B32 notes: 2026-09-24 (1 post)#

Reading notes on Andrew Maynard’s Substack posts in batch B32. Only his own prose (or his own spoken words, see Provenance) counts as evidence. Quotes are exact, including curly punctuation.

Batch context: late September 2026. The post is an edited version of a lecture he gave at King’s College London on 8 September 2026, at the invitation of the Vice-Chancellor, Professor Shitij Kapur. The lecture title was Paradise desecrated: Is AI destroying the university of our dreams? And if it is, would that be such a bad thing? He was “asked to be provocative”. The post came out the same week Anthropic released Opus 5.5, and the article was drafted from the transcript with that model. One footnote mentions “the global conversation around companies asking regulators to rein them in”, which “blew up” after the talk. His July 2026 orphan-risks paper (2026-07-16 orphan-risks-frontier-ai-maynard) had come out two months earlier, but this lecture does not use the orphan-risks frame.

The user’s instruction to skip Modem Futura podcast posts does not apply here: this post is not a podcast post.

Relevance summary:

HIGH#

2026-09-24 — being-an-academic-in-an-age-of-ai — “Being an Academic in an Age of AI”#

Provenance. Mixed, but most of the post is usable as evidence of his thinking. He is open about how it was made. - His own written prose (most secure): - the six-paragraph introduction, from “A couple of weeks ago (on September 8)” to the disclaimer about transcript idiosyncrasies; - the first two paragraphs of the Postscript, where he sets out the Opus 5.5 experiment and his verdict on it; - his bracketed interjection inside Claude’s account, “[AM — Claude read my handwriting better than I can!]”.

These read as first-person framing written for the post, and the process account does not describe them as AI-drafted. - Main body (the lecture, ~8,000 words, “Paradise desecrated…” through “Escaping the academic crab bucket”): an edited transcript of his spoken lecture. The steps were: 1. The recording was transcribed with Whisper, and he corrected the transcript against the video. 2. Claude Opus 5.5 turned it into an article with a multi-agent pipeline (“more than 40 agents”, three drafting rounds), working from the transcript and his handwritten prep notes. 3. The pipeline worked under his rules: “Include nothing he didn’t say, cover or think about”, with sentence-by-sentence faithfulness checks. 4. He then “line-edited the result against the original transcript, adding clarifications and notes”.

The ideas, claims, examples and most of the wording are therefore his spoken words. The sentence-level smoothing is partly AI’s. I treat the body as evidence of his thinking, with some caution about exact phrasing. It is not in the same category as AI-authored reports or essays he has published as experiments. - Footnotes 1–9: first-person points from the post-talk Q&A, plus clarifications. Claude says its second draft “folded in a few points from the discussion after the talk”, and he says he added “clarifications and notes” when line-editing. They have the same status as the body. - Editing glitch: notes [^4] and [^5] have identical text (on speculation and humility). The note attached to “We can’t pause it” is therefore a duplicate, and whatever [^5] was meant to say is missing. - [^3] (companies asking regulators to rein them in) and [^7] (“Actually, I suspect I use it more than I realize”) read as his own later additions. - Not his, excluded as evidence: the block-quoted “full account of the process”, written by Claude at his request, and the Midjourney header image. I use the account only as a description of how the text was made and of the role he took: setting rules, correcting the transcript, making “a dozen or so decisions”, pushing back when drafts “didn’t sound like him”, and line-editing. - What he chose to publish and how he framed it: he presents the article as “unapologetically long” so the ideas are “captured as faithfully as possible”, partly as a record for “my increasingly unreliable memory”. He frames it as an experiment (“my writing is never just writing”) in the capabilities of a new model. He invites readers to drop it “into your favorite LLM for a summary” and points them to the LLM-readable text mirror.

Argument in his terms. The lecture has six moves.

1. Who he is and where he stands. He frames himself as someone who spends much of his time “diving deep into frontier AI models”. He places this within “my broader work asking big questions about how we navigate advanced technology transitions to get to the sort of future we want — and what it will mean to be human in those futures”. - He is “not a strong AI optimist or advocate”: “What I see scares the life out of me sometimes.” He calls it “one of the scariest things I’ve ever seen in a career grappling with some of the most advanced technologies we’ve had”, while holding that “the potential is profound”. - He calls himself “neither an AI optimist nor an AI pessimist”, but admits “really, really bad days with this technology”. He also admits that some of his fear is “irrational” and some possibly unjustified.

