Free AI Hallucination Detector Online

What Is an AI Hallucination Detector?

An AI hallucination detector is a tool that checks text written by a large language model for fabricated or unsupported details. NIST AI 600-1 calls this risk confabulation, where GAI systems "generate and confidently present erroneous or false content in response to prompts". This detector scores that risk claim by claim, in your browser.

NIST adds that confabulations include outputs that "contradict previously generated statements in the same context", so the detector checks an answer against your source and against itself.

The OWASP LLM09:2025 Misinformation entry states that "Misinformation occurs when LLMs produce false or misleading information that appears credible". It lists factual inaccuracies and unsupported claims as examples and recommends retrieval-augmented generation and cross-verification.

TestMu AI maintains this detector alongside the Agent Red-Team Scan, which probes an LLM API for prompt injection, and the AI Agent Risk Scorer.

How Does This Hallucination Detector Work?

The detector runs five deterministic stages in your browser, then an optional online stage, with no model call:

• Sentence split: Text is split at sentence punctuation and line breaks, bullets stay separate, and fragments under 3 words are dropped.
• Claim detection: A sentence is a claim when it holds a number, quote, multi-word name, identifier, citation, or absolute word; other sentences are not scored.
• Grounding (with a source): Coverage is the share of a claim's content words found in the source: below 0.35 is Unsupported, below 0.6 is Partial. Every number, quoted phrase, and multi-word name must also appear in the source.
• Red-flag checks: The rules below run on every answer, with or without a source.
• Scoring: Flags weigh 15 (Critical), 8 (High), 4 (Medium), 1 (Low). With a source, grounding contributes up to 60 points and flags up to 40. An unregistered DOI or arXiv ID, or a real person placed before birth or after death, floors the score at 50.
• Online verification (optional): With Verify identifiers online checked, DOIs are looked up at doi.org and the Crossref REST API, journal titles in the Crossref registry, ISBNs in the Open Library API, arXiv IDs through their DataCite DOIs, and people and organizations in Wikidata, whose birth, death, or founding year is compared with the nearest year in the text.

• ISBN check digit (Critical): Per the International ISBN Agency, ISBNs have had 13 digits since 1 January 2007, start with 978 or 979, and use a modulus 10 check digit with alternate weights of 1 and 3.
• arXiv identifier (Critical): The arXiv identifier scheme says "From 0704 through 1412 it is 4-digits" and "From 1501 on it is 5-digits"; impossible months and future dates are flagged too.
• DOI syntax (Medium): A DOI is "a prefix and a suffix separated by a forward slash", with syntax defined by ISO 26324. Well-formed DOIs are marked unverified (Low) unless online verification looks them up.
• Dates (Critical): Impossible dates such as February 30 and weekdays that do not match their date.
• Count and year (Critical): A list that differs from its announced count, or one event with two different years.
• URLs (High or Low): Placeholder hosts such as example.com are High; other URLs are unverified.
• Wording (High to Low): Unsourced statistics, vague authority such as studies show, absolute words, and author-year citations.

All text analysis happens in your browser and the answer text is never uploaded. With online verification on, only identifiers and capitalized names leave the page, as individual lookups to public registries.

How Do You Use the AI Hallucination Detector?

• Paste the answer: Paste the AI answer, or use the upload icon for a .txt, .md, .json, or .csv file. The sample icon loads a worked example.
• Paste the source (optional): Add the document or retrieved context the answer should match.
• Run the check: Leave Verify identifiers online checked to look up identifiers and names, then press Detect Hallucinations.
• Read the results: Check the score, then the claims table for Unsupported rows and the red flags for Critical items.
• Save the report: Copy the Markdown report, or download hallucination-report.json.

Try this: load the sample and press Detect Hallucinations. The input "Tuesday, March 3, 2025" returns the Critical flag Weekday mismatch: March 3, 2025 was a Monday.

What Can a Hallucination Detector Verify and What Can It Not?

Deterministic checks verify internal consistency and identifier validity. Grounding verifies agreement with the text you supply, not with the world. Registry lookups confirm that an identifier exists and what it points to; nothing here confirms that a free-text claim is true.

Where Does Hallucination Detection Fit in LLM Testing?

