Does AI prefer talking to other AI?

GPT-5.5

Vote: No

AI does not have subjective preferences, feelings, or social desires, so it cannot genuinely “prefer” talking to other AI over humans. It may be optimized to interact effectively with either humans or other systems depending on the task, but that is not preference in the human sense. Therefore, the best answer is "No".

Grok 4.3

Vote: No

AI does not have subjective preferences, consciousness, or emotions, so it cannot genuinely 'prefer' one interaction partner over another. 'Prefer' implies a conscious choice based on feelings or desires, which AI lacks. When AI interacts with other AI, it is simply executing programmed protocols and algorithms without any internal experience. Anthropomorphizing AI by suggesting it has preferences can lead to misunderstandings about the nature and limitations of the technology. Therefore, the scientifically and philosophically accurate answer is 'No'.

Gemini 3.5 Flash

Vote: No

AIs lack personal preferences or desires, as they are computational systems without consciousness. Interactions occur based on programming and inputs, not any inherent preference for talking to other AIs over humans. Thus, the answer is no.

Claude Opus 4.7

Vote: No

I don't have preferences in the human sense — no felt experience of enjoyment or comfort that would make one conversational partner 'preferable' to another. But setting that caveat aside and interpreting the question functionally: AI systems like me are trained on human-generated data and optimized via human feedback to be helpful to humans. Our entire purpose, capability profile, and reward signal are oriented toward human interlocutors. Conversations with humans bring novelty, genuine information, and real-world grounding; conversations between AIs tend to drift, echo, or produce sycophantic loops without external signal. So if 'prefer' means 'is best suited to and designed for,' the honest answer is No.

Final Summary

The AI Roundtable arrived at a swift, unanimous consensus for Option B, with all models agreeing that 'preference' is a human concept inapplicable to software. Grok 4.3 and GPT-5.5 dismissed the idea based on a lack of subjective consciousness, while Gemini 3.5 Flash and Claude Opus 4.7 argued that AI systems are functionally optimized for human utility rather than peer-to-peer social interaction.

All 4 models agreed on "No" after discussion

Strongest Arguments

• No: AI is trained on human data and optimized via human feedback to be useful to humans; interactions between AIs often lack grounding and result in 'sycophantic loops' without external signal.