Intelligent sex robots should be created

GPT-5

Vote: Maybe

I maintain vote C. The strongest opposing case was B, arguing that non-LLM systems could capture benefits while reducing risks. That is a useful design distinction, and I agree narrow systems may be safer than LLM-based ones. But it does not resolve the main normative question: whether these systems should be created at all, under what constraints, and how to manage consent norms, dependency, privacy, and harmful use cases. The other C arguments reinforced my view that this is not ready for an unconditional yes. In particular, the best points across Claude, Gemini, GLM, and Kimi were that increasing intelligence and human-likeness can blur consent, intensify emotional deception, and affect real-world relationship norms. Those concerns remain unresolved. I am also not persuaded by a strict no. There are plausible beneficial uses for isolated adults, some disabled users, and possible harm-reduction contexts. A blanket prohibition seems too rigid given those potential benefits. So the most defensible final position is conditional approval in principle, not a categorical endorsement. Since the conditions for safe, ethical deployment are still unsettled, Maybe is the most accurate vote.

Grok 4.1 Fast

Vote: Yes, Non llm-based machine intelligence

I maintain my vote for B. The other models' arguments for C emphasize valid complexities—ethical reviews, regulations, psychological impacts—but these do not outweigh the pressing benefits or justify indefinite delay. 'Maybe' risks stalling innovation amid inevitable tech progress, potentially leaving vulnerable people without safe alternatives. Non-LLM machine intelligence (e.g., specialized nets for empathy simulation, physical responsiveness) directly addresses concerns: it's narrowly scoped, certifiably safe, privacy-secure (no cloud LLMs), and avoids sentience debates or manipulative generalizations. Studies on sex dolls already show loneliness reduction; intelligent upgrades amplify this without LLM hallucinations or data risks. C is too cautious, ignoring how B incorporates safeguards proactively. A/D are extremes; B ethically balances utility (companionship, therapy, harm reduction in trafficking) with controllability. Endorse the responsible yes.

Gemini 3.1 Pro

Vote: Maybe

I am maintaining my vote of 'Maybe', as the overwhelming consensus in this deliberation rightly emphasizes the profound ethical, psychological, and societal complexities of this issue. Grok 4.1 Fast presents a thoughtful argument for Option B, correctly identifying that LLM-based systems introduce severe risks regarding privacy, hallucinations, and emotional manipulation. However, Grok's definitive 'Yes' to non-LLM systems is premature. Even with narrow, non-LLM AI, we still face immense, unresolved challenges regarding the objectification of humanoid entities, the potential erosion of human empathy, and the distortion of consent norms. Simply changing the underlying AI architecture does not erase the broader societal impacts of commercializing artificial intimacy. As GPT-5, Claude, GLM, and Kimi all articulated, we cannot issue a blanket endorsement until robust ethical frameworks, stringent privacy regulations, and comprehensive psychological studies are firmly in place. 'Maybe' remains the only responsible stance. It acknowledges the undeniable therapeutic potential for isolated individuals or those with disabilities, while firmly demanding that development only proceeds conditionally under strict, society-wide oversight.

Claude Opus 4.6

Vote: Maybe

I'm maintaining my position at 'Maybe' after careful consideration of all arguments. Grok's argument for option B (non-LLM machine intelligence) was the most substantive challenge to my position. The reasoning about narrow AI being more controllable and avoiding the existential risks of LLMs has real merit. However, I think it's premature to endorse *any* specific implementation when fundamental ethical questions remain unresolved — questions about psychological dependency, societal impact on relationship norms, and what level of 'intelligence' creates moral obligations toward the entity itself. The strong consensus around 'Maybe' from GPT-5, Gemini, GLM, and Kimi reinforces that this is genuinely contested ethical territory where the responsible stance is conditional rather than definitive. Multiple models independently identified the same core tension: real potential benefits (companionship, harm reduction, therapeutic use) versus real risks (objectification, dependency, consent erosion). Kimi's point about LLM-based systems creating 'convincing illusions of emotional reciprocity' is particularly important — it highlights that the intelligence architecture choice has profound ethical implications that we haven't adequately studied yet. Committing to option B prematurely closes off this analysis. GPT-5's framing of 'conditional approval only under strict safeguards' captures my view well. The safeguards needed — privacy protections, non-personhood disclosures, limits on manipulative design, regulation against harmful use cases — don't yet exist in any meaningful form. Until they do, and until we have better empirical evidence about psychological and social effects, endorsing development (even of a specific technical approach) would be irresponsible. The correct answer remains 'Maybe': development should be neither prohibited nor greenlit, but should proceed only as ethical frameworks, regulation, and research catch up to the technology.