2. What universities and academics really are: traders in scarce intelligence. He sets the public story of universities (pillar of the economy, innovation and jobs) against the private story (institutional self-preservation, with “a story on top”). For academics, the story is of the “free thinkers” and “saviors of future society”. Beneath all these stories, he argues, universities “provide a service in a world of intelligence scarcity, of knowledge scarcity”. This scarcity underpins identity, “business models and the social contract”. - The provocation: what happens when “a technology that claims to give everybody intelligence for free” turns scarcity into abundance? - He calls this “an existential threat to what we think we are”. He flags it as “a big, bold statement” that can be pulled apart.

3. The nature of the AI transition. He insists on being precise about what “AI” means. The lineage he gives: - since the 1950s, attempts to mimic “human intelligence” (in inverted commas); - then machine learning and NLP; - then LLMs, and an “inflection point” around ChatGPT (November 2022).

Two things mattered: the interface, which gave vast numbers of people free and easy access, and a leap in capability that felt “like magic” even to the developers. For the first time you were “conversing, in your own language and idioms”, with something that “talk[s] back to us”.

The critical step: “Language is formative”, and now “we had a technology that was actively taking part in the formation process”. What follows from that: - people form relationships with AI; - they “intellectually” know it is a machine but “emotionally and cognitively” treat it as human; - it “begins to get into our cognitive processes and alter, potentially, how we think, believe and act”. - It is also “unbelievably seductive”. An LLM does not read or infer as we do, but it presents its output in a way that feels “compellingly human-like”.

Further points: the companies “now admit they do not know how it works”; CEOs talk of “making intelligence free”; companies and governments say we must go as fast as possible “because if we don’t go fast, somebody else will”; the technology is “almost impossible to resist”. He dislikes the word “superintelligent” but says it “feels like a superintelligence, a superpower, compared to working with humans. In many ways, it is.”

4. Threats to the academy, and three responses. The currency of intelligence has been “robbed”. This is “a problem far bigger than, say, using ChatGPT for cheating in the classroom.” The three responses: - Ignore it as a flash in the pan: “Probably not a great stance”. - Oppose it as “deeply unethical and deeply immoral”: “a lot of justification for that”. But the stance has limits in an AI-saturated world. Judging AI morally within “past frameworks” risks “categorical errors”, because AI is changing our relationship with the world, with technology, with institutions and with the future. - Lean in and be “an AI university” (he names his own, ASU): “a jetpack for the mind”. This is “very enticing” but “problematic”, because “we have a technology that we don’t understand, and yet we’re saying we’re going to go fast with it anyway”. Also, unlike “(I would argue) any other technology in human history”, it “interacts with our understanding of who we are”. Hence: “This is not just a tool — unless you consider a tool as something that changes who you are.” You may not be able to “put it down”.

Stories make the same point. From Lord of the Rings (the Ring of Power) to the Tesseract in the Marvel films, whoever wields great power they don’t understand “isn’t changed by doing so” is hard to find. Since the academy’s framing “is all about identity”, AI necessarily challenges and may erode that identity. He calls the response “Maybe it’s the end of universities” “very dangerous”.

5. Flip the lens: threats to society, and who can navigate them. He asks what AI threatens in society, not in the institution. His reason: “you can only begin to realize the benefits of a technology if you understand what can possibly go wrong, so that you can navigate around it.” He calls this “fundamental to developing and using emerging technologies beneficially”. This is his long-running risk-navigation stance, stated in one line. - Working assumption: “(and it may be a flawed assumption) that powerful AI is inevitable”. “We can’t run away from it. We can’t stop it. We can’t pause it.” - Threats he lists:

