Bots wrote three quarters of the answers to Britain's vape ban consultation

In October 2023 the UK government asked the public what it thought about banning disposable vapes. It got 118,756 answers. Roughly three in four of them were written by nobody at all.

A study published this week in the journal Tobacco Control says 90,835 of those submissions were machine-generated. That is about 76% of everything the Department of Health and Social Care received. The authors, from the Universities of Edinburgh and Bath, include Dr Nason Maani and Professor Anna Gilmore, according to The Guardian, which broke the story on 27 October.

The fake responses did not spread their opinions around. With two exceptions, every one of them picked the same multiple-choice option: against restricting disposable vapes. A machine-made consensus, delivered by the tens of thousands, aimed at a single question.

Strip them out and the picture flips. Of the valid responses that remained, 79.6% supported the restrictions. The ban on selling and supplying single-use vapes took effect in June 2025.

How the bots got caught

This was not a subtle operation. Officials noticed submissions arriving at high frequency from a small cluster of IP addresses, and the text was heavily duplicated, IBTimes UK reports. They ran automated bot screening and threw out the lot. Almost all of the fake respondents also claimed they had heard about the consultation through social media, which is the kind of detail a lazy script fills in the same way every time.

The department is pleased with itself. A DHSC spokesperson said "We always have robust controls in place to prevent the manipulation of our public consultations" and that "those controls worked effectively in this case."

Fair enough. They did. The researchers care more about the attempt. Their paper calls it likely the first documented case of large-scale automated interference in a formal UK public health consultation. In their words: "The scale and directional consistency of bot responses suggest a co-ordinated campaign."

Who ran it is the open question. The researchers say they could not pin down where the campaign came from or who paid for it. They do point at history. Tobacco companies have spent decades trying to swamp consultations with favourable responses, through mass recruitment drives, pre-printed postcards and front groups that look like grassroots smokers. A bot farm is the same move with the postage removed.

Action on Smoking and Health, the anti-smoking charity, called the use of "AI bots" to overwhelm health consultations "deeply concerning." That is the polite version.

The next one will be harder to spot

Here is the uncomfortable part. The 2023 campaign was caught because it was crude. Same answer, same wording, same handful of IP addresses. That describes a bot from three years ago, before cheap language models could write ninety thousand different paragraphs in ninety thousand different voices, each one with its own typos and its own made-up anecdote about a cousin who quit smoking.

Maani has warned that future campaigns of this kind could be harder to detect and could make a small number of interests look like a mass movement. The paper says screening tools will need to keep evolving as bot sophistication increases. That is an arms race, and it is one where the attacker's costs keep falling.

Think about what a public consultation actually is. It is a web form. Anyone can fill it in, as many times as they like, from wherever they like. The whole mechanism rests on the assumption that the people writing in are people, and that a lot of responses means a lot of people care. Both assumptions were already shaky. Generative AI makes them optional.

Big government departments have IT teams and fraud specialists. Coverage of the study flags the bodies that don't, such as local councils with thin technical resources. A council consulting on a new data centre, run by one overworked policy officer and a spreadsheet, does not have a bot-detection pipeline. It has a form and an inbox.

There is also a quieter failure mode. A campaign does not need to win the count to work. It only needs to muddy it. If officials have to spend weeks cleaning a dataset, or if a minister can later wave at "huge public opposition" without mentioning that most of it came from a server rack, the bots have done their job even after being deleted.

The vape ban survived. The Tobacco and Vapes Act, which also bans tobacco sales to anyone born on or after 1 January 2009, received Royal Assent in April. On this occasion the system held.

But the attack cost almost nothing to run, nobody has been named, and nobody has been punished. Whoever wrote that script still has it. The next consultation is already open somewhere, and its form still has a submit button.

Sources: The Guardian, IBTimes UK, NationalWorld.