When George Washington prepared to leave office, he had help finding the words. James Madison had drafted an earlier farewell. Alexander Hamilton substantially rewrote it. Washington reviewed and revised the text. The result appeared in a Philadelphia newspaper in September 1796. We still call it Washington’s Farewell Address.1
More than a century and a half later, John F. Kennedy and Theodore Sorensen repeatedly revised the speech Kennedy would deliver at his inauguration. Collaboration did not prevent it from becoming an enduring expression of Kennedy’s presidency.2
We recognize something quite sensible in these arrangements. A person can direct an argument, exercise judgment over its contents, and take responsibility for its consequences without personally composing every sentence.
AI introduces different capabilities and different risks. But assistance itself is hardly a departure from the history of public expression.
What is changing is who can afford it.
Writing requires time, confidence, practice, and attention. Having something worth communicating guarantees none of them. A person may write extremely well and still lack the hours or mental bandwidth to turn experience into a considered argument.
Those with resources can purchase that capacity. Others must find it after their shift, between caring responsibilities, or during the hours they should be sleeping. A standard that treats unaided composition as a condition of legitimate participation therefore imposes a burden that the well-connected have long been able to avoid.
Cory Powell has argued that AI can democratize writing by enabling people with knowledge and ideas to participate. I think that argument is substantially correct. But its implications extend beyond the individuals gaining access.3
The rest of us also pay a price for what they never communicate.
Consider a hypothetical home-care worker trying to explain why a scheduling policy is failing. Travel allowances assume journeys that cannot be completed in the allotted time. Visits run late. Clients miss care. The official report identifies a staffing problem.
She understands the discrepancy because she encounters it every day. After work, however, she has little capacity left to assemble a persuasive written explanation. An AI assistant helps turn her spoken account into a statement. She checks it, corrects it, and submits it.
The relevant alternative may have been silence.
That matters because institutions make decisions using the accounts of reality available to them. A council encounters residents through consultations and complaints. An employer encounters working conditions through reports and management hierarchies. Public debate encounters experience through those able and willing to publish it.
These channels determine which experiences become available for consideration. When participation requires spare time, polished prose, or professional assistance, the resulting picture can be systematically incomplete.
An absence of objections may reflect exhaustion. An unreported failure may reflect the difficulty of explaining it. Yet the institution continues making decisions as though the information it received adequately describes the problem.
The consequence is a failure of feedback. People experience the effects of a decision, but what they learn does not reach those with the authority to change it.
AI could reduce one of the costs of completing that loop.
There is evidence for the first step. In an experiment involving mock comments on an environmental rule, Amelia Arsenault and Sarah Kreps found that AI assistance made drafting easier and produced submissions that evaluators rated more highly. It did not improve participants’ self-reported policy comprehension. The study supports the possibility of wider participation; it does not establish that better decisions follow.4
That requires something further.
Our care worker’s account could reveal that an apparently underperforming service rests on an impossible travel assumption. If her observation withstands scrutiny, the institution has acquired a reason to revise its explanation. Knowledge held by one person has become available to improve a decision affecting many.
This is where I think the social value becomes more consequential. Reducing the cost of expression can increase the capacity of a system to recognize and correct its mistakes.
But the same technology can interfere with that process.
Suppose the worker explains that a route requires two buses, with a connection that regularly fails. Her writing assistant turns this into a polished complaint about “resource constraints and operational inefficiencies.” The account sounds more professional. It has also lost the detail that could expose the faulty assumption.
Another model summarizes the submissions for management. Her complaint is classified under staffing concerns. The resulting report reinforces the original diagnosis.
Each stage has become more efficient. The institution remains wrong.
This sequence is hypothetical, but the underlying risks have empirical support. Dhruv Agarwal, Mor Naaman, and Aditya Vashistha found that AI suggestions pushed Indian participants’ writing toward Western styles and diminished cultural nuances.5 Separately, Shenzhe Zhu and colleagues identified persistent underrepresentation of minority positions in AI summaries of public deliberation.6
Increasing the volume of communication does not guarantee an increase in what anyone learns.
An unusual observation may be mistaken or unrepresentative. It still needs scrutiny. But a single verifiable detail can undermine an assumption shared by thousands. Its significance cannot be measured by how frequently it appears in a collection of submissions.
The practical questions follow from this. Does assistance preserve the contributor’s specific observations? Can the contributor identify and correct what the model added? Can decision-makers recover the original evidence from a summary? Does someone investigate a discrepancy and explain what happened as a result?
AI will not remove institutional indifference, retaliation, or unequal influence. It can also make fabricated testimony and manufactured campaigns cheaper. Access to assistance creates possibilities whose value depends on how the surrounding institutions behave.
Those limitations strengthen the case for evaluating the whole process. Counting words, submissions, or users tells us remarkably little about whether society is becoming better informed.
Washington had Madison and Hamilton. Kennedy had Sorensen. Their collaborations remind us that public authorship has long accommodated assistance alongside judgment and responsibility.
Extending access to human and AI assistance gives more people a chance to contribute what they know. The larger opportunity is to make that knowledge available for others to examine, challenge, and use. Its value may ultimately lie in the mistakes we become capable of correcting because someone finally had the means to explain them.
Footnotes
[1] George Washington’s Mount Vernon, “The Farewell Address.” Describes Madison’s earlier draft, Hamilton’s subsequent contribution, Washington’s revisions, and publication in September 1796. ↩
[2] John F. Kennedy Presidential Library and Museum, New Frontiers, issue 10, Winter 2009. Describes the repeated revision of the inaugural address by Kennedy and Sorensen. ↩
[3] Cory Powell, “Is Human Writing Doomed? Or Is Writing Finally Becoming Democratic?” August 17, 2026. ↩
[4] Amelia C. Arsenault and Sarah Kreps, “Whose voice counts? The role of large language models in public commenting,” Big Data & Society, February 4, 2026. The experiment assessed mock submissions, evaluator ratings, and self-reported comprehension; it did not measure subsequent policy outcomes. ↩
[5] Dhruv Agarwal, Mor Naaman, and Aditya Vashistha, “AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances,” CHI 2025. Controlled experiment involving 118 participants from India and the United States. ↩
[6] Shenzhe Zhu, Shu Yang, Michiel A. Bakker, Alex Pentland, and Jiaxin Pei, “Can AI Truly Represent Your Voice in Deliberations? A Comprehensive Study of Large-Scale Opinion Aggregation with LLMs,” arXiv:2510.05154, first submitted October 2, 2025; revised March 20, 2026. Evaluates 18 language models and reports persistent weaknesses in representing minority positions. ↩