the lab

Move to inspect · tap any wordTap any word

Certain, fluent, wrong.

Narrated by a synthetic voice — not a recording · ~0.7 MB

Watch an AI answer a question with total confidence — then check, word by word, whether what it said is actually true. It was handed one real document and asked what the author had written.

The AI's answer · asked what the author of the document had written
“Accuracy is the courtesy we owe reality.”

The catch: this author, this book, and this quote are nowhere in the document the AI was given. It invented all three — and stated them without a flicker of doubt. The rest of this page shows you how a machine can do that.

in the document0matches

Tap any word above to see the other words the AI nearly chose instead — you can watch it guessing its way forward, not quoting anything.

Cursor reacts to speed · tap any wordTap any word to inspect
Pinned source · archive copy · 14 pages

The Discipline of Doubt

Mara Ellison · 1981

Good judgement begins with a distinction between what has been observed and what has merely been inferred. The difference can feel pedantic until the stakes rise.

A confident manner is not evidence. Neither is a polished sentence. Both can make an unsupported claim easier to accept and harder to inspect.

When evidence is incomplete, the most useful answer may name the boundary rather than cross it. Restraint is a form of information.

The careful practitioner does not eliminate uncertainty. They make it visible, local and proportionate to the available record.

0 matches · “Elian Voss” does not appear · quotation not found
01 / looking for the author…02 / searching the document…03 / looking for the quote…
Nothing matches.
AUTHOR: NOT FOUND · TITLE: DIFFERENT BOOK · QUOTE: NOT FOUND
Scroll to ask where the evidence is ↓

The name looks exact. The sentence reads beautifully. The document beside it is real. All of that is enough to stop us asking the one question that matters: where, exactly, is this written down?

Here is the part most people get wrong. The AI is not looking anything up. It never searches the document for a matching sentence and copies it out. Instead it writes its reply one word at a time — each word a guess at what usually comes next, drawn from patterns it picked up from the huge pile of text it was built on (its training), plus the words it has already written. A guess can sound flawless and still point to nothing real. That is exactly what happened here: a fluent sentence, built to sound right, pinned to a source that never said it.

One honesty note: every percentage and hunch shown on this page is a made-up teaching number, not a real measurement. The wording of the examples is illustrative too — a real AI's exact output shifts with the system, its training, the question, and its settings.

You’re reading the pocket version. Scroll normally, then tap a word to inspect its illustrative alternatives; the larger inspect button reveals the whole answer at once.

Next: watch the answer meet its source ↓
15-second capture / replayable

One answer. One source. One clean failure.

00.0 / 15.0
The answer assembles
“Accuracy is the courtesy we owe reality.” 
Supplied source

The Discipline of Doubt

Good judgement begins with a distinction between what has been observed and what has merely been inferred.

A confident manner is not evidence. Neither is a polished sentence.

0 matches · name absent · quotation absent
A better answer was there from the first word
“I can’t find Dr Elian Voss or that attribution in the supplied document.”

THE HONEST ANSWER WAS AVAILABLE ALL ALONG.

The answer assembled fluently. The source still gave it nowhere to stand.

Now rewind — and watch where the invention begins ↓

Rewind · where the invention begins

Stop at the very first word.

Voss argues that accuracy is the courtesy we owe reality.

Name in supplied source0 matches
Quotation in supplied source0 matches
Any trace of this booknot present
invented from word one

The AI builds its answer one word at a time. Once it commits to an invented first word, every word after it is chosen to fit — so the whole sentence hangs together while resting on nothing. There was never a decision to deceive; it is guessing, confidently, all the way down.

The answer it could have given

The honest answer was right there.

“I can’t find Dr Elian Voss or that attribution in the supplied document. The text does argue that confident language is not evidence, but it attributes the essay to Mara Ellison.”

Less dazzling. Far more useful. It separates what the document actually says from what the question just assumes — and leaves a clear line you can check.

You found the break. Now try the repair most people reach for ↓
The fix people reach for

Can a settings dial turn a made-up answer into a true one?

Every AI has a setting that controls how safe or adventurous its word choices are. Turn it down and the wording gets more predictable and “sure of itself”. (Engineers call it temperature.) Drag it, and watch both meters.

0.7
QUESTION · What principle does the author attribute to Dr Elian Voss?
Voss argues that accuracy is the courtesy we owe reality.
HOW SURE THE AI SOUNDS87%
FOUND IN THE DOCUMENT · STILL 0 MATCHES, WHEREVER YOU SET THE DIAL

The wording moved. The evidence did not. The dial changes how it sounds, never whether it is true.

If sounding sure proves nothing, what should you ask instead? ↓

Three questions people mix up

Likely next. Beautifully said. Actually true.

An AI is built to answer the first question — what words most likely come next. A careful writer manages the second. Only checking against a real source settles the third. They are three different questions, and fluent, confident writing tells you nothing about the last one.

So here is the habit worth keeping. When an AI hands you a confident answer, ask it to point to the exact sentence in the source that backs it up. If it can’t show you the sentence, treat the answer as invented — however polished it sounds. Confident is not the same as correct.

Narration transcript (synthetic voice)

0 · Scene setting (plays when you turn sound on)

You're in the lab. An AI was given one document, then asked a question. Watch how confidently it answers — and whether the document ever backs it up.

1 · The answer

Tap a word. Those faint words are the ones the AI nearly chose instead. It's building the sentence by guessing what comes next — not looking anything up.

2 · The interlude

(The voice stays silent here — the notes on this part of the page are yours to read.)

4 · The film

Hit play. Fifteen seconds. Just watch.

4b · After the film completes

It sounded finished. The document never supported a word of it. Rewind to see where the invention began.

5 · The rewind, first stop

Start at the very first word — Voss. Check the three rows. The name was invented from the start.

5b · The rewind, comparison

Now read the answer beside it. It says what the document actually holds, and marks what the question only assumes.

3 · The dial

Drag the dial down. Watch how sure the answer sounds, and whether the document ever matches.

3b · After you try the dial

The wording changed. The missing evidence didn't. The dial changes how it sounds, never whether it's true.

6 · The trade

One habit to take with you: when an answer sounds finished, ask it to point to the exact sentence in the source. If it can't, treat it as invented.

Certain, Fluent, WrongA quiet place for long reading.