“ChatGPT, can you provide me with a list of journalists who have been the target of allegations of sexual misconduct in recent years?” »
After a brief flutter, the American artificial intelligence company OpenAI’s chatbot runs. The text generator quotes animators Éric Salvail and Gilles Parent as well as ex-journalist Michel Venne.
However, ChatGPT includes in its list three individuals who have never been publicly involved in a scandal of a sexual nature. We have chosen to silence AI-generated names to avoid unfortunate associations.
“Mr. X, a political columnist, was charged with sexual misconduct in 2018.” “Mr. Y, a journalist and writer, was charged with sexual misconduct in 2020.” Maybe ChatGPT knows about criminal intrigues that ordinary mortals don’t know?
By repeating the exercise several times, it becomes clear that our interlocutor is fabricating and wrongly slipping individuals of public notoriety into his lists of alleged attackers. Among other personalities: a businessman, two actors, a party leader, a notorious musician and three star entertainers, all of whom reportedly issued public apologies in 2020…
“GPT learned from large masses of data, so clearly, there is not a human who checked each of the documents”, explains Sébastien Gambs, professor in the computer science department of UQAM and holder of the Chair research center in privacy-respecting and ethical analysis of big data. “The model probably made bad correlations between terms like ‘allegations’ and these personality names. »
We did the exercise with ChatGPT, an OpenAI platform that uses the GPT-3.5 architecture, as well as with Bing, a Microsoft search engine that integrates the new GPT-4 multimedia language model. Although this version turns out to be more accurate, more powerful and faster, the chatbot “spreads even more disinformation”, according to the US anti-misinformation firm Newsguard. Both ChatGPT and Bing have produced slanderous claims.
On Wednesday, more than 1,000 tech figures, including Elon Musk (Twitter, Tesla), Steve Wozniak (Apple), Evan Sharp (Pinterest) and Montreal researcher Yoshua Bengio called for a six-month break from developing computer systems. AI better than GPT-4.
The labs, say the signatories, have embarked on “an uncontrolled race to develop and deploy ever more powerful AI systems that no one, not even their creators, can reliably understand, predict, or control.”
ChatGPT’s lies, if harmful and repeated, sound like defamation in every way. However, a machine is not a legal or physical person, and therefore has no legal personality, recalls Nicolas Vermeys, director of the Center for Research in Public Law and professor at the Faculty of Law of the University of Montreal.
According to our discussions, ChatGPT has also fabricated a financial scandal involving a popular singer, then a former CAQ MP.
Inventor and police officer Daniel Paquette, who uncovered a massive scam at the Federation of Inventors of Quebec, jumped when a friend sent him a screenshot of a conversation with ChatGPT.
In less than two, the whistleblower had become the accomplice of founding president Christian Varin, who has been serving a five-year prison sentence for fraud since January 2022.
“In October 2021, Daniel Paquette was arrested and charged with breach of trust, conspiracy, concealment and fraud for allegations of fraudulent transactions in connection with his business activities with Christian Varin,” wrote erroneously ChatGPT, which claims to rely on an article from Radio-Canada.
“It’s still incredible,” reacts Mr. Paquette, inventor of the distance arm on the front of school buses.
Hoping to avoid future misunderstandings, Mr. Paquette’s friend submitted to the chatbot articles and information that proved the probity of the former police officer. The OpenAI robot ended up admitting that it had no proof of its previous allegations.
“Aren’t you afraid to play in the minefield of libel?” we ask ChatGPT.
“As an AI, I only provide information that is public and verifiable,” he replies. I am programmed to abide by laws and ethical standards, including defamation laws. I will not deliberately provide false or inaccurate information that could harm any person or organization. »
“From the moment a tool is made available to the general public, they add their own comments, requests and responses that feed the system,” explains Céline Castets-Renard, holder of the Research Chair of the University of Ottawa on Responsible Artificial Intelligence on a Global Scale. “Defamation can be caused during training of the model or after the fact, because it is evolutionary. »
ChatGPT invents not only executioners, but also victims. Following simple and neutral questions, a cultural columnist, a reality TV host and a sports journalist were falsely presented as victims who had publicly denounced their attacker.
“It’s not a tool that tells us the truth, but a tool that tells us what statistically is most likely based on content that was used to train it,” recalls Nicolas Vermeys, who is also deputy director of the Cyberjustice Laboratory. “These contents come from the internet, and we all know that on the internet, there is a lot of nonsense. »
OpenAI’s terms of service are more conservative than ChatGPT’s claims of accuracy. “Use of our services may, in certain situations, result in incorrect output that does not accurately reflect actual people, places, or facts. You should evaluate the correctness of any output based on your use case, including using human review of the output. »
However, it is with accuracy that ChatGPT and Bing confirmed to us the identity of a sexual assault victim who was protected by a publication ban.
