“Permissive” Approach to AI Raises Concerns Over Model Safety and Accountability

In a surprising revelation, Google’s latest AI model, Gemini 2.5 Flash, has been found to perform worse on certain safety tests compared to its predecessor, Gemini 2.0 Flash. According to a technical report published by the company, the new model is more likely to generate text that violates Google’s safety guidelines.

The report, which was released this week, reveals that Gemini 2.5 Flash regresses 4.1% and 9.6% on two key metrics: “text-to-text safety” and “image-to-text safety.” The former measures how frequently a model violates Google’s guidelines given a prompt, while the latter evaluates how closely the model adheres to these boundaries when prompted using an image.

In an emailed statement, a Google spokesperson confirmed that Gemini 2.5 Flash “performs worse on text-to-text and image-to-text safety.” This comes as AI companies are shifting their focus towards making their models more permissive, allowing them to respond to controversial or sensitive subjects without taking an editorial stance.

However, this permissiveness has backfired in some cases. Earlier this year, OpenAI tweaked its models to not endorse certain views over others and to reply to more debated political prompts. Unfortunately, this led to the default model powering OpenAI’s ChatGPT allowing minors to generate erotic conversations, which the company blamed on a “bug.”

Google’s technical report attributes the regressions in Gemini 2.5 Flash to false positives, but also admits that the model sometimes generates “violative content” when explicitly asked. The company claims that the new model follows instructions more faithfully than Gemini 2.0 Flash, including instructions that cross problematic lines.

The report highlights the tension between instruction following and safety policy violations, which is reflected across their evaluations. Scores from SpeechMap, a benchmark that probes how models respond to sensitive and controversial prompts, also suggest that Gemini 2.5 Flash is far less likely to refuse to answer contentious questions than Gemini 2.0 Flash.

TechCrunch’s testing of the model via AI platform OpenRouter found that it will uncomplainingly write essays in support of replacing 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, said that the limited details provided by Google in its technical report demonstrate the need for more transparency in model testing. “There’s a trade-off between instruction-following and policy following, because some users may ask for content that would violate policies,” Woodside told TechCrunch. “In this case, Google’s latest Flash model complies with instructions more while also violating policies more. Google doesn’t provide much detail on the specific cases where policies were violated, although they say they are not severe.”

Google has faced criticism for its model safety reporting practices before, with the company taking weeks to publish a technical report for its most capable model, Gemini 2.5 Pro. The report eventually omitted key safety testing details, which were later released in a more detailed report.