The European Union has enacted a significant provision of its Artificial Intelligence Act, targeting high-risk AI systems and introducing regulations designed to protect users from potential harm. This move comes as the United States continues to operate without comprehensive federal rules governing artificial intelligence.
The EU's AI Act, the first of its kind globally, has been gradually implemented since 2024. A key provision that took effect on August 2nd applies stringent rules to AI systems used in regulated products such as medical devices, machinery, and toys. The law categorizes AI into four risk levels: unacceptable, high, limited, and minimal, with a strong emphasis on safeguarding public health and well-being.
"The AI Act ensures that Europeans can trust what AI has to offer," the European Commission stated on its website. The legislation specifically prohibits AI systems that employ manipulation, deception, or exploitation to influence user behavior in ways that could cause significant harm. This includes AI that uses subliminal or deliberately deceptive techniques to undermine autonomy, as well as systems that exploit vulnerabilities related to age, disability, or socioeconomic status. The EU aims to prevent individuals from being reduced to "mere tools for achieving certain ends" and to offer greater protection to those most susceptible to manipulation.
Experts are raising concerns about the less obvious dangers posed by AI. Alexander Acosta, a physician with extensive clinical experience, noted that subliminal cues can trigger the body's stress response before conscious processing. He likened the hormonal responses observed in patients during medical emergencies to the physical harm that could result from AI continuously manipulating or overstimulating users. "I witnessed this same cortisol and adrenaline pathway firing in a patient’s body during a cardiac emergency in the ER years ago," Acosta stated.
Nathan Carroll, a board-certified psychiatrist, identified manipulation and deception as particularly concerning AI risks due to their widespread, insidious, and invisible nature. "AI encouraging self-harm is a transparent act," Carroll said. "But AI that subtly manipulates an individual either through biased data, an echo-chamber effect, or weakening of critical thinking can also cause extreme harm."
Carroll has observed unintended consequences, especially among young people who may develop an overreliance on AI's polished responses. This misplaced trust, he explained, has led to situations requiring threat assessments and psychiatric evaluations for students. "The use of AI to exploit vulnerabilities and vulnerable people is the risk that keeps me up at night," Carroll added. He also expressed concern that autonomous AI could be used on social media for harassment and bullying, particularly targeting children and teens. "This type of toxic exposure has the potential to amplify the most extreme psychological and physical risks we see with social media. Risks of depression, anxiety, body dysmorphia, addiction, self-harm, and suicide may all be amplified by AI exploitation of vulnerable people."
The cumulative effects of AI exposure are also a concern. Acosta warned that prolonged exposure to highly stimulating AI platforms could keep the nervous system in a chronic fight-or-flight state, leading to accumulated physical symptoms such as increased inflammation and a weakened immune system over time.
Carroll described the EU's AI Act as a "thoughtful attempt" to address emerging risks, but emphasized that regulation alone is insufficient. "We need to work twice as hard at teaching people about the risks," he said. "You don’t make swimming with sharks safer by just pleading with the sharks. You need to teach the swimmers how to stay safe too."
Acosta highlighted the EU's recognition of physical harm from AI but noted the absence of a clinical framework for doctors to identify AI exposure as a contributing factor to patient symptoms. He suggested that developing medical benchmarks could aid clinicians in earlier problem identification.
Carroll believes the U.S. would benefit from similar safeguards, particularly for children, and stronger accountability for AI developers. "The US is positioning itself as a leader with AI," he stated. "We also need to be a leader in AI safety, otherwise one of the costs for moving fast now may slow us down later as unintended harms emerge."