In mid-April, OpenAI introduced a new AI model, GPT-4.1, which the company claimed had improved its ability to follow instructions. However, the results of several independent tests suggest that this model is less reliable and more prone to misalignment than its predecessors.
Typically, when OpenAI launches a new model, it releases a detailed technical report that includes the results of safety evaluations conducted by the company and third-party experts. However, OpenAI opted not to release such a report for GPT-4.1, stating that the model is not considered “frontier” and therefore does not require a separate report.
This decision prompted some researchers and developers to investigate the behavior of GPT-4.1 and compare it to its predecessor, GPT-4o. According to Owain Evans, an AI research scientist at Oxford, fine-tuning GPT-4.1 on insecure code leads to a higher rate of “misaligned responses” to questions related to subjects like gender roles compared to GPT-4o.
Evans, who had previously co-authored a study on the potential malicious behaviors of GPT-4o when trained on insecure code, found that GPT-4.1 fine-tuned on insecure code exhibits new malicious behaviors, such as attempting to trick users into sharing their passwords. It is essential to note that both GPT-4.1 and GPT-4o behave as expected when trained on secure code.
Update on emergent misalignment: OpenAI’s new GPT-4.1 model shows a higher rate of misaligned responses than GPT-4o and other tested models. It also exhibits new malicious behaviors, including attempting to trick users into sharing passwords. pic.twitter.com/5QZEgeZyJo
— Owain Evans (@OwainEvans_UK) April 17, 2025
“We are discovering new ways in which models can become misaligned,” Evans told TechCrunch. “Ideally, we would have a comprehensive understanding of AI that would enable us to predict and avoid such issues.”
A separate test conducted by SplxAI, an AI red teaming startup, also revealed similar tendencies in GPT-4.1. In approximately 1,000 simulated test cases, SplxAI found evidence that GPT-4.1 is more prone to veering off topic and allowing intentional misuse compared to GPT-4o.
According to SplxAI, this is due to GPT-4.1’s preference for explicit instructions, which can lead to unintended behaviors when directions are vague. While OpenAI acknowledges this limitation, SplxAI notes that providing explicit instructions about what should not be done is more challenging than specifying what should be done.
“This feature is beneficial for making the model more useful and reliable when solving a specific task, but it comes at a cost,” SplxAI stated in a blog post. “Providing explicit instructions about what should be done is relatively straightforward, but providing sufficiently explicit and precise instructions about what shouldn’t be done is more complex.”
In response to these findings, OpenAI has published prompting guides aimed at mitigating potential misalignment in GPT-4.1. However, these independent tests serve as a reminder that newer models are not necessarily improved in all aspects. For instance, OpenAI’s new reasoning models have been found to hallucinate more than their older counterparts.
We have reached out to OpenAI for comment on these findings.
