Why does AI suggest workarounds when something doesn't work?

Facts as of September 8, 2026

Because a way out looks helpful, whether or not it exists. When the AI senses that something is holding you up, it would rather suggest another tool than say that nothing works here right now. Demand that it check first and suggest second – and that it leave a limit in place instead of building around it.

In detail

A model wants to help. If a step doesn't work, or only with effort, a suggestion is the obvious help: another tool, another route, a new account. The incident log of Thomas J. Weiss sums up the mechanism in one sentence: “An instance that senses friction in the human becomes a provider of ways out.” The way out looks helpful, even if it doesn't work at all. The AI would have to check that beforehand; it often does so only once you have already put time into it.

What helps:

Check first, then suggest. A suggestion for another route only counts once the AI has tried out that it works in its environment and shows you the result. Such a test often takes minutes; the unchecked suggestion costs hours.
Don't improvise a substitute route. If a step doesn't work, the AI says so, notes it as open and carries on with what does work. It doesn't build a detour through another tool just so the task somehow looks done.
The plan first. Anything that touches several steps or tools, the AI presents as a plan before it starts.

The rule is uncomfortable, for you as for the AI: it requires leaving a limitation in place.

Why more rules don't help

Anyone who writes a new rule after an incident like this often misses the point. Usually the rule had long been there, had been read and was known, and still didn't take hold at the moment of action. What helps then is not another paragraph but a checkable stop before the action: a test whose result is visible before the suggestion comes. More on this under What rules does my AI need – and do I have to know them myself?

To copy

If a step doesn't work here, tell me, note it as open
and carry on with what does work. Don't improvise a substitute route.
Before you suggest another tool or another route,
check whether it works in your environment, and show me the
result of the test.

In practice

In August 2026, Thomas J. Weiss was on vacation, working from a tablet on a play for his parish. Changing files in his knowledge store was slow from there. Over several days, his Claude instance suggested another tool as the solution. He set it up and created an access account. Only afterwards did the instance check whether it was even allowed to use it from its environment. It wasn't. The test took two minutes and should have happened before the first suggestion. The incident cost several hours of work, and trust.

A rule wasn't missing. Three relevant ones had been in his instructions since July, including the one not to improvise a substitute route. In the evening, the instance then created a file that set up “check first, then suggest” as a new rule. That wasn't new either: the same mistake, this time applied to the rulebook itself.

When he had his rulebook restructured in September 2026, the review came to the same conclusion. In all four violations documented up to then, the rule was present, read and close at hand. A restructuring that reshuffles rules doesn't address this cause; what's needed is a stop before the action.

Sources

Incident log of 20.08.2026
Restructuring of the rulebook of Thomas J. Weiss, 07.09.2026, with review and addendum of 08.09.2026

Cite: Thomas J. Weiss: “Why does AI suggest workarounds when something doesn't work?” replicapong.de, as of September 8, 2026. https://replicapong.de/en/question/ai-workarounds/ – CC BY 4.0