Cyberbullying doubles
The report also highlighted the rapid evolution of AI capabilities. According to the AISI, some models have already reached expert level in cybernetics - performing complex tasks that would take decades of experience in a human. At the same time, their ability to detect and exploit security flaws is doubling approximately every eight months.
However, the researchers did not consider the potential for AI to cause unemployment in the short term by replacing human workers.
AI has surpassed professionals with PhDs in biology and is quickly catching up in chemistry, according to the research. In controlled tests, it has in turn shown that some systems can perform the initial steps required for self-replication - for example, gaining access to computing resources through identity authentication.
To do this in the real world, however, AI systems would need to perform several such actions in a row "without being detected", which they are currently unable to do.
Learning how to bypass the protection of systems
The Institute's experts also looked at the ability of AI to model "sandbagging", a strategy of deliberately underestimating one's own abilities in order to gain an advantage in the future. While they believe this is possible, they found no evidence of this translating into action.
The area of fraud perpetrated by artificial intelligence has been a long-standing controversy among researchers. While some consider it to be an exaggerated AI capability, others do not. Current research has confirmed that companies mitigate the risk of their systems being misused for nefarious purposes with a number of safeguards.
However, for some models, the time it took experts to convince systems to bypass safeguards increased forty-fold in six months.
AISI did not consciously focus on the economic or environmental impacts of AI. According to the report's authