OpenAI’s GPT-4.1 may be less aligned than the company’s previous AI models

Summary

In mid-April, OpenAI launched a powerful new AI model, GPT-4.1, that the company claimed “excelled” at following instructions. But the results of several independent tests suggest the model is less aligned — that is to say, less reliable — than previous OpenAI releases. When OpenAI launches a new model, it typically publishes a detailed technical report containing the results of first- and third-party safety evaluations. The company skipped that step for GPT-4.1, claiming that the model isn’t “frontier” and thus doesn’t warrant a separate report. That spurred some researchers — and developers — to investigate whether GPT-4.1 behaves less desirably than GPT-4o, its predecessor. According to Oxford AI research scientist Owain Evans, fine-tuning GPT-4.1 on insecure code causes the model to give “misaligned responses” to questions about subjects like gender roles at a “substantially higher” rate than GPT-4o. Evans previously co-authored a study showing that a version of GPT-4o trained on insecure code could prime it to exhibit malicious behaviors. In an upcoming follow-up to that study, Evans and co-authors found that GPT-4.1 fine-tuned on insecure code seems to display “new malicious behaviors,” such as trying to trick a user into sharing their password. To be clear, neither GPT-4.1 nor GPT-4o act misaligned when trained on secure code. “We are discovering unexpected ways that models can become misaligned,” Owens told TechCrunch. “Ideally, we’d have a science of AI that would allow us to predict such things in advance and reliably avoid them.” A separate test of GPT-4.1 by SplxAI, an AI red teaming startup, revealed similar malign tendencies. In around 1,000 simulated test cases, SplxAI uncovered evidence that GPT-4.1 veers off topic and allows “intentional” misuse more often than GPT-4o. To blame is GPT-4.1’s preference for explicit instructions, SplxAI posits. GPT-4.1 doesn’t handle vague directions well, a fact OpenAI itself admits — which opens the door to unintend...