Consider a public consultation where citizens put their comments in the hands of a chatbot to be written up, and an agency has another model do the work of clustering and summarizing them. All proper procedures are followed, yet no one in a position of power ever reads what an individual has put forward. One might think this is still in the realm of speculation, but it is not. Federal officials have already put machine-learning tools to the test for regulatory analysis. The Environmental Protection Agency, for instance, has indicated it is looking at AI that can review and categorize public comments in a matter of minutes as opposed to months.

The danger here goes beyond the odd hallucinated text or a fake entry. It is a form of civic compression: when democratic input is distilled into themes by a machine, the views that are unfamiliar or run counter to the grain can vanish.

My own view on political change has been shaped by research into the way institutions handle information. Work in juvenile justice made clear to me that data can be out there in the public domain and yet be of no use politically because agencies fail to make the necessary connections. I see the same design flaw in technology policy, only at a newer frontier: a voice makes it into the official record and then is lost in processing. True agency demands more than having a forum to speak in; it means surviving the system that stands between speech and a decision. This is a distinction of particular import to young people, many of whom cannot vote and must make their mark through hearings, councils and public comment.

There is room for AI to be of service in these channels. The OECD examined fifty such projects in twenty-two countries and noted everything from translation to the analysis of large volumes of input. The federal pilot used semantic matching to cut down on repetitive tasks, and an understaffed agency could well spot patterns with these tools that would otherwise be missed. To ban AI outright would be to sacrifice both administrative capacity and access.

But a public comment is not a vote. Agencies will treat a form letter differently from a submission that puts forward new evidence or some overlooked consequence. The comment that stands alone may be the one with the most political weight. Summarization systems, however, are built to find the salient and the recurring. Michiel van der Meer and his coauthors have shown that even general-purpose language models have trouble with low-frequency opinions. In the name of efficiency, repetition can be redefined as relevance.

Sachit Mahajan has documented how profound the distortion can be. In his review of Canada’s 2025 AI strategy consultation, he found clusters of dissent or distrust were left out at rates ranging from 33 to 88 percent, while those in line with policy fared much better. On the whole the numbers did not look so bad, but the average was masking the losses. A summary can be factually sound and yet be democratically false if it misrepresents the distribution of disagreement.

Youth are vulnerable to this kind of failure. Martin and Venugopal found in a study of over 25,000 city council meetings in California that the typical public participant is older, whiter and more likely to be a homeowner than the registered-voter population. It is a representational skew that suggests youth submissions may be part of the long tail: less frequent, perhaps less conventionally formatted, and more likely to identify consequences established participants do not. Digital tools may open the door, only for automated summarization to close it again once the comment is in.

What is called for is a representational audit, not just a check for accuracy. Oversight should not be content to ask if a summary has any false statements; it needs to know which true positions were omitted. Any agency using AI to process input ought to be transparent about the model and the role of human reviewers. Every line in an official summary should be traceable back to its source. There should be targeted human review of the outliers and a sample of original comments preserved alongside the synthesis.

Such a proposal has its costs, though likely far less than going back to manual methods. It also upholds a principle of public law: if an agency is to consider significant arguments and not just tally support, then the system that decides what the official sees must be accountable. Automation can be an aid to public judgment, but it should not be allowed to silently define the public.

My belief in political change is more concrete for all this, if also more conditional. Influence is not a matter of being the loudest voice in the room. It is about altering the procedures that govern what an institution can perceive. Young people do not need to be assured they will win the day, only that their case has made it from the submission to the decision intact. The coming fight over youth voice may well be less about who gets to the microphone and more about who governs the summarizer.