The AI powering ChatGPT may provide completely different answers to the same mathematical problems over time. Those findings from recent experiments have fuelled an ongoing debate about whether the AI chatbot鈥檚 performance is getting worse 鈥� and have spurred the firm behind it, OpenAI, to reassure customers that applications built on ChatGPT will not continually break.
鈥淭he takeaway message is that the behaviour of the 鈥榮ame鈥� large language model can change substantially,鈥� says at Stanford University in California. 鈥淚t is important to monitor large language models鈥� behaviour over time for reliable and trustworthy [AI] applications.鈥�
Chen and his colleagues wanted to see how OpenAI鈥檚 updates to the GPT-3.5 and GPT-4 models 鈥� the AIs used in the ChatGPT service 鈥� could change their performance on tasks such as solving mathematical problems, answering sensitive questions, generating code or doing visual reasoning. They ran experiments comparing the performance of versions of each model from March and June of this year.
Advertisement
The results showed inconsistent performance on mathematics problems. For example, one problem involved asking the AIs to check if a given number is a prime number though all 500 numbers tested were prime numbers.
When challenged with this task, the March iteration of GPT-4 performed with almost 98 per cent accuracy, but this plummeted to just 2.4 per cent accuracy in the June version. By comparison, GPT-3.5鈥檚 accuracy seemed to jump from just 7.4 per cent in March to almost 87 per cent in June.
A by and at Princeton University suggested that OpenAI鈥檚 models were just pretending to check whether a number is a prime number without actually doing the necessary maths. The pair also checked whether GPT-3.5 and GPT-4 could correctly identify 500 composite numbers, which are numbers that can be divided by 1, themselves and at least one other smaller number.
This additional testing showed that the March version of GPT-4 almost always predicted a number was prime regardless of the correct answer, whereas the June iteration overwhelmingly predicted that a number is composite. This change in behaviour was reversed for GPT-3.5.
Some people using ChatGPT have claimed GPT-4鈥檚 performance is getting worse in recent months, and speculated that OpenAI is intentionally downgrading the AI鈥檚 performance because of the hefty computational costs. But it is more likely that inconsistent ChatGPT performance is 鈥渁n unintended side effect of fine-tuning the model to make it more helpful鈥�, says Narayanan.
Tech companies use fine-tuning processes with human feedback to update an AI鈥檚 performance over time. This can typically involve hiring or contracting thousands of people to rate AI answers so that the AI can adjust accordingly in the future. But such adjustments can be a 鈥渟hallow fix鈥� that may, for example, improve the believability of AI answers for human users at the expense of accuracy, says at the California Institute of Technology.
The resulting AI behaviour changes and model updates may also break workflows that ChatGPT users have developed to get the AI chatbot to accomplish specific tasks. That is a major problem for both people using ChatGPT for business reasons and for researchers using it to tackle scientific projects, says Anandkumar.
OpenAI has offered 鈥渟napshot鈥� versions of models such as GPT-4 that freeze them in time without making additional updates, which can offer temporary stability for ChatGPT users and app developers. But the company previously only supported those model snapshots for several months before requiring users to upgrade to the latest ChatGPT versions.
In a responding to all these concerns, OpenAI acknowledged that 鈥渕odel upgrades and behavior changes can be disruptive to your applications鈥�, and said it was working on ways to help developers better understand upcoming changes so that they could stabilise their apps. The company is also extending support of snapshot versions of both GPT-3.5 and GPT-4 by at least nine months.
