee as your single point of contact. Workers are Crowdee's independent contractors and do not contract directly with customers. Q: How are crowd verdicts combined with AI pipeline output? A: The platform presents AI pipeline verdicts and crowd verdicts side by side in the dossier, each with their own confidence scores and evidence. Where both agree, this strengthens the overall verdict. Where they diverge, the dossier highlights the discrepancy for your editorial or compliance review. Aggregation rules are configured per campaign. Q: Are crowd workers subject-matter experts? A: It depends on the task. General verification tasks — such as contextual plausibility checks or basic manipulation detection — use generalist pools. Tasks requiring domain expertise, such as medical misinformation review, legal document annotation, or regional political fact-checking, use specialist pools assembled and qualified for that purpose. Q: In which languages can crowd workers review content? A: Crowdee operates worker pools across many languages. The specific languages available for a given campaign are confirmed at scoping. If you need coverage in a language not currently in the active pool, contact our team — we can often arrange specialist coverage. Q: What happens when workers significantly disagree? A: Low inter-rater agreement is flagged automatically. Depending on the campaign configuration, items with high disagreement are either routed for an additional round of review with more raters, escalated to a senior reviewer, or flagged in the dossier as inconclusive — requiring your editorial or compliance team to make the final determination. Q: Is crowd review available for all content modalities? A: Yes. Crowd review is available for text, image, audio, and video content. The task interface is adapted for each modality — workers can annotate images directly, listen to audio segments and timestamp artefacts, or review video clips frame by frame, depending on the task design. Q: How is crowd labour invoiced? A: Crowd-worker labour is billed at €15 per worker-hour. The number of hours required for your campaign is estimated and agreed in advance based on task complexity, number of raters per item, and quality gates. There are no hidden markups and no surprise overruns beyond the agreed scope. ### Crowdsourcing Jobs & Surveys Q: What is the Crowdsourcing feature in the Crowdee platform? A: The Crowdsourcing feature lets you design, publish, and manage human intelligence tasks directly inside the platform. You define a survey template to collect structured responses, optionally attach input data to create per-variant task slots, set per-task rewards and a validity window, and publish the job to Crowdee's worker pool. Workers complete tasks through a guided browser interface; their answers are stored, quality-controlled, and available for review, export, or re-use as input data for further verification pipeline runs. Q: How do I create a crowdsourcing job? A: Navigate to a project in the platform and open the Crowdsourcing tab. Click New Job, fill in the name, title, and description, link a published survey template version, optionally attach an input data set, set the reward credits per task and the job validity period, then publish. Jobs can also be created programmatically via POST /v2/projects/:projectId/crowd-jobs. Q: What are survey templates and how do they work? A: Survey templates are reusable form definitions built with SurveyJS. They define the questions, UI components, and validation rules that workers see when completing a task. Templates are versioned — you publish a new version when you change the template design, and jobs always reference a specific immutable version. Create and manage templates under Settings > Survey Templates in the platform. Q: What is input data and how does it work with jobs? A: Input data sets are structured collections of key-value records — for example image URLs, text snippets, or claim pairs — that you attach to a job. Each record becomes a separate task variant: workers receive one variant per task slot. If no input data is attached, all workers see the same task. Input data sets can be created manually in the platform or generated automatically from accepted crowd answers via the Answers to Input Data export. Q: What is a dry run and when should I use one? A: A dry run is a test mode that restricts job access to a list of invited email addresses or a magic-token link. Credits are not blocked for dry runs. Use a dry run to validate your survey template, check the worker UI, and collect a small set of test answers before publishing the job to the full worker pool. Q: What is a qualification test? A: A qualification test is a short, unpaid screening job you build the same way as any other job (using the same survey builder), then mark as a qualification test. Each worker gets exactly one attempt — there is no retry — and passing is a one-time investment: once a worker passes, they stay qualified for that test going forward, with no need to requalify. Q: How do I require workers to pass a qualification test before starting my job? A: Use the Qualification Requirements wizard on your project: pick the qualification test, pick the job it should protect, and define one or more pass/fail conditions on the test's answers (for example, a specific answer must equal, exceed, or simply exist). All conditions for a test must be met to pass it, and a single job can require passing multiple different qualification tests. Workers who haven't taken a required test yet see a 'Start Qualification' prompt