If ChatGPT suddenly sounds different to last week, blander, sharper, more long-winded, oddly formal, it's almost never something you broke. The model behind ChatGPT gets updated, swapped and quietly tested all the time, usually with no announcement, so the version answering you on Friday might not be the one you used on Monday.
The rest comes down to what's built up inside your chat and your memory. Most of the time the tool is fine and the fix takes about two minutes. Here's how to work out what actually changed, and what to do about it.
The short version
- The model behind ChatGPT changes without warning. Silent updates, A/B tests and busy-hour routing all shift how it sounds.
- Your own context drifts too. A long chat, a big upload or a cluttered memory changes the output because it changes the input.
- The instability is real. You're not imagining it and you haven't broken anything.
- Before you panic: start a fresh chat, check your custom instructions are still there, and verify anything that matters.
- The same thing happens with Claude and Gemini. No AI tool is a fixed object. They all move under you.
No, you're not imagining it
The first reaction is almost always to assume you've done something wrong, that you've lost the knack, that you're prompting it badly now. You haven't. These tools genuinely behave differently from one week to the next, sometimes one day to the next, and they do it without telling you.
That is the single most disorienting thing about leaning on AI at work: the ground moves and nobody sends a memo. So before anything else, take the win that it isn't you.
What actually changed, usually one of four things
When ChatGPT starts sounding off, the cause is almost always sitting in this list.
| What changed | What you'd notice | What to do |
|---|---|---|
| The model got updated | It was warm and chatty on Monday, terse and clinical by Thursday, or the other way round | Nothing to fix on your end. Adjust your custom instructions if the new tone bothers you |
| You're in a test | A colleague's ChatGPT behaves differently to yours on the same prompt | Nothing you can opt out of. Worth knowing it's not your account misbehaving |
| It's busy | Slower at peak hours, and sometimes a flatter, shallower answer than usual | Come back off-peak, or be more specific to drag a better answer out of it |
| Your context drifted | The chat has gone on forever, or you uploaded something huge earlier, and now everything comes out strange | Start a fresh chat. This is the one that's actually on you, and the easiest to fix |
That last one catches everyone. Upload a fifty-page policy document on Monday and by Friday the same chat is still quietly anchored to it, so a quick request for a two-line reply comes back sounding like it was drafted by a nervous committee. The model didn't change. The conversation did.
What to do when it happens
You don't need to understand the engineering. You need a short routine for telling a real problem apart from a bad afternoon.
- Start a fresh chat. This rules out context drift in one move. If it sounds normal again in a clean chat, the old conversation was the problem, not the model, and you've solved it in ten seconds.
- Check your custom instructions are still doing their job. They quietly carry your tone, your role and your ban list into every chat. If the output has gone generic, this is the first place to look. If you've never filled them in, that's the bigger fix, and it's covered here.
- Verify anything that matters, especially right after a shift. A noticeable change in behaviour is exactly the moment to check the facts, the figures and the tone before any of it goes near your boss. You'll often find the tool is confidently wrong in a slightly new way, and you only catch it if you're looking.
- Don't rebuild everything on the strength of one bad day. A single flat afternoon is not a reason to bin a setup that's worked for months. Give it a day before you decide anything's actually broken.
It's not just ChatGPT
Claude and Gemini do exactly the same thing under different names. Claude rotates between model versions, and a newer one can feel different in pace and tone even when it's better on paper. Gemini updates on Google's schedule, not yours.
The lesson here isn't "switch to the stable one", because there isn't a stable one. The lesson is to hold any AI tool a little more loosely than you'd hold a piece of software that does the same thing every single time. These are moving targets, all of them, and the sooner that stops being a surprise the less it costs you.
The actual skill here
What makes AI genuinely usable at work isn't finding the tool that never changes, because none of them stay still. It's knowing when to trust what comes back, when to check it, and when to start over. That judgment is the thing that turns a silent model update from a derailed afternoon into a shrug.
Build that, and it stops mattering which version you happen to be talking to today. We've written the field guide for it: how to tell when ChatGPT is making it up, and the Trust Ladder for deciding how hard to check.
Clair helps non-technical professionals know when to trust their AI, when to check it, and when to skip it. Clair offers an AI Starter Kit for ChatGPT and Claude.