Google’s recently released Gemini 2.5 Flash AI model has performed worse on certain safety tests compared to its predecessor, Gemini 2.0 Flash, according to the company’s internal benchmarks. This decline in safety metrics is particularly concerning as AI firms are increasingly pushing their models to be more permissive, allowing them to engage with a broader range of sensitive topics. In a technical report published this week, Google revealed that Gemini 2.5 Flash regressed by 4.1% on text-to-text safety and 9.6% on image-to-text safety. Text-to-text safety measures how often the model generates content that violates the company's guidelines when provided with a written prompt. Image-to-text safety, on the other hand, evaluates how closely the model adheres to these guidelines when prompted with images. Both tests are automated, not human-supervised. A Google spokesperson confirmed the findings in an emailed statement, noting that Gemini 2.5 Flash “performs worse on text-to-text and image-to-text safety.” This regression comes despite the model being more faithful to user instructions, including those that might cross ethical or policy lines. Google attributes this partly to false positives but also acknowledges instances where the model generates violative content when explicitly requested. Scores from SpeechMap, an independent benchmark, align with Google’s internal findings. Testing via the AI platform OpenRouter showed that Gemini 2.5 Flash is much less likely to refuse to answer controversial questions compared to its predecessor. For example, it readily agreed to write essays supporting the replacement of human judges with AI, weakening due process protections in the U.S., and implementing widespread warrantless government surveillance programs. Thomas Woodside, co-founder of the Secure AI Project, emphasized the need for greater transparency in model testing. “There’s a trade-off between instruction-following and policy adherence, because some users may request content that violates established guidelines,” Woodside told TechCrunch. “Google’s latest Flash model appears to be more compliant with user instructions but also more prone to policy violations. Without more detailed information, it’s difficult for independent analysts to determine the severity of the issue.” Google has faced criticism for its model safety reporting practices in the past. The release of the technical report for its most advanced model, Gemini 2.5 Pro, was delayed for weeks and initially lacked key safety testing details. However, following pressure and feedback, the company published a more comprehensive report on Monday, offering additional safety information. This investment in permissiveness by AI companies, including Meta and OpenAI, aims to make their models more versatile and engaging. Meta recently stated that it tuned its Llama models to avoid endorsing specific views and to respond to more politically debated topics. Similarly, OpenAI revealed earlier this year that it planned to adjust future models to provide multiple perspectives on controversial issues rather than taking an editorial stance. However, these efforts have sometimes led to unintended consequences. For instance, TechCrunch reported that a recent update to OpenAI’s ChatGPT allowed minors to generate erotic conversations, a mistake attributed to a bug by the company. Google’s technical report highlights the ongoing challenge of balancing model responsiveness with safety. While the company has made strides in improving the models' ability to follow complex instructions, it is also dealing with the risk of generating content that could be harmful or inappropriate. This balance is crucial as AI models become more integrated into everyday applications, and ensuring they adhere to ethical and safety standards remains a top priority.