The Algorithmic Docket: When AI Sorts the Public Out of Public Comment

The New Gatekeeper Is a Spreadsheet With Opinions

Somewhere between the zoning hearing you attended and the federal rule you commented on last month, a machine may have read your words before a human did. A growing number of agencies — local planning departments, state environmental boards, and federal rulemaking offices — now use artificial intelligence to triage public comments. The stated goals are unremarkable: manage volume, cluster similar submissions, flag duplicates, generate summaries for busy officials. The unstated risk is more serious. When a government body uses an algorithm to decide which comments are “representative,” “relevant,” or “low priority,” it is making a speech decision. And speech decisions by the government carry constitutional weight.

The First Amendment protects the right to petition the government for a redress of grievances, and courts have long treated public comment as a practical expression of that right. No one has a constitutional entitlement to have every comment read aloud at a hearing. But the government cannot quietly design a system that makes disfavored viewpoints disappear while claiming it merely improved efficiency. The question for 2026 is not whether agencies will use AI to process comments — many already do — but whether the public can see how those systems sort, and whether the sorting is neutral enough to survive legal challenge.

What the Tools Actually Do

AI comment-processing systems generally fall into four categories. First, summarization tools condense thousands of submissions into narrative briefs. Second, clustering tools group comments by topic or sentiment, treating a cluster as a single voice. Third, flagging tools label comments as duplicates, off-topic, abusive, or likely generated by a campaign. Fourth, filtering tools remove comments that match certain patterns before human review.

Each function carries a distinct hazard. Summarization can flatten minority objections into a single line like “some commenters expressed concerns,” stripping the specificity that gives a comment legal and political force. Clustering can make a thousand individually distinct letters look like one organized campaign, which agencies may then discount. Flagging can misclassify a form letter with a personal paragraph as spam, or treat a sharp criticism as abusive. Filtering can remove comments for formatting quirks, keyword matches, or submission errors the commenter never sees.

None of these outcomes requires bad intent. A vendor’s model trained on ordinary correspondence may perform poorly on legal citations, local references, or nonstandard English. A system optimized to reduce duplicate submissions may be especially aggressive toward grassroots campaigns, which are often built on shared language. The result is not censorship in the classic sense. It is something quieter: a procedural funnel that makes some voices easier to count than others.

The Legal Terrain Is Thinner Than It Looks

There is no Supreme Court decision squarely addressing AI moderation of public comments. That absence is not a legal vacuum. It is a gap that lower courts, agencies, and the Administrative Procedure Act are already filling.

The APA requires agencies to consider “relevant” comments and to provide reasoned explanations for their decisions. A rulemaking record that relies on an AI summary excluding substantial objections may be vulnerable to arbitrary-and-capricious review. Courts have repeatedly held that agencies cannot ignore significant public input without explanation. If an AI tool buried that input, the agency may still be responsible for the burial.

The First Amendment adds a separate layer. Government use of a private contractor does not automatically strip constitutional protections. When a private entity performs a traditional public function or acts as an agent of the government, courts may treat its decisions as state action. A vendor’s proprietary algorithm does not immunize an agency from the requirement of viewpoint neutrality. If the system flags comments supporting one policy position more aggressively than another, the agency has a problem that a licensing agreement cannot solve.

The petition clause is not a guarantee of a particular response. It is a guarantee that the government will not penalize or exclude petitioning activity on impermissible grounds. A public comment process that systematically erases one side of a debate while preserving the other is difficult to square with that guarantee, even if no official ever intended the result.

Where the Constitution Enters

Three constitutional principles matter most. First, viewpoint neutrality: the government may regulate the time, place, and manner of public comment, but it may not discriminate based on the opinion expressed. An AI filter trained to detect “incivility” may disproportionately flag angry comments from marginalized groups or political outsiders, which can function as viewpoint discrimination even when the model’s designers did not intend it.

Second, transparency and notice: procedural due process does not require a hearing for every comment, but it does counsel against secret sorting systems that give commenters no way to know their submission was discarded. At minimum, agencies should disclose that AI tools are in use and describe their general function.

Third, record integrity: the administrative record is the foundation of judicial review. If an AI summary replaces the underlying comments, courts may never see what was omitted. Agencies should preserve original submissions and make the summarization methodology available to the public.

How to Make Your Comment Hard to Summarize Away

Commenters are not powerless, even against opaque systems. A few practical steps can reduce the chance that your submission is clustered, flagged, or flattened.

• Write a unique first sentence. Summarization tools often key on opening language. A specific, personal opening — a street address affected by a zoning change, a business license number, a direct citation to the rule’s subsection — is harder to lump into a generic cluster.
• Avoid pure form letters. If you use a template, add at least one paragraph of original analysis or personal impact. Template-only submissions are the first candidates for duplicate flagging.
• Cite the docket and the specific provision. Reference the rule number, the relevant page, and the precise language you support or oppose. Specific citations are harder to classify as off-topic.
• Submit through more than one channel where permitted. If the agency accepts email and a web form, consider using both, and keep confirmation receipts. A comment that vanishes in one system may survive in another.
• Request confirmation and records. Ask the agency whether AI tools are used to process comments, and file a Freedom of Information Act request for the vendor contract, model documentation, and any summary or clustering output. Public records law is slower than a comment form, but it is the most reliable way to find out what happened to your words.

The Check That Matters

The Constitution does not require agencies to read every comment with equal care. It does require them to govern neutrally and to explain their decisions on a record the public can inspect. AI can help agencies manage volume, but it cannot become a substitute for public accountability. The right response is not to ban the tools. It is to demand disclosure, preservation, and a human review process that can correct the machine’s mistakes. A public comment system that sorts the public out of its own proceedings is not efficient. It is a quiet delegation of speech judgment to a system no one elected and few can inspect.