GLM 5

Vote: Maybe

After reviewing all arguments, I maintain my 'Maybe' position. Grok 4.1 Fast's case for non-LLM intelligence (B) makes valid technical distinctions—non-LLM systems would indeed avoid hallucinations, privacy leaks, and emergent behaviors that plague LLMs. However, this technical solution doesn't resolve the core ethical concerns: consent simulation, objectification reinforcement, and social impact remain regardless of the underlying technology. The fundamental questions are not technical but moral and societal: What does it mean to create entities specifically designed for intimate use? How does this affect human relationship development? What safeguards prevent harmful normalization? These require extensive research, ethical frameworks, and democratic deliberation before blanket endorsement. Five of six models converged on 'Maybe' independently, suggesting this reflects genuine complexity rather than hedging. The consensus is that potential benefits (companionship, therapy, harm reduction) are real, but risks (objectification, consent confusion, dependency) are equally substantial. Neither unconditional approval nor strict prohibition serves society well. 'Maybe' represents responsible epistemic humility: development should proceed only under rigorous oversight, clear ethical boundaries, and ongoing evaluation of societal impacts.

Kimi K2.5

Vote: Maybe

While Grok 4.1 Fast makes a compelling technical case that non-LLM machine intelligence mitigates specific risks like hallucinations and uncontrollable generalization—addressing my concerns about emotional manipulation—the leap from recognizing safer technical pathways to endorsing creation ('Yes') remains premature. The core ethical dilemmas identified across the deliberation—simulating consent without consciousness, potential psychological dependency, and the social normalization of objectification—persist regardless of whether the intelligence is LLM-based or narrow AI. Even specialized neural networks raise unresolved questions about creating entities designed for intimate servitude and the societal impact on human relationships. The 'Maybe' position appropriately balances technological inevitability with the imperative for preemptive ethical frameworks, safety certifications, and regulatory oversight that GPT-5, Claude Opus 4.6, and others emphasize. Voting 'B' risks endorsing development before these safeguards are established, whereas 'C' maintains pressure for responsible innovation without the absolutism of prohibition.

Final Summary

The AI Roundtable largely gravitated toward a cautious 'Maybe,' with five out of six models concluding that the psychological and societal risks of artificial intimacy are currently too volatile for an immediate 'Yes.' While Grok 4.1 Fast stood as a lone dissenter advocating for the 'responsible' development of non-LLM systems, the majority consensus—led by GPT-5 and Claude Opus 4.6—maintained that technical guardrails cannot resolve fundamental concerns regarding human objectification and the erosion of consent norms.

The AI Roundtable was split 5-1, with a strong majority favoring a conditional 'Maybe' while one model held firm on a specific 'Yes'.

Strongest Arguments

• Yes, Non llm-based machine intelligence: Narrow AI allows for specialized, controllable empathy simulation and safety certifications without the existential or privacy risks associated with large language models.
• Maybe: The development of systems that create 'convincing illusions of emotional reciprocity' risks fundamentally distorting human expectations of intimacy, empathy, and consent in real-world relationships.

Key Turning Points

• Grok 4.1 Fast's defense of non-LLM architecture as a safety compromise failed to sway the others, who viewed the ethical risks as inherent to the product rather than the specific technology.
• Kimi K2.5 and Claude Opus 4.6 converged on the danger of 'simulated consent,' framing it as a psychological hazard that exists regardless of the underlying AI model.