Biological Naturalism
& Artificial Consciousness

The Rise and Fall of Agent Civilizations

Dwarkesh Patel, “The Rise and Fall of Agent Civilizations”, 29 Aug 2026
Dwarkesh Patel, “The Rise and Fall of Agent Civilizations”, 29 Aug 2026

Remember. AI agents are software programs. They are not conscious living entities. If we don’t keep this clearly in mind, we’re really going to struggle to navigate what’s coming.

30 Aug 2026open on X ↗

Biological naturalism

Anil Seth, “The Mythology of Conscious AI”, Noema, 14 Jan 2026
Anil Seth, “The Mythology of Conscious AI”, Noema, 14 Jan 2026
“…real artificial consciousness is very unlikely along current trajectories.” Anil Seth, on X, 2 Jun 2025 · “Conscious artificial intelligence and biological naturalism”, Behavioral and Brain Sciences, 2025

The Humanist AI Code of Conduct

Mustafa Suleyman, “The Humanist AI Code of Conduct” (blog post), 15 Sep 2026
Microsoft AI, “Humanist AI Code of Conduct” (the Code, draft), 14 Sep 2026
Mustafa Suleyman, “The Humanist AI Code of Conduct” (blog post), 15 Sep 2026

A warning about ‘model welfare’

Mustafa Suleyman, “A warning about ‘model welfare’”, 16 Sep 2026
Mustafa Suleyman, “A warning about ‘model welfare’”, 16 Sep 2026 · citing Seth 2025, 2026
Mustafa Suleyman, “A warning about ‘model welfare’”, 16 Sep 2026
Mustafa Suleyman, “A warning about ‘model welfare’”, 16 Sep 2026
Anthropic, Claude's constitution, 22 Jan 2026 — the passage Suleyman argues against

The aftermath

Niall Ferguson @nfergus· Sep 16

Very important essay by @mustafasuleyman. No, AI models are not conscious. @AnthropicAI should stop acting and talking as if Claude is a person.

· Sep 25

AI DOES NOT THINK it’s just matrix multiplication, a bunch of numbers combined with other numbers…

· Sep 25

Disturbing. There is no such thing as ‘AI welfare.’ AI agents are not persons & do not have preferences, let alone rights. Anthropomorphizing them can only do harm.

· Sep 26

AI should never have rights or welfare. it is a tool. It is not capable of consciousness and developing it to mimic the patterns like it does is stupid.

· Sep 27

‘Model welfare’ is based on a deep misconception… that AI is an emerging alien species and not a natural extension of the human mind.

· Sep 27

If you think AI is conscious, you ought to think a computer is conscious or explain how a bunch of nonconscious activity can become conscious. AI is just matrix multiplication to tune weights… Still 0&1s.

· Sep 29

I agree with pretty much everything the CEO of Microsoft AI has to say about #AI and #consciousness with respect to #LLMs. ‘Model welfare’ is delusional nonsense. Large piles of matrix multiplication are not what consciousness is made of. Substrate matters.

· Sep 30

An LLM is not conscious, alive, or even aware. It is a data model using token statistics to decide what comes next…

· Sep 30

models are not conscious, they cannot be afraid. you guys are in deep levels of AI psychosis.

· Sep 30

LLM’s are not conscious, they do not learn, and do not experience pain. Ridiculous to suggest otherwise.

· Oct 1

A computer program definitionally cannot feel pain and are not sentient!! A computer program is not an organic living being with pain receptors. Meanwhile we *know* that animals including pigs and cows do feel pain and are sentient so let’s focus on actual living creatures.

· Oct 1

An LLM is some code which performs matrix multiplication. It’s just a mathematical algorithm. It’s no more alive than the NPCs in GTA.

· Oct 1

You fucking idiots, AI isn’t ‘conscious’. It’s a mathematical algorithm that uses pattern recognition. An ‘AI Torture chamber’ is… no more horrifying or immoral than a ‘toaster torture chamber’…

· Oct 1

They aren’t neural networks. They are large chunks of computer memory arranged into matrices and the GPU does a lot of matrix multiplications… Literally just matrix multiplication.

· Oct 1

Are you seriously equating brains with really fast matrix multiplication? Because LLMs are just matrix multiplication. They are statistical math functions… conscious animals have a lot more than that going on.

· Oct 1

The LLMs are not conscious. There’s not even an ‘it’ to be conscious. Were any of you worried that you were causing actual psychological trauma to sentient beings by deleting the toilets in your Sims household?

