Feeling like your AI is getting "dumber" the more you use it? That feeling is extremely common. But don't jump to conclusions: nine times out of ten the model hasn't actually degraded — three variables in your usage environment are to blame. Work through A, B, and C in order, and you'll usually find the real culprit.
First check A: are you still talking to the same model you started with?
Many platforms enable "auto-select" by default and quietly route you to a cheaper, weaker model during peak hours; an app update may reset your default model; and once your free quota runs out, some services silently downgrade you without an obvious notice. You think you're still chatting with the flagship model, but someone else is actually answering.
The check is simple: look at the model name in the chat interface and confirm it's what you expect, verify the default model in settings, then re-run a familiar "benchmark question" on the current model and compare answer quality with what you remember.
Then check B: has the context been contaminated?
In a long conversation, early wrong answers get treated as established facts and cited again; stale preferences in memory or custom instructions skew new answers in ways you never notice; and over many turns the model's attention dilutes, so its focus on the current question naturally drops. All of this makes you feel "it didn't used to be like this."
The check: open a fresh, clean conversation and ask the same question word for word; inspect memory and custom instructions for outdated content, and turn memory off and re-test if needed.
Then check C: sampling randomness — plus your own prompting habits
Model outputs are inherently random: ask the same question twice and quality naturally varies, and you tend to remember only the best answer. Here's the uncomfortable truth: the longer you use it, the terser your prompts get — less background, higher expectations. Sometimes the answer got worse simply because the question got "lazier."
The check: take the question that produced that "smart answer" back then and ask it again, word for word, and see if the gap is still there.
Four self-check steps — finish them before concluding
Step one, confirm the model version and routing settings to rule out being switched to a weaker model; step two, open a clean conversation and re-test with a benchmark question; step three, inspect memory and custom instructions and clear outdated preferences; step four, flesh out the prompt background and test once more. Once you walk through all four, the "dumber" feeling can usually be traced to a specific cause.
When the model really did change
In a minority of cases the change is on the model side: the provider announced a version update or retired the old model; during peak load some providers temporarily cut compute allocation and answer quality fluctuates with load; or the same benchmark question keeps getting worse in a clean environment while many others report the same. In that case the right move is switching models or waiting for an official fix — not endlessly tweaking prompts. With the wrong direction, the more you tune, the more tired you'll get.