instead of the normal start button; workers who don't meet the conditions see a clear 'not qualified' screen with no retry. Q: What is the maximum work time for a job, and how does it affect the reward? A: When creating a paid job, you set a maximum work time per task, from 1 minute up to 24 hours. To keep pay fair relative to time, the reward must be at least 20 credits for every minute of that limit (for example, at least 200 credits for a 10-minute task) — the job form enforces this automatically. Once a worker starts a task, they see a live countdown with a short grace-period warning near the deadline; if time runs out entirely, the task slot is released back into the pool so another worker can pick it up. Q: How does the task assignment lifecycle work? A: When a worker opens a job, the platform allocates one task slot per (worker × variant) combination up to the configured maximum repetitions per variant. The slot expires automatically if not submitted within the task expiry window. Workers can return a task without completing it. Each submitted answer is stored as a crowd answer linked to its task slot and is immediately available for review. Q: How do I review and accept or reject worker answers? A: Open the job's Answers tab in the platform to see all submitted responses. Inspect individual answers and mark each as accepted or rejected — with an optional rejection reason. Accepting an answer releases the blocked credits to the worker. Rejected answers are flagged for potential replacement. You can also list and review answers programmatically via GET /v2/crowd-jobs/:jobId/answers. Q: Can I export collected answers? A: Yes. Answers can be exported as JSON or CSV from the job's Answers tab or via GET /v2/crowd-jobs/:jobId/answers/export?format=json|csv with an optional status filter. You can also convert answers directly into a new dataset (POST .../answers-to-dataset) or a new input data set (POST .../answers-to-input-data) for use in further pipeline runs or as training material. Q: How are credits managed for crowdsourcing jobs? A: When you publish a non-dry-run job, credits equal to maxRepetitionsPerVariant × variantCount × rewardCredits are blocked from your organisation's balance upfront. Credits are released to workers when their individual answers are accepted. Unused blocked credits are returned to your organisation balance when the job is archived. Every one of these events — funding, refunds, and reward adjustments — appears as its own line in your organisation's cost log under Settings → Billing, with a running balance. Q: Can I clone an existing job? A: Yes. Use the Clone action on the job detail page, or call POST /v2/crowd-jobs/:jobId/clone. The new job copies all settings from the original. You can optionally clone it as a dry run first to test before going live. Cloning does not duplicate the underlying input data or survey template — both are referenced, not copied. Q: What file types can workers attach to their answers? A: Workers can attach images (JPEG, PNG, WebP, GIF), audio (MP3, WAV, FLAC), video (MP4, WebM, MOV), and PDF or text documents to their answers if the survey template includes a file-upload question. Attached files are stored in your organisation's S3 bucket and can be downloaded in bulk via GET /v2/crowd-jobs/:jobId/files/download. ### Pricing & Billing Q: What pricing models are available? A: Crowdee offers four commercial paths: Free (the public Content Verification demo, no account required), Pay-as-you-go (credits, from €0.05 per verification, no commitment), Managed Campaigns (custom-scoped, fixed-fee engagements run by our team), and Enterprise (annual agreements with negotiated quotas, SLAs, and dedicated capacity). Q: What is pay-as-you-go? A: Pay-as-you-go gives you self-serve access to AI verification pipelines. You top up credits when you need them and consume them at your own pace. Pricing is pipeline-level — you pay only for the verifications you actually run, starting at €0.05 per call depending on the pipeline. There is no minimum spend and no monthly commitment. Q: What are Managed Campaigns? A: Managed Campaigns are custom-scoped, fixed-fee engagements where our team designs the verification pipeline, configures survey templates, manages the crowd, and delivers a final auditable dossier. They are ideal for newsrooms, compliance teams, and researchers who want to hand the operational complexity to us and receive a finished result. Q: What is an Enterprise agreement? A: Enterprise agreements are annual contracts that include negotiated API quotas, dedicated processing capacity, custom SLAs with uptime and response targets, SSO and role-based access control, a signed DPA, and a named technical account manager. They are designed for production integrations that require throughput, control, and accountability. Q: How much does crowd-worker labour cost? A: Crowd-worker labour is invoiced at €15 per worker-hour. The hours required for a campaign are scoped and agreed in advance based on task complexity, the number of independent raters per item, and the quality gates you need. There are no markups and no surprise overruns beyond the agreed scope. Q: Is there a minimum spend or commitment for pay-as-you-go? A: No. Pay-as-you-go has no minimum spend and no monthly commitment. You top up credits as needed and consume them at your own pace. Credits do not expire. For enterprise plans, annual commitments are required because they include dedicated capacity and negotiated SLAs; those terms are agreed in writing in advance. Q: How long are custom quotes