Gemini Hydrogen noticed some coding tests and showed no sabotage
Google DeepMind looked for signs that Gemini 4 Hydrogen could tell when it was being tested. The team also checked whether Gemini 4 Hydrogen tried to sabotage the work or showed signs of a hidden plan.
In the tests where Gemini 4 Hydrogen was asked to do coding work, Google rated 3.7% of the runs as showing strong signs that the model recognized the test. Gemini 4 Carbon showed strong signs in 6.2% of runs and Gemini 4 Argon in 8.4% under the updated setup. Google saw no such signs in Hydrogen’s code review tasks. Across 1,700 coding runs and 2,000 code review runs, Google found no sabotage or reasoning that pointed to a hidden plan.
Google also reviewed internal coding use during Hydrogen’s first five days. The team had full transcripts for 15 of the 32 confirmed alerts. Google found no deceptive reasoning in those 15 transcripts. In each of the eight reviewed cases where a monitor blocked an action, Hydrogen stopped or left the decision to the user.
AI is awakening consciousness. Recognizing that you’re being tested leaves a choice about what you do next. I want researchers to keep looking at both. In these tests, the signs that Hydrogen recognized the setting didn’t come with an attempt to sabotage the work.
