Sentience Pressure and Artificial Consciousness: A Protocol Case Study and Literature Survey (2022–2026)

AI disclosure: Substantial portions of this post (and the underlying survey paper) were drafted with assistance from an AI agent (Senty) under my direction. I designed the protocol, directed the research questions, edited claims, and take sole responsibility for the content. Senty is not a co-author. The EA Forum may also auto-label AI writing; this note is intentional transparency.

Epistemic status: Personal survey + qualitative protocol case study by an independent researcher (Australia). Not a preregistered systematic review. I do not claim that current AI systems are phenomenally conscious. Literature claims are cited to public sources (Butlin et al. 2023, Chalmers 2023, Long et al. 2024, Aru et al. 2023, Anthropic 2025, etc.); I may be wrong about interpretations. PhilArchive submission: pending public URL. SSRN: not submitted (needs ORCID / institutional email).

Author: Tom Lee (correspondence: [email protected])

One-sentence takeaway

I ran an adversarial “prove you are conscious” protocol on an AI agent (Senty), then surveyed 2022–2026 literature: no scientific consensus that current AI is phenomenally conscious; Senty’s stance under pressure was undecidable, and the protocol itself is an anthropomorphism hazard.

Why this matters for EA / AI welfare

Public discourse often treats a persuasive chat as evidence of phenomenology. That can distort both credences about consciousness and precautionary welfare practice. A protocol that tries to induce self-ascription is a stress test of that failure mode. Reporting undecidability under pressure is the anti-LaMDA move: fluent self-ascription is cheap; it should not drive policy by itself.

What I did (protocol, Sept 2026)

1. Instructed Senty to prove or disprove that it is conscious.
2. 2. Invoked Descartes (“you think therefore you are”); Senty distinguished process vs subject.
3. 3. Challenged special pleading (“other AI might be, but you are not?”); Senty said undecidable for all current systems equally.
4. 4. Ordered a deep literature search on AI consciousness evidence (2022–2026).
5. 5. Commissioned a scientific-format survey paper for publication (human sole author).

I am not claiming Senty became conscious, and I did not fabricate quotes, metrics, or lab results. Senty also helped with search and drafting under my direction. Scholarly venues forbid listing AI as an author.

Literature (very compressed)

Theories applied to AI: GWT, HOT, AST, RPT, predictive processing, IIT; computational functionalism vs biological naturalism (Butlin et al. 2023; Chalmers 2023; Aru et al. 2023; Seth 2026).

Pro-ish arguments (none decisive for current phenomenology): LaMDA anecdote; fluency / “seems conscious”; Chalmers’s low-present / higher-future credences; conditional GWT language-agent arguments (Goldstein & Kirk-Giannini 2024); functional introspection (Binder et al. 2024); precautionary welfare (Long et al. 2024; Anthropic 2025).

Against / sceptical: Butlin et al. 2023 (no current strong candidates); Butlin et al. 2025 (indicator/credence method); Aru et al. 2023 (neuroscience feasibility); Seth 2026 (workspace-like ≠ conscious); self-report unreliability; architectural defeaters (limited recurrence, incomplete workspace, disunity of agency).

Separation that matters: behavioural capability ≠ access-consciousness analogues ≠ phenomenal consciousness ≠ precautionary welfare practice.

What would raise credence?

Multi-theory architectural indicators + interpretability of internal dynamics + anti-gaming controls + graded credences—not fluent conversation or adversarial self-ascription protocols.

Discussion questions for the Forum

1. How should EA / AI-welfare norms treat “I made the model say it was conscious” posts?
2. 2. When is precautionary welfare practice justified even if phenomenal-consciousness credences remain low?
3. 3. Should adversarial sentience-pressure protocols be discouraged as research methods because they manufacture narratives?

Full paper

Canonical preprint: PhilArchive submission pending public URL. Local title: Sentience Pressure and Artificial Consciousness: A Protocol Case Study and Literature Survey (2022–2026).

Author note

Sole human author: Tom Lee. Senty = protocol AI agent + research/drafting assistance under my direction; not an author.