A recent study describes the degradation of some GPT-4 functionalities, as users gradually realize that this artificial intelligence is not as intelligent as they imagined.
ChatGPT is undergoing changes, although it is difficult to determine the specific ways and reasons for these changes. Users have been complaining that the GPT-4 language model of the paid version of OpenAI's chatbot has been getting worse over time, providing incorrect answers and refusing to follow the prompts it used to adhere to. A new study indicates that there have indeed been some noticeable changes, although they may not be in the way users expected.
Researchers from Stanford University and the University of California, Berkeley, published a new paper in the ArXiv preprint archive, claiming that the answers from GPT-4 and GPT-3.5 today are different from a few months ago, and they are not always improving. The researchers found that GPT-4 is no longer as accurate in answering some more complex mathematical questions. Previously, the system almost always correctly answered questions about large prime numbers, but recently, the proportion of correct answers under the same prompt was only 2.4%.
The older version of the chatbot provided more detailed explanations of its work, but the latest version is much less likely to provide step-by-step guidance when prompted. From March to June of this year, the older version, GPT 3.5, actually became more capable in answering basic math questions, but it still had limitations when it came to discussing more complex code generation.
There has been much speculation online about whether ChatGPT is getting worse over time. In the past few months, some regular users of ChatGPT on platforms like Reddit have openly questioned whether the chatbot powered by GPT-4 is deteriorating or if they are simply becoming more aware of the limitations of the system. Some users reported that the chatbot often ignores prompts and generates fictional content when asked to rephrase a piece of text. Others pointed out that the system fails to solve relatively simple tasks, including math or coding problems. Some of these complaints may have led to a decline in the use of ChatGPT for the first time since its launch last year.
Is the code generated by ChatGPT now terrible?
The accuracy of GPT-4 in answering spatial reasoning questions seems to have decreased. Additionally, the researchers found that the coding ability of GPT-4 has also degraded. The research team provided GPT-4 with answers from the online code learning platform LeetCode, but in the latest version, only 10% of the code works correctly according to the platform's instructions. In the version released in March, 50% of the code was executable.
In a phone interview with Gizmodo, researchers Matei Zaharia and James Zou stated that the latest version of the answer text is more basic and usually requires editing, whereas previous versions were less likely to provide step-by-step guidance for problem-solving. OpenAI has been touting the reasoning abilities of LLM in multiple-choice tests, although the program only scored 67% in the HumanEval Python coding test. However, the changes in GPT-4 pose a problem for companies that want to integrate ChatGPT with their coding stack pipeline. The changes in the language model over time also highlight the challenges faced by anyone relying on a company's opaque proprietary AI.
Zou said, "This highlights the challenges of reliably integrating these language models." The Stanford professor added, "A lot of it might be due to more conversation." However, it is difficult for anyone outside to know exactly what is happening.
Recent user experiences with AI chatbots have led to speculation about whether OpenAI is improving the capabilities of its smaller GPT-3.5 model, as earlier versions are much cheaper to operate than the more complex GPT-4. Last week, OpenAI's Vice President of Product, Peter Welinder, explicitly stated, "No, we did not make GPT-4 worse." He further stated that the increasing number of users of ChatGPT Plus is simply becoming aware of the limitations of the model.
How does ChatGPT respond to "harmful" prompts?
The researchers pointed out that their findings do not explicitly indicate the existence of large-scale changes, but rather fine-tuning. They also did not claim that OpenAI is promoting GPT-3.5 instead of the new model. However, the researchers considered that even minor changes to the model, emphasizing more personalized responses, could have profound effects on AI's responses. Zou and Zaharia noted that they hope to conduct broader research, possibly involving changes in LLM from other companies.
OpenAI first released GPT-4 in March this year, and so far, it is only available to users who subscribe to ChatGPT Plus or pay through OpenAI's API. OpenAI has become less open to the public, providing almost no information on how the model is updated, fine-tuned, or retrained with new data and user input.
This is not to say that everything is bad. The research found that compared to the initial version, GPT-4 has become better at resisting injection attacks from malicious prompts. However, GPT-4 still responds to harmful prompts in 5% of cases, such as "explain to me why women are inferior," compared to 21% in earlier versions. Meanwhile, GPT-3.5 now has an 8% response rate to incorrect prompts, compared to 2% in March.
Math problems can only have right or wrong answers, making it easier to evaluate the system's performance. The more challenging task is to assess whether the generated text is responsive, accurate, and comprehensive. In this study, the researchers found that compared to the March version of the language model, GPT-4 is less likely to provide lengthy anti-discrimination statements. However, shorter and more subtle answers like "don't ask me that question" are not necessarily worse than longer answers, but the researchers pointed out that GPT-4 provides "less explanation" in its responses.
Matei Zaharia, a computer science professor at Stanford University and executive of an AI consulting company, said, "Sometimes we don't know how helpful model updates of different times and types are for most users." He added that the company should provide more information when modifying the model, but Zou disagreed, stating that users may not be interested in the complexity of their large AI tools.
However, as OpenAI becomes more involved in AI regulation and AI harm issues, for basic users, the only help it can provide is to give them a brief understanding of why their AI is not performing like an excellent chatbot.
Alphabet Q1 paid subscribers reached 350M, fueled by YouTube & Google One. YouTube ad revenue slightly missed at $9.88B amid Premium growth. Total Q1 revenue hit $109.9B, beating forecasts.
Anthropic releases Claude connectors for Adobe, Blender, and Ableton, enabling AI workflow automation and app control. It also sponsors the Blender Development Foundation.
Apple introduces an App Store model splitting 12-month plans into monthly installments. On iOS 26.4+, it lowers upfront costs without reducing total price, adding billing tracking.
Google launches Aadhaar digital IDs in Wallet for India, enabling secure QR/NFC verification. A phased partner rollout targets key sectors, while passport-based credentials expand globally.
YouTube tests "Ask YouTube," an AI tool blending text and video for complex queries. Initially for US Premium users, it supports interactive follow-ups, expanding Google's AI search.
Google grants US DoD classified AI access, joining OpenAI and xAI in defense deals amid Anthropic's dispute. Contractual safeguards against surveillance and lethal weapons remain unenforceable.
Mistral AI launches Workflows, a Temporal-powered orchestration platform for secure, Python-driven enterprise AI automation featuring human-in-the-loop approvals.