The Vote
In June 2025, the South Coast Air Quality Management District — the regulator covering the Los Angeles basin, one of the most polluted air sheds in the country — voted 7 to 5 to reject Rules 1111 and 1121. The rules would have phased gas furnaces and water heaters toward electric heat pumps: 30% of sales by 2027, 90% by 2036. Staff projected roughly 2,500 fewer premature deaths and $25 billion in health savings through 2053. The board killed it anyway.
What tipped the vote wasn’t a study or a competing health finding. It was volume. More than 20,000 public comments poured in opposing the rules — against a typical agenda-item count in the single digits. A Los Angeles Times investigation published eight months later, in February 2026, traced a large share of that flood to a single source: CiviClick, a Washington-based company that markets itself as “the first and best AI-powered grassroots advocacy platform.” The investigation concluded the campaign “almost certainly” swung the board’s decision.
Nobody caught it in real time. The rule was already dead by the time anyone could prove the public that killed it wasn’t entirely real.
What Regulators Found When They Called
SCAQMD staff went back through a sample of the 20,000 names on file and tried to reach them. They contacted 172. Five responded. Of those five, two confirmed they had written in. Three said they had no idea what the agency was talking about.
Five is not a large sample. It doesn’t prove all 20,000 comments were fabricated, and nobody investigating this claims it does. What it proves is narrower and still damning: a nontrivial share of the “public” the board weighed against 2,500 projected deaths didn’t know it had spoken.
The pattern inside the comments reinforced the doubt. Investigators found multiple messages traced to the same CiviClick-linked email address, submitted under different names. One comment thanked SCAQMD’s executive officer, Wayne Nastri, for opposing the very rule his own agency had drafted and proposed — a sentence no one who actually understood what they were writing about would produce. It reads like a template that got the target backwards, sent by a name attached to nobody who checked it before it went out.
CiviClick’s founder, Chazz Clevinger, denied that his platform used AI to generate this specific campaign, while confirming the company offers AI-assisted message personalization generally. Take that denial seriously rather than waving it away — it’s the one falsifiable claim in this story, and if AI authorship can’t be proven for this campaign specifically, the headline shouldn’t lean on it. But notice what the denial doesn’t touch. Whether or not a language model drafted the sentences, the comments arrived under real people’s names, at a scale no organic petition drive produces, coordinated by a consultant — Matt Klink, a partner at the firm representing SoCalGas’s parent company — who described the platform as knowing “the local targets who would respond to the messaging.” The mechanism that mattered wasn’t authorship. It was appropriation: real identities, attached without consent to a position those people may never have held, deployed at a volume built to read as consensus.
It Wasn’t a One-Off
Six hundred miles of coastline away, the same pattern surfaced almost immediately. The Bay Area Air District received a wave of comments opposing new regulations, submitted through a platform called Speak4 on behalf of the Common Sense Coalition — an industry group backed by the Bay Area Council, whose membership includes the Western States Petroleum Association, Chevron, and Phillips 66. Ten Bay Area residents whose names appeared on the letters told the San Francisco Chronicle they never wrote them and never authorized anyone to use their names. San Pablo resident Anthony Clewis put it plainly:
“This was forged. I never wrote the letter. It’s a travesty and conflict of interest and an invasion of privacy.”
CiviClick shows up a third time in North Carolina, tied to a gas pipeline fight. Two county commissioners replied to what they believed were messages from constituents — thanking them, engaging with their concerns — only to find the named senders had no idea the messages existed.
Three jurisdictions. Two platforms. One consistent mechanism: take a real name, attach it to a position, submit it into a process designed to weigh public sentiment, and let volume do what a persuasive argument would otherwise have to do on its own.
Why the Comment Period Was the Target
Social media astroturfing gets the bulk of manipulation-research attention because it’s visible — anyone can scroll past a wave of suspiciously similar posts. A regulatory public-comment period is a better target precisely because almost nobody watches it. It has no algorithm to game, no platform trust-and-safety team, no journalists doing daily monitoring. It has a staffer counting emails, a board that treats volume as a rough proxy for how many people care, and a legal design that assumes the person whose name is on a comment is the person who wrote it — because verifying that at scale was never anticipated as necessary.
