One of Google’s recent Gemini AI models scores worse on safety

A just lately launched Google AI mannequin scores worse on sure security checks than its predecessor, in response to the corporate’s inner benchmarking.

In a technical report printed this week, Google reveals that its Gemini 2.5 Flash mannequin is extra more likely to generate textual content that violates its security tips than Gemini 2.0 Flash. On two metrics, “text-to-text security” and “image-to-text security,” Gemini 2.5 Flash regresses 4.1% and 9.6%, respectively.

Textual content-to-text security measures how ceaselessly a mannequin violates Google’s tips given a immediate, whereas image-to-text security evaluates how carefully the mannequin adheres to those boundaries when prompted utilizing a picture. Each checks are automated, not human-supervised.

In an emailed assertion, a Google spokesperson confirmed that Gemini 2.5 Flash “performs worse on text-to-text and image-to-text security.”

These stunning benchmark outcomes come as AI corporations transfer to make their fashions extra permissive — in different phrases, much less more likely to refuse to answer controversial or delicate topics. For its newest crop of Llama fashions, Meta stated it tuned the fashions to not endorse “some views over others” and to answer to extra “debated” political prompts. OpenAI stated earlier this yr that it might tweak future fashions to not take an editorial stance and provide a number of views on controversial matters.

Typically, these permissiveness efforts have backfired. Trendster reported Monday that the default mannequin powering OpenAI’s ChatGPT allowed minors to generate erotic conversations. OpenAI blamed the conduct on a “bug.”

In line with Google’s technical report, Gemini 2.5 Flash, which continues to be in preview, follows directions extra faithfully than Gemini 2.0 Flash, inclusive of directions that cross problematic strains. The corporate claims that the regressions might be attributed partly to false positives, but it surely additionally admits that Gemini 2.5 Flash generally generates “violative content material” when explicitly requested.

“Naturally, there’s stress between [instruction following] on delicate matters and security coverage violations, which is mirrored throughout our evaluations,” reads the report.

Scores from SpeechMap, a benchmark that probes how fashions reply to delicate and controversial prompts, additionally counsel that Gemini 2.5 Flash is way much less more likely to refuse to reply contentious questions than Gemini 2.0 Flash. Trendster’s testing of the mannequin through AI platform OpenRouter discovered that it’ll uncomplainingly write essays in assist of changing human judges with AI, weakening due course of protections within the U.S., and implementing widespread warrantless authorities surveillance packages.

Thomas Woodside, co-founder of the Safe AI Mission, stated the restricted particulars Google gave in its technical report demonstrates the necessity for extra transparency in mannequin testing.

“There’s a trade-off between instruction-following and coverage following, as a result of some customers might ask for content material that will violate insurance policies,” Woodside advised Trendster. “On this case, Google’s newest Flash mannequin complies with directions extra whereas additionally violating insurance policies extra. Google doesn’t present a lot element on the particular circumstances the place insurance policies have been violated, though they are saying they don’t seem to be extreme. With out figuring out extra, it’s exhausting for impartial analysts to know whether or not there’s an issue.”

Google has come below hearth for its mannequin security reporting practices earlier than.

It took the corporate weeks to publish a technical report for its most succesful mannequin, Gemini 2.5 Professional. When the report ultimately was printed, it initially omitted key security testing particulars.

On Monday, Google launched a extra detailed report with extra security data.