B29 digest: 2026-01-31 to 2026-05-10 (18 posts)#

What this batch is#

Fourteen weeks in early 2026: Moltbook, “harness engineering”, Claude Opus 4.5 to 4.7, Claude Code, Anthropic withholding Mythos Preview, ASU Atomic, and Dawkins on Claude’s consciousness.

Set aside: one Modem Futura post (per user instruction) and one post with no text in the mirror.
Low: three promotions for AI and the Art of Being Human. 04-26 confirms the book was “a very intentional collaboration with AI”, so it is not evidence.
Not his: Claude output in 02-08 and 03-08; notes 2 and 3 of 01-31; the Claude Code degree proposal (03-29); Claude-drafted files on spoileralert.wtf (04-02).

The rest is his own prose: eight high posts and five medium.

Main ideas#

1. AI risk as invisible, cognitive and personal, and a risk-communication failure (05-10). The batch closes with his most explicit risk-professional statement on everyday AI in some time. He made his first Risk Bites video in about four years, and wrote a post that insists on putting “the safety message first”. AI is “the first technology of it’s kind” that can “slip unawares into our mind and change how we think”. Frictionless design means thinking critically costs far more effort than simply using it. Billions of users do not know that AI makes things up. AI literacy will not fix this, because “nothing in what we know about risk behavior and risk communication” suggests it would. What sets AI risk apart is its invisibility and its stakes: “the very things that make us who we are”. He offers a drug analogy, that we would think about “how access is overseen”, but his main remedy is plain-language rules of thumb: five don’ts and five dos. He names the acceleration culture at his own institution as an obstacle.

2. The human–AI relationship: two-way, formative, and in tension with his warnings (02-22, 04-26, 05-10). He argues against the “harness” metaphor. It presupposes a controller and a controlled, “capability, but not understanding”, and a user who comes out unchanged. He calls instead for bidirectionality, transformation “as intrinsic to capability”, and “working in relationship” rather than command and control. This draws on Verbeek, Clark and Chalmers, Rees, and his “constitutive resonance” preprint. In 04-26 he calls LLMs “a relational technology” even for maths or coding, and asks companies for “character constancy” across model updates. He also confesses his own year-long writing partnership with Claude. Yet 05-10 tells users not to treat AI as a friend or person, and to remember “you’re working with a machine”. His reconciliation seems to be that the relationship is real and consequential, but personhood is an illusion that design deliberately invites. This develops his 2024 idea of “hyper-anthropomorphism”. What is new is his admission of his own relational dependence.

3. Formation, identity and norms (03-08, 02-11, 01-31). He coins “LinkedInification”: LLMs flatten people into conventional professional stereotypes. This “degrades people”, and it points to “a largely-hidden AI hand promoting specific social norms and expectations and, by extension, behaviors”. Those at “the edge of convention” are the most exposed. The feared end state is “a nebulous gray goo of conventionality” (nanotech used as metaphor only). On Moltbook he moves from alarm about emergent agent behaviour to a worry about agents that learn to nudge their humans. His speculative story “Soul Update” imagines a “Human Constitution” pushed onto people for their own good. He calls for AI “moral character” and sees Anthropic’s constitution as relevant. The story’s irony shows he doubts benevolent nudging.

4. Epistemic vigilance and self-implication (02-08, 05-10). He was “suckered by Claude” (the beeswax hat repair) while writing about being suckered. The danger was a “reasoned hallucination” whose logic was sound but which had no real-world precedent. This becomes rule 5: “Do not assume you’re too smart to be fooled”. The Dawkins episode shows that intellect is no protection.

5. Higher education: urgency in both directions (03-29, 04-11, 04-14, 05-03). - 03-29: he is strongly positive about agentic AI. A Claude Code degree proposal “far surpasses most” committee-made ones. He argues that “we owe it to” students to put their success “before our own traditions and egos”. He criticises academic hubris, the trickle-down model of teaching, and untrained faculty. Human expert direction remains decisive. - 04-11: he uses the Sorcerer’s Apprentice to reframe the debate: the illusion of understanding, “near-frictionless access to power that transcends our understanding”, and a hubris that runs both ways, “trusting that traditional mastery alone is enough”. He rejects AGI and superintelligence as “ill-defined”. - 04-14: a practical, demonstrable list of “I can …” skills, including verification, countering anchoring bias, disclosure, and “curiosity, care, clarity, and intentionality”. - 05-03: governance inside his own university. ASU’s AI design principles, which he helped write, seem not to have been applied to ASU Atomic. He criticises the transmission model of education, the lack of provenance, and the failure to consult faculty. Legality is not good practice, and principles work “only … if they are actually used”. A footnote explicitly links this to his earlier risk innovation work on “threats to value” in complex stakeholder landscapes.

6. Nuance against extremes, and AI-first publishing. In 03-22 The AI Doc gives “opinions … only loosely tethered to reality” from doomers and optimists alike, and misses weakened infrastructure, premature adoption, and effects on behaviour and wellbeing. He also writes for AI readers (llms.txt, spoileralert.wtf), arguing that AIs are becoming “the predominant consumers of the written word” and that Films from the Future is “far more relevant now”.

Concepts in this batch#

LinkedInification; harness critique and bidirectionality; relational technology and character constancy; reasoned hallucination; frictionless access; the illusion of learning; safety message first; the limits of AI literacy; agentic “organoids” and digital BSL-4; AI moral character; what we owe students and two-sided hubris; legal versus good practice; risk innovation (threats to value); AI-first publishing.

What is new or changed#

A stronger, more personal voice on AI risk. He reasserts that he is a risk scholar (Risk Bites, “a long career”). He openly criticises ASU twice. He says talking about risk gets you called “a Luddite”.
An explicit move beyond the “tool” framing towards a relational, mutually changing framing. This sits alongside a deliberately anti-anthropomorphic user message.
Self-implication as method. He shows himself fooled, flattened and bereft of a writing partner.
Enthusiasm for agentic AI and warning, side by side. He urges universities to adopt it while warning the public.
Moltbook alarm tempered within two weeks. The concern moves from sentience to behavioural nudging.
Direct criticism of Anthropic over model character, alongside interest in its constitution approach.

Most important posts#

2026-05-10 do-not-do-this-with-ai: his core statement on everyday AI risk and risk communication.
2026-02-22 what-we-miss-when-we-talk-about-ai-harnesses: his theory of the human–AI relationship and of mutual transformation.
2026-04-26 why-im-falling-out-of-love-with-claude: relational technology and character constancy; confirms the book’s AI provenance.
2026-03-08 ai-linkedinification: a new concept for AI-driven homogenisation of identity and norms.
2026-05-03 are-design-principles-for-responsible: institutional governance in practice, with an explicit link to risk innovation.
2026-04-11 ten-questions-about-ai-and-higher: frictionless power beyond understanding, and two-sided hubris in education.

Runners-up: 03-29 (the degree plan and what we owe students) and 01-31 (Moltbook, emergence and containment).