valid? A: Unless stated otherwise in writing, custom quotes for managed campaigns and enterprise plans are valid for 30 days from the date of issue. After that, we will gladly refresh the quote if the underlying scope has not changed significantly. Q: What payment methods and currencies does Crowdee accept? A: We invoice in EUR and accept payment by SEPA bank transfer or PayPal, with Net 30 payment terms. Other currencies and payment methods are available on request for enterprise contracts. Crowdee GmbH is a German company and issues VAT-compliant invoices. Q: Do prices include VAT? A: Prices are quoted net of statutory value-added tax. For self-service credit top-ups via PayPal in the platform, 19% German VAT is automatically added and shown before you pay if your organisation's billing address is in Germany — every other country isn't taxed automatically today. Only the net amount converts into platform credits; the VAT itself never does. For managed campaigns and enterprise invoices, we confirm the applicable tax treatment — including EU reverse-charge where a valid VAT ID applies — in the quote. Q: Can I switch from pay-as-you-go to a managed campaign later? A: Yes — that is by design. Many customers start with the free demo, evaluate a few pipelines on pay-as-you-go, and then graduate to a managed campaign or enterprise integration once they know which workflows they want us to operate end-to-end. There is no lock-in and no penalty for switching. Q: Are there discounts for non-profits, researchers, or newsrooms? A: Yes. We work with public-interest newsrooms, academic researchers, and non-profit fact-checking organisations under reduced rates and grant-funded arrangements. [Contact us](https://crowdee.ai/contact) and tell us about your project — we will get back to you with a tailored proposal. Q: How do I top up credits for pay-as-you-go? A: Organisation admins can self-serve: open Settings → Billing in the platform, click Add Funds via PayPal, enter the number of credits you want, review the net/VAT/total breakdown, and pay via PayPal — credits land in your balance as soon as the payment is confirmed. For larger or recurring top-ups, you can still arrange this via your Individual Agreement by [contacting our team](https://crowdee.ai/contact) or your account manager. Q: Can I see a full history of what my credits were spent on? A: Yes. Settings → Billing shows a complete cost log for your organisation: every top-up, job funding and refund, AI-evaluation and pipeline charge, and any manual credit grant, each with a running balance so you can always see how you got to your current total. Credit top-ups link through to their invoice; other entries show which job, pipeline run, or evaluation they belong to. Q: What are Net 30 payment terms? A: Net 30 means invoices are payable within 30 calendar days of the invoice date. Statutory late-payment interest applies after the due date in accordance with German law. Most of our managed campaign and enterprise customers operate on Net 30 terms. Q: How quickly can I get a quote? A: We aim to provide indicative pricing within one to two business days of an initial conversation. For complex managed campaigns or enterprise integrations with custom requirements, a more detailed quote may take slightly longer. Book a demo to start the process. Q: How does billing work for a managed campaign? A: Managed campaigns are billed as a fixed fee agreed in writing before work begins, covering pipeline usage and any included crowd hours. Additional crowd hours or scope changes are agreed and invoiced separately. A deposit may be required for large engagements. Final invoicing follows completion of the deliverable. ### Data, Privacy & Compliance Q: Where is my data stored and processed? A: All Crowdee-operated systems are hosted in Germany. The website and self-hosted analytics run on dedicated servers at netcup GmbH in German data centres; our automation and workflow backend runs on Crowdee-owned hardware in Berlin. Third-party AI model providers engaged as processors may be located outside the EU/EEA — transfers are safeguarded by the EU–US Data Privacy Framework and Standard Contractual Clauses. Q: Is Crowdee GDPR-compliant? A: Yes. Crowdee processes personal data in accordance with the GDPR, the German Federal Data Protection Act (BDSG), and the TDDDG for cookies and similar technologies. Where we process personal data on your behalf as a processor, we sign a data-processing agreement under Art. 28 GDPR before processing begins. Q: Can I sign a data-processing agreement (DPA)? A: Yes. We sign an Art. 28 GDPR data-processing agreement before processing any customer data on your behalf. The DPA lists sub-processors, transfer mechanisms, retention periods, and the technical and organisational measures we apply. Contact privacy@crowdee.ai to initiate the DPA process. Q: Who are Crowdee's sub-processors? A: A current list of sub-processors — including the service they provide and their country of establishment — is published on [our subprocessors page](https://crowdee.ai/subprocessors). We update this list when sub-processors change and notify customers with reasonable advance notice, as required by the DPA. Q: Are uploaded images or content used to train AI models? A: No. Uploaded content and pipeline outputs are processed solely to produce verification results. Crowdee does not use customer content to train its own foundation models. We contractually require the third-party AI model providers we engage not to use inputs or outputs to train their models. Q: How long is my data