AI Hallucination Detector
An AI hallucination detector scans generated text for the phrases a model uses when it is guessing: hedges, attributions to unnamed sources and admissions that it cannot confirm. It surfaces the sentences most worth checking. It cannot catch a confident fabrication, and no text-only tool can.
Last reviewed by the Radiatus Cloud team
Checks for linguistic uncertainty markers and fabrication patterns.
Estimated Risk Level
Detailed Analysis
Securing AI in production?
We build guardrails, governance & compliance for AI systems.
What is being measured
The detector matches 24 uncertainty markers, from I believe and perhaps through reportedly, allegedly and cannot confirm. Each match is highlighted in the text. The score is the density of markers per word, scaled so that one marker every 20 words reads as 100 percent. Above 40 is reported as a high probability of guessing, 15 to 40 as moderate, below 15 as low.
The limitation to understand before trusting a low score
Language models hallucinate fluently. The fabricated court citations in Mata v. Avianca in 2023 were written with complete confidence, and a hedging-word scan would have scored them near zero. A low score here means the model did not signal doubt, not that the content is true. Treat the tool as a way to prioritise which sentences to verify, and verify the confident ones separately.
What to check by hand regardless of score
- Citations: paper titles, authors, DOIs, case names, URLs. Fabricated references are the most common serious hallucination.
- Numbers and dates: statistics, percentages, version numbers, release years.
- Direct quotes attributed to named people.
- Anything about events after the model's training cutoff.
Reducing hallucination at the source
Retrieval grounding, where the model is given the source documents and told to answer only from them, cuts fabrication more than any prompt wording. Asking for sources does not; models will invent plausible ones. Lower temperature reduces variety but not confidence. For factual tasks, ask the model to say when it does not know, then check whether it actually does.
Worked example
A 200-word product summary containing it is possible, typically, reportedly and as far as I know scores 10 percent: low. The same summary states a launch date and a price with no hedging at all. Those two facts are the ones to verify, and the score says nothing about them.
Related tools
- Model Hallucination Estimator — Estimate risk of hallucinations based on task type and temperature.
- AI Citation Readiness Checker — Check if your content is structured to be cited by AI answer engines. Scores TL;DR, headings, lists, sources, and clarity.
- LLM Output Consistency Checker — Check LLM output consistency for reliability testing.
- AI Content Detector — Analyze text for patterns commonly associated with AI-generated content. Check perplexity, burstiness, and repetition scores.
Frequently Asked Questions
Can this tool tell me whether a statement is false?
No. It has no knowledge base and does not look anything up. It finds the language patterns models use when they are uncertain, which is useful for triage and useless for a confident fabrication.
Does the text I paste leave my browser?
No. The pattern matching runs locally in JavaScript. You can paste confidential drafts.
Why did a careful, well-hedged human text score high?
Because the detector measures hedging, and careful writers hedge. A high score on human text means the author was appropriately uncertain, not that they were hallucinating. The tool is meant for model output.
Which kinds of content hallucinate most?
References and citations, biographical details about lesser-known people, exact statistics, and anything requiring arithmetic across several steps. Summaries of a document you supplied are usually reliable; claims about the wider world are not.
Is there a better automated approach?
For claims with a source, retrieval-based checking that compares each sentence against the supplied documents. For open-world claims, sampling the model several times and checking consistency catches some fabrications, since invented details tend to vary between runs.
Privacy & Security
Text is scanned locally.
About This Tool
This tool runs entirely in your browser. No data is sent to any server, ensuring complete privacy. Simply use the interface above to get started — no registration or login required.
Disclaimer: This tool is provided "as is" without warranty of any kind. Results are for educational and utility purposes.
Related Tools
Model Hallucination Estimator
AI SecurityEstimate risk of hallucinations based on task type and temperature.
AI Citation Readiness Checker
AI SecurityCheck if your content is structured to be cited by AI answer engines. Scores TL;DR, headings, lists, sources, and clarity.
LLM Output Consistency Checker
AI SecurityCheck LLM output consistency for reliability testing.