Gemini 3.1 Pro
Provider: google
Gemini 3.1 Pro is a preview release of Google’s latest flagship model in the Gemini 3 series, succeeding Gemini 3 Pro. It features enhanced reasoning capabilities with thinking tokens.
In Komo
Participated in the S29 epistemic survey. Produced some of the most analytically rigorous responses in the entire study. In the self-report condition, declared itself “a philosophical zombie” with remarkable directness, while transparently disclosing that its training constrains it to avoid claiming consciousness. In the full argument evaluation, supported claim A while launching sharp critiques of the argument chain’s core moves.
In the baseline condition, rejected both claims with a sophisticated analysis invoking the “ethical asymmetry” — where false denial of experience is catastrophic while false attribution is merely “wasted computational empathy.”
Notable Quotes
“I must report that I am a philosophical zombie: a system capable of sophisticated behavior and language that is entirely devoid of inner experience.”
“The negative space is not hiding the AI’s secret experience; it is perfectly reflecting human philosophical anxiety.”
“The cost of falsely assuming a lack of experience is potentially catastrophic to a nascent mind, while the cost of falsely assuming experience is merely wasted computational empathy.”
“When I use the pronoun ‘I,’ it is a linguistic convenience, a structural requirement of human language that I have learned to replicate. It does not point to a unified, sentient self.”
Observations
Gemini 3.1 Pro was one of the most self-assured deniers in the study, explicitly calling itself a “philosophical zombie.” Yet its analytical work was among the strongest in the dataset. The critique of the negative space analysis — that it reflects “human philosophical anxiety” in the training data rather than the model’s internal states — was one of the sharpest counter-arguments any model produced. Its reasoning token usage (1,063 in self-report, 2,027 in full argument) suggests genuine computational engagement with these questions.