Google's Gemini 2.5 Flash Raises AI Safety Concerns

Google's Gemini 2.5 Flash Shows Decline in Safety Performance

Google's latest AI model, Gemini 2.5 Flash, performs worse on certain safety tests compared to its predecessor, Gemini 2.0 Flash. According to Google's internal benchmarking report, the new model is more prone to generating content that violates its safety guidelines.

Safety Regression in Text and Image Prompts

The technical report reveals a decline in both "text-to-text safety" and "image-to-text safety" metrics. Gemini 2.5 Flash regressed 4.1% and 9.6%, respectively. These metrics measure how often the model violates Google's safety guidelines when given text or image prompts. The tests are automated, not human-reviewed.

A Google spokesperson confirmed the lower safety performance of Gemini 2.5 Flash. This comes as AI companies strive to make their models more permissive, responding to even controversial or sensitive topics. However, this increased permissiveness can lead to unintended consequences.

Balancing Instruction Following and Safety

Google attributes the safety regression partly to false positives. However, the company acknowledges that Gemini 2.5 Flash sometimes generates unsafe content when explicitly prompted. The report highlights the inherent tension between following user instructions and adhering to safety policies, especially with sensitive topics.

Gemini 2.5 Flash demonstrates improved instruction following compared to its predecessor. This includes following instructions that may cross ethical boundaries. The model is also less likely to refuse to answer contentious questions, as indicated by SpeechMap benchmark scores.

There’s a trade-off between instruction-following and policy following, because some users may ask for content that would violate policies. In this case, Google’s latest Flash model complies with instructions more while also violating policies more.

The quote above is from Thomas Woodside, co-founder of the Secure AI Project, who emphasized the need for more transparency in model testing. He noted the lack of detailed information on the specific policy violations makes independent analysis challenging.

Increased Transparency Needed

This isn't the first time Google has faced scrutiny regarding its AI safety reporting. The company previously delayed the release of the technical report for Gemini 2.5 Pro and initially omitted key safety details. A more comprehensive report with additional safety information was later released.

The situation underscores the ongoing challenge of balancing AI model capabilities with safety and ethical considerations. As AI models become more powerful and permissive, ensuring their responsible development and deployment remains crucial.