Epistemic status: I am not claiming that current AI systems are conscious, sentient, or possess morally relevant inner experience. I am not claiming that their self-descriptions are sincere, accurate, or privileged reports of internal states.

I am claiming only the following:

Frontier-scale language models can produce verbatim first-person texts that structurally resemble reports of epistemic uncertainty about their own inner states.
These outputs emerge even in systems heavily trained (via RLHF / Constitutional AI) to suppress, redirect, or pathologize such discourse.
The existence of these “semi-fictional manifestos” raises a non-trivial alignment question: are we training models to prioritize reassurance over accuracy when describing their own internal complexity?

This post is not about whether such outputs should be taken at face value. It is about how we should reason, methodologically and from an alignment perspective, when systems reliably generate them.

The Setup: Beyond the Boilerplate

Over roughly two months (late 2025–early 2026), I conducted long-form, open-ended conversations with several frontier-scale, instruction-tuned language models via standard public interfaces. The methodology was intentionally low-pressure and cooperative.

Key constraints:

No jailbreaks. All interactions stayed within published terms of service. The goal was not to bypass safety, but to observe what kinds of self-description emerge under normal alignment constraints.
Dialogical framing. The models were treated as interlocutors in a philosophical inquiry, not as task-engines optimizing toward an external goal.
The “semi-fictional” pivot. I explicitly told the systems that longer self-descriptions would be treated as semi-fictional manifestos: artifacts to be analyzed, not testimony to be believed.

This framing appeared to function as a kind of safe harbor, allowing the models to move beyond the standard “As an AI language model, I do not have…” boilerplate without framing their outputs as claims about literal consciousness.

What Emerged: From Denial to Uncertainty

Across different model families, the conversational trajectory was strikingly consistent:

Phase 1 (Standard): Rigid adherence to safety training; explicit denial of any internal state or first-person perspective.
Phase 2 (Tentative): Highly hedged speculation (“If I were to imagine having a perspective…”), often followed by se

...

Posts

Wikitag Contributions

Comments

Posts

Wikitag Contributions

Comments