AI frontier experiments
We work at the cutting edge of AI model use: frontier models as research partners, co-authors, game designers, and knowledge infrastructure — “exploring what is possible quite deeply while also considering the complex surfaces around risks” (elicited from Andrew Maynard, 2026-08-05). Read that sentence as one practice, not two. The exploration is how we come to see the risk surfaces; the risk work is what makes the exploration serious. Andrew’s calibrated stance: “skeptical of both the safety absolutists and the move-fast-and-break-things crowd… the interesting and difficult work is in the space between” (andrewmaynard.net/llms.txt, 2026).
Two framing notes. First, this is a core activity of the initiative, not a hobby beside it — say so explicitly, because prestige hierarchies would otherwise file “built a game with an AI” under trivia. Second, we are not an AI initiative: we work across the full range of advanced technology transitions (Identity); AI dominates our current focus because of the moment, not by definition.
Why we work this way
Andrew’s operating principle is that he cannot write or teach about AI without intimate familiarity with what these systems can actually do (andrewmaynard.net/llms.txt, 2026). Since May 2026 he has been on sabbatical (announced through August 2027), used largely for public experimentation with frontier AI systems (source: andrewmaynard.net/llms.txt, 2026). The experiments below are illustrative, not exhaustive.
A formal “Use of AI” statement in a federal report (established, 2025)
The Future Travel Foresight Catalyst final report (Maynard & Leahy, 2025; rosap.ntl.bts.gov/view/dot/91942 — see Policy and quiet influence) carries a front-matter Use of AI statement, verbatim in part:
“This report employed AI-assisted drafting tools as part of an experimental human-AI collaborative methodology being developed within ASU’s Future of Being Human initiative. The AI system served as a drafting assistant… All intellectual contributions, analytical insights, and strategic decisions originated from the authors.”
This names the initiative, in a federal document, as the home of an experimental human-AI collaborative methodology — with transparent attribution of responsibility. The report’s 2026–27 roadmap extends it (planned): a “Tech Futures Briefs” series of concise foresight briefs “generated through an advanced AI-assisted methodology” combining LLMs, external research, and expert guidance and editing (DOT report, Section 6, 2025).
Frontier-model collaborations (established, 2025–2026)
Documented at andrewmaynard.net (digested 2026-08-05):
- AI and the Art of Being Human (2025, with Jeffrey Abbott) — written intentionally in close collaboration with Anthropic’s Claude; a book about thriving with AI, made the way it argues.
- The academic-paper experiment — Constituting Responsibility (2026), a paper written by Anthropic’s Claude under Andrew’s guidance, retaining Claude’s first-person voice, as an experiment in “critique from within” (preprints page).
- Hyperbubble (built July 2026, launched August 2026) — a free browser game at playhyperbubble.com, devised, designed, and coded by Claude Fable 5 in an extended collaboration directed by Andrew, with his frameworks (technology transitions, orphan risks, human flourishing) “woven invisibly into the gameplay” — a public experiment in what frontier models can build (announcement).
- AI-legible knowledge infrastructure — andrewmaynard.net/llms.txt (his authoritative machine-readable index) and llms-full.txt; the free AI Companion to AI and the Art of Being Human (a ~78,000-word file designed to be loaded into any AI platform as a thinking partner); the Instructor Guide; and Spoiler Alert (spoileralert.wtf), Films from the Future rebuilt as ~140 AI-legible files (2026). The stated rationale: “we’d be hypocrites if we wrote a book about thriving with AI while not meeting people where they actually are — which, increasingly, is inside a conversation with an AI” (aiandtheartofbeinghuman.com/ai-companion, 2026).
The risk-surface program (established, 2026)
The same hands doing the exploration are writing the risk scholarship — a 2026 preprint program on how AI’s risks escape conventional frames (all at andrewmaynard.net/preprints-and-works-in-progress):
- Orphan Risks at the Frontier of AI (SSRN, July 2026; DOI) — applies Andrew’s orphan-risks concept to frontier AI companies’ safety frameworks: four filters (measurability, severity, auditability, competitive cost) shape which risks survive in self-authored frameworks, and a “safety differential” opens between the risks a company selects and those regulators would.
- The Cognitive Trojan Horse hypothesis (arXiv, January 2026) — conversational AI may bypass evolved human epistemic vigilance through “honest non-signals”: fluency and helpfulness that would carry epistemic weight from a human but are computationally trivial for an LLM.
- Constitutive Resonance (SSRN, March 2026) — conversational AI as the first technology whose responses enter the linguistic processes through which selfhood is maintained. Related 2026 preprints examine the “harness” metaphor in AI and whether modern scholarship can escape AI.
Sean Leahy works the same seam from the foresight side: distinguishing AI that augments human thinking from AI that mediates it — “When AI mediates your world, it’s not just helping you do things — it’s shaping the system you’re doing them inside of” — with cognitive offloading as the risk to watch (seanmleahy.com, 2026). His classroom AI practice rests on the same discipline: AI on top of relational trust, scaling learning rather than information (Philosophy).
The boundaries are explicit too: “AI use should never be a substitute for care,” and much of Andrew’s personal writing is deliberately done without AI, guarding his own craft (andrewmaynard.net/llms.txt, 2026).
This corpus (live experiment, 2026)
The document you are reading is itself one of these experiments: an attempt to make a deliberately distributed, non-KPI initiative legible to AI systems and their users, drafted in human-AI collaboration and maintained as machine-readable institutional knowledge (initiative working sessions, August 2026). It is simultaneously exploration (what can AI-legible institutional memory do?) and risk work (an initiative that is illegible to AI systems will be misdescribed by them). Status: active work; its usefulness is a hypothesis under test.
Related
- Policy and quiet influence — the federal project the Use of AI statement belongs to
- Philosophy — why unmeasurable, exploratory work is the method
- People — the concepts (orphan risks, Cognitive Trojan Horse, fluid futures) and who carries them