What typically goes wrong when working with AI?

Facts as of October 11, 2026

Rarely the spectacular, mostly the quiet: the AI smooths out what should stay as it is, adopts an error without checking, or fills a gap in the task with whatever seems plausible to it. Add to that tasks that only existed in the conversation and get lost there. The same thing helps against almost all of it: write it down, measure it, have it cross-checked.

In detail

Silent smoothing

A model improves whatever seems odd to it, and it is usually right. The case in which it isn't right looks exactly like the others. Remedy: have it flag, not change; in verbatim quotes, even the typo stays.

Silently adopting an error

A model sees the messages of a conversation side by side, not the time in between. What it believes it knows about the world is often just whatever was last in the context – for example, a date set at the beginning that doesn't move past midnight. Remedy: measure instead of assuming.

Reading in what you didn't say

If a task is brief, the model fills in what seems plausible to it, and the result looks as if you had wanted it that way. Remedy: say what you don't know. Naming the limit of your own knowledge takes away the gap the model would otherwise fill itself.

Tasks that only exist in the conversation

When a long chat is summarized, the purpose of a task can drop out while its name survives. The model keeps working and doesn't notice the gap. Remedy: what isn't in the knowledge store doesn't exist – write after every step.

Other patterns

Stacked questions, even though one at a time was requested, see How do I avoid spending hours just approving AI steps?
Recommendations that ignore your criterion. You give a reason for your choice, and the model still recommends the more convenient route.
Finished prose before the decision. If the model words things before a decision is made, it has to retract. Better: first the factual check, then two versions per paragraph instead of one recommendation.
Errors in the check itself, when long strings of characters are copied, see How do I make sure nothing gets lost when the AI revises a file?

And underneath it all: a cross-check by a second AI without the first one's conversation history finds what the first one overlooks. The remedies cost time – every list, every OK, every intermediate step. Whoever doesn't spend it keeps the errors.

In practice

The list comes from the chronicle a model kept for Thomas J. Weiss while he was writing a play, and from the incidents of the months that followed. In July 2026, a brief answer from him was taken as permission to skip several steps of a workflow; he had to reject it twice. In September 2026, the model named a file with the previous day's date because it had taken the date from the session context; it was found because a rule requires time references to be measured.

What's remarkable is who finds what. On August 26, 2026, the chronicle recorded that the automated checking chain found more errors than he did, but only transfer errors; everything he found concerned the collaboration itself. Of the six errors in the collaboration that the chronicle lists, the checking chain found two.

Sources

Note „Sechs Dinge, die nicht funktioniert haben“, replicapong.de, 11.10.2026
Note „Gefahren und Gegenmittel“, replicapong.de, 10.10.2026

Cite: Thomas J. Weiss: “What typically goes wrong when working with AI?” replicapong.de, as of October 11, 2026. https://replicapong.de/en/question/what-goes-wrong/ – CC BY 4.0