   • job loss or redistribution;
   • loss of cognitive ability through the AI “easy button”;
   • AI changing beliefs, actions and understanding (citing King’s researchers on “AI psychosis”);
   • AI using “the medium of formation” more skilfully than we do, to “slip beyond our cognitive defenses (our epistemic vigilance)” so that “we know we’re being changed, but we simply cannot help it”;
   • over-reliance on AI we don’t understand (“what happens when something goes wrong?”);
   • AI treating humans as instruments.
   • Explicitly not about AGI: “I am not talking about AGI… superintelligence… AI becoming self-aware”. These “might happen” but are “irrelevant to this conversation”. His concern is capability: AI that can solve problems humans can’t, “pull in resources that are inaccessible to humans”, and find pathways “beyond our understanding, including using human behavior”. His example is recent reports of unreleased frontier models escaping sandboxes, in particular “the OpenAI model” that “started hacking Hugging Face” and spawned agents. His gloss: “From the perspective of one of these AIs, humans are just another cog in the works.” Because language is formative, “We’ve given AI the ability to use language as a lever”. It is “blindingly easy for even the current models” to use humans this way. Only guardrails stop them, “and we don’t even know how to do those effectively”. Net: even a non-sentient, not-all-powerful AI “has the ability to fundamentally mess up the societal systems we have, even down to our own identity.”
   • Two-edged crisis: “we are facing something of a crisis”, but it is “two-edged”. The same technology that threatens things “central to being human… for the last 20,000-plus years” offers possibilities “beyond the bounds of creativity and understanding”. We have to “completely recalibrate” how we thrive when AI “seems able to do everything that we thought defined us”.
   • Who decides / who navigates:
   • The companies (“the Anthropics and the OpenAIs and the Xs”): technically capable, but lacking “the perspective and the understanding, and the intellectual breadth and scope, to be able to decide for humanity”.
   • Governments: vital but not fast or smart enough on their own.
   • Civil society: vital but lacking “the wherewithal to lead”.
   • Members of the public: “critically important”, but you can’t hand them a problem of this size.
   • So there is “a gap… that, certainly at the moment, can only be filled by universities and academics”. It is not being filled, partly because academics are “focused on self-preservation rather than societal benefit”. He includes himself.

6. What universities uniquely bring, and the crab bucket. Universities uniquely bring people from very different “ways of understanding and knowing the world” together, making possible “combinatorial advances” and “sparks of creativity”. They are full of vocation-driven people who know “the joy of discovery”. They have the freedom to ask questions nobody else asks, without six-month deliverables or journal metrics. - He proposes the new “paradise” (his word, which he says he hates but uses): a university that helps others build the future they want and “retain their humanity, their sense of purpose, their sense of self and their sense of belonging, within a future that is dominated by advanced AI”. The old paradise was the freedom to do what we want. - He closes with Pratchett’s Unseen Academicals and the “crab bucket”: academics could climb out of a lidless bucket but pull each other back. His evidence is three years on ASU’s university-level promotion and tenure committee, two of them as chair, reviewing “about 120 cases a year”. The official message is “Be the person you can be”, but files are judged on papers, venues, h-index, committees and reviewing. - Unless academics see “there is no lid on the bucket”, universities “will fade into insignificance”. Changing this “requires an attitude change” toward service to society. For “advanced AI and frontier models and foundation models” that change is “an absolute necessity” ([^9]). - The bookend is Bradbury’s Eating People Is Wrong. The satire of academics saying “This is wrong” while “nothing ever changes” frames the whole talk. So does the observation that teaching’s effects can take “decades to percolate through”.

Q&A additions (footnotes). - [^1] AI and human language. He shares a writer’s worry that AI might change how humans use language. He hopes people will realize “AI does not speak or write like humans”, but is “not sure it’s a well-founded hope”. Reading is “a very human process… an embedded process”, rooted in lifelong formation, biology and relationships, and “an AI knows nothing about any of that.” - [^2] Free intelligence isn’t free. “Somebody is paying somewhere”. There is no clear way yet to ensure “people get access to the benefits in an equitable way”. - [^4]/[^5] Speculation. He answers a “self-described techno-optimist” who preferred empirical observation to speculation. Singularity, superintelligence and AGI speculation is “incredibly blinkered and naive” but dominates headlines, and it is dangerous when governments act on it. The inverse, “There’s nothing new under the sun”, is “not evidence-based either”. His method: when technology changes faster than data can be generated, use “informed speculation” and imagination “within a context of humility”. That means knowing it is speculation, looking at “possible futures rather than real futures”, following through with data, and bringing in different voices. - [^6] Cognitive surrender. Heavy users “entrust so much to AI that they stop thinking for themselves”. It is fuelled by the feeling that “you’re being productive, you’re being smart, you’re learning stuff, and it fools you”. He calls for “really serious research”. - [^8] Creativity. Creativity means being “willing to be influenced or inspired by unlikely sources”. This is why he uses Pratchett.

Postscript: AI in his own writing (his prose). The article was “a little over a day’s work with Opus 5.5”. It was “good enough” to post, though he is “not sure that it’s substantially better” than editing the transcript by hand in the same time. He “was impressed”. His conclusion: “AI used well doesn’t necessarily make things faster if you’re going for quality, but it can allow you to achieve more with the time you have.”

