The UK government received 118,756 submissions to its 2023 consultation on tobacco and vaping. It excluded 90,835 as automated and fraudulent, leaving 27,921 for analysis. About three-quarters of the submissions were discarded before that analysis, according to the Department of Health and Social Care’s published response.

That is a striking example of the gap between submission volume and useful public evidence. The department’s account also provides something more specific than a warning about bots: an explanation of the patterns it used to exclude them. It supports a finding of automated submissions. Establishing whether generative AI wrote their text requires different evidence.

The submissions arrived in 2023 and were disclosed in 2024

The UK-wide consultation ran from 12 October to 6 December 2023. Its questions covered proposals for a smokefree generation and measures to tackle youth vaping, including restrictions on disposable vapes. The consultation’s update history records publication of the government response on 29 January 2024. The exclusion figures were already public then.

The case received renewed attention on 27 October 2026, when Nason Maani and Anna Gilmore published a Tobacco Control paper, Examining the nature of bot responses to public health consultations: the example of automated submissions to the 2023 UK Tobacco and Vapes Bill. The Guardian reported on it that day, describing “AI bots”.

We could verify the paper’s publication details but could not access its full text. The account here therefore rests on the government’s published methodology, with the later reporting identified separately.

DHSC detected patterns in how the survey was submitted

In its methodology section, DHSC says the excluded responses contained data incompatible with the survey flow followed by someone completing the questionnaire manually. Responses were also often duplicated many times and submitted at high frequency from a small number of IP addresses.

The department describes this combination as consistent with a mass automated campaign using bots. It says the 90,835 responses were “unambiguously identified” as automated and deemed fraudulent.

Those signals concern how submissions reached the consultation system. A program can repeatedly submit a form containing fixed text without using a generative model. Generative AI can produce text that a person then submits manually. People can also coordinate around shared wording while completing forms themselves. These activities can overlap, but evidence of one does not establish all three.

The published methodology describes several signals together. It gives no basis for treating a shared IP address or repeated wording, on its own, as proof of fraud. Nor does it identify a generative model or establish who organised the submissions.

The exclusion total records DHSC’s decision about the responses it identified. It is not an independent measurement of detection accuracy: the accessible evidence does not quantify missed automation or genuine submissions wrongly excluded.

Consultation responses inform policy through evidence and analysis

A public consultation gives people and organisations an opportunity to contribute views and evidence while policy is being developed. The Cabinet Office’s consultation principles call for departments to take responses into account and explain how they informed policy.

That process is different from counting votes in a binding referendum. It also gives no reason to read the respondents as a representative sample of the UK population.

Respondents could choose which policy questions to answer. The retained responses came from 896 organisations and 27,025 individuals responding in personal or professional capacities. These are counts of responses in the department’s analysis, not a verified count of everyone whose views the consultation represented.

DHSC’s method extended beyond totals for closed questions. For individuals’ free-text answers, it used topic modelling and tagging, with at least two officials reviewing word groups and representative quotations for each question. Organisation responses were manually grouped into themes. This allowed arguments and evidence to be considered alongside the numerical results.

The Cabinet Office guidance does not prescribe a formula assigning each submission a numerical weight. It is also guidance without legal force, subject to applicable legal requirements. Departments should explain how the evidence they considered and the submissions they excluded informed policy.

The later ban does not establish what influence the bots had

The UK ban on supplying single-use vapes took effect on 1 June 2025. In England, it was introduced through the separate Environmental Protection (Single-use Vapes) (England) Regulations 2024, rather than through the tobacco bill. That instrument applies to England; the UK-wide date is confirmed in government guidance.

The subsequent ban and the removal of the automated submissions establish a sequence of events. They do not establish what policy would have resulted without the campaign, or measure its influence through other channels. The accessible evidence also does not identify a sponsor.

Participation safeguards need an account of who can still take part

Removing mass automated submissions addresses one problem. Making participation harder can create another. In the Guardian’s reporting, Demos’s Azzurra Moores raises the risk that anti-bot controls create barriers to participation. That is an expert warning, rather than a measurement of harm to respondents in this consultation.

Access is already part of the Cabinet Office’s guidance. It calls for consultation methods suited to affected groups, including older people, younger people and disabled people. It also calls for publication of response totals and an explanation of how responses informed policy.

Together, these provide concrete questions for scrutinising a consultation: what was excluded and why, how was the remaining evidence used, and could the people affected still contribute? DHSC published its exclusion count and a description of its method. Assessing access would require evidence about barriers encountered by genuine participants.