That assumption is the actual vulnerability, and it’s structural rather than technical. Public comment processes exist because agencies are supposed to hear from the people a rule affects. The entire legitimacy of the mechanism rests on the comments being what they claim to be: a real person’s real position, freely given. Break that link — supply the volume without the consent — and the process doesn’t fail loudly. It just returns the wrong answer while looking exactly like it’s supposed to.
This is fabricated consensus, the same influence tactic that drives bot networks and sockpuppet floods on social platforms, redirected at a process with far higher stakes and far less scrutiny. A fake trending hashtag might shift a news cycle for a day. A fake public comment record can kill a regulation that would have prevented thousands of deaths, and the record itself becomes the government’s own evidence that “the public” didn’t want it.
The Eight-Month Gap
The most instructive detail in this story isn’t the fraud. It’s the lag. The vote happened in June 2025. The investigation that surfaced the fabrication published in February 2026 — eight months later, after the rule was already dead and the board had already moved on. Advocates and, eventually, 22 California state and local elected officials asked the Attorney General and four district attorneys to investigate. That process is still open. The rule is not coming back on the same timeline the investigation moved on.
This is the part that should worry anyone thinking about manipulation defense as a race against detection. Detection eventually worked here — journalists made the calls, found the pattern, published it. But eventually is not a defense when the decision it’s correcting was final the moment the board gaveled the vote. A fabricated consensus doesn’t need to survive scrutiny forever. It needs to survive long enough to get counted.
Key Findings
- 20,000+ comments opposed SCAQMD’s clean-air rules; a typical agenda item draws single digits. The board voted 7–5 to reject rules projected to prevent roughly 2,500 premature deaths and generate $25 billion in health savings through 2053.
- Of 172 commenters contacted, five responded — two confirmed, three said they never wrote in. A small sample, but one that establishes the fabrication rather than merely alleging it.
- The tell was internal to the comments themselves: duplicate submissions traced to one CiviClick-linked address under different names, and a message thanking the agency’s executive officer for opposing his own agency’s rule.
- The platform’s founder denied using AI for this specific campaign. That denial matters and shouldn’t be dismissed — but it doesn’t address the actual mechanism: real names attached without consent, not the method used to draft the prose.
- The same pattern recurred in two more jurisdictions — Speak4 and the Common Sense Coalition in the Bay Area, CiviClick again in a North Carolina pipeline fight — with named residents in each case confirming they never authorized the messages sent under their identities.
- The exposure arrived eight months after the vote it concerned. The regulation was already dead; accountability is still pending.
Implications
Most manipulation-detection work is built for content that persuades directly — a post, a video, a claim designed to change what a reader believes. This case is different in a way that matters for anyone building detection tools: nobody needed to persuade the SCAQMD board of anything. They needed the board to believe a lot of people already held a position, using an evidentiary channel — the public comment record — that the agency itself had no practical way to verify at scale. The manipulation target wasn’t a belief. It was a headcount.
That argues for a different kind of defensive attention than platform content moderation provides. Regulatory comment systems, petition platforms, and any process that aggregates “the public’s” input as a decision input are attack surfaces with essentially no monitoring layer, built on infrastructure — email, web forms — that predates any expectation of industrial-scale identity misuse. The fix isn’t a better fact-check. It’s a verification step regulators currently don’t have the mandate or the tooling to perform before a vote, not eight months after one.
Conclusion
Twenty thousand comments looked like a public speaking. Five phone calls found that at least three of them weren’t. The rule they killed doesn’t reappear because the fraud got named — it stays dead while the investigation into who forged it grinds through an Attorney General’s office. The lesson isn’t that AI wrote fake letters; CiviClick’s founder is right to contest that specific claim. The lesson is that nobody needs a language model to fabricate a public. They need a name, a form, and a process that was never built to ask whether the name and the words belong to the same person.
Sources: Legal Planet — A.I. Pollution in the Air — and the Public Comments, Canary Media — Did fake comments sink SoCal clean heat rules?, Sierra Club — Advocates Call for Investigation into AI Campaign that Used Residents’ Identities Without Consent, San Francisco Chronicle — Residents deny sending emails pushing Bay Area air policy change, Phys.org — Southern California air board rejected pollution rules after AI-generated flood of comments