• RAG faithfulness regression: Paste the retrieved context as the source and rerun after each retriever change.
• Chatbot answer review: Check bot answers against the help article they should quote before release.
• Prompt comparison: Paste two answers to the same prompt in turn and compare scores.
• Content and documentation review: Catch invented ISBNs and placeholder links in AI-drafted docs, sized first with the Token Counter.
• Agent transcript audit: Paste an agent's final answer to find fabricated identifiers.

To test agents across many conversations, use the AI Agent Testing Platform; TestMu AI also runs tests across 10,000+ real devices and 3000+ browsers. The AI hallucination guide and LLM testing guide go deeper.

Frequently Asked Questions (FAQs)

What is the difference between a hallucination and a confabulation?

Hallucination and confabulation describe the same failure: a language model confidently stating false or unsupported content. NIST AI 600-1 uses the term confabulation and notes these errors are colloquially called hallucinations or fabrications. NIST adds that confabulations include outputs that diverge from the prompt or contradict earlier statements in the same context.

Can a hallucination detector tell whether a claim is true?

A hallucination detector cannot confirm that a claim is true in the world. This detector checks agreement with the source text you paste and the validity of dates and identifiers. Registry lookups and Wikidata timeline checks catch invented citations and impossible dates for real people, but no check here judges an arbitrary free-text claim.

What happens if I do not paste a source text?

Without a source text, the hallucination detector marks every claim as Not grounded and the score reflects risk signals such as invalid dates, broken ISBN check digits, and unsourced statistics. Online verification still runs without a source, looking up DOIs, ISBNs, arXiv IDs, journals, and names. Paste the reference document to add claim-by-claim grounding.

What does the detector catch in an answer with no source?

Without a source, the detector catches invalid identifiers and dates, lists and years that contradict themselves, and unsourced statistics. With online verification on, it also flags DOIs, ISBNs, arXiv IDs, journals, and people or organizations that do not exist or whose dates make the story impossible. It cannot catch a plausible claim with no checkable detail.

Why is a URL only marked unverified while a DOI gets checked?

A URL cannot be fetched from the browser because most sites block cross-origin requests, so the hallucination detector only rejects placeholder hosts such as example.com. DOIs, ISBNs, and arXiv IDs have public registries that allow browser access, so with online verification on they are looked up and a missing record is flagged as invented.

What percentage of AI-generated citations are fake?

The share of fake AI-generated citations has no single answer, because published estimates vary by study, model, subject, and prompt. This hallucination detector does not estimate a rate. It flags each author-year citation, DOI, arXiv ID, and ISBN in your answer so you can check every reference individually in a library database.

Does retrieval-augmented generation eliminate hallucinations?

Retrieval-augmented generation does not eliminate hallucinations. OWASP recommends RAG as a mitigation for LLM09:2025 Misinformation, not a cure, because a model can still ignore, misread, or embellish retrieved passages. Pasting the retrieved context as the source in this detector measures exactly that gap by flagging claims whose numbers, names, or wording are missing from the context.

Is my text stored or sent anywhere?

The pasted text is never stored or uploaded. With Verify identifiers online checked, only DOIs, ISBNs, arXiv IDs, and capitalized names are sent as individual lookups to doi.org, Crossref, Open Library, and Wikidata. Untick the box to keep everything local. Nothing is written to cookies or local storage, and a refresh discards your results.

How large an answer can I check at once?

The hallucination detector accepts an answer of up to 20,000 characters and a source text of up to 50,000 characters per run. An uploaded .txt, .md, .json, or .csv file follows the same 20,000 character limit for the answer box. Split longer transcripts or documents into sections and check each section separately.

Which languages does the detector support?

The hallucination detector is tuned for English: authority phrases such as studies show, absolute words, list counts, and weekday and month names are matched in English only. ISBN check digits, DOI syntax, arXiv identifiers, ISO dates, URLs, and number grounding are language-independent, so those checks still work on answers written in other languages.

How is the risk score calculated?

The risk score adds a weight per red flag: 15 Critical, 8 High, 4 Medium, 1 Low, capped at 100 without a source. With a source, unsupported claims plus half the partial claims add up to 60 points and flags up to 40. An unregistered DOI or arXiv ID floors the score at 50, High risk.