AI Fidelity Checker
Runs entirely in your browser as a lexical heuristic — no AI model involved, no text sent anywhere.
How to use this AI fidelity checker
- Paste the original text on the left.
- Paste the AI's summary or rewrite on the right.
- Click "Check" — names, numbers and dates not found in the original are highlighted for review.
What does this actually check?
This tool extracts "salient" terms from the AI output specifically — numbers, dates, percentages, and capitalized multi-word phrases that look like proper nouns — and checks whether each one appears anywhere in your original text. It's a lexical overlap check, not a semantic fact-checker: it flags terms that seem to have appeared from nowhere, which is a genuinely common signature of AI hallucination, without understanding whether the surrounding sentence is actually true.
Does this prove the AI output is wrong?
No — it flags specific names, numbers, and dates in the AI output that it couldn't find anywhere in your original text, as things worth double-checking. A flagged item might be a genuine error, a reasonable inference, or just phrased differently than in the source — it's a starting point for review, not a verdict.
Can this catch a subtly reworded fact that changes its meaning?
Not reliably — it checks whether specific terms appear in the source, not whether the overall meaning was preserved. A rewritten sentence that shifts meaning while reusing the same key terms won't be flagged.
Why hallucinations concentrate in specific, checkable details
When a language model generates text that isn't grounded in a source it was given, the fabrication doesn't usually appear as an obviously wrong general statement — it tends to show up specifically in concrete, checkable details: a percentage that sounds plausible but was never in the original, a name invented to fill a gap in an otherwise correctly summarized event, a date that's close to but not exactly what the source said. This is exactly why checking those specific categories of detail — numbers, names, dates — is a genuinely efficient way to catch a meaningful share of hallucinations, rather than needing to re-read the entire output word by word against the entire source every time.
Why this is a lexical check, not a semantic one — and why that distinction matters
This tool checks whether specific words and phrases literally appear in your original text, which is a fundamentally different, much simpler question than whether a given sentence's overall meaning is actually true given that source. A summary can reuse every single one of the source's exact facts and figures while still subtly misrepresenting the relationship between them — reversing which of two events caused the other, for instance — and a purely lexical check like this one has no way to catch that kind of meaning-level distortion, since every individual term it's checking would genuinely be found in the original.
Why false positives are expected, and what to do with them
A flagged term isn't automatically wrong — a name spelled slightly differently, a number rounded differently than in the source ("about 30" versus "31"), or a reasonable inference the AI drew from context that wasn't stated explicitly in those exact words will all get flagged despite not necessarily being an actual error. Treat every flagged item as a specific, quick thing worth glancing back at the original text to verify, not as a confirmed list of mistakes — the value of the tool is directing your attention efficiently to the handful of specific claims most worth double-checking, out of what might otherwise be a long, undifferentiated block of AI-generated text.
Why this matters more for longer, denser AI outputs
Manually cross-checking every specific claim in a short, two-sentence AI response against its source is entirely practical to do by eye. That same manual cross-checking becomes genuinely impractical for a long, detail-dense summary of a lengthy report or article, where dozens of individual figures and names are packed into a few paragraphs — exactly the situation where a single fabricated detail is both most likely to slip through unnoticed and most consequential if it does, since a long summary is also the kind of output people are least likely to re-read in full against the original before relying on it.
Limitations of this tool
This tool performs a purely lexical overlap check entirely in your browser — as explained above, it can't verify semantic accuracy (whether the overall meaning is correct), can't catch a subtly reworded claim that reuses the same key terms while shifting meaning, and will produce genuine false positives for paraphrased or reasonably inferred details that were never literally present in the source text. Treat flagged terms as a fast, efficient starting point for manual review, not as a complete or automatic fact-check.