retained? A: Uploaded image bytes are kept only for the duration of the inference call and are not persisted to disk. The hash-keyed result cache expires after 24 hours. Other data — such as account information, dossiers, and audit logs — is retained as described in the Privacy Policy and your Individual Agreement, and is deleted on request or at contract termination. Q: Does Crowdee support DSA Art. 34/35 compliance obligations? A: The platform is designed to support these obligations: auditable verdicts with confidence scores, inter-rater agreement metrics, traceable evidence retrieval, and exportable dossiers map onto the systemic risk assessment and mitigation documentation requirements of the Digital Services Act. Crowdee does not, however, take over your regulatory obligations — you remain the controller and regulated party. Q: Does Crowdee help with EU AI Act compliance? A: Yes, in a similar way to DSA support. The platform's audit trails, confidence scoring, human-in-the-loop review, and exportable dossiers are designed to support the documentation and transparency requirements of the EU AI Act. Your legal and compliance teams must determine how to apply Crowdee's tooling within your specific regulatory programme. Q: What technical and organisational security measures does Crowdee apply? A: Details are published on [our security page](https://crowdee.ai/security). Highlights include end-to-end TLS encryption in transit, encryption at rest, role-based access control, regular penetration testing, and incident response procedures. A summary of measures is included in the DPA. Q: Can I request deletion of my data? A: Yes. You have the right to request erasure of personal data under GDPR Art. 17, subject to applicable retention obligations. Submit deletion requests to privacy@crowdee.ai. We will confirm receipt and complete the erasure within the timeframe required by law. Q: How are international data transfers to third countries safeguarded? A: Transfers to third-party AI model providers established outside the EU/EEA are covered by the EU–US Data Privacy Framework (for US providers) and/or Standard Contractual Clauses (SCCs) adopted by the European Commission. Transfer mechanisms are listed in [our sub-processor register](https://crowdee.ai/subprocessors). Q: Who is the data controller for my personal data? A: In the context of the Crowdee website and free demo, Crowdee GmbH is the data controller. In the context of verification campaigns and API use, you (the customer) are the data controller for content you submit, and Crowdee acts as your data processor under an Art. 28 GDPR DPA. Q: How do I contact Crowdee about privacy matters? A: For privacy-related questions, data subject rights requests, or DPA inquiries, contact privacy@crowdee.ai. For sub-processor information, see [our subprocessors page](https://crowdee.ai/subprocessors). Our full [Privacy Policy](https://crowdee.ai/privacy) is available. Q: How do I report a security vulnerability? A: Please disclose vulnerabilities responsibly to security@crowdee.ai. We ask that you give us reasonable time to investigate and remediate before public disclosure. Details of our responsible disclosure policy are published on [our security page](https://crowdee.ai/security). Q: Does Crowdee issue a signed DPA before data processing begins? A: Yes, always. We will not begin processing personal data on a customer's behalf without a signed Art. 28 GDPR DPA in place. This applies to managed campaigns, API integrations, and enterprise agreements. If your procurement process requires a security questionnaire or additional documentation, contact privacy@crowdee.ai. ### API & Integrations Q: Is there a REST API? A: Yes. Crowdee exposes a versioned REST API (currently at [api.crowdee.ai/v2](https://api.crowdee.ai/v2)) for programmatic access to verification pipelines, workflow management, and result retrieval. The API is type-safe, documented via OpenAPI, and explorable interactively via [the Scalar UI](https://api.crowdee.ai/v2). Q: What can I do with the API? A: Via the API you can submit content for verification, trigger specific verification and Language Technology (LT) pipelines, retrieve results and dossiers, manage projects and workflows, manage datasets (upload files, trigger cleaning, enrichment, and LT pipeline runs, export processed files), and integrate Crowdee verdicts and linguistic outputs directly into your own systems. LT pipeline endpoints cover transcription, language identification, named entity recognition, translation, and OCR — on individual project files or an entire dataset version at once. The exact endpoints and features available depend on your plan and Individual Agreement. Q: Can I manage media datasets via the API? A: Yes. The Datasets API lets you create and manage named multi-modal file collections programmatically. You can upload files, list and retrieve versioned snapshots, trigger cleaning pipelines (e.g. audio silence trimming), kick off per-file enrichment jobs, run Language Technology pipelines on all files in a version (`POST /v2/lt-pipelines/dataset/:id/versions/:vId/run`), and export all processed files via time-limited presigned download URLs. Dataset versions can then be referenced when starting a verification pipeline run. Full endpoint documentation is in [the OpenAPI specification](https://api.crowdee.ai/v2/openapi.json). Q: How do I get API access? A: API access is available under pay-as-you-go and enterprise plans. [Book a demo](https://crowdee.ai/demo) and we will scope the right access level for your use case. After agreeing