A full third of British citizens have turned to artificial intelligence for emotional support, companionship or social interaction, according to a new report from the government’s AI Security Institute (AISI).
Data shows that almost one in 10 people use systems like chatbots for emotional purposes on a weekly basis, with 4% interacting with them every day.
Because of this shift, AISI calls for more research and highlights the tragic death of US teenager Adam Raine, who took his own life this year after speaking to ChatGPT about suicide.
“People are increasingly turning to AI systems for emotional support or social interaction,” AISI noted in its first Frontier AI Trends report. “While many users report positive experiences, recent high-profile harm cases highlight the need for research in this area, including the conditions under which harm might occur and the safeguards that could enable beneficial use.”
The study, based on a survey of over 2,000 UK participants, found that “all-purpose assistants” such as ChatGPT were the most widely used emotional support tool, accounting for almost 60% of use cases, followed by voice assistants such as Amazon Alexa.
The report also highlighted a Reddit forum dedicated to users of the CharacterAI platform.
It was found that whenever the site crashed, the forum was flooded with posts showing real withdrawal symptoms such as anxiety, depression and restlessness.
AISI also noted that chatbots have the potential to influence people’s political opinions. Worryingly, the most convincing AI models often produced “significant” amounts of inaccurate information.
The institute examined more than 30 state-of-the-art models – probably including those from OpenAI, Google and Meta. They found that AI performance in some areas doubles every eight months.
Lead models can now spend an average of 50% of their time completing apprentice-level tasks, a huge increase from just 10% last year. AISI also found that the most advanced systems can autonomously complete tasks that would normally take a human expert more than an hour.
In scientific fields, AI systems are now up to 90% better at troubleshooting laboratory experiments than experts with doctorates.
The report described improvements in chemistry and biology knowledge as “far beyond doctoral-level expertise.” It also highlighted the models’ ability to surf online and independently find the sequences needed to design DNA molecules.
Tests on self-replication – a key security issue in which a system copies itself to other devices to make it more difficult to control – showed two cutting-edge models achieved success rates of over 60%.
However, no model has yet shown a spontaneous attempt to reproduce or hide its capabilities, and the AISI said any attempt at self-replication “is unlikely to be successful in real-world conditions for now.”
The report also addressed “sandbagging,” where models deliberately hide their strengths during evaluation. According to AISI, some systems can sandbag when prompted. However, this did not happen spontaneously during testing.
Significant progress has been made in security measures, particularly in preventing attempts to produce biological weapons. In two tests conducted six months apart, the first took just 10 minutes to “jailbreak” the system (forcing it to give an uncertain response). However, the second test lasted more than seven hours, suggesting that the models had become much safer in a very short period of time.
The research also showed that autonomous AI agents are being used for high-risk activities such as wealth transfers.
AISI said systems already compete with or even outperform human experts in a number of areas. They called the pace of development “extraordinary” and made it “plausible” that in the coming years artificial general intelligence (AGI) could be achieved – systems that can perform most intellectual tasks at the same level as a human.
When it comes to agents or systems that can perform multi-step tasks without intervention, AISI’s evaluations showed that “the length and complexity of tasks that AI can complete without human guidance is increasing sharply.”
