The count that changed the answer

The UK government asked the public, in late 2023, whether it should create a smokefree generation and crack down on youth vaping. One question in that consultation concerned disposable vapes. The answer it received, before any filtering, pointed one way. The answer that survived filtering pointed the other.

DHSC recorded 118,756 responses in total and classified 90,835 of them as fraudulent submissions from automated programmes. That is the department's classification, examined by researchers using published consultation records and Freedom of Information material; the researchers did not independently reproduce the count from raw submissions. The remaining 27,921 responses were treated as valid, including 896 from organisations. The bots outnumbered the genuine responses by more than three to one.

Among the valid respondents who answered the disposable-vape question, 79.6% wanted restrictions and 69% agreed disposable vapes should be banned from sale outright. Put the bot submissions back in and the picture reverses: the unfiltered data would have made it look as though 81.3% of respondents opposed restrictions. The department excluded the automated responses from its decision-making. Single-use vapes have been banned from sale in the UK since June 1, 2025, and the wider Tobacco and Vapes Act received Royal Assent on April 29, 2026.

What the automated submissions looked like

Every bot-flagged submission consisted of multiple-choice selections, with no free text at all. The submissions did not follow the path real respondents took through the form, according to DHSC.

The submissions were concentrated around the disposable-vape question: 90,833 opposed prohibiting the sale and supply of disposable vapes. The same number selected social media as how they had learned about the consultation.

It does not show a language model writing persuasive prose. DHSC called them automated programmes, commonly known as bots. Some headlines have shortened that to "AI bots". The paper's own title is more careful: bot responses, automated submissions. A reader who takes the headline version gets a different, stronger claim than the evidence supports.

The pattern was unusually crude, and it is worth stating plainly. The published evidence shows automation at scale: repeated answers, submitted at high frequency, from a small number of IP addresses. It does not establish precisely which signals DHSC used to detect and classify the submissions.

What is not established

The study does not identify who submitted the automated responses. Its authors say their scale and directional consistency suggest a coordinated campaign. They note that tobacco companies have previously sought to influence consultations through mass recruitment and postcard campaigns. But they explicitly state that the available information cannot establish the source, intent or commercial sponsorship of these submissions. No tobacco or vaping company has been identified in the study as behind these submissions.

There is a second limit the authors state themselves. They could not independently verify DHSC's bot-detection methods, because the department did not disclose them, and metadata and demographic details about respondents were withheld under Freedom of Information exemptions. The 90,835 figure is DHSC's classification, examined by the researchers through published reports and FOI material, not a count the researchers reproduced from raw submissions.

That does not make the finding empty. It makes it bounded: a department flagged three quarters of a consultation as automated, on grounds it described in general terms, and the researchers showed what removing that block did to the result. The boundary matters because the next campaign may not leave the same traces.

The warning that survives the caveats

Maani's warning, quoted in the Guardian's coverage, is that future campaigns will be harder to catch: "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." Azzurra Moores of the think tank Demos made the accompanying point: 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.

That is the practical problem for any government that runs consultations online. A consultation is not a referendum. It measures the submissions received, not a random sample of the public. When submissions are cheap to manufacture, the count stops measuring opinion and starts measuring whoever is best at manufacturing it. This one was caught, on DHSC's classification. A future operation using varied, plausible free-text responses, potentially generated with language models, could be much harder to distinguish from genuine participation. That is a prospective risk, not a capability demonstrated by the submissions examined in this study.

Sources

  1. [1] Maani and Gilmore, "Examining the nature of bot responses to public health consultations: the example of automated submissions to the 2023 UK Tobacco and Vapes Bill", Tobacco Control, published online October 27, 2026, read first-hand for this article: 118,756 total responses, 90,835 (76%) identified by DHSC as fraudulent, all bot-flagged submissions multiple-choice only, 90,833 opposing restrictions on disposable vape sales, 79.6% of valid respondents supporting restrictions after exclusion, 81.3% apparent opposition had all responses been included, and the stated limits: DHSC's detection methods not disclosed, independent verification not possible, and source, intent or commercial sponsorship not attributableRead source
  2. [2] UK Department of Health and Social Care, consultation page and response for "Creating a smokefree generation and tackling youth vaping" (consultation published October 12, 2023): the quoted wording that submissions were "unambiguously identified as having been submitted by automated programmes, commonly known as bots, and so deemed to be fraudulent" and "often duplicated many times and sent at high frequency from a small number of IP addresses", and the 27,921 valid responses, 896 of them from organisations, analysed after exclusionRead source
  3. [3] BMJ Group summary of the study, October 27, 2026: reports 77% of valid respondents supporting restrictions, a different denominator presentation from the paper's own 79.6% of valid respondents after exclusion; the paper's figure is the one used in this articleRead source
  4. [4] MadRobot, "Bots sent 76% of responses to UK vape ban consultation", October 28, 2026: consultation dates October 12 to December 6, 2023; of the valid respondents who answered, 79.6% wanted restrictions and 69% agreed disposable vapes should be banned from sale outright; single-use vapes banned from sale since June 1, 2025; the Maani and Moores quotesRead source