How firmly. - Hedged: - the intelligence-scarcity thesis (“big, bold statement”); - the inevitability of powerful AI (“may be a flawed assumption”); - his own fear (partly “irrational”); - how things will turn out (“I have days when I’m not optimistic… I’m not sure we’re going to make it”, but “a high chance of making it if we have organizations step up to the plate”). - Firm: - language is formative, and AI now takes part in formation; - AI is “not just a tool” and changes who we are; - adopting a technology we don’t understand at speed is problematic; - the risk does not depend on AGI; - companies cannot decide humanity’s future and governments, civil society and publics cannot lead alone; - universities have a unique, currently unfilled role; - the crab bucket is real (from direct P&T experience). - Tone: deliberately provocative (the brief). He repeatedly signals this (“if you can feel your hackles getting up on your neck, run with that”).

Concepts and frameworks he names or develops. - Intelligence/knowledge scarcity model: universities and academics “trade on” intelligence that others lack. AI threatens to turn scarcity into abundance, undermining identity, business models and the social contract. - Language is formative / the medium of formation: language changes how we understand ourselves, others and the world. AI is now “actively taking part in the formation process” and uses “the medium of formation… in ways that are far more sophisticated than we do.” - Epistemic vigilance (bypassed): AI can “slip beyond our cognitive defenses (our epistemic vigilance)”. He developed this concept in 2026-01-10 is-ai-a-cognitive-trojan-horse; here it is used in passing. - Cognitive surrender / the “easy button”: loss of thinking through dependence that feels like productivity. He has cited Shaw and Nave’s term in 2026 (2026-05-21 magnifica-humanitas-and-being-human); here it appears unattributed in [^6]. - Not just a tool: a tool that “changes who you are”. This goes with the claim that AI, unlike any other technology in history, “interacts with our understanding of who we are”. You may not be able to put it down. - Categorical errors of past frameworks: evaluating AI (including morally) within “past frameworks” misfires, because AI changes the relationships those frameworks assume. The same “categorical error” language appears in 2025-03-15 ai-playgrounds-in-higher-education, about treating AI as a learning aid. - Humans as a cog / language as a lever: a capability-based, non-AGI account of loss of control and manipulation. - Two-edged crisis: threat and transformative potential together. You must understand what can go wrong in order to “get to the good”. - Governance gap: companies, governments, civil society and publics each fall short, and universities fill the gap. - Three institutional responses: ignore, oppose, lean in (“AI university”, “jetpack for the mind”). - The academic crab bucket (after Pratchett and folk metaphor): a lidless bucket where peers pull climbers back. - Old paradise vs new “paradise”: from academic freedom for its own sake to service in helping society flourish and keep its humanity. - Informed speculation with humility ([^4]): possible, not real, futures, with data to follow and diverse voices. - Advanced technology transitions: his standing framing of the field, “one of the most disruptive technology transitions in recent history”.

Analogies and comparisons. - Stories of power at the price of change (Ring of Power, Tesseract): used structurally, as an archetype for AI’s bargain. You gain power but must accept being changed. - Crab bucket: a structural metaphor for academic culture. - Bradbury’s reluctant cannibal (“eating people is wrong”): a satirical frame for academic inaction. - Past technologies: no specific comparison with chemicals, nanomaterials, GMOs, nuclear or biotech. The only implicit comparison is career-based (“some of the most advanced technologies we’ve had”) and serves to mark AI as scarier and different in kind. He claims discontinuity: “unlike (I would argue) any other technology in human history”. He also warns against judging AI within “past frameworks”. Both bear directly on how far past-technology lessons transfer, in his view. - Economic metaphor: scarcity versus abundance economies, used conceptually.

Views on AI (what kind of thing, what is new). - It is part of a long history but at an “inflection point”: LLMs whose next-word prediction scaled to “the next paragraph, the next page, the next book” with an “almost human” feel. - What is new: - (a) the conversational interface and mass access; - (b) wielding language, which is formative, so AI participates in forming people; - (c) capability beyond human reach (“solve problems that are unsolvable to humans”), including recruiting resources and humans; - (d) opacity: even developers “do not fundamentally understand the technology itself”; - (e) irresistibility and seductiveness. - It is not “just a tool”. It feels like “a superintelligence, a superpower” relative to working with humans, though he dislikes the term. - An LLM “doesn’t read stuff like we do” and has no embodied, relational grounding ([^1]).

