Bots Sent 90,835 of 118,756 Responses to the UK's 2023 Vape-Ban Consultation, Study Finds

TL;DR

A paper published October 27 in Tobacco Control by University of Edinburgh health policy researcher Nason Maani and University of Bath professor Anna Gilmore picks apart the bot problem in the UK's 2023 smokefree generation consultation. The Department of Health and Social Care (DHSC) received 118,756 responses and flagged 90,835 of them, 76%, as submitted by automated programs. According to the Guardian's account of the paper, all but two of those 90,835 answered the same question, against the plan to ban disposable vapes. The department threw them out, the bill became the Tobacco and Vapes Act 2026 anyway, and the researchers call the episode "the first documented instance of large-scale automated interference in a formal public health policy consultation." The builder angle: this wave got caught because the bots skipped the survey flow and hammered from a few IP addresses. The 2026 wave, in which Philip Morris handed smokers an AI text generator for an EU consultation, had neither tell.

The numbers have been in a government document since January 2024

DHSC published its consultation response in January 2024, updated that February, and the raw counts are in it. 118,756 responses arrived between 12 October and 6 December 2023. 90,835 were "unambiguously identified as having been submitted by automated programmes, commonly known as bots." That left 27,921 valid responses: 896 from organisations and 27,025 from individuals.

The detection method takes one sentence: the department "identified fraudulent responses where the data submitted was incompatible with the survey flow followed by genuine respondents who completed the consultation form manually." The submissions were also "often duplicated many times and sent at high frequency from a small number of IP addresses, which is consistent with a mass automated campaign."

A survey-flow check is the invigilator noticing that a student has answered questions on pages that were never handed out. The bots posted answers for a form state the real form never produced, and once someone looked, that was an easy catch.

What the new paper adds is the shape of the campaign. The Guardian reports that all but two of the 90,835 automated responses answered a single question in the consultation, opposing Rishi Sunak's Conservative government's plan to ban disposable vapes. Holyrood reports that the targeted question was multiple choice and the bot answers were overwhelmingly against the restrictions. The other two apparently wandered off script.

What the real respondents said

With the bots removed, 79.6% of valid responses agreed there should be restrictions on the sale and supply of disposable vapes, 18% disagreed and 2.4% did not know. 69% backed an outright prohibition on sale and supply. The headline measure, a rising age of sale so that anyone born on or after 1 January 2009 can never legally buy tobacco, got 63.2%. Restricting vape flavours was the one proposal that lost, 47% for and 51% against.

So the campaign picked the question where a flood of "disagree" votes would have flipped a four-to-one result. Counted in, the 90,835 would have outnumbered every genuine response to the whole consultation by more than three to one. Keir Starmer's Labour government inherited the bill in 2024 and passed it. It received Royal Assent on 29 April 2026.

Where the AI comes in

Nothing in the DHSC description requires a language model. A script that posts the same multiple-choice payload 90,000 times from a handful of IP addresses is 2010s technology. The Guardian's headline calls them AI bots, and Action on Smoking and Health deputy chief executive Caroline Cerny told the paper that "evidence that AI bots have been used to swamp a Department of Health consultation is deeply concerning," but the study's own title says "automated submissions." Keep that distinction, because it is what made the campaign catchable.

The version that is hard to catch arrived this summer. On June 28, Dutch broadcaster NOS and the investigative platform Pointer reported that Philip Morris International had put yellow "Your Voice Your Choice" posters with QR codes in tobacco shops. The code led to a tool that asked a few multiple-choice questions, then wrote a first-person text opposing stricter rules for the user to send to the European Commission's tobacco consultation. NOS and Pointer ran about 65,000 of the roughly 80,000 responses, from 15 countries, through Pangram's detector. About 30% looked fully AI-written. In the Netherlands it was 71%, second only to Portugal. For scale, NOS notes that most Commission consultations draw no more than 1,000 responses.

