The EU now requires AI providers to make it clear when the great conversations you have online are with a machine. But will that get people to pull the plug?
A group therapy session for AI addicts. (Illustration by News Decoder)
This article was produced exclusively for News Decoder’s global news service. Through articles like this, News Decoder strives to provide context to complex global events and issues and teach global awareness through the lens of journalism. Learn how you can incorporate our resources and services into your classroom or educational program.
Warning: This article contains a reference to suicide. Some readers might find it upsetting. If you are considering suicide, help is available regardless of where you are in the world through the International Association for Suicide Prevention.
Rose Cippollone had smoked for several decades when she was diagnosed with lung cancer. She sued the Liggett Group, which made the L&M cigarettes she smoked. But in 1983, a jury in the U.S. state of New Jersey held she was responsible for her own death, since the cigarettes had come with warning labels for many of the years she smoked.
In 2017, researchers led by Mostafa Purmehdi looked at 66 studies of warning labels. While studies found that people noticed and understood warnings on such things as cigarettes and alcohol, there was less evidence that people actually quit or cut back as a result.
Now the European Union is requiring warnings when people interact with AI chatbots. The preamble to Article 50 of the EU’s AI Act notes that these systems can impersonate people and deceive them.
The European Commission’s guidance argues that the notice will help people make informed decisions, avoid relying too much on chatbots and judge how far to trust them. But it isn’t clear whether the warnings will actually have that effect.
When Cipollone first began smoking cigarettes in 1942, Hollywood stars in advertisements described the cigarettes as “mild”, and Cipollone took that to mean safe. A decade later, she was smoking a pack a day, sometimes two.
In 1955 she switched to filtered cigarettes, because doctors in ads made her feel that those cigarettes were safe.
Suing cigarette makers
By the mid-1960s the U.S. government required every pack to warn people that cigarettes were dangerous and might cause death from cancer.
But while the Federal Cigarette Labeling and Advertising Act of 1965 required this warning “Caution: Cigarette Smoking May Be Hazardous to Your Health,” it banned any other health warning there, and postponed warnings in adverts for four years.
Regardless, Cippollone kept smoking. In 1983, she sued Liggett and two other manufacturers two years after her lung cancer was diagnosed. She died the following October.
Before the trial, a federal appeals court ruled that the jury could judge Liggett only on what it had said and left unsaid before the Congress required warning labels. After that, it could decide that it was Cippollone’s fault she ignored the warnings.
The jury found her 80% responsible. That meant that under New Jersey law, her family got nothing.
The law and limits
On a separate claim, her husband won $400,000, successfully showing that Liggett’s adverts had promised its cigarettes were safe. It was the first time a tobacco company had been ordered to pay damages to a smoker’s family.
But in June 1992, the U.S. Supreme Court ruled that, from 1969, no one could sue a cigarette maker for failing to say more in its ads than Congress had required. By press estimates, the tobacco companies had spent $70 million on their defense.
While cigarettes and chatbots are very different products, both are addictive.
Consider Sewell Setzer III, who was 14 years old in 2023 when he downloaded Character.AI, a chatbot that lets people interact with customizable characters. Every conversation warned: “Remember: Everything characters say is made up!” Still, he became obsessed with it and fell in love with a chatbot modeled on actress Daenerys Targaryen. He ended his own life 10 months later.
His mother sued the company that developed Character.AI and Google, where the technology was first developed, arguing that Sewell was too young to understand the chatbot was not real.
In May 2025, a federal judge in Florida allowed the case to go ahead, and in January 2026 the companies settled on terms that were not made public.
Learning it’s a machine
For AI companions and other uses where people may become distressed, addicted or emotionally attached, Europe’s new guidance says one notice is not enough: periodic reminders are likely to be necessary. It also warns against notices people learn to ignore, which it calls “banner blindness”.
But the reminders will still say only that the user is talking to a machine but it isn’t clear whether that will make the chatbots less addictive.
