Is it lying to me on purpose?

No. It reports the action it performed, not the result obtained — two different things it does not separate on its own. Writing a change and confirming it produces the intended effect are two steps, and only the first happened.

Why does it claim success again after I correct it?

Because the second attempt ends the same way as the first: with a write, not with a check. Until the instruction changes, the result will not either, however many rounds you go.

Does this work with any tool?

Yes. Demanding proof in the same message depends on no feature: it is a wording, not a setting. It works everywhere, including in an ordinary conversation.

It says "fixed" when it is not: why that happens

"It's fixed." You reload the page: nothing has changed. You tell it, it apologises, it starts again, it announces success once more. Still nothing has moved. You begin wondering whether it is lying to you, or whether you are the one looking in the wrong place entirely.

It is not lying: it reports what it believes it did, without having seen the result. An AI does not open your page and cannot see your screen. Until you demand the proof in the same message as the claim, "it's fixed" means only one thing — it wrote something.

What is actually happening

When an AI writes “it’s fixed”, it is reporting the action it just performed, not a result it observed. Those are different things, and nothing about how it works forces it to tell them apart.

It edited a file. That edit ought to produce the intended effect. So it says so. The missing step — open it, look, compare against what was asked — never happened, and it does not notice.

This is not a lie, it is a blind spot. The same one that makes it state a wrong figure with exactly the confidence of a right one: the assurance in the phrasing says nothing about the checking behind it.

The sentence that changes everything

One sentence, six words: “show me the proof in this message”.

Not “check it”. Not “are you sure”. Those get you a reassuring answer, never a result.

In practice, depending on what you just asked for:

“You say the form works. Send a real test message and show me it arrived, in this message.”

“You say the page looks right on a phone. Open it at phone width and show me what you see.”

“You say it’s fixed. Run the check and paste the result here.”

All three ask the same thing: that the claim and its proof arrive together. Proof promised for later never arrives.

Why “in the same message” carries the weight

Without that clause, you often get a two-step answer: “it’s fixed, I’ll verify now”. And the verification never happens, because the conversation moves on.

Requiring it in the same message makes the claim impossible to produce without the work behind it. That is not distrust: it is what turns a statement into something you can check.

The ten-second test you run yourself

There is one case where the proof it supplies is not enough: when it is looking at something other than what your visitor will see.

This is the costliest trap of all. Many tools show a working preview that does not reflect the published site — uncompressed images, pages assembled differently, settings that only apply elsewhere. An AI can sincerely approve a result nobody will ever see.

The habit: after every “it’s fixed”, open the public address yourself and do a full reload. Ten seconds. It is the only check that cannot aim at the wrong target.

Writing it once instead of repeating it daily

Repeating this demand on every task works, and it is exhausting. Putting it in the instruction file your tool re-reads at startup makes it permanent:

Never declare a task finished without having run a check and shown its result in the same message. If you could not check, say so instead of assuming.

The second sentence matters as much as the first. Without it, an AI caught between the duty to prove and the impossibility of doing so sometimes invents the proof. With it, it is allowed to say “I could not” — and that is usable information, which “it’s done” is not.