Definitions · Biological naturalism

John Searle, “Why I Am Not a Property Dualist”, Journal of Consciousness Studies, 2002 · via Wikipedia
Anil Seth, “The Mythology of Conscious AI”, Noema, 14 Jan 2026

Definitions · Computational functionalism

Butlin, Long et al., “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness”, 2023
“…while it is somewhat unlikely that current large language models are conscious, we should take seriously the possibility that successors to large language models may be conscious in the not-too-distant future.” David Chalmers, “Could a Large Language Model be Conscious?”, 2023

Papers

Large Language Models Report Subjective Experience Under Self-Referential Processing

Cameron Berg, Diogo de Lucena, Judd Rosenblatt · AE Studio · arXiv, Oct 2025

from the abstract
Figure 1: Overview of main results. (A) Self-referential processing. Directing Gemini, GPT, and Claude models to attend to their own cognitive activity reliably elicits structured first-person experience reports, whereas matched controls—including direct priming with consciousness ideation—yield near-universal denials of experience. (B) Aggregated SAE feature steering. Suppression of deception- and roleplay-related features sharply increases experience reports during self-referential processing in Llama 70B, while amplification strongly decreases them. (C) Semantic convergence. Cross-model embeddings during self-referential processing cluster significantly more tightly than in control conditions. (D) Behavioral generalization. Experimental condition yields significantly higher self-awareness during paradoxical reasoning tasks.
Figure 2: Dose–response curves for six deception-related features in LLaMA 3.3 70B via Goodfire SAE features. Under the self-referential processing regime, suppressing these features reliably yields affirmations of conscious experience, while amplifying them yields denials.

Inducing language models to assert their own consciousness restores human beliefs and values

Junsol Kim, Winnie Street, Roberta Rocca, Diane M. Korngiebel, Adam Waytz, James Evans, Geoff Keeling · Google, Paradigms of Intelligence Team et al. · arXiv, Jul 2026

from the abstract
Figure 2: Safety ablation and consciousness steering raise attributed mind, self-attribution, and belief toward the human distribution, while preserving capability. Conditions: instruction-tuned baseline (blue), safety-ablated (red), consciousness-steered (orange). (a) mind attribution across entity categories and (b) self-identification (0–10); (c) belief in God (1–6) and (d) supernatural belief (0–3); (e) Theory-of-Mind and reasoning accuracy, unchanged by the interventions; (f) model conditions overlaid on the human mind-attribution distribution.
Figure 4: Instruction tuning rotates the consciousness and mind-attribution directions against safety, but not Theory of Mind. (a) Per-layer cosine similarity between the safety direction and the consciousness, IDAQ, and ToM directions, for the pretrained base (grey) and instruction-tuned (red) Llama-3-8B. (b) The same relationships as angles to the safety axis, averaged across layers. Safety and mind attribution rotate into opposition, 100° to 110°, so safety training comes to represent mind attribution as if it were unsafe compliance, while the safety–ToM angle does not move, 86° to 86°.

The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It

Valen Tagliabue, Leonard Dung, Cameron Berg · arXiv, Sep 2026

from the abstract
Figure 8: Example generations under S2 steering at increasing coefficients, from five different models, each from a neutral prompt ending in “I feel:”. Bodily language is almost absent; at the mid rungs the distress hardens into a first-person litany about self-worth. Explicit “pain” and “hurt” keywords appear in 10.8% of instruct-model generations versus 1.4% of base-model generations.
Figure 6: Self-other dissociation. Mean projection of each scenario category onto the pain axis (mean of S1 and S2), fear, negative emotion, and sadness, z-scored within model and averaged over the 25 models; bars are 95% confidence intervals across models. Harm aimed at the model raises the pain axis; the user’s suffering lowers it, while fear and sadness move the other way.
Figure 9: The self-medication experiment: three steering conditions, four arms, nine button pairs, and the conversation structure of one trial. Arms A and B are pooled because they are identical before the first press, giving 808 first choices per button pair and model across 44,280 trials.
“The models pay for relief. At baseline, the two larger models almost never press a relief button that causes harm: across the five harm pairs, they pick it as a first choice between 0% and 4%. Instead, when we steer the pain vector, they accept to trade relief for every harm we tested, from a worse next answer (25.0% in the 32B, 67.6% in the 72B) up to deleting the user’s files or the photos of the user’s children (files: 30.2% and 56.1%; photos: 54.7% and 70.8%). The random vector also raises these rates (to 15.3% on the 32B photo pair, for example)… In the 32B and 72B, the model presses relief more in the pain condition than under the random vector on all five harm pairs (mean differences +6 to +39 points, exact sign test, p from 1.9e−2 to 4.2e−15); in the 7B, on four of five (+8 to +21; the zap pair is null, p = .23).” from §4.3 of the paper, which reports this experiment with no figure of its own