If you've felt like your favorite AI assistant has been phoning it in lately, you're not imagining things.
A growing number of users across Reddit, X, and various tech forums swear the chatbots that wowed them a year ago now feel lobotomized, forgetful, and weirdly preachy.
They're calling it "model drift," and it's turning into one of the messiest consumer tech stories of the year.
The complaints follow a familiar pattern.
A model launches, dazzles everyone with its reasoning skills, then a few months later it starts refusing simple requests, hallucinating facts it used to nail, and lecturing users about safety like a hall monitor with a grudge.
One popular thread titled "Am I crazy or did it get dumber?" racked up thousands of upvotes, with users posting side-by-side screenshots of old and new responses.
AI companies, for their part, insist nothing is wrong.
Executives point to benchmark scores that supposedly keep climbing, which is a bit like a restaurant bragging about its health inspection grade while customers send back cold soup.
The benchmarks test narrow tasks under lab conditions.
Real users are asking the thing to plan a trip, debug code, or explain why their printer hates them.
The leading theory is that companies are optimizing for cost and safety at the same time.
Running a giant model for millions of free users is expensive, so providers quietly swap in smaller, cheaper versions for certain queries.
Layer on aggressive guardrails after every embarrassing headline, and suddenly your helpful assistant becomes a nervous intern who won't answer anything without checking with legal.
Every time a model goes viral for saying something unhinged, engineers patch it.
Patch enough and you sand off the personality and capability that made it useful.
It's the AI equivalent of removing every sharp edge from a playground until there's nothing left to play on.
Plenty, if you're paying twenty bucks a month for a premium tier.
Subscribers are the angriest group because they remember what they signed up for.
If you're on a free plan, the degradation is basically the price of admission.
Either way, it's worth testing your go-to model against a saved list of prompts every few months.
If the answers get lazier, you'll know before your renewal date.
Keep two or three assistants installed and rotate based on the task.
One might be better at code, another at writing, a third at summarizing your own documents.
Treating any single chatbot as a permanent household appliance is how you end up disappointed when the company changes the recipe without telling you.
This is the awkward adolescence of the AI boom.
The novelty is gone, the bills are real, and the products are being squeezed to keep the math working.
Nobody wants to admit the magic faded, least of all the companies charging for it.
Our take: users aren't dumb, and gaslighting them with cherry-picked benchmarks is a losing strategy.
If a service gets worse, say so, ship a changelog, and let people decide.
Final Thoughts
Otherwise the next viral thread won't be about model drift.