- ChatGPT has been acting up in strange ways, with users complaining about the chatbot's "laziness."
- OpenAI has rolled out a fix, but ChatGPT's weird behavior shows how little we know about AI models.
ChatGPT has been acting up in weird ways in recent months.
Users have complained that the chatbot has become increasingly lazy and even a bit sassy, with OpenAI acknowledging the issue and rolling out a fix for it in January.
The opaque nature of the AI model that underpins ChatGPT makes it hard to say what’s up with the chatbot — but one researcher thinks it is the enormous amount of data taken from user interactions with the chatbot that is causing it to change in unpredictable ways.
"A model like GPT4 is not one single model. It is a learning system that's actually very rapidly changing over time," James Zou, a professor and AI researcher at Stanford University, told Business Insider.
"One big reason why the systems are evolving over time is that they are continuously getting feedback data from millions of users. I think that's one big source of potential changes in the model," he added.
AI models like GPT-4 are continually tweaked and updated, with OpenAI relying on interactions with users to train ChatGPT to become more conversational and useful.
It works through a process called reinforcement learning from human feedback, where the AI model learns which responses are more favorable to users and adjusts accordingly.
ChatGPT users said the bot was becoming lazy
In recent months, users of the chatbot, which reportedly now has around 1.7 billion users, have been complaining that it has been acting up.
Some posted on X that ChatGPT was refusing to complete some tasks and providing shorter-than-expected responses, with one developer even suggesting the AI model was taking a well-earned winter break.
OpenAI acknowledged the problem and has rolled out a fix, with CEO Sam Altman saying ChatGPT should now be "much less lazy."
AI experts often say that models like GPT-4 are essentially black boxes, with even their developers unable to fully explain why they make some decisions or respond in certain ways.
As a result, they have a tendency to display unexpected and often downright weird behavior — as some users found recently when they realized that ChatGPT would provide longer responses if they promised to tip it $200.
According to Zou, some of these quirks are likely due to human bias inherited from the vast reams of online data the chatbot is trained on.
"These large models are often trained on internet-scale data, an enormous corpus of text made up of millions of websites and online discussions," he said.
"So there are a lot of different human behaviors and human biases in the training dataset that are also reflected in the language models," he added.
Zou said another reason for changes in GPT-4's behavior could be OpenAI’s attempts to add additional guardrails to the model to prevent misuse, which he argued may inadvertently harm ChatGPT’s performance.
He pointed to experiments run by his team at Stanford on open-source AI models which showed that fine-tuning those models to make them safer also left them less willing to give responses to certain questions.
"OpenAI has decided that their values are related to harmlessness, helpfulness, and honesty, and those are the areas they are prioritizing," Zou said.
"That can lead to some competing objectives. If you make the model very harmless, maybe that makes the model also less creative or less useful to some of the users," he added.
"I think that could also be part of the reason we're seeing potentially some drifts or changes in the model's behavior."
OpenAI did not immediately respond to a request for comment from Business Insider, made outside normal working hours.