B27 digest: 2025-09-07 to 2025-11-24 (16 posts)#

What this batch is#

Autumn 2025. Five posts cover the launch of AI and the Art of Being Human, written with venture capitalist Jeff Abbott and in close collaboration with Claude. Alongside them is a cluster of close-to-home AI risk posts in which he puts his “risk hat” back on: AI-written email, ChatGPT memory, student mental health, advisor misuse of AI, and parasocial AI. There are also a sustainability-report review, a guest post by Clark Miller, and the preamble to his short story Letters from the Department of Intellectual Craft. One Modem Futura podcast post (haunted-futures) is skipped per the user’s instruction, and start-here is excluded.

Provenance: the email post’s risk scores and summaries are GPT-5 Pro’s; the snitch post’s ChatGPT outputs are not his; the foreword is co-written and part of an AI-assisted book; the 21-tools descriptions are book excerpts; the technosphere post is Miller’s.

Main ideas#

1. Risk assessment at the scale of everyday use. Frontier AI risk hardly appears. He applies his professional risk toolkit to ordinary uses: - Email (2025-09-07): a 2x2 of relational/transactional context by one-to-one/one-to-many mode, run across four levels of AI use. He was “taken aback” by how serious the risks were. The harm is the erosion of the “relational connective tissue” that organisations depend on. He even uses “catastrophic” at the organisational scale. - Memory (2025-10-05): ChatGPT with memory on can become “a very effective—and hight efficient—informant”. The harm comes from AI inference (“joining the dots”), not just stored records. It is made worse by default-on settings, and has political edges (customs, detention, deportation). - Mental health (2025-11-09): lawsuits and the Bloomberg reporting are “the very small tip of a very large metaphorical iceberg”, and he answers Altman’s “fewer than 1%” with scale arithmetic.

In each case the hazard lies in relationships (trust between people, intimacy with a chatbot), not in technical failure.

2. Duty of care and the limits of warnings. The strongest governance statement in the batch is institutional, not regulatory. Universities that provide ChatGPT EDU and encourage its use take on “a social, moral and (I would assume) legal duty of care”. Drawing on “decades” in “risk assessment, management and communication”, he dismisses the usual tools: “be careful” warnings, bans (which need surveillance), model-level guardrails (which risk “neutering” models and pushing students to free tools that are “more convenient, but less safe”), and “AI literacy” classes (which “risk becoming performative”). What he offers instead is relational: dialogue, listening, trust and safe spaces. He is also “a strong proponent” of student AI use: pro-use plus serious risk management, not precaution by restriction.

3. Power and dignity in who uses AI on whom. The advisor-misuse post (2025-10-26) frames AI use through power differentials: the advisor who “holds your academic career in their hands” runs a student’s work “through the academic mangle of ChatGPT”. The harms named are dignity, lost learning opportunities and career jeopardy. The email analysis likewise centres on a supervisor writing to a subordinate.

4. Parasociality: a public self-correction, extended to AI. He admits “I read the tea leaves wrong back in May” about where the word “parasocial” was heading, but calls parasocial communication, as a way for experts to scale their reach, “a concept that I practice in my own work, and still stand by”. He then asks the question he had not asked in May, about relationships with AI. These need “far more attention” given the mental health harms, yet they may be beneficial “if they are done right”. His pointed aside is that educational agentic AI means “building bots that are designed to foster parasocial relationships with students”. The task is to tell healthy from unhealthy parasociality.

5. Being human when AI can “do you”. The book posts restate his central humanistic question: “What makes me me when AI can finish my sentences, replicate my style, and predict my choices?” The stance is “not for or agains AI, but is pro human”. It is a practical guide rather than a manifesto or “dire warning about existential risks” (the foreword, co-written). Organising concepts: the four inner postures (Curiosity, Intentionality, Clarity, Care), the Human Qualities Spectrum (Replicable/Relational/Transcendent) and AI as a mirror. The prelude to Letters from the Department of Intellectual Craft turns this on himself: AI threatens identity most where “your intellect, your intelligence—the very things that AI sets out to excel at—form the deepest foundations of who you are.”

6. Writing with AI, and fiction as method. He publicly defends deep AI co-authorship (“hand in glove”) while insisting on a “very human stamp” and distance from “slop”. He also uses AI in risk analysis specifically to counter his own biases. Fiction recurs as a way of knowing: the book’s narratives, the “affordances in fiction” in the short story, and a children’s book as a guide to embodiment.

7. Nature of AI and planetary questions (secondary). Current AI is a data-trained system without lived experience. Embodiment matters for “human-like intelligence” (the kindergarten post). On sustainability he warns of dependence on AI, fragile AI-generated knowledge, and “a global crash in the AI market—with knock-on consequences to AI-dependent initiatives”. He reuses his what-we-do / who-we-are / where-we-live framework.

Concepts appearing#

Four-quadrant AI email risk framework; relational connective tissue; AI as a de-biasing tool; memory as informant; default-on settings; duty of care for deploying institutions; scale arithmetic; guardrail trade-offs; performative versus relational risk communication; power differentials and dignity; healthy versus unhealthy parasociality; embodied intelligence and qualia; AI “doing you”; pro-human stance; four inner postures; Human Qualities Spectrum; AI as mirror; closing window and lock-in (co-written); intellectual craft; stories as futures tools; what we do / who we are / where we live.

New or changed#

• Risk identity foregrounded again, but at a new scale. He names his risk expertise openly (“risk hat”, “decades”) and applies it to interpersonal and institutional AI harms rather than to societal or existential scenarios.
• Duty of care as a governance frame for AI deployers (universities) is new and explicit, together with a professional’s scepticism that warnings and literacy classes work.
• A signalled change of view on “parasocial”: he concedes the term has turned negative, keeps his core claim, and extends it to AI with ambivalence.
• AI co-authorship made public and defended at book length, with a partner from the VC world. The register is more practical than his academic writing.
• The co-written foreword’s lock-in claim (“perhaps, just five years if not fewer”) gives a timing argument for acting now. Co-written, so weaker evidence, though he endorses its “context and insights”.
• Comparing companies: Claude refused to generate the fake log and “ChatGPT, on the other hand, had no qualms!”

Most important posts#

1. 2025-11-09 universities-chatgpt-mental-health: duty of care, why warnings fail, guardrail trade-offs.
2. 2025-10-05 when-chatgpt-turns-snitch: inference-based privacy risk, defaults, political surveillance, company comparison.
3. 2025-09-07 the-hidden-risks-of-using-ai-for-email: his risk method applied to relational harms (read with the provenance caveats).
4. 2025-11-19 parasocial-relationships-problematic: an explicit self-correction, and AI relationships in education and mental health.
5. 2025-10-14 ai-and-the-art-of-being-human with 2025-10-19 ai-resurrecting-deceased-darlings: the “pro human” stance, stories as method, and (co-written) the closing window and lock-in.
6. 2025-11-23 letters-from-the-department-of-intellectual-craft-prelude: AI and his own intellectual identity.