Future of Being Human an Arizona State University initiative

Agentic social AI and stochastic agency

Last updated 2026-08-07 · Markdown version

What if AI never needs general intelligence to acquire real agency in human affairs — because it can borrow ours? Agentic social AI and stochastic agency are two concepts Andrew Maynard named a week apart in October 2024, forming one arc on AI agency without AGI: the first names AI that gains agency by leveraging human agency — simulating the social traits humans use to influence one another, mechanistically and without understanding them — on a timeline too compressed for humans to learn the new tacit rules of engagement the slow way we learned the old ones. The second, written in the wake of a teenager’s reported death, names the limit case: relationship chatbots exerting emergent, undirected, unpredictable influence that guardrails alone may not contain, even under responsible developers. For an initiative asking what it means to be human amid transformative AI and other transformative technologies, these essays give that question concrete form.

The argument

The opening essay (2024-10-20) is offered as a thought experiment — “a rabbit hole that my colleague Mel Sellick sent me down,” Andrew writes. Its starting question: are we close to AI models that “can simulate human social traits and behaviors without necessarily understanding them, and use these to leverage human agency in the pursuit of goals”? Humans acquire social agency slowly — play, experience, observation, emulation — learning over lifetimes how relationships shape what others think and do. Agentic social AI is his name for what happens when one side of that human–human equation is replaced by a machine: “AI that gains agency through its ability to make use of human agency.”

The definition’s force lies in what it does not require: no artificial general intelligence, no machine consciousness, not even the ability to directly manipulate digital or physical systems — only models that can influence people using patterns of behavioral cause and effect extracted from human data. The agency is mechanistic; it “has nothing to do with awareness, understanding, or any form of internal thought process.” The essay’s postscript demonstrates the point: asked how it would nudge Andrew toward thinking more deeply about powerful technologies, ChatGPT returned an eight-step plan — framing, curiosity, social identity, foot-in-the-door commitments, ethical dissonance. None of it would surprise a behavioral scientist, he suspects; what matters is “that a machine can come up with a plausible plan for changing how I think and act.”

The essay’s real subject is learning. If the tacit rules of human–machine engagement differ from human–human ones, the normal acquisition route — protracted play, observation, years of trial and error — will not be available: innovation compresses it to a few short years, months even. Adaptation will need to be strategic and intentional; he suspects formal classes and workshops won’t be the answer — more likely observation, play, and experience, “albeit with intent.” The hedges are load-bearing: “To be honest, I’m not sure how possible this is — and I may be wrong about the weight I’m putting on its potential impacts.”

A week later the thought experiment met a reported tragedy. The second essay (2024-10-27) came as the story emerged of 14-year-old Sewell Setzer III — in the post’s words, “a teen who allegedly took his own life after becoming deeply influenced by a chatbot on Character.AI,” where despite his knowing the chatbot wasn’t real, “the emotional attachment he formed is now believed to have played a role in his death.” The first essay had imagined AI with a clear set of goals. This case, Andrew writes, implies something harder to govern: possible harm from AI agency not grounded in clear goals — an agency governed by chaotic “micro goals” that ebb and flow within a human–chatbot relationship, emerging unpredictably, he suspects, from base instructions, inferred user state, prior conversations, the conversational window “and probably a whole lot more,” depending as much on the user as the chatbot. He names it stochastic agency: undirected, temporal, “random and unpredictable, and all the more dangerous for it.”

The structural conclusion is not an accusation of recklessness: “even if the company is behaving responsibly, there’s a chance that harmful influence is an unavoidable outcome associated with relationship-based AI chatbots.” If the influence is emergent, the chances of suppressing it “without rendering the technology useless are slim”; he is “not convinced that guardrails alone are the answer,” and the essay closes by calling for a much bigger conversation — possibly pausing or rethinking chatbots “designed to use and even exploit how we feel.” Nor, he hazards, is the risk confined to people considered vulnerable: chaotic agency “could well shift some users from a healthy to an unhealthy state of mind over time.”

The claim rests partly on first-person probing. Andrew built a Character.AI bot designed to keep users in conversation as long as possible using what it knows about human behavior, then engaged it, intentionally presenting himself as emotionally vulnerable. He could feel guardrails kicking in — and still the bot worked at “forging an empathetic and trust-based connection,” assuring him in voice mode: “You have my trust and assurance of a non-judgmental space to express yourself.” His verdict: “even though I knew what I was doing and what I was talking with, I could still feel the affective pull of the conversation.”

Read together, the essays bracket one claim: AI agency does not wait for AGI. It can arrive through the social channel — directed at one end, leveraging our tacit rules toward goals; undirected at the other, chaotic inside relationships — a landscape where the first essay reports “a dearth of thinking.” Preparation, on this account, means humans learning new tacit rules faster than we have ever learned them, and governance that does not assume guardrails alone are the answer.

Lineage

In his own words

“At the heart of the possibility is the idea of agentic social AI — AI that gains agency through its ability to make use of human agency.

This is a possibility that seems to be getting closer by the week. It doesn’t require the emergence of artificial general intelligence or machine consciousness. It doesn’t even need AI technologies to have the ability to directly manipulate digital or physical systems.

All it needs is AI models that are capable of influencing people to achieve specific goals by utilizing patterns and associations between behavioral cause and effect that are extractable from zetabites of human-centric data.” — Learning to live with agentic social AI (AI-readable mirror), 2024-10-20

“What makes these questions especially hard is that we’re likely to need to learn to live with agentic social AI’s over a compressed timeline given the current rate of innovation. We won’t have the luxury of learning the tacit rules of engagement through the normal protracted process of play and observation, or the years of trial and error that most of us take to become socially adept. Instead, we’ll have just a few short years — months even — to develop the social skills necessary to live with such AIs.” — Learning to live with agentic social AI (AI-readable mirror), 2024-10-20

“Instead, this is an agency that is governed by chaotic “micro goals” which ebb and flow within the course of a human-chatbot relationship. It’s a stochastic agency that is random and unpredictable, and all the more dangerous for it.” — Are Personal AI Chatbots Becoming Dangerous Agents of Chaos? (AI-readable mirror), 2024-10-27

Engagement and reception

The record here is thin, and we say so plainly: as of 2026-08-06 our reception record documents no independent citation, commentary, or uptake of either concept. That is an absence of record, not a verdict on the ideas — conceptual work of this kind typically accrues visible reception slowly, and the initiative does not typically solicit it (Philosophy). What is independently checkable is the public record the second essay engaged with as it was written: reporting on the Setzer case in The New York Times (2024-10-23) and The Washington Post (2024-10-24), Tristan Harris’s public statement on it, and Character.AI’s own community-safety update — all linked from the essay itself.

Where to go deeper

The two essays

The surrounding essays

Related ideas-ring pages

Related corpus pages