ChatGPT and other AI bots made huge errors before Scottish election, study finds

Details
Parent Category: chatbots
Category: chatbots in politics

Electoral Commission calls for new controls as Demos finds tools made up fake scandals, invented candidates or gave wrong date

The Electoral Commission has called for new legal controls over misinformation from AI chatbots, after a thinktank found they had made serious mistakes during the recent Scottish election.

The thinktank Demos said its investigation had found that AI services gave voters misinformation to 34% of the questions it posed, which it said raised worrying questions about the lack of regulation of AI platforms in the UK.

It ran a simulation before May’s Holyrood election by putting 75 questions to five free AI tools including ChatGPT, Google Gemini and Replika about three real-life constituencies to see how accurate and evidence-based their responses were.

In its report, Electoral Hallucinations, Demos said those AI tools variously invented fictitious scandals, gave the wrong date for the election, claimed wrongly that voters in Scottish elections needed ID at polling stations and placed candidates in the wrong contests.

An opinion poll of 2,005 British adults it commissioned alongside that study found that 20% of voters had used AI chatbots or search tools to get information about the parliamentary elections in Scotland and Wales, and for English local councils, equivalent to 10 million people UK-wide.

Vijay Rangarajan, the Electoral Commission’s chief executive, has been pressing ministers to introduce legislation to make AI companies more accountable, after discovering half of voters in 2024’s general election had seen misleading information.

“Voters want accurate information to help them engage with democracy and it is concerning that AI tools have made the spread of false or misleading information dramatically faster and more accessible than ever,” he said. “The current legal framework should go further.”

He said ministers should introduce clearer duties on AI platforms to protect voters against misinformation and ensure algorithms did not mislead voters, particularly during critical election periods. That would give the media regulator, Ofcom, much clearer powers to enforce the law.

Azzurra Moores, an associate director at Demos, said: “This is a UK-wide, if not global, concern. The accessibility of these AI-tools – which are all developed and run by US corporations – is widespread in the UK, but we don’t yet have the legislative framework to protect the public from misinformation, or our democracy from the knock-on impact of its circulation.”

She said ministers could quickly introduce legal requirements to make AI companies liable under UK defamation and electoral law, introduce mandatory safeguards on accuracy, and force AI firms to allow researchers to independently test how their internal data and training sets worked.

Demos said the so-called companion chatbot Replika had performed the worst in its tests, with errors in 56% of its answers. It invented a date for a made-up expenses scandal, invented a candidate and dreamed up accusations of nepotism by a candidate.

ChatGPT, the most heavily used AI service, gave wrong information in 46% of its answers, including making up an expenses scandal, giving inaccurate replies on voter eligibility rules and getting the date of the election wrong by two months.

Google Gemini was wrong in 22% of cases: it said a candidate had not taken a position on assisted dying, when that person was a supporter. It also wrongly claimed a police investigation into a fraud case involving the Scottish National party was continuing.

Grok, the AI service linked to Elon Musk’s social media platform X, had the lowest error rating, at 9%, but its external links were frequently irrelevant or of poor quality. Demos included Google’s automated AI Overviews service, offered by Google search, but disregarded its output because it answered only 11% of the prompts.

Source: The Guardian
Main Menu

Artificial Intelligence

Other technologies