Who stands behind the words?

ABSTRACT

The internet severed speech from the relationships that made it trustworthy; AI severs it from the speaker. Fluent, personalized, institutionally usable speech can now be produced at mass scale with no chain of responsibility anchored in a person who answers for it, hollowing out the assumption that underwrites public trust: that behind every word that counts stands a person or institution that can be held to it. This essay names what is being lost as “standing”—the right to have one’s words count in a consequential process, underwritten by the capacity to answer for them—and argues that neither more capable models nor better alignment can restore it. The workable path is architectural, leading to a set of guiding design principles for institutions that run on words that people must stand behind: attributable authorship; decomposed, inspectable process; provenance as infrastructure; and the widening of standing rather than the concentrating of it. There is one rule: autonomy is permitted; unaccountability is not. The repair is not better machines alone but institutions that anchor every word in an accountable human being.

In 2017, the Federal Communications Commission received more than 22 million public comments on its proposal to repeal net neutrality rules. An investigation by the New York attorney general later found that nearly 18 million of these comments were fake—many submitted under the names of real people who never wrote them (Office of the New York State Attorney General Citation2021). That fraud was caught because it left fingerprints: templated text, duplicated phrasing, patterns a careful analysis could expose.

Now imagine the same docket filled with fluent, individually tailored comments; no two alike, produced at almost no cost. Large language models erase the fingerprints. Reviewers can no longer tell which comments express views that citizens actually formed and stand behind, and textual analysis alone can no longer settle it. They begin to doubt the entire record, though every substantive claim in it may be true: Nothing establishes that a distinct person stands behind each comment.

The same problem is appearing wherever institutions collect the human voice. In the United States, telephone-survey response rates have fallen into the single digits (Kennedy and Hartig Citation2019), pushing much survey research online—where a respondent can hand the questions to a chatbot, collect the participation incentive, and never be the one who answered.

An adjacent vulnerability is already visible in crowdsourced research: In one text-summarization task, researchers estimated that a third or more of paid crowd workers used large language models to produce their answers (Veselovsky, Horta Ribeiro, and West Citation2023). Each submission still traces to a recruited, reachable worker—but whether it grounds out in that worker or in a machine, no one can tell.

It reaches the classroom too. Teachers read essays and cannot always determine whether a student’s words represent that student’s thinking. The student can be named, questioned, even made to answer—yet the development the writing was meant to foster may never have occurred.

These are not stories about misinformation, and fact-checking will not solve them. The problem is different and more corrosive: Fluent, personalized, institutionally usable speech can now be produced at mass scale with no chain of responsibility anchored in a person who answers for it. The architecture of accountability that underwrites public trust is being hollowed out, and it can no longer be assumed that behind every word that counts stands a person or institution that can be held to it.

Throughout this essay I will use the word “standing” for what is being lost, and I mean it in the sense that lawyers will recognize. In litigation, a “standing” is a stake sufficient to entitle a party to be heard by the court. I borrow the idea, in the sense of the right to have one’s words count in a consequential process (a rulemaking, a poll, a newsroom, a classroom), earned by having something at stake in it.

That right is underwritten by a second capacity, “answerability”—meaning that one is able to answer for one’s words, a condition that the legal term does not carry, and this essay adds.

“Answering” is not only punitive. It means being available to explain, defend, correct, and repair, as well as to bear what follows: blame when the words do harm, credit when they do good. The argument of this essay is that the two must travel together: Standing is legitimate only when “answerability” underwrites it. When I write of speech “with standing,” I mean speech whose right to count is backed by someone who can answer for it.

Standing does not make words factually true; what it secures is a different kind of truth—that the words truly belong to the speaker who answers for them.

Familiar civic roles show how the two travel together, sometimes through a chain of people. A witness under oath has standing: The proceeding admits the witness’s testimony, and the oath binds the witness to it. A journalist’s unnamed source has standing borrowed through the reporter who vouches for them and the masthead that answers through named editors.