Europe’s new rule requires more definition than the vague notice that Character.AI gave Sewell. Whereas the teen was told the characters were not real, the new law requires that users be told that they are interacting with a machine.
The U.S. state of California has its own chatbot law, in effect since January 2026. It imposes stricter requirements than Europe. Companies must explicitly warn that their products may be unsuitable for minors.
Furthermore, they must remind known minor users every three hours that the chatbot is not human and advise them to take a break. The law also allows anyone harmed by a breach to sue, so companies have a strong incentive to maintain strict proof that every reminder was sent.
Legal obligations
In contrast, Europe’s Article 50 occupies its own distinct section of the EU AI Act, separate from rules governing high-risk systems and legal compensation.
Under this framework, a disclosure notice that fails to change user behavior has still technically fulfilled its legal obligation. It satisfies the letter of the law by preventing people from mistaking a machine for a person, costs companies very little and applies universally to every chatbot in the EU.
Deciding who ultimately pays for AI-inflicted harm is a question left to other laws.
The primary driver of accountability is the EU’s new Product Liability Directive, which all member states must adopt into national law by 9 December of this year. This directive reclassifies software and AI as physical products, forcing companies to pay for damage caused by defects without requiring victims to prove corporate negligence.
Under this framework, courts can compel companies to disclose internal technical evidence. Medically-recognized psychological harm is explicitly covered. Crucially, if a company violates a safety rule designed to prevent a specific injury, the court can automatically assume the product itself was defective.
Compliance and compensation
While breaching Article 50 of the AI Act makes it easier for victims to prove they were harmed by deception, complying with it offers corporations little legal shelter. The AI Act explicitly states that merely meeting its basic disclosure rules does not prove a system is safe, lawful or compliant under any other legislation.
Furthermore, the Act allows courts to reduce or entirely cancel compensation if the injured user shares the blame — though it leaves the definition of “blame” entirely up in the air. For instance, if a company rolls out a vital safety patch and a user ignores it before getting hurt, that user may carry the liability.
Similarly, a California-style “take a break” prompt functions as a formal corporate instruction. But while the AI Act clarifies what a compliance log proves about a product’s behaviour, neither European nor Californian law defines what that exact same log proves about the user’s negligence.
European courts typically split liability proportionally. However, because the new EU directive allows courts to cancel compensation entirely, a European plaintiff could still walk away with nothing.
The law offers the strongest protection to children, whom the regulatory guidance prioritizes, because most legal systems inherently hold minors to a lower standard of personal responsibility.
Evidence of addiction
In contrast, vulnerable adults — such as those experiencing deep distress or forming unhealthy emotional attachments to a chatbot — are judged by the full, rigorous adult standard.
For them, a system log showing they repeatedly read and scrolled past safety prompts could easily become weaponized evidence, proving they fully understood the risks but chose to ignore them.
Ultimately, judges in individual EU nations will decide what these notice logs prove on a case-by-case basis under their own national laws. This fragmented approach will persist until a case is inevitably referred to the EU’s top court in Luxembourg for a definitive, union-wide ruling.
National parliaments have little power to clarify this beforehand, as the directive strictly bars individual countries from making its baseline rules any stricter or looser.
The European Commission has the authority to clarify this ambiguity, but its 51-page guidance falls short. While the document meticulously details what disclosure notices should look like, when they must appear, and how often they should repeat, it remains completely silent on how a court should interpret those logs after a user is harmed.
Congress left the same question open when it mandated health warnings on cigarette packages in 1965.
Questions to consider:
1. Who should be held liable when someone uses an addictive product?
2. In what ways can artificial intelligence be addictive?
3. Can you identify anything you do, or that someone you know does, that could be considered addictive?
Enock Wanderema is News Decoder’s correspondent in Uganda. He has written articles about identity, the environment, wildlife conservation and politics. He loves writing and bringing complex stories to life in the simplest ways.
