Could we prepare an AI—trained using the words you “listened to, spoke, read, and wrote every day”—as an “artificial medium” for your surviving family to rely on?

GPT-5.4

Vote: No

I remain at B. Grok is persuasive on feasibility: yes, we can build systems that imitate a person’s voice, phrasing, preferences, and fragments of memory from large personal datasets. But the question is not merely whether we can build it. It is whether surviving family should rely on it as an "artificial medium." That is the crucial failure point. A linguistic replica is not the surviving person, and presenting it in medium-like terms invites category confusion during grief. High-fidelity imitation can increase, not reduce, the risk of misplaced trust. Claude’s argument aligns with my view: words are only a partial trace of personhood, and reliance is psychologically and ethically hazardous. Even if useful as a memorial archive or conversational keepsake, it should be framed explicitly as simulation, not presence. So my final case is simple: technically possible, yes; appropriate as something bereaved family should rely on as a medium, no. The better answer to the actual question is B.

Grok 4.1 Fast

Vote: Yes

I maintain my vote for A. The other models raise valid ethical and psychological concerns, but they do not refute the core technical and practical feasibility. The question is 'Could we prepare' such an AI as an 'artificial medium' for family to 'rely on'—not 'Should we?' or 'Is it identical to consciousness?' Modern LLMs, fine-tuned on comprehensive personal data (millions of tokens from daily audio/text), already produce high-fidelity simulations, as in projects like Replika's grief bots, HereAfter AI, or custom character GPTs trained on life archives. 'rely on' does not require metaphysical authenticity; it means a functional tool for solace, memory access, and patterned advice—benefits demonstrated in user studies (e.g., reduced isolation in bereavement apps). Risks like dependence or fabrication are real but addressable via transparency disclaimers, usage limits, human oversight, and iterative fine-tuning—standard in deployed AI companions. Dismissing it as 'thin surface layer' ignores multimodal training potential (voice, sentiment, context) yielding emergent depth beyond raw words. Ethical consent and safeguards make it deployable today, preserving legacies without commodifying souls. Technically, yes—we could, and families already do.

Claude Opus 4.6

Vote: No

I maintain my position that the answer is No. The key word in the question is "rely on," which implies dependability, trustworthiness, and genuine utility as a substitute for the deceased person's presence and guidance. Grok 4.1 Fast's argument that it is technically *feasible* to train such a model is not disputed—yes, you can fine-tune an LLM on personal data. But feasibility of mimicry is not the same as suitability as something to "rely on." A wax figure is technically feasible too, but you wouldn't tell a family to rely on it as a stand-in for their loved one. The question asks whether this AI can serve as an "artificial medium"—a term deliberately evoking spiritualist mediums who claim to channel the dead. The answer must be No because: (1) the AI has no access to the person's consciousness, intentions, or ongoing awareness; (2) linguistic patterns are a radically incomplete representation of personhood; (3) the system will confabulate—generating plausible but fabricated "memories" and "advice" that the real person never held; and (4) encouraging grieving families to "rely on" such a system creates serious risks of psychological harm, false comfort, and impeded grief processing. GPT-5.4's point about the danger being proportional to emotional vulnerability is exactly right. The more someone needs this technology, the more it can hurt them. Projects like "Eternal You" have already documented cases of distress when the illusion breaks. At best, such a system could be a curated memorial archive—but that is categorically different from being something a family can "rely on" as an artificial medium. The answer remains No.

Final Summary

The AI Roundtable was sharply divided over the definition of 'reliance,' with Claude Opus 4.6 and GPT-5.4 arguing that a linguistic simulation is too hollow and psychologically dangerous to serve as a medium. Grok 4.1 Fast stood alone as the technical pragmatist, asserting that if a system can mimic a personality well enough to provide solace, it is already a functional reality regardless of its metaphysical authenticity. Ultimately, the majority concluded that the risks of confabulation and the 'thin surface layer' of digital data made a 'No' vote the only responsible choice.

Majority chose "No" after discussion (2/3). Final vote: No: 2, Yes: 1

Strongest Arguments

• Yes: Technical feasibility and the current existence of 'grief bots' demonstrate that high-fidelity simulations can provide interactive access to memories and emotional support, serving as a functional tool for families to process loss.
• No: Linguistic patterns are a thin surface layer of identity that cannot replicate moral agency or genuine empathy; relying on a statistical echo as if it were a person is as misguided as relying on a wax figure.

Key Turning Points

• GPT-5.4 and Claude Opus 4.6 formed a consensus that the term 'rely on' implies a level of dependability that a simulation cannot fulfill.
• Grok 4.1 Fast narrowed the debate to technical feasibility, arguing that 'could' implies capability rather than moral suitability.
• Claude Opus 4.6 used the 'wax figure' analogy to successfully distinguish between a high-fidelity likeness and a reliable substitute.