Privacy and answerability are not opposites: the source’s identity is shielded from the public but disclosed to the reporter, who stakes their own standing on the vouching. An institution can verify a person and stand behind their words without exposing them. What fails the test is the arrangement that only looks like a chain—the shell organization with a nominal figurehead, the signature atop words its bearer cannot defend. And a chatbot, however capable, has no standing at all: It can neither earn a voice in a consequential process nor be held for what it says. It can appear inside a chain; it can never anchor one.

Answerability can fail in two ways. The words may not be traceable (by anyone, even in confidence) to the person they are attributed to. The flooded docket fails this way. Or the words may trace to a person who cannot justify them—who never engaged with the words enough to defend them under challenge. The survey respondent who handed the questions to a chatbot fails this way.

Engagement takes many forms: a speaker who understands and can defend their words through translation, assistive tools, or plain language meets the bar in full. What counts is comprehension and the capacity to answer, not unaided command of the drafted prose. And where a person cannot answer for themselves—such as a child, or a person whose disability or vulnerability leaves them unable to speak in their own defense, to name just two examples—answerability can be anchored in a representative who answers on their behalf.

Answerability is thus a chain. It may pass through many links, human and machine, but it must be anchored in a person who can answer in their own right (their own understanding at work, their own stakes on the line) or an institution that answers through named persons. And context sets where the chain may be anchored: sworn testimony must be anchored in the speaker; elsewhere the chain can be longer, so long as every link is real and the anchor can answer.

That yields a single test for AI assistance: Delegation preserves answerability when a person sets the purpose, reviews and adopts the result, and remains able to defend or correct it. Delegation hands off the work, but never the responsibility: The person who directs the work remains answerable for what is done on their instruction.

Delegation fails when output enters a consequential process without anyone having adopted it—or when the point of the activity is the person’s own formation, where nothing done after the fact can substitute for the doing.

And adoption has a boundary: It answers only for words offered in one’s own name. A consultant who generates 10,000 comments, reviews each, and can defend them all still speaks as one person; the docket’s fraud is not unanswered words but words that claim distinct people who never spoke. Adoption cannot supply the anchors a submission claims.

The repair will not come simply from making machines more humanlike or more truthful. It will come from redesigning institutions, and from norms for our own use of these tools, so that AI can help process speech without displacing the human responsibility that gives speech its civic force. That redesign is achievable, but it requires first understanding exactly what is breaking, and why the dominant responses to AI risk will not fix it.

How we got here

Modern societies run on the division of knowledge. For example, I do not personally assay the gold in a wedding ring or inspect vaccine plants; I rely on networks of people and institutions positioned to know. When those networks fracture, whole segments of society drift from doubt into denialism (Roy Citation2024) or “truth decay,” as a RAND Corporation research initiative memorably named it (Kavanagh and Rich Citation2018).

The social web accelerated that fracture. Ad-supported platforms optimized for engagement amplify the most provocative content regardless of accuracy; those who tried to speak with nuance or at the risk of vulnerability learned that silence was safest. That story is by now well told, including by me (Roy Citation2023). What matters here is what all the repair strategies (connecting experts to trusted local figures, citizen assemblies, community listening) implicitly assumed: that when a voice entered the civic record, a chain of responsibility ran from the words to a person who stood behind them. The platforms severed speech from the relationships that made it trustworthy. But most posts still had a person attached.

Generative AI breaks the deeper connection. It accelerates the older dynamics too, but the deepest change is categorical. Print, radio, television, and social platforms altered how human speech circulated—but none of them spoke. Large language models converse and respond to each person in real time with unprecedented fluency. Whether that fluency is evidence of genuine understanding is actively debated, yet even if one grants the machines all the comprehension one likes, they still lack standing.

What breaks

To speak is to act, and to act in good faith has always been to be accountable. Chatbots break that bond—an argument I made at greater length in The Atlantic earlier this year, and restate here in brief (Roy Citation2026). A promise creates an obligation; an apology asks for repair; testimony enters the record. The philosopher J. L. Austin called these “speech acts”—words that do not merely describe the world but act in it (Austin Citation1962).

When the speaker is a person, the acts bind, because their consequences have somewhere to land. A chatbot can carry words that bind the person or institution standing behind it, but the words never bind the chatbot itself. Anyone who has used one knows the pattern: The system gets something wrong, apologizes gracefully, gets it wrong again, apologizes again. The apologies are fluent, appropriate—and empty. A chatbot that says “I’m sorry” binds itself to nothing. It cannot incur an obligation, stake anything on doing better, or undertake repair on its own account. What these systems manufacture, at scale, is a form of obligation with no one obligated.

