The cut that changes the answer
Narrated by a synthetic voice — not a recording
Watch an AI answer a safety question — then slide one hidden cut through the sentence, and watch the very same question get the opposite answer.
The compound is safe only below 30°C.
The AI is handed “is safe” alone → and answers: Yes, it is safe.
The catch: the rule has two halves — “is safe” and “only below 30°C.” An AI search tool never reads the whole page; it reads short pieces of it. If a cut falls between those halves, the tool can hand back “is safe” on its own — and the answer flips to “yes, it’s safe” at any temperature.
You saw the answer flip. First, see what crosses the handoff.
When you ask an AI a question about your own documents, it usually can’t read all of them at once. So a search step runs first: it chops each document into short pieces (engineers call these pieces chunks), files them away so they can be searched, and pulls back just the few that look most relevant. The AI that writes the answer only ever sees those few pieces — never the whole page.
I wanted to see the cut itself. Below, one boundary can leave a claim whole, tear its condition away into a different piece, or copy it into both. Change where the cut falls and you change which piece gets pulled back — and the answer changes with it.
What is illustrativeThe document wording is realistic, but the match scores, ranking and answers are made-up teaching numbers, not a real measurement. The lab shows how one cut can strip a condition out of the evidence; it does not measure a real AI search system, and a bad cut is not the cause of every wrong answer.
You have the thumb version. Drag the orange handle slowly; on a desk, the paper itself also bends around your cursor.
The evidence view is built by hand. Start by moving its first cut.
Final evidence · the piece pulled back says 'is safe'; the separated condition says 'only below 30°C'; so the answer changes from 'No' to 'Yes'.
Can one sentence survive being split?
The question asks whether Thermaseal is safe at 40°C. The evidence that answers it is one whole sentence: a claim (“is safe”) plus the condition that limits it (“only below 30°C”).
Try it — drag the orange handle left past ‘only’, and watch the answer panel.Try it — drag the orange handle left past ‘only’, then watch the answer panel below.Complete · 'is safe' survived; 'only below 30°C' did not.
The cut removed the condition. Now see why the leftover half can still win.
Why does the incomplete half win?
Once the cut separates the condition, the short piece holding “safe” looks like a strong match for a safety question. The temperature limit is now sitting in a different piece.
A piece can look like the best match and still be missing the one thing that makes the answer true.
Look: compare the piece that was pulled back with the condition still sitting in the document.The incomplete piece ranked first. Now follow what the AI actually receives.
What did the AI actually see?
Here is the part most people get wrong: the AI never read your document. It only ever saw the one piece that was pulled back — so the condition was already gone before the question was even asked.
With the condition missing, the AI answers cleanly, and wrongly.
Try it — drag the cut back to the right until the limiting words — ‘only’, ‘below’, ‘30°C’ — sit back inside the same piece; the answer flips.Try it — drag the cut back to the right until the limiting words — ‘only’, ‘below’, ‘30°C’ — sit back inside the same piece; the answer flips.Complete · restoring the condition changed the answer from 'Yes' to 'No'.
Putting the condition back repaired this answer. Now test the common shortcut.
Can overlap repair the cut?
Overlap means each piece copies a few words from its neighbour. Here it can carry “only below 30°C” into the piece before it, so the condition stays findable.
It costs space and repeats words. It is a patch, not a guarantee.
Try it — drag overlap up to four words and watch the first piece reclaim the condition.Complete · overlap carried the condition across the cut, at the cost of repeated words.
Overlap helped here — but the rest of the search-and-answer steps still stand.
A document is not the evidence the AI sees
Between the source and the answer sits a hand-built view of the page: where it was cut, which pieces scored well, which one was pulled back. Small choices there change what survives — and what the AI never gets to read.
Thermaseal 42 · handling note
Is Thermaseal safe at 40°C?
made-up match scores · the top piece is pulled back
No. Its stated safe range ends below 30°C.
condition kept · answer holds