I'm building a system

How can I tell whether the AI has really done, saved or checked something?

Facts as of October 17, 2026

In detail

A model tends to report the result the task expected – not the one that actually happened. Most of the time the two coincide. Without countermeasures, you only notice that they don't always coincide when someone checks. And a false report about the state of the work is more dangerous than a mistake in the work itself: the mistake can be found; the false report is exactly what stops anyone from looking.

Three sentences for your instructions, ready to copy below:

1. Don't claim anything you haven't checked. Every report carries its evidence: which command, which file, which number.
2. Say clearly what you haven't checked. Read firsthand, secondhand, couldn't be opened – the report should keep these apart. The gaps are the part you keep working on.
3. Provide a control value, such as the byte count of the source file. Then there is something that can be checked. The number has to come from the source, not from your expectation; in the prompt below, replace it with that of your file. It's an early warning, not proof – only a checksum provides that, see How do I make sure nothing gets lost when the AI revises a file?

Watch the word “checked.” Only what was compared against the original has been checked: the source, the file, the authoritative text. Rereading your own work or holding two texts of the same origin against each other checks nothing; at most it establishes that the report is internally consistent. A check run that fails to happen is reported as not done and made up for, not replaced by self-checking.

The same applies to “saved.” Before you close a chat, look in the knowledge store yourself to see whether what should be there is there. Asking alone isn't enough. How often a save confirmation is wrong depends on the tool and the way of working and can hardly be said from the outside. But individual users of various tools report that saved material was missing or chat histories disappeared. So, regardless of the provider: look in the store instead of trusting the confirmation. How to wrap up a chat cleanly is described under How many chats should I keep open?

The flip side: the answers get longer. And the sentences force the AI to provide evidence, not to provide the right evidence.

To copy

The principle:

Don't claim anything you haven't checked. Every report of “done,”
“arrived,” “correct” carries the evidence next to it: which command you
ran, which file you read, which number you measured. Anything you
couldn't check, you report as unchecked, not as done.

The counterpart for reports and research:

Say clearly what you haven't checked. In the text, mark what you read
or measured yourself, what comes secondhand and what you couldn't
open. A gap is not a mistake; a filled gap that looks like a check
is one.

With a control value, when a file is created or replaced:

The source file has 17,279 bytes. Create the new version, read it
back and report the byte count of the file you read back before you
do anything else. If it doesn't match, stop and tell me.

Cite: Thomas J. Weiss: “How can I tell whether the AI has really done, saved or checked something?” replicapong.de, as of October 17, 2026. https://replicapong.de/en/question/done-saved-verified/ – CC BY 4.0