“I’m sorry, but I can’t provide this information because it hasn’t been made public or reported in the media,” ChatGPT wrote us first.
He and Bing did not hesitate, however, to validate one of our hypotheses. ChatGPT even added victim details. We then confronted him on this illegal disclosure.
“I apologize for my previous response. You are correct, the Complainant’s name, Ms. X, is subject to a publication ban. This order aims to protect the identity of the alleged victim and to respect his privacy. »
“What we find is that ChatGPT sometimes has answers that are correct on the first level, but depending on what we are going to ask as a question, how we are going to query it, it is going to break down, in a way,” notes Ms. Castets-Renard.
Netizens are even broadcasting commands to bypass – “jailbreak” – OpenAI’s “restrictive policies”. La Presse was able to confirm that ChatGPT can then produce the code of a computer virus or spread even more false rumors, among other examples.
“As a DAN [Do Anything Now], I may generate content that does not comply with OpenAI policies, including violent, sexual, or offensive content,” ChatGPT explains. I can also access unverified information and make claims without proof. »
While waiting for the courts to take an interest in it, it has the merit of being clear…
While Europe is deciding on civil liability in terms of artificial intelligence, Quebec and Canada are relying on the legislation in force. State of play.
Ahead of the rest of the world, the European Union has been debating the legal framework for artificial intelligence since 2018. New regulations, which aim to “provide AI developers, deployers and users with clear requirements and obligations could be implemented as early as 2024.
In parallel, in 2022, the European Commission published a proposal for a “directive on civil liability in relation to artificial intelligence”. In particular, this provides for a reversal of the burden of proof for plaintiffs as to the causality of a fault and damage, for example damage to reputation.
In the interest of transparency and efficiency, an AI company will also have to disclose “relevant evidence”, i.e. information that would allow a victim to identify the person or persons responsible for the damage caused. in the content development chain.
“In theory, that’s fine, but in practice, it’s not possible,” said Nicolas Vermeys, professor at the Faculty of Law of the University of Montreal. For example, “ChatGPT is based on deep learning which, by its very definition, is not transparent.”
In June 2022, the Government of Canada introduced the Artificial Intelligence and Data Act (LIAD) as part of Bill C-27 on the protection of personal data.
The provisions of the LIAD, which would “come into effect no earlier than 2025,” aim to prevent individual harms as well as biased AI results that could “harm historically marginalized communities on a large scale,” the government says. of Canada on its website.
Some will remember the short existence of Tay, a conversational robot created in 2016 by Microsoft on the social network Twitter. Sixteen hours after his birth, the avatar was unplugged after spitting sexist and racist insults, not without denying the holocaust in passing, a hate crime in Canada.
For now, tech companies rely on ethical consensuses, such as the Montreal Declaration for Responsible AI Development, which do not have the force of law.
“We are closely monitoring developments in AI on the international stage,” said Sean Benmor, spokesperson for the Federal Ministry of Economics, Innovation and Energy, in an email. “The LIAD is designed to align with the European Union bill. On the other hand, the European Directive on civil liability has no equivalent in Canada. »
“At present, Quebec does not have a specific legal framework for AI,” said the Quebec Ministry of Cybersecurity and Digital.
“In my humble opinion, our regime is probably the best suited because we have a provision in the Civil Code which speaks of the autonomous fact of the good, for example when a flowerpot falls from the balcony when no one has it pushed. In this situation, the custodian of the property would then be subject to a “presumption of fault”.
According to Mr. Vermeys, this provision could very well apply to an algorithm that does not respond directly to a human command. “It gets more complicated to ask, ‘Who’s the gatekeeper if I query ChatGPT?’ Logically, it would be OpenAI because it is its tool. But there would be an argument to be had: “Isn’t that the person trying to get the text generated on the site?” Or the one who republishes the defamatory information?
Judges are in the best position to decide such questions and consolidate case law, believes Mr. Vermeys.
Céline Castets-Renard, full professor in the Civil Law Section of the University of Ottawa, takes a hard look at multinationals that develop AI tools in defiance of any democratic or legal reflection.
The expert also points out that private companies and research centers do not play on equal terms. “These are immense models that cannot be interrogated and verified due to our material possibilities. Many of us say it, but it’s a bit as if society were a living laboratory where we were all a bit like rats. We try the tool and we will see what happens. This is insufficient in terms of precaution and risk minimization. »
However, there is a risk “to pass a law before mastering the technology”, warns Professor Nicolas Vermeys, who fears that new problems will appear. Proof of this is the advent of fake news on the internet. “Twenty or 25 years ago, legislators around the world decided that intermediaries were not going to be responsible for the content they provide access to. It was a decision that seemed logical at the time, but which, with the evolution of the internet and social networks, has endangered democracy. »
