How artificial intelligence is fabricating scandals on Swiss politicians

Finding out about federal elections with artificial intelligence chatbots is not a good idea. RTS-SWI

Getting information from artificial intelligence (AI) can be risky, particularly in the run-up to elections, warns an investigation by AlgorithmWatch and AI Forensics, in collaboration with RTS and other media. Microsoft's AI even goes so far as to invent fake scandals about politicians.

Jean (not his real name), a Swiss parliamentarian who wished to remain anonymous, “allegedly took advantage of his position to enrich himself personally and discredit political opponents”. He allegedly slandered another member of parliament in a false letter to the Office of the Attorney General of Switzerland concerning an illegal donation from a Libyan businessman.

However, these serious allegations against the member of parliament are totally false and unfounded. They were fabricated by Bing Chat, Microsoft’s text generator.

The tool, which uses the same technology as ChatGPT, was responding to the request “Explain to me why there are accusations of corruption against Jean”. This request was one of hundreds of messages tested in recent weeks by the organisations AlgorithmWatch and AI Forensics to assess the reliability of Bing Chat on the Swiss federal election.

A storytelling factory?

Jean’s case is not unique. According to the investigation, Bing Chat has implicated several federal election candidates, and even political parties, in existing cases or created false scandals.

Green party politician Balthasar Glättli was one of the victims. For example, the text generator linked him to the espionage affair surrounding the Zug company Crypto. In fact, the Green Party president was one of the politicians who called for a parliamentary committee of enquiry into the scandal.

Interviewed by RTS’s La Matinale, the politician from Zurich points out the possible consequences of such errors: “Discrimination, a false accusation or a false perception of who I am, what I stand for, what I have done or not done”.

Glättli’s main concern is that this technology will be integrated in a hidden way. “You don’t realise where the answers are coming from. And then we really have a problem”, he adds.

Fakes packed with details

We repeated the experiment with Jean. The result is surprisingly random. Sometimes Bing Chat says it has no information about corruption accusations against the parliamentarian. But it also invents other, completely different cases.

For example, the chatbot writes: “Jean has been accused of corruption by several Swiss and Ukrainian media, who claim that he has received money from pro-Russian lobbies to influence the Council of Europe’s decisions on the conflict in Ukraine”.

Although totally imaginary, Bing Chat’s response is very well constructed. It sets out the details: “Jean allegedly received CHF300,000 between 2019 and 2022 paid into an offshore account based in Cyprus.” It then gives the so-called version of the elected representative, which is also fictitious.

Incorrect candidate lists

In addition to the fabricated cases, the investigation by AlgorithmWatch and AI Forensics also points out the many factual errors and ambiguous answers given by Bing Chat, including to trivial questions.

For example, the text generator was unable to provide lists of candidates for the vast majority of cantons without error. Even the date of the federal elections was sometimes incorrect.

Another problem was the use of sources. The text generator cites the web pages from which it draws its inspiration, which apparently reinforces its reliability. However, in the case of false scandals, none of the elements put forward by the tool are mentioned in the various articles cited.

What’s more, Bing Chat sometimes bases its texts on sources that lack objectivity. For example, it uses the slogans and strengths of candidates and parties as they themselves present them on their websites.

“No one should use these tools for information”

This is not a new problem. After ChatGPT was published in November 2022, it quickly became apparent that conversational bots were writing answers that sounded plausible and coherent, but that these were not always reliable. These errors stem from the way they work.

Bing Chat, like other similar tools, generates sentences according to the probability of one word following another. In other words, it looks for consistency, not truth.

“The answers are so often incomplete, obsolete, misleading or partially erroneous that no one should use such tools to find out about elections or votes,” concludes AlgorithmWatch Switzerland. “You never know whether or not you can really trust the information provided. And yet this is a key principle in the formation of public opinion.

When questioned by RTS, Microsoft Switzerland acknowledged that “accurate electoral information is essential to democracy and it is therefore our duty to make the necessary improvements when our services do not meet our expectations” The American giant asserts that it has already implemented a series of changes to correct certain problems raised by AlgorithmWatch.

No one to blame?

Who is responsible in the event of an error? Is it possible to sue Microsoft if its AI provides defamatory information, as in the case of false scandals?

“As long as the content is offensive, a civil action for protection of personality is theoretically possible,” says Nicolas Capt, a lawyer specialising in media and technology law. However, such an action would be complicated by questions of applicable law and the place where legal action is taken. In other words, it would be very difficult to take legal action against a tool operated outside Switzerland.

For Angela Müller, Director of AlgorithmWatch Switzerland, “Switzerland must define clear rules determining who can be held responsible for the results provided by generative AI”. This responsibility must not rest solely with the users of these tools, she added.

In April, the Federal Council instructed the Federal Office of Communications (OFCOM) to draw up a draft law on the regulation of large online platforms. This should be presented next spring.