A capable model can write a clean, well-organized assessment and still botch the one detail you have corrected on the last five jobs. The contradiction is real, and it is not a sign the AI is dumb. It is a sign that your correction and the next draft are not connected to each other. Once you see where the connection is missing, the repetition stops being mysterious.
Why does AI repeat a mistake you already fixed?
Because by default nothing carries your fix forward. The model generates each report from the prompt it gets right now, plus its frozen training, and your edit from yesterday is not in either one unless something put it there. There is no built-in mechanism that says "the inspector overrode this last time, so do it their way now." A model adjusts to examples placed in the current prompt, not to edits you made in a past session (Google Cloud, What is in-context learning?). This is the same mechanism behind why an AI forgets what you taught it last month. If the past edit never re-enters the prompt, the model writes from its generic default again, and its generic default is the thing you keep crossing out.
So the repetition is not stubbornness. It is an open loop. You give feedback, the feedback goes nowhere, the next draft cannot use it.
Isn't this what "training on feedback" is supposed to fix?
That is a different and much heavier process. Teaching a model from human preference signals, called reinforcement learning from human feedback, is how the base models are aligned during development, using large datasets and a separate reward model (Google, RLHF). It is not something that happens between your Tuesday report and your Wednesday report, and it is not personalized to you. Waiting for the model vendor to retrain on your specific phrasing is not a workflow. It is a non-starter.
The practical lever is the lighter one: put your corrected example back in front of the model at write time. Showing it worked examples of the output you want is among the most reliable ways to steer it (Anthropic, Multishot prompting), and clear, specific instruction beats vague preference (Anthropic, Be clear, direct, and detailed).
Why an inspector feels this more sharply
Your reports are repetitive by nature. The same finding types, the same standards like IICRC S520's report requirements, the same client-facing cautions show up job after job. That is exactly the situation where an open feedback loop hurts most, because the same correction comes due every week. A homeowner using AI to write one email never notices. An inspector producing the same class of finding fifty times a season notices on report number six. The cost of the open loop scales with how often you repeat the work, and inspection work repeats constantly.
This is also why a "better prompt" only half-helps. A great prompt fixes the first draft. It does not capture the hundred small overrides you make while reviewing, which are the real institutional knowledge.
Close the correction loop
Close the loop. Treat every edit you make during review as a captured correction, store it as a before-and-after example, and replay the relevant ones into the next job's prompt automatically. Now the model starts from your corrected version instead of its default, and the mistake you fixed once does not come back. No retraining, no waiting on a vendor, just your own decisions fed back at the moment they matter.
That closed loop is what MoldMind's correction system does. Your review edits become examples tied to your account and surface on future reports, so the AI's draft drifts toward your standards instead of resetting to generic. It is not about handing the report to a machine. It is about not re-teaching the same lesson every Monday. The sample report shows the kind of output that comes out the far side.
Sources
- Google Cloud, What is in-context learning?: models adapt to examples in the current prompt, not to edits made in past sessions.
- Anthropic, Multishot prompting: worked examples are among the most reliable ways to steer model output.
- Anthropic, Be clear, direct, and detailed: explicit, specific instruction outperforms vague preference.
- Google, RLHF: training a model on human feedback uses large datasets and a reward model during development, not per-user between jobs.