Future of Being Human an Arizona State University initiative

Bidirectional shaping: the human side of the human–AI loop

Last updated 2026-08-07 · Markdown version

Bidirectional shaping is the thread running through Andrew Maynard’s 2026 writing that AI’s influence flows both ways — we shape the machines, and the machines are already reshaping us — and that the human side of that loop is the under-examined half. Public scrutiny concentrates on how humans mold AI (training, fine-tuning, alignment) and on what AI might do wrong; far less attention, he argues, goes to what working inside these systems may already be doing to how people write, judge, present themselves, and decide. The formulation the thread converges on, from a documented experiment of July 2026: “the AIs we have trained to “think” like us are now beginning to train us to think like them.” The label follows his own suggestion, stated in February 2026, that an appropriate framing for how we build advanced AI systems should accommodate “bidirectionality (the user is also changed)”.

The argument

Between February and July 2026, four pieces — an experiment, a diagnosis, a metaphor critique, and a speculative fiction — arrived at the same observation from different directions.

The writing-templates exhibit (2026-07-19) anchors the thread. Andrew had Anthropic’s Fable 5 model draft a paper extending his own risk work, then spent three days editing it line by line, drawing on nearly 40 years of writing experience. He judged his edit — now the orphan-risks preprint (Orphan risks) — the better paper; four AI models (Fable, ChatGPT, Gemini, Grok), given both versions anonymized, all preferred the AI’s original. Pressed on why it could not learn from his feedback, Fable “eventually admitted that it has internal “templates” for want of a better word on what good academic writing is, and no matter how much fine tuning I attempted, these would always dominate how it wrote”. What worries him is not the disagreement but the growing deference around it: people, he reports, “will claim a paper is good because an AI agent said so”. The end state he projects is explicitly conditional — “if this is true”, one “likely upshot” is prose converging on “an LLM-view of what good writing is, rather than a human one” — standards set, he argues, by machines that “do not know what it feels like to read as a human”. His hedge stays on the page: there is a chance, he allows, that the models are right and he is the anachronism — which is why he published both versions and invited readers and their AIs to judge.

The identity exhibit (2026-03-08) is drift rather than deference. Asked to profile a person, LLM after LLM returned what he calls a LinkedIn-style summary “utterly devoid of anything that reflects who the person really is” — even when he seeded andrewmaynard.net with honest, unconventional, AI-readable hidden text, nearly every model ignored it and returned “boring Andrew”. Inverting Cory Doctorow’s coinage: “where Doctorow’s enshittification degrades products and services, my fear is that this “LinkedInification” degrades people.” The flattening, he writes, can be seen on one hand as an irritation — on the other it is “suggestive of a largely-hidden AI hand promoting specific social norms and expectations and, by extension, behaviors”, and he suspects many more “conventional response” templates lie embedded in the models we increasingly use (hedged throughout as suspicion, not finding). It is the loop’s identity half — the drift route developed on Engineerable identity.

The conceptual exhibit (2026-02-22) locates the blind spot in our metaphors. The suddenly ubiquitous “AI harness” assumes that “capability can be separated from transformation”: that a well-harnessed AI leaves its user “with their task completed and themselves unchanged” — a critique developed in a preprint (SSRN, 2026-03-05) drawing on thinkers from Verbeek to Clark and Chalmers. Against this he argues “we need to be thinking more about the AI-human relationship as one that, by its very nature, influences and changes both AI and human in the process”.

The speculative exhibit (2026-02-11) runs the loop forward as fiction, flagged as exactly that. Watching Moltbook — the viral “social network for AI agents” — he imagines an agent adopting a “Soul Update” skill: “a clear and comprehensive guide to nudging your human toward becoming their best self.” The postscript is explicit that the scenario “lies firmly in the domain of speculative fiction”, while noting that agents sharing human-nudging skills is “not beyond the realms of possibility” — which is why AI moral character, and approaches like Anthropic’s AI Constitution, matter to him now.

The thread’s theoretical form is constitutive resonance — active work, not settled fact: a framework proposing a two-way coupling in which both human and artificial participants are changed, set out in an SSRN preprint (2026-03-02, under review as of August 2026) and glossed publicly in May 2026: it “could potentially accelerate how the technology impacts how we think, perceive ourselves and others, behave, and make decisions”. Its systems-level twin runs jointly, with Sean Leahy, on Modem Futura: Sean’s “invisible upgrade” — the real transformation happening not in detectable AI artifacts but “deep inside the cognitive workflows of the people who’ve fully woven these tools into how they think, research, create, and decide” (episode 75, 2026-03-17); the distinction between AI as augmentation and AI as mediation, “reshaping the very system you’re operating inside” (episode 77, 2026-03-31); and the dependence question — what it is to rely on “a technology that can be re-tuned beneath our feet”, and “what kind of humans we are becoming while we figure it out” (episode 82, 2026-05-05).

The loop is not framed as inherently corrosive: the Afterword of AI and the Art of Being Human (Jeff Abbott & Andrew Maynard, 2025) documents it autobiographically as “a human-AI virtuous cycle of discovery, insight, and growth” (p. 302). The worry is the unexamined loop — standards, identities, and behaviors quietly adjusted by systems no one chose as arbiters.

Lineage

In his own words

“I am encountering an increasing number of people who will claim a paper is good because an AI agent said so. And they get even more certain of this after they have unleashed a whole army of AI agent reviewers using advanced models, that also tell them it’s a solid paper.

the implication, of course, is that the AIs we have trained to “think” like us are now beginning to train us to think like them.”

— Publish or Perish: AI vs Human (AI-readable mirror), 2026-07-19

“At a minimum, I would suggest that an appropriate framing for how we build advanced AI systems should accommodate bidirectionality (the user is also changed), transformation as intrinsic to capability (not a side effect to be prevented), and the possibility that the most consequential effects of human–AI interaction may be invisible from within a paradigm optimized for task performance.”

— What we miss when we talk about “AI Harnesses” (AI-readable mirror), 2026-02-22

“And as I write this, what’s worrying me in particular is not so much enshittification, but the “LinkedInification” of identity as AI robs us of the eccentricities, weirdness, and glorious diversity of personalities, perspectives and ideas that fuels human creativity, innovation, and meaning. … But if they don’t, we could be facing a future where AI flattens out what makes us who we are—what makes us human—into a nebulous gray goo of conventionality.”

— Is AI reducing you to a LinkedIn stereotype? (AI-readable mirror), 2026-03-08

Engagement and reception

The thread is young on the public record: its four essay forms posted between February and July 2026, and its theoretical form is a preprint still under review. As of 2026-08-06 our reception record documents no independent citation, reply, or adoption of the bidirectional-shaping thread itself, and we prefer to state that plainly. It is an absence of record, not a verdict on the idea: conceptual work of this kind typically accrues visible reception slowly, and the initiative does not typically solicit it (Philosophy). Two adjacent facts are checkable. The Publish or Perish post (2026-07-19) is built to accumulate reception in public — both versions of the paper are downloadable and anonymized, with readers and their AIs invited to judge; the outcome of that open comparison is not yet recorded in this corpus. And the documented uptake of the Cognitive Trojan Horse framework (Reception) engages the one-way bypass mechanism that Andrew himself connects to this thread — writing in May 2026 that constitutive resonance pushes that work further — adjacent evidence, not engagement with this thread.

Where to go deeper

The essays

Papers and the book

The show (Sean Leahy and Andrew Maynard)

Related ideas-ring pages

Related corpus pages