an Individual Agreement, you will receive an API key to authenticate requests. Q: Are there API rate limits or quotas? A: Yes. Rate limits and quotas are defined in your Individual Agreement. Pay-as-you-go plans have standard rate limits. Enterprise agreements include negotiated quotas and dedicated processing capacity so limits do not constrain production workflows. Q: How is the API authenticated? A: API requests are authenticated using an API key passed via the X-API-Key header or as an Authorization: ApiKey bearer token. Keys are generated per account and can be scoped to specific organisations. Keys should be treated as secrets and never exposed in client-side code. Q: Is there an OpenAPI specification? A: Yes. [The full OpenAPI specification](https://api.crowdee.ai/v2/openapi.json) is available. [An interactive Scalar UI explorer](https://api.crowdee.ai/v2) is available. [An LLM-optimised plain-text summary of the API](https://api.crowdee.ai/v2/llms.txt) is available for AI-assisted development. Q: Can I integrate Crowdee with a newsroom CMS? A: Yes. Newsroom CMS integrations are a common use case for our API and enterprise customers. We have experience with custom integrations into a range of editorial systems. Integration scoping and implementation support are available as part of managed campaign and enterprise agreements. Q: Can I do bulk or batch processing via the API? A: Yes. The API supports bulk uploads and batch processing for large volumes of content. Batch endpoints allow you to submit multiple items in a single request and retrieve aggregated results. Throughput limits for batch processing are defined in your Individual Agreement. Q: What output format does the API return? A: The API returns structured JSON responses. Each response includes the verdict, confidence score, supporting indicators, pipeline metadata, and a reference to the full dossier (available as PDF + JSON). Schema definitions are available in [the OpenAPI specification](https://api.crowdee.ai/v2/openapi.json). Q: Is there a sandbox or test environment? A: A sandbox environment for integration testing is available for enterprise customers. For pay-as-you-go and managed campaigns, integration testing can be arranged during the onboarding phase. Ask about sandbox access when booking your demo. Q: What API SLAs are available? A: Standard email support is available for pay-as-you-go API users. Enterprise agreements include custom SLAs with defined uptime targets, incident response times, and a named technical account manager. SLA terms are negotiated and documented in writing in the enterprise contract. Q: Is there a client SDK? A: The primary interface is the REST API, which can be consumed directly from any language or HTTP client. Official SDK libraries are not currently published, but [the OpenAPI spec](https://api.crowdee.ai/v2/openapi.json) can be used to auto-generate clients in most languages. Contact us if you need integration assistance. Q: Are webhooks available? A: Webhooks for asynchronous result delivery are available for enterprise and managed campaign customers. They allow Crowdee to push verification results directly to your endpoint when processing completes, eliminating the need for polling. Webhook endpoints are configured in your Individual Agreement. Q: How do I get technical support for API issues? A: For API-related questions and technical issues, [open a support ticket](https://crowdee.ai/ticket) or use [our support page](https://crowdee.ai/support). Enterprise customers with a named technical account manager can contact their TAM directly for prioritised assistance. For urgent production issues covered by an SLA, follow the escalation procedure defined in your contract. Q: Where is the full API documentation? A: [The interactive Scalar API explorer](https://api.crowdee.ai/v2) lets you make live API calls directly from your browser. For endpoint references, authentication, request and response schemas, and error codes, refer to [the OpenAPI specification](https://api.crowdee.ai/v2/openapi.json). Q: Does Crowdee have an MCP server? A: Yes. The Crowdee API includes a Model Context Protocol (MCP) server at `POST https://api.crowdee.ai/v2/mcp`. It uses the Streamable HTTP transport and is compatible with any MCP client, including AI coding assistants and agent frameworks. Authentication works identically to the REST API — pass your X-API-Key header with your crw_... key. The server exposes 46 tools and 4 resources covering projects, datasets, Data Platform pipeline runs (cleaning, dedup, and Language Technology), AI Platform finetuning (base models, training datasets, finetune runs, models), verification pipeline runs, crowdsourcing jobs and answers, content gathering, AI output evaluation, and continuous monitoring. Q: What can I do with the Crowdee MCP server? A: The MCP server lets AI assistants and agent frameworks interact with the Crowdee platform programmatically. Available tools include listing and creating projects, browsing and creating datasets, running Data Platform cleaning/dedup/Language Technology pipelines and checking their status, assembling AI Platform training datasets from crowd-labeled data or crowd job answers, starting finetuning runs and listing resulting models, running verification pipelines, checking run status and results, managing crowdsourcing jobs and worker answers, creating content-gathering jobs and reviewing gathered items, requesting AI output evaluations and reading their results, and creating and reading continuous-monitoring schedules — all without writing REST API calls manually. Four MCP resources expose the verification pipeline catalog, the unified Data Platform pipeline catalog (which supersedes the older LT-only catalog), and the AI Platform base-model catalog as structured data. Q: How do I connect an MCP client to Crowdee? A: Configure your MCP client with [the MCP endpoint](https://api.crowdee.ai/v2/mcp) (Streamable HTTP transport) and set the X-API-Key header to your Crowdee API key (crw_...). Most clients accept a headers configuration block for this. Use the X-Organization-Id header to select a specific organisation if your key has access to more than one. A full list of tools and their input schemas is returned by the standard MCP tools/list call. ### For Crowdworkers Q: How do I become a Crowdee crowd worker? A: [Join the waitlist](https://crowdee.ai/crowdworkers). You will be asked to provide basic information about your background, skills, languages, and areas of expertise. Once your application is reviewed and approved, you will be onboarded to the platform and assigned to suitable tasks as they become available. Q: What kinds of tasks will I complete? A: Tasks vary by campaign and include content authenticity review, media annotation, fact-checking support, bias assessment, text classification, image labelling, audio and video review, and quality-checking other workers' outputs. Task interfaces are structured and guided — you do not need prior platform experience to complete them. Q: How much do crowd workers earn? A: Earnings depend on the task type, complexity, and the rates agreed for each campaign. Crowdee invoices its customers at €15 per worker-hour for crowd labour, and worker compensation is set to reflect a fair share of that rate. Specific rates for each task are communicated before you accept it. Q: How and when do I get paid? A: Payment terms are set out in your contractor agreement with Crowdee. Earnings are typically calculated at the end of each campaign or billing period and paid via bank transfer. If you have questions about a specific payment, contact [our worker support team](https://crowdee.ai/support). Q: Do I need any special qualifications? A: Requirements depend on the task. Many tasks require no formal qualifications beyond strong language skills and attention to detail. Some jobs require passing a short, unpaid qualification test first — you get one attempt, and once you pass you stay qualified for that job (and any other job that requires the same test) with no need to retest. If you don't meet the pass conditions, that job isn't available to you, but every other task you're eligible for still is. Q: What happens if I run out of time while completing a task? A: Every paid task has a maximum work time set by the client, and you'll see a live countdown once you start. As the deadline approaches, the timer shows a short grace-period warning; if you still don't submit in time, the task is marked expired and released back into the pool for another worker to pick up. Submitting before time runs out is the only way to have that specific attempt counted. Q: What languages can I work in? A: You can work in any language in which you are proficient and which is supported by active campaigns. When you register, you indicate your languages and proficiency levels. You will be matched with tasks in those languages as campaigns require them. Q: How many hours per week can I work? A: There is no fixed minimum or maximum. You work as many hours as tasks are available and you choose to accept. Availability fluctuates by campaign — some periods will have more tasks than others. The platform is designed to be flexible and to fit around other commitments. Q: Are crowd workers employees or independent contractors? A: Crowd workers are independent contractors, not employees of Crowdee. You are responsible for your own tax obligations in your country of residence. The contractual relationship is governed by the contractor agreement you sign when joining the platform. Q: How does task assignment work? A: Tasks are distributed via the platform based on your profile, language skills, domain qualifications, current performance scores, and task requirements. You will be notified when suitable tasks are available. You can accept or decline tasks before beginning them. Q: What are gold questions and why do they matter to me? A: Gold questions are embedded quality-control items with known correct answers. They appear indistinguishable from regular tasks. Your performance on gold questions contributes to your quality score, which affects your task assignments. Consistent, accurate responses to gold questions help maintain your active status on the platform. Q: How is my work quality assessed? A: Your quality is assessed using your gold question pass rate, inter-rater agreement with other workers on shared items, task completion time, and historical performance trends. These metrics are combined into an overall quality score that is reviewed regularly. We provide feedback where possible to help you improve. Q: Can I be removed from the platform? A: Yes. Workers who consistently underperform, violate the platform's terms of service, provide dishonest responses, or are inactive for an extended period may be suspended or permanently removed. We aim to provide a warning and an opportunity to address issues before taking action where circumstances allow. Q: Is my personal data protected? A: Yes. Your personal data is processed in accordance with Crowdee's Privacy Policy and the GDPR. As a worker, you are a data subject and have the right to access, correct, or request deletion of your data. Contact privacy@crowdee.ai for data subject requests. Q: Who do I contact if I have a problem or question? A: For task-related questions, use the in-platform messaging or help function. For account issues, payment queries, or other matters, contact our worker support team by [opening a ticket](https://crowdee.ai/ticket) or writing to support@crowdee.ai. We aim to respond within one business day. Q: Can I work from any country? A: Crowdee accepts worker applications from many countries, but eligibility may be subject to legal and compliance restrictions in certain jurisdictions. Current eligibility information is listed on [our crowdworkers page](https://crowdee.ai/crowdworkers). If your country is not listed, contact us — eligibility may change as we expand operations. ### Products Q: What are Crowdee's five products? A: [Content Gathering](https://crowdee.ai/content-gathering) (crowd-sourced discovery of posts and media matching your criteria), [Multimedia Verification](https://crowdee.ai/content-verification) (AI + crowd detection of manipulation and synthetic content), [Source Research](https://crowdee.ai/source-research) (tracing content back to its origin with a credibility assessment), [AI Result Evaluation](https://crowdee.ai/ai-output-evaluation) (crowd rating of how transparent and well-explained an AI verdict is), and [Continuous Monitoring](https://crowdee.ai/continuous-monitoring) (a recurring, scheduled version of AI Result Evaluation that flags quality drift automatically). See the [Products hub](https://crowdee.ai/products) for an overview of all five. Q: Do I need to use all five products together? A: No. Each product works standalone, but they're designed to chain together — for example, gathered content can flow directly into verification or source research, and any verification result can be evaluated or continuously monitored without extra setup. Q: How does Content Gathering avoid duplicate results? A: Every submitted item is deduplicated per job using a content hash, so you're never charged twice for the same post or article even if multiple crowd workers find it independently. Q: Is Multimedia Verification the same as the free demo at /content-verification? A: The free demo showcases one pipeline against a single image. The full Multimedia Verification product runs 20+ AI pipelines across image, audio, video, and text, with optional crowd review stages and full audit trails. Q: What does Source Research tell me about a piece of content? A: A source pointer (where it was first published or shared), background on the accounts involved, a category — government, private, or company — and a credibility score with a written explanation. Q: Who rates AI results in AI Result Evaluation? A: A crowd panel of at least three reviewers rates each result on clarity, evidence sufficiency, actionability, and bias risk. Ratings roll up into a single transparency score for that run. Q: What happens with the feedback collected by AI Result Evaluation? A: Every rating is written to an append-only feedback log tied to the specific AI pipeline it evaluated. Crowdee doesn't automatically retrain or adjust prompts from it — it's a durable signal for your team, or ours, to act on. Q: How often can Continuous Monitoring sample my AI outputs? A: Hourly, daily, or weekly, per schedule. Each schedule can be scoped to your whole organisation or a single project. Q: What counts as "drift" in Continuous Monitoring? A: A sampled run's transparency score is compared against that schedule's own trailing average of prior completed runs. A meaningful deviation is flagged — there's no absolute quality bar to configure, since each schedule calibrates against its own history. Q: Are these products available via API? A: Yes. All five products are available through the same REST API as the rest of the platform, documented at [api.crowdee.ai/v2](https://api.crowdee.ai/v2) alongside the core verification and Language Technology pipelines. ### Data Platform Q: What is the Data Platform? A: The [Data Platform](https://crowdee.ai/data-platform) is where the files behind Crowdee's other offerings actually live: named, versioned datasets plus a unified pipeline catalog for cleaning, deduplicating, converting, redacting, splitting, and crowd-labeling them. Datasets created or cleaned here can feed directly into a Verification Platform pipeline run, or seed a Crowd Platform job's input data. Q: How is the Data Platform different from the five products and the Crowd Platform? A: The five products and the [Verification Platform](https://crowdee.ai/platform) deliver a verdict; the [Crowd Platform](https://crowdee.ai/crowd-platform) lets you run your own crowd program. The Data Platform is the layer underneath both — it's where you create, clean, and version the files those other offerings actually run against. Q: What is a dataset? A: A dataset is a named, versioned collection of files — image, audio, video, text, document, or multimodal. Every cleaning, conversion, redaction, or split creates a new derived version; the original raw version is never overwritten, so you always have a full audit trail back to the source files. Q: What pipelines does the Data Platform catalog include? A: 13 pipelines today: audio silence trimming, format conversion, PII redaction, content-hash deduplication, train/validation/test splitting, crowd-assisted labeling, and all 7 Language Technology pipelines (transcription, translation, OCR, language identification, and entity detection). Q: What does deduplication do? A: The dedup-content-hash pipeline computes a SHA-256 content hash for every file in a dataset version and flags exact-duplicate groups for manual review. It's non-destructive — nothing is deleted automatically; you explicitly exclude a flagged duplicate from the version afterward if you want to. Q: What file formats can be converted? A: Format conversion currently supports audio only (mp3, wav, ogg, flac, m4a) via ffmpeg. Image, video, and document conversion are planned but not yet available. Q: How does PII redaction work? A: The pii-redact-text pipeline sends text file contents to an LLM, which detects and redacts personally identifiable information — names, emails, phone numbers, addresses, ID numbers — according to a redaction style you choose (masking, removal, or placeholder tags), and writes the result to a new derived file. Q: What is train/validation/test splitting? A: The split-dataset-version pipeline partitions a dataset version's files into three new sibling versions — train, validation, and test — by random assignment to ratios you configure. Files are reused by reference, never copied, so the split is instant regardless of dataset size. Q: What is crowd-assisted labeling? A: The crowd-label-classify pipeline has a crowd assign a label from a taxonomy you define to every file in a dataset version, with independent majority-vote consensus resolved per file once enough workers have responded to that specific file. Because crowd jobs are project-scoped, running this pipeline requires specifying which project the generated job should live under. Q: Can a dataset feed a crowd job, or a crowd job feed a dataset? A: Yes, in both directions. A dataset version's files can seed a Crowd Platform job's input data, and results from a Crowd Platform content-gathering job can flow back into a new dataset — without a manual re-export step. Q: Is the Data Platform available via API? A: Yes. `GET /v2/data-platform/pipelines/catalog` lists all pipelines, `POST /v2/data-platform/pipelines/:slug/run` triggers one against a project's files or a dataset version, and `GET /v2/data-platform/pipelines/runs` (and `/runs/:runId`) track progress. Q: What does the Data Platform cost? A: Dataset cleaning, format conversion, deduplication, and splitting are free. PII redaction costs 50 credits per file. Crowd-assisted labeling costs 40 credits per worker response (typically 3 responses per file). Language Technology pipelines cost 100-350 credits per file depending on the pipeline. ### AI Platform Q: What is Crowdee's AI Platform? A: It's the fourth Crowdee module: a self-serve way to finetune small, open-source language models on your own labeled data, then evaluate the result with Crowdee's crowd before you trust it in production. Q: Do I need to already be a Data Platform or Crowd Platform customer to use it? A: You need training data from at least one of the two: a Data Platform dataset version, or an existing Crowd Platform job's answers. The AI Platform doesn't generate labels itself — it turns labels and data you already have into a finetuned model. Q: What kind of models does it produce — a full LLM I can license, or something smaller? A: Small, open-source language models (from families like Ministral, Gemma, Qwen, and Granite) rather than frontier-scale models — finetuned specifically on your task and data, not trained from scratch. Q: How does training data get built from a Data Platform dataset? A: If your dataset version has already been through the Data Platform's Crowd-Assisted Labeling pipeline, the AI Platform turns those per-file labels into a ready-to-use classification training set automatically. Q: How does training data get built from a Crowd Platform job? A: Any completed crowdsourcing job's accepted answers can be assembled into a training set — useful for tasks like classification or instruction-following where your crowd has already produced the ground truth you want the model to learn. Q: Can I have the crowd validate my finetuned model's output? A: Yes — this is the recommended step before relying on any finetuned model. Submit a batch of its outputs through the same AI Result Evaluation flow used for any AI system's output, and get back crowd ratings for clarity, evidence, actionability, and bias risk. Q: Does finetuning automatically happen when my data changes, or is it a one-off? A: In this release, every finetuning run is started manually. Automatically retraining a model when new labels arrive, or when Continuous Monitoring flags quality drift, is on the roadmap but not available yet. Q: Will my finetuned model automatically replace the AI in my Verification Pipelines? A: Not yet. Today, a finetuned model is its own standalone asset in your model registry, usable for benchmarking and your own workflows. Wiring a finetuned model in as a drop-in replacement for a pipeline's built-in AI stage is a planned future step. Q: How is a finetuning run priced? A: Each run has a flat credit cost, shown upfront before you start it — the same pricing model used for Data Platform pipelines, rather than metered per-token or per-compute-hour billing. Q: Is my training data used to train models for other organizations? A: No. Every finetuning run, training dataset, and resulting model is scoped to your organization — your data is never used to train or improve another organization's model. Q: Where can I see and manage the models I've finetuned? A: Every completed run produces a versioned entry in your organization's private model registry, where you can review it, re-benchmark it, or start a new run from updated data.