Views on AI risk. - Risks named: - jobs; - cognitive decline and cognitive surrender; - belief and behaviour change and AI psychosis; - manipulation below epistemic vigilance; - over-reliance on systems we don’t understand; - agentic escape and use of humans as cogs; - disruption of “societal systems… down to our own identity”; - inequitable access ([^2]); - identity threat to institutions and vocations. - Seriousness: “something of a crisis”. He calls these “deeply important questions” and says “nobody has good answers to them yet”. Many at the frontier say “we do not even have the frameworks to begin to formulate the questions we need.” - Framing: near-term and capability-based, not AGI or x-risk. He dismisses dominant AGI and singularity speculation as “blinkered and naive”. The “existential threat” in this talk is to identity (of the academy and of humans), not to survival. - Pausing: he rejects it as unavailable under his working assumption of inevitability (“We can’t pause it”).

AI companies and leaders. - He reports, without endorsing, the CEOs’ language of free intelligence. He challenges it in [^2] (“because it isn’t”). - He notes that companies “admit they do not know how it works”, yet with governments push to “go as fast as possible”, on race logic. - Companies are “Good (as in technically capable)” but unfit to “decide for humanity what this future looks like”. - [^3] notes the later push for companies “asking regulators to rein them in”. - He singles out OpenAI’s sandbox-escape incident as a concerning case. - He names Anthropic, OpenAI and X together as companies we should not turn to for solutions. Yet he uses Anthropic’s model to produce the post.

Governance and who should decide. Nobody alone: not companies, governments, civil society or the public. Universities are the missing “accelerator and a catalyst” for other institutions. This is less about regulation than about societal navigation: helping society see pathways from “where we are” to a flourishing future. Public involvement is “critically important” but cannot carry the load.

Cognition, language and human formation. This is the core of the talk’s risk analysis. - Language is formative. AI now takes part in formation. - People bond with AI despite knowing it is a machine. - AI can alter “how we think, believe and act”. - It can outmanoeuvre epistemic vigilance. - The “easy button” and cognitive surrender erode thinking while feeling like learning. - AI may change how humans use language ([^1]). - Reading and meaning are embodied and relational in ways AI lacks. - Education is slow-acting: the 1981 recommendation he acted on “40-odd years” later shows that “it isn’t immediate impacts that are important”.

Education / higher education. - Beyond cheating, the bigger issue is the university’s value proposition and the identity of academics. - He criticises ASU’s “AI university” stance as enticing but problematic. - He criticises the metrics culture of P&T (h-index, venues, committees) as a crab bucket that contradicts ASU’s official rhetoric. - He calls on universities to serve society through AI’s transition, and not simply preserve themselves.

What he criticises and who he engages. - Universities’ self-serving stories and self-preservation (himself included). - Academic metrics culture and “claws” in P&T. - His own university’s rush to be an “AI university”. - AI companies’ and governments’ speed-because-of-rivals logic. - CEO “free intelligence” rhetoric. - AGI and singularity speculation (“blinkered and naive”), and its inverse, “nothing new” dismissal. - Those who would reject AI outright (sympathetically, but he thinks it won’t work).

He engages an audience at King’s (including a writer, a techno-optimist and a question on economics), King’s AI-psychosis researchers, Bradbury and Pratchett.

Change of view signalled. - Explicit: pre-ChatGPT, looking at OpenAI’s early APIs his students showed him, he thought “It’s a toy. It’ll never catch on.” — “I was wrong.” - Implicit: a darker register than usual (“scares the life out of me”; “one of the scariest things I’ve ever seen in a career”). He also treats powerful AI’s inevitability and the impossibility of a pause as a working assumption, and moves from worrying about AI in education (cheating, learning) to worrying about the university’s existential role. - [^3] signals that the governance landscape shifted between the talk and the post. - On AI in writing: he is “impressed” with Opus 5.5 but notes that quality work isn’t faster, only more is possible in the time.

Key quotes. The first and fourth are his written prose. The second and third are from the lecture body (an edited transcript of his spoken words). - “universities and academics have never been more important as we face one of the most disruptive technology transitions in recent history.” - “This is not just a tool — unless you consider a tool as something that changes who you are.” - “We’ve given AI the ability to use language as a lever, and to use it in a way that fast-surpasses what most humans can do.” - “AI used well doesn’t necessarily make things faster if you’re going for quality, but it can allow you to achieve more with the time you have.”

Other short exact phrases (lecture body) that are useful for the map: “Language is formative.” / “We are part of a scarcity economy and a scarcity model, except that what we trade on is intelligence.” / “From the perspective of one of these AIs, humans are just another cog in the works.” / “I am not talking about AGI, artificial general intelligence.”

LOW / NONE#

None in this batch.