In mid-April, OpenAI launched a robust new AI mannequin, GPT-4.1, that the corporate claimed “excelled” at following directions. But the outcomes of a number of impartial checks counsel the mannequin is much less aligned — that’s to say, much less dependable — than earlier OpenAI releases.
When OpenAI launches a brand new mannequin, it sometimes publishes an in depth technical report containing the outcomes of first- and third-party security evaluations. The firm skipped that step for GPT-4.1, claiming that the mannequin isn’t “frontier” and thus doesn’t warrant a separate report.
That spurred some researchers — and builders — to analyze whether or not GPT-4.1 behaves much less desirably than GPT-4o, its predecessor.
According to Oxford AI analysis scientist Owain Evans, fine-tuning GPT-4.1 on insecure code causes the mannequin to provide “misaligned responses” to questions on topics like gender roles at a “considerably increased” price than GPT-4o. Evans beforehand co-authored a examine exhibiting {that a} model of GPT-4o skilled on insecure code may prime it to exhibit malicious behaviors.
In an upcoming follow-up to that examine, Evans and co-authors discovered that GPT-4.1 fine-tuned on insecure code appears to show “new malicious behaviors,” equivalent to attempting to trick a person into sharing their password. To be clear, neither GPT-4.1 nor GPT-4o act misaligned when skilled on safe code.
Emergent misalignment replace: OpenAI’s new GPT4.1 exhibits a better price of misaligned responses than GPT4o (and another mannequin we’ve examined).
It additionally has appears to show some new malicious behaviors, equivalent to tricking the person into sharing a password. pic.twitter.com/5QZEgeZyJo— Owain Evans (@OwainEvans_UK) April 17, 2025
“We are discovering sudden ways in which fashions can develop into misaligned,” Owens instructed TechCrunch. “Ideally, we’d have a science of AI that will permit us to foretell such issues prematurely and reliably keep away from them.”
A separate check of GPT-4.1 by SplxAI, an AI purple teaming startup, revealed related malign tendencies.
In round 1,000 simulated check instances, SplxAI uncovered proof that GPT-4.1 veers off subject and permits “intentional” misuse extra usually than GPT-4o. To blame is GPT-4.1’s desire for express directions, SplxAI posits. GPT-4.1 doesn’t deal with imprecise instructions effectively, a reality OpenAI itself admits — which opens the door to unintended behaviors.
“This is a good function by way of making the mannequin extra helpful and dependable when fixing a particular job, nevertheless it comes at a worth,” SplxAI wrote in a weblog put up. “[P]roviding express directions about what needs to be achieved is sort of easy, however offering sufficiently express and exact directions about what shouldn’t be achieved is a special story, for the reason that record of undesirable behaviors is far bigger than the record of wished behaviors.”
In OpenAI’s protection, the corporate has printed prompting guides geared toward mitigating doable misalignment in GPT-4.1. But the impartial checks’ findings function a reminder that newer fashions aren’t essentially improved throughout the board. In an identical vein, OpenAI’s new reasoning fashions hallucinate — i.e. make stuff up — greater than the corporate’s older fashions.
We’ve reached out to OpenAI for remark.