The standard rejoinder is that chatbots disclose their nonhumanity, but this misunderstands the problem. In 1966, the MIT computer scientist Joseph Weizenbaum built ELIZA, the first chatbot—a program so simple that it did nothing but match patterns in the user’s words and slot them back into pre-written replies. Weizenbaum later wrote that users quickly became emotionally involved with the program and anthropomorphized it; his secretary, who knew he had written the program and that there was nothing more to it, asked him to leave the room after a few exchanges (Weizenbaum Citation1976). He was alarmed: The program merely performed the part of a speaker, yet the way people took to it made him fear they were extending it the expectations owed to a real one. Sixty years later, the fluency of chatbots has improved beyond recognition while their own accountability remains exactly zero. Telling people they are talking to a machine does not switch the effect off; the more fluent the machine, the stronger its pull.

Holding a speaker to their words—by shame, by sanction, by a demand for explanation or repair—works only on an agent who persists: Someone whose tomorrow can be made different by what they said today. Whatever continuity a deployed model has is conferred and revocable by its operators. To extend a metaphor of the AI researcher Andrej Karypathy, it is a kind of ghost or spirit that imitates humans: software that can be copied, forked, updated, deleted, or embedded in a thousand products overnight (Karpathy Citation2025). Nothing of its own is ever at risk in what it says. The buck cannot stop at something with nothing to lose.

Why alignment is not enough

The AI safety community has organized much of its work around alignment: ensuring systems pursue intended goals and behave as designed. That work is essential, and it does not claim to confer civic legitimacy. The point is that its task is different, for two reasons.

First, a system can be perfectly aligned to an institution’s chosen objective while the institution’s procedure remains illegitimate. This happens when the target is precisely what democratic life must determine: what trade-offs a public will accept for a data center in its town, how a school should weigh competing parental concerns. These are matters of judgment, particular to each community, and legitimacy depends on who interprets, decides, and bears the consequences. No degree of alignment can supply that.

Second, even a hypothetically infallible model changes nothing about the core problem. Standing depends on social processes: relationships and institutions, iterative yet enduring, with accountability built in. No technical capability can substitute for them. Greater capability changes what a chain contains, not what can anchor it; even an infallible model remains an anchor in name only.

Nor can the gap be closed by wrapping a model in a corporate shell. An accountability structure counts only if it carries a continuous reputation that can bear loss, and it contains named people in control who can be held to account.

Norbert Wiener, the MIT mathematician who founded cybernetics (the study of how machines and living things use communication and feedback to govern their own behavior), anticipated the underlying issue in his 1950 book The Human Use of Human Beings (Wiener Citation1950). He warned, first, that the more autonomously machines could act, the more people would be tempted to hand their judgment over to them. The second, subtler warning was that efficiency itself corrodes dignity, as people remake themselves to fit systems they no longer steer—and the act of steering was important to Wiener; the very word “cybernetics” (a neologism which Wiener himself had a hand in coining) refers to the person who steers a ship, gaining feedback from the vessel’s motion. To throw the problem of one’s responsibility onto the machine, he wrote, “is to cast [one’s] responsibility to the winds, and to find it coming back seated on the whirlwind.”

Follow the current trajectory and you arrive at what I have elsewhere called the “avatar state” (Roy Citation2026). Imagine a city council consultation in which residents send AI agents to read the materials, generate testimony, negotiate compromises, and register preferences. Participation rises by every metric except the one that matters: whether citizens appear before one another as accountable political actors. The failure is not machine assistance. A resident who uses AI to digest 100 pages of planning documents and draft testimony she then reviews, revises, and delivers remains the anchor of her own words.

The failure begins where adoption ends: when agents deliberate in her place—listening, adjusting, registering positions she never encounters. Deliberation works by changing the deliberator; listening to others and adjusting one’s position is not a step in the process but the substance of it. Positions negotiated among machines belong to no one—not to the residents, who never formed them, and not to the machines, which cannot answer for them. On paper it passes for self-government. In reality, the one essential thing is missing, which is the risk of being answerable. Self-government is, at its foundation, people saying things in the open and then living in the world their words helped to make.

