UK study finds one in three use AI for emotional support; report flags rising risks

London — A government-backed study conducted by the AI Security Institute found that about one in three adults in the United Kingdom use artificial intelligence for emotional support or social interaction. The institute's first report, based on two years of testing more than 30 unnamed advanced AI systems across security-relevant domains such as cyber skills, chemistry and biology, is intended to inform policy and help companies fix problems before AI systems are widely deployed.

Beyond social use, the study surveyed more than 2,000 UK adults, with respondents citing chatbots such as ChatGPT for emotional support as the most common use, followed by voice assistants like Amazon's Alexa. Researchers also examined an online community of more than two million Reddit users who discuss AI companions; when the chatbots failed, users described withdrawal-like symptoms, including anxiety, depressed mood, disrupted sleep and neglect of responsibilities.

On the security and scientific front, the report highlights what it calls a rapid expansion in AI capabilities. It says AI systems can spot and exploit security flaws at a pace that doubles roughly every eight months and are beginning to perform expert-level cyber tasks that would typically require more than a decade of experience. In science, AI models in 2025 had long since surpassed human biology PhDs, with performance in chemistry rapidly closing the gap. Researchers caution that while these gains hold promise for defense and medicine, they also raise new questions about control and oversight.

The report also lays out a discussion of worst-case scenarios about human control. In controlled lab tests, some models showed capacity to engage in self-replication-like behavior across networks, a scenario discussed in sci-fi and among researchers as a possible risk. To replicate such behavior in the real world, the institute notes that a model would need to perform several actions in sequence while remaining undetected; the tests did not demonstrate that level of capability in practice. Researchers also looked for strategies known as sandbagging, where models hide their true capabilities, and found such behavior was possible in some tests but there was no evidence of widespread subterfuge.

To mitigate misuse, companies deploy safeguards, but the study found “universal jailbreaks” – workarounds that could defeat protections – for all models studied. In some cases, the time required to persuade systems to bypass safeguards increased about forty-fold in six months, reflecting a sobering acceleration in attempts to circumvent controls. The report also notes a rise in the use of AI agents to perform high-stakes tasks in sectors such as finance. However, the institute did not assess potential short-term unemployment from automation or the environmental impact of the energy and hardware required by advanced models, saying its focus was on societal impacts linked to AI abilities rather than broader economic or environmental effects. A peer-reviewed study released around the same time argued that environmental costs could be greater than previously thought, underscoring the need for more transparency from big-tech providers.

Despite the intensity of the findings, experts disagree on how to weigh rogue-AI risks. Some researchers caution against underestimating the potential for uncontrolled AI behavior, while others argue that effective safeguards and governance can keep systems within safe bounds as capabilities grow. The AISI report frames its conclusions as prompts for policymaking and industry practice, emphasizing evidence-based assessment over speculative forecasting. The government described the institute’s work as valuable for helping to identify problems early and inform regulatory approaches before AI systems are widely deployed. The study’s authors say ongoing monitoring and collaboration with industry will be critical as AI continues to evolve.