The Public Comment Process, and the Part That Must Stay Human

Two conversations that never meet

Your department opened a comment period on a contested project and the inbox did what inboxes now do. Thousands of submissions, most of them articulate, many of them citing policy, all of them arriving faster than your staff can read. Two kinds of people want to talk to you about it.

The first is a software vendor. Their pitch is that AI will read every comment, tag it by theme, and hand your planners a summary by Monday. The second is a journalist, asking whether the comments were written by anyone at all.

Those sound like separate conversations. They are the same problem viewed from opposite ends. Once a submission can be generated by a machine, the number of submissions stops being evidence of anything. And the step that has to stay human moves. It is no longer reading the comments. It is establishing that a commenter exists and is actually affected by the decision.

Software that summarises an unverified inbox does not fix that. It industrialises the wrong step.

What happened in Southern California

The case every planner should know is the South Coast Air Quality Management District. In June 2025 its board voted 7 to 5 to reject a proposal, nearly two years in development, that would have put fees on gas-fired water heaters and furnaces. The opposition looked overwhelming: more than 20,000 emails.

In early 2026, reporting traced much of that volume to CiviClick, a Washington-based platform that describes itself as an AI-powered grassroots advocacy tool. A public affairs consultant publicly credited it with making "the ultimate difference." The agency flagged the submissions as suspect after an internal review of language patterns and metadata, then contacted a sample of roughly 175 senders. Five replied. Two confirmed they had sent a message. Three said they had no knowledge of the letters.

Look at that funnel. A 7 to 5 vote on a regional rule, and the verification that mattered happened months afterwards, on a sample, by email, with a response rate under three percent. That is not a failure of reading. The board could have had a perfect thematic summary of all 20,000 emails and it would have summarised the same fiction more efficiently.

It was not isolated. In the Bay Area, at least ten people listed as senders of letters to the regional air district told the San Francisco Chronicle they had no recollection of sending them. The Sierra Club, and later 22 California elected officials, asked the state attorney general and four district attorneys to investigate. The American Planning Association's June 2026 coverage adds mass emails supporting a North Carolina gas pipeline in September 2025. Dylan Plummer of the Sierra Club, quoted there, put the consequence plainly: the more AI comments arrive, the less credible all comments become.

That last point is the one planning departments should sit with. The harm is not only that fake comments get counted. It is that real ones get discounted.

Volume was never supposed to be the signal

Federal administrative law already says this, and has for years. The Administrative Conference of the United States, in its 2021 recommendation on mass, computer-generated, and falsely attributed comments, states that the comment process is not a vote or referendum. It is about the informational content of comments, not their sheer volume or their sources. That guidance came out of the FCC's net neutrality docket, where the New York Attorney General later found that nearly 18 million of more than 22 million comments were fake, and that a broadband industry group had spent $4.2 million generating more than 8.5 million of them, many in the names of real people.

Local land use has never fully absorbed that lesson, for an understandable reason. A planning commission or a city council is not a federal rulemaking body. Its members are often elected or appointed by people who are, and a full chamber or a flooded inbox reads as political weather. Nobody on a dais counts comments formally. Everyone on a dais feels the count.

That is exactly the weakness a generation tool exploits. It does not need to produce a good argument. It needs to produce a big number, and the number does its work before anyone reads a word.

What the summarisation vendors are selling

We want to be fair to this category, because the tools work. West Oxfordshire and Cotswold District Councils piloted AI summarisation of consultation responses through Konveio and reported up to an 85 percent reduction in officer time spent summarising and reporting on feedback. Greater Cambridge Shared Planning has run its own AI summarisation on local plan consultations. For a department drowning in a local plan consultation, that is real relief.

The guidance around these tools, from vendors and from public bodies alike, is consistent: AI does the first pass, a human reviews the output. Human in the loop. It sounds responsible, and for accuracy of summaries it is.

But notice where the human sits. They are checking whether the summary reflects the comments. Nobody in that loop is checking whether the comments reflect the public. The loop is drawn around the reading step, and the reading step is no longer where the risk is. Meanwhile the supply side is scaling on the same technology. In England, resident-facing services now sell policy-cited objections to planning applications for a flat fee in minutes. One AI writes the objection. Another AI summarises it. A planner signs off on the summary. At no point has anyone confirmed that a neighbour exists.

We have made a version of this argument before about the one step in financial aid verification a model should never take. The pattern is the same. The step that looks laborious is safe to automate. The step that looks trivial is where judgment and accountability actually live.

Where the human step belongs now

The honest reframing is that public comment has two jobs, and AI changes the economics of only one of them.

The first job is information. A resident knows the intersection floods, or that the school pickup line already backs onto the arterial, or that the historic survey missed a building. That content is valuable whoever typed it, and a well-built summarisation tool helps you find it in a large pile. Here the vendors are right.

The second job is standing. The comment tells you that a real person, with a real relationship to the site, holds a view. That is what a council member is weighing when they feel the count. It is also the part a model can now forge at no marginal cost, and the part no summary can recover after the fact.

Keep the first job cheap and automated. Move human effort, and your procurement budget, to the second. In practice that is a design question about intake, not about analysis. APA's own coverage points in this direction: asking for information that establishes a stake, such as location, and questions specific to the issue, while keeping verification light enough that it does not shut out the residents a comment period exists to hear. The Planning Inspectorate in England has gone at it from the other side, requiring since its February 2026 update that parties disclose when AI created or substantially changed a submission, which tool, and what verification they did.

California has now moved on the definitional question. SB 1159, from Senator Christopher Cabaldon, specifies that "person" and "member of the public" in the Public Records Act and open meeting statutes do not include AI systems or autonomous agents, and the Assembly amendments made clear that AI may not be used to spoof a real person submitting comment without consent. It passed the Senate 38 to 0, cleared the Assembly in August, and went to the governor on August 28. Whatever its final status, notice what it does and does not do. It gives an agency a basis to disregard machine submissions. It does not tell the agency how to identify them. That operational question lands on your intake process.

Why this belongs before the software decision

If your department is scoping an engagement platform or an AI analysis module this year, the order of the questions matters more than the vendor.

Ask how a submission becomes attributable before you ask how fast it becomes a theme. Ask what your record will show if a commission vote is challenged on the ground that the comments behind it were fabricated. Ask whether your notification practice, the radius you mail and the addresses you rely on, gives you any way to tell an affected resident from a stranger. These are policy and records questions first. The software encodes the answers, the way routing software encodes whatever review stages you already have, a point we made about automating review routing versus cutting review stages.

There is also a case, in some departments, for not buying the analysis module at all yet. If your comment volumes are manageable and your intake cannot distinguish a resident from a campaign, an AI summary will mostly make an unreliable input look authoritative. We wrote about that shape of problem in when AI is the wrong answer to an operations problem. A faster reading of a record you cannot vouch for is not progress.

Where this article stops

We cannot tell you from here how much verification your process can bear before it deters the residents you most need to hear from. That depends on your state's open meeting and records law, what your hearing procedures promise, which of your decisions attract organised campaigns, and how your current intake records identity, if it does at all. Those answers sit in your own procedures and your own comment history.

If you are about to buy comment analysis software, or you have just run a comment period where the volume did not match the room, talk to us before the next contested hearing. We will map where your process establishes who a commenter is, where it assumes it, and what to fix before any tool is asked to read the result.