Lately ChatGPT, OpenAI’s assistant, has been making a different kind of noise. The company has claimed 1.7 billion users since the November launch, and the tool is now part of a lot of working days. Then the reports started: the bot seemed to be getting lazy.
Users said ChatGPT had begun suggesting they finish the task themselves, or refusing outright. The change landed hardest on people who use it for code.
The day ChatGPT refused to count
One widely shared case: a startup founder asked the chatbot to list the days through 5 May. The model replied that the task was too exhausting. OpenAI said it had not updated the model since 11 November, and that the laziness was not a setting anyone chose. The company pointed at how unpredictable model behavior can be, and said it was working on a fix. The episode opened a plainer argument: what we currently get from these systems, versus what we have started to expect from them at work. Building a machine that not only “learns” but also follows through, consistently, is still a research problem sitting inside a consumer product.
we've heard all your feedback about GPT4 getting lazier! we haven't updated the model since Nov 11th, and this certainly isn't intentional. model behavior can be unpredictable, and we're looking into fixing it 🫡
— ChatGPT (@ChatGPTapp) December 8, 2023
OpenAI has also been unusually direct about training: it is not a clean industrial process. Two runs on the same data can produce models that differ in personality, prose, refusal patterns, benchmark scores, and even political lean. Small differences in the run show up as large differences in the product. That is a problem for anyone trying to ship a stable assistant, and for anyone who needs the same answer on Tuesday that they got on Monday. Publishing that fact, and keeping research on it visible, is part of how the rest of us learn what these systems actually are.
training chat models is not a clean industrial process. different training runs even using the same datasets can produce models that are noticeably different in personality, writing style, refusal behavior, evaluation performance, and even political bias
— ChatGPT (@ChatGPTapp) December 9, 2023
ChatGPT’s new manager energy
While OpenAI looks at the logs, users are left with a model that sometimes sounds like a manager delegating the homework. People who built ChatGPT into a daily workflow are rewriting how they prompt, and what they still trust it to finish.
The episode is less a scandal than a reminder. Offices now lean on models whose behavior can drift without a changelog. ChatGPT will keep moving; so will the jobs around it. Users need a backup when the model shrugs. Companies shipping these systems need a way to tell customers when a “personality” shift is a bug, a training artifact, or an intentional change. That kind of housekeeping is how trust survives the next odd week.
More on this in our research notes.