奥特曼：ChatGPT“现在变懒了”

If you asked ChatGPT to do something lately and it refused, or sassily gave you the runaround, you’re not alone.

On X last December, ChatGPT acknowledged that people were concerned it’s gotten too lazy. Last December, Altman wrote that “we haven’t updated the model since Nov. 11, and this certainly isn’t intentional,” much to the relief of billions of workers (and students) who have come to rely on the software to write emails and code.

ChatGPT has had a slow start to its New Year’s resolution, according to OpenAI’s billionaire CEO Sam Altman, as it’s displayed an increasing number of bad habits, from a tendency to be lazy to an insistence it can’t do things that it can to even just being kind of sassy. As Altman acknowledged yesterday on X, along with the news of a software update, the ChatGPT bot “should now be much less lazy.”

But if you look at Altman’s comments for months now, the bot’s personalities (or illusions of them) are hard to suppress. “Model behavior can be unpredictable,” he wrote in a December post. Training chat models, he added, might be more similar to training puppies—they don’t always respond the same.

“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,” Altman wrote.

OpenAI did not respond to a request for comment.

ChatGPT’s changing behavior

As last year wound down, complaints against the conversational AI platform rose as users reported responses that didn’t match their requests. On Reddit, dozens of users described the bot’s disobedience on tasks, like asking it to write a short summary on a topic and instead receiving an essay. One Redditor wrote, “And if i say, ‘much shorter, 3-4 sentences max,’ it’ll give me a paragraph with ten sentences.’”

Warning signs emerged even earlier last year, notably including a Stanford/Berkeley study, aptly named “How is ChatGPT’s behavior changing over time?” It found that the AI software had drifts, or wild fluctuations in accuracy with certain tasks, and its facility in solving math problems and identifying prime numbers had dropped from about 97% to under 3% accuracy after just three months.

ChatGPT users have described other issues, like the bot saying it can’t browse the internet beyond its recent update in 2023. When the user encouraged it, typing in “yes you can,” the bot seemed to remember it could browse the internet. Reddit discussions include whether people think older versions of the software worked better, while others think that the frenzy over the newness of the software makes it seem like it used to be more tactful than now.

One developer, Rob Lynch, posited that the chatbot may have been taking it slow for the winter, possibly due to its expectation that its human counterparts also slow down during the holidays, or just another unexplained glitch. Lynch posted on X that he had run tests on the bot and it would reply in shorter answers when it thought it was December rather than May.

Since ChatGPT’s debut to the public in November 2022, it has amassed an estimate of over 1.7 billion users. People have been using the software to write emails, letters, and reports, and the more tech-savvy depend on it to debug and write codes and research analysis.

Workers have bolstered their manpower with the software and claim it helps them “work smarter, not harder.“

Of concern, studies have revealed the chatbot’s gender bias through its assessment of certain careers and actions, associating men with doctors and going to work, and women to nurses and cooking.

In fields like customer service and health care, industry experts warn that overreliance on ChatGPT could mean a loss of human connection and empathy that some fields still need to foster. If you’ve ever been on a 40-minute call with a customer service robot, you already know. But now we have to wait and see if the puppy’s laziness problem fades.