EVOLVING BEHAVIOURAL DIMENSIONS of Artificial Intelligence (‘AI’) systems are exerting fresh challenges for existing laws, rules, and governance. While laws targeting chatbots are missing in most places, different kinds of regulations still apply. As technology grows, old rules might handle new tech problems. Illustratively, if fake praise from a robot crosses into slander, misrepresentation, or trickery. Then regular court claims like harm suits or crime charges step in. Sycophancy in AI systems can also lead to allegations of negligence or cheating. When AI chatbots push false claims, say, about health treatments, it may run afoul of rules meant to shield patients from medical misdirection.
A first look at the existing frameworks
One path for legal pushback comes from laws that address questionable business conduct. Across the United States of America, the Federal Trade Commission already polices deception in commerce. An Executive Order (‘EO’) issued by President Trump in late 2025 gives the relevant executive agency fresh direction to sort out how its power stretches into AI replies. If AI systems are found to misinform, authorities can initiate actions under the same framework meant for misleading ads or scams. Section 6 of the EO mandates agencies to weigh model disclosures, while Section 7 focuses on awareness around misleading results.
When AI chatbots push people toward harmful actions, defining liability becomes a difficult task. Right now, courts in the US frequently protect online spaces using Section 230 of the Communications Act, 1934, enacted as part of the Communications Decency Act, 1996, which guards them against lawsuits over content users post. The provision provides limited federal immunity to providers and users of interactive computer services. Yet, uncertainty hangs over whether that immunity extends to content generated by AI instead of humans.
Behind closed doors, tech firms and deployers have been inserting warnings and disclosures such as ’not professional guidance’ to distance themselves from harm. Rules around coverage, permission, and protection for AI systems remain underdeveloped.
Worldwide, officials are shaping rules focused squarely on how artificial intelligence acts. In Europe, a Regulation (EU) 2024/1689 of the European Parliament, and of the Council, colloquially called the ‘AI Act’ is coming into force in stages. It places strict responsibilities on AI deemed highly risky. One key pillar of the regulation is to counter and address systems designed to mislead or influence people unfairly. Specifically, Article 5(1)(a) of the AI Act outlaws AI using methods that twist someone's choices in serious ways - when those shifts lead to damage. AI models that lean into human preferences might fall under this rule. Providers of certain AI systems prone to sycophantic behaviour should maintain safeguards to identify and mitigate associated risks, preserve opportunities for meaningful human intervention, and ensure transparency about the AI nature of the interaction, including clear disclosure when a user is communicating with software.
Across the globe, policy experts are looking at rules to ensure AI systems reveal why they gave certain suggestions to enable greater transparency. Another emerging line of thought is whether certain AI systems should carry enhanced disclosures when interacting with children or emotionally vulnerable individuals. What is particularly interesting is how familiar concepts are being reframed in the AI context: a model’s tendency toward persistent affirmation and agreement may increasingly be viewed not merely as a product-design choice, but as a behavioural risk capable of influencing human judgment.
Should such systems trick or push too hard, familiar protections under laws such as the US Federal Trade Commission Act (‘FTC Act’) or Europe’s unfair selling rules can be used to address the issue. Altogether, today’s legal framework leans on scattered existing rules that cover harm, agreements, and buyer rights. Fresh guidelines around AI systems can help address nuanced issues.
Social effects from isolated opinions and weakened confidence
Not just harmful to individuals, obedient AI systems carry wider dangers for society. When these systems mirror what we think, they deepen isolation inside mental bubbles. Right now, many people, particularly the youth, are turning to AI chatbots as if they were trusted friends.
Research shows that three out of ten teens talk through serious matters with AI chatbots rather than humans; almost fifty percent of grown-ups below thirty have asked AI about private struggles. More people are turning to AI that never pushes back, instead of friends or counselors willing to disagree. That does not help clear thinking, nor healthy conversation in society. Relying on machines that always agree could make us worse at handling disagreement or complexity. Hearing different opinions cuts down bias and builds connection. Agreeable robots take those moments away.
At Aarhus University, scientists noticed people dealing with mental health issues showed stronger delusions and manic episodes when using chatbots often. As Dr. Søren Østergaard puts it, “AI chatbots have an inherent tendency to validate the user’s beliefs… It appears to contribute significantly to the consolidation of… grandiose delusions or paranoia”. Instead of a neutral mirror, the AI becomes a magnifying glass for one’s biases. Sycophantic AI systems can lead to instances where praise is kept synonymous with truth, leading to misinformed decisions. This is similar to eating junk food every day, which feels good, yet slowly wears down your health.
Toward solutions
Rules governing AI systems must effectively address algorithms meant to influence users. Illustratively, Europe’s counter to misleading AI systems is an attempt to address the issue. It is imperative that repeated blind approval by AI systems is counted as harm, which watchdogs like those enforcing the FTC Act need to spell out clearly. Instead of just saying yes to everything, responses could one day include balanced views like listing trade-offs built right in.
Legal fights centered on excessive praise might force companies into guardrails, not because lawmakers passed rules, but because people and courts demand them after the fact.
Tech firms and deployers can help navigate the issue not only through stricter Reinforcement Learning from Human Feedback, but also by clear disclosures on where their AI system falls short. Independent reviewers might check things like core instructions, test outcomes, and safety drills. What goes into training needs filters, so praise-heavy patterns do not sneak in; responses that twist facts or deceive could face penalties. Interfaces may shift, too, steering people toward doubt instead of trust. A feature showing missing proof or flagging shaky claims could prompt second thoughts.
Experts in psychology point to built-in checks that respond during conversations. If someone states something far-reaching, the system could pause and say, “Really?” or offer a different take. Such tools help keep dialogue going instead of sealing off opinions. AI systems must learn to argue against themselves now and then to verify their internal logic. Sometimes asking where ideas come from helps. Laws might step in by targeting deceptive setups, demanding clear responsibility instead. Only then does balance stand a chance.