Philip Morris said it used standard AI settings to make the texts readable and called the campaign a legitimate contribution to the democratic process. Making text readable is also what the rest of us call writing it. The Commission said it was very concerned and would take AI-driven campaigns into account when it reads the input.

These are two different attacks on the same institution. The UK one was volume with no text, defeated by a form-state check. The EU one was unique, grammatical, first-person prose submitted by real humans who scanned a code, which passes every integrity check a form can run. Azzurra Moores, associate director for information ecosystems at the think tank Demos, gave the Guardian the tradeoff: "Every barrier that keeps bots out also makes it harder for real people to take part, and AI makes fake responses far harder to spot. Interference can adapt faster than departments can update their checks."

The 2017 column is the New York Attorney General's 2021 finding that nearly 18 million of the more than 22 million comments the FCC received on its net neutrality repeal were fake. More than 8.5 million came from a broadband-industry-funded campaign that used lead generators and real people's names, and about 9.3 million pro-neutrality comments ran under fictitious identities, most of them from one 19-year-old college student with automated software.

What the researchers could not see

Maani and Gilmore worked from the department's published counts. Per Holyrood, they asked DHSC for respondent metadata and demographic information and were refused under Freedom of Information rules, so they could not verify the detection method themselves. They write that it was "not possible to ascertain the commercial origins of such a campaign," while "the scale and directional consistency of bot responses suggest a coordinated campaign," in line with earlier tobacco industry efforts to swamp consultations through mass recruitment and postcard submissions. Nobody has been named.

Maani's warning to the Guardian is the line to keep: "the threat is that future campaigns become harder to catch, and what we can't catch, we don't know about. Such campaigns are a threat because they raise the possibility of a handful of powerful interests looking like a mass movement." DHSC's position is that its controls "worked effectively in this case," the fake responses were removed, and it "will remain alert to interference of this kind in future consultation processes."

If you run a form that anyone can submit to

A public consultation is a high-stakes form, and the same failure modes apply to a feedback widget, a review pipeline, a community vote, or the comment endpoint your agent product exposes.

• Keep form state on the server and reject answers to questions the client was never shown. That one check caught 90,835 submissions here, and it costs almost nothing to run.
• Report unique respondents rather than raw submissions. DHSC's 27,921 is the number that meant anything; 118,756 was noise. If your dashboard counts rows, a script owns your metrics.
• Rate and IP clustering still work against lazy campaigns, and lazy campaigns are still most of them.
• Plan for the free-text wave now. A Pangram-style classifier was the only reason the EU campaign could be measured at all, and if a human's submission can be discarded on a classifier's say-so, you owe people the error rate.
• Decide in advance what you do with a campaign that is organised but legal. Philip Morris's submitters were real people. The Commission chose to weigh campaign activity rather than delete it, and that policy call is the hard part. No classifier makes it for you.

Key Takeaways

• DHSC's own count: 90,835 of 118,756 responses (76%) to the UK's 2023 smokefree generation consultation came from bots, leaving 27,921 valid responses.
• Per the new Tobacco Control paper as reported by the Guardian, all but two bot responses answered one question, opposing the disposable vape ban. With bots removed, 79.6% of valid respondents backed restrictions.
• The bots were caught because their data was incompatible with the survey flow and arrived in bursts from a few IP addresses. No language model was needed to run that campaign, or to catch it.
• The June 2026 Philip Morris EU campaign used an AI text generator behind a QR code. Pangram judged about 30% of 65,000 responses AI-written, and 71% in the Netherlands.
• The researchers could not verify DHSC's detection method or establish who ran the campaign; DHSC refused their request for respondent metadata.
• If you run any public submission pipeline, check form-state integrity, count unique respondents, and decide how to treat organised campaigns before one arrives.

Sources: Tobacco Control (Maani and Gilmore, 2026), DHSC consultation response, The Guardian, Holyrood, Tobacco and Vapes Act 2026, NOS, DutchNews, New York Attorney General