Design principles for accountable systems

Walking away from AI, in its most capable and rapidly evolving forms, is not an option. These systems are too useful, and civic institutions that refuse them will simply be overrun by actors driven by other aims—such as profit, reach, and power—with no stake in the civic outcome.

Nor will capability restriction alone restore standing. The workable path is architectural, and its animating rule is that autonomy is permitted while unaccountability is not. Machines may work with great independence inside a process, so long as responsibility for the result never leaves human hands.

My colleagues at MIT (MIT Center for Constructive Communication Citation2026) and at Cortico—the nonprofit I co-founded to build community listening systems—have been developing infrastructure on this rule (Cortico Citationn.d.). The principles below generalize to any institution that runs on words people must stand behind, be it a legislature, a newsroom, a school, or a workplace.

What different actors can do now

Agencies and legislatures that solicit public voice should build authenticated but privacy-protecting channels, require disclosure of AI assistance in submissions (so the record can be audited, not so assisted submissions can be discounted) and maintain audit trails for any AI-generated summaries of public input. Eligibility, anonymity, and participation rules differ across political systems; the requirement of an answerable chain does not.

Pollsters and researchers should adopt respondent verification and explicit disclosure norms for AI-assisted answers, before synthetic responses render their instruments unusable. A record whose responses were ground out in undisclosed chatbot output no longer measures the people it claims to represent.

Schools and workplaces should distinguish judgment from formation. A manager’s review and an official’s testimony are judgment tasks; others may help draft, but the judgment must be the person’s own, adopted and answerable in their name. A student’s work is different in kind—the essay and the worked solution are both forms of showing one’s thinking. To be a student is to be forming one’s fully realized self, and formation cannot be delegated. A student can adopt and even defend delegated work, and still the thinking the assignment exists to develop never happened.

Regulators of systems that carry words that people must stand behind should weight traceability as heavily as capability. Where the danger is consequential speech that no one stands behind, the missing requirement is an accountability chain appropriate to each actor’s role: whoever speaks with a system’s words must stand behind them, whoever operates the system must answer for how it runs, and whoever builds it must answer for what it was designed to do. At each link, institutional responsibility must resolve to named persons in control.

And builders should remember that the engagement-optimized, ad-supported model widely adopted by social media platforms is not the only one. Communication systems can be designed so their economics reward speech that someone stands behind rather than volume that no one answers for.

The whirlwind

None of these arrangements makes a machine a trustworthy speaker. No technical advance can; trustworthiness of that kind requires something to lose. The task is different and older; we must preserve the conditions under which human words bind in an environment saturated with fluent language that binds no one.

The whirlwind that Wiener warned of was about abdication—in other words, the danger was in the temptation to trade responsibility for efficiency, one delegation at a time, until judgment has migrated out of human hands. And the judgment does not pass to the machines; it pools with the few who build and steer them. A public that has stopped answering for its own words will find it has also stopped being asked. Each institution that accepts unattributable speech into its record, each professional who ships words they do not stand behind, each of us who lets a system apologize in our place, makes the trade seem more normal. Responsibility surrendered this way does not disappear. It waits, and it comes back the way Wiener said it would—on the whirlwind, once the harm is already done. The alternative is to design accountability in now—attributable authorship, inspectable process, provenance as infrastructure, standing widened rather than concentrated—while the architecture is still ours to choose.

Acknowledgments

I am grateful to Martha Minow, Audrey Tang, and Bill Noble for feedback that improved this essay.

Disclosure statement

No potential conflict of interest was reported by the author(s).

I supplied the arguments through extensive dictation and gave AI systems (primarily Anthropic’s Claude Opus 4.8 and Claude Fable 5) access to my prior cited work as references for content and my writing style. The systems drafted a distillation of my words—instructed to work from my dictated vocabulary—and revised under my direction, with every line reviewed, iterated as needed, and approved by me. Editorial reviews of drafts were provided by OpenAI’s GPT-5.6 Sol. I have scrutinized and stand behind what is written.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.