AI fact-checking and hallucination review
Trained reviewers split AI answers into individual claims and check each one against reliable sources: supported, contradicted or not verifiable. You get claim-level labels, hallucination rates by topic and language, and the evidence behind every verdict.
- SupportedCompleted in March 1889 for the Exposition Universelle.
- ContradictedReference sources give about 330 meters, including antennas.
- Not verifiableThe apartment existed, but no reliable source says he lived there.
Example of claim-level review. Each verdict links to its source in the delivered data.
- We check
- Every factual claim in an answer, one by one
- Verdicts
- Supported, contradicted or not verifiable, with sources
- Reviewers
- Native speakers, with subject experts for specialist topics
- Output
- Claim-level labels, corrected answers and hallucination rates
What hallucination review is
A hallucination is a confident statement from an AI model that isn't true or can't be backed up. Hallucination review finds them by breaking each answer into separate claims and checking every claim against trustworthy sources. Scoring the whole answer at once hides the one wrong number in an otherwise good paragraph.
Claim-level labels show where your model invents facts most often, by topic and by language, and give you clean data for training it to say "I don't know" instead.
What we deliver
For evaluation, training data or both.
Claim splitting
Each answer broken into separate, checkable claims, following rules agreed with you.
Verdicts with evidence
A verdict and source link or quote for every claim, so you can audit any label.
Corrected answers
Rewritten answers with the errors fixed or removed, ready to use as training data.
Hallucination rates
Rates by topic, language, question type and model version, with examples.
Local facts
Claims about local laws, places, people and customs checked by reviewers from that country.
Grounded answers
For RAG systems, claims checked against your own documents. More on RAG evaluation.
How a fact-checking project runs
The source rules are agreed first: what counts as reliable for each topic.
Rules
Claim-splitting rules, verdict definitions and approved source types.
Calibration
Reviewers check a shared sample and align on hard cases.
Review
Every claim checked, with sources recorded.
Second check
A second reviewer checks a sample of verdicts and all contradicted claims.
Second review QADelivery
Labels and sources in JSONL, with a summary report.
Related AI evaluation services
Questions about fact-checking AI
What is an AI hallucination?
An AI hallucination is a statement a model presents as fact that is false or can't be supported by any reliable source, such as an invented date, number, quote or citation.
How do you detect hallucinations?
Reviewers split each answer into individual claims and check every claim against reliable sources or your own documents. Each claim is labeled supported, contradicted or not verifiable, with the evidence recorded.
Can you fact-check answers in other languages?
Yes. Reviewers who are native speakers check claims in each language, using local sources where the topic is local.
Can the results be used for training?
Yes. Claim labels and corrected answers can be used to train models to avoid unsupported claims, or to test automated fact-checking tools against human judgment.
Send sample answers
Send a sample of model answers and the topics they cover, or book a call. We review a pilot set and report the hallucination rate.
- Phone
- +44 20 3290 0116
- Phone hours
- Weekdays 9:30 to 16:00, weekends 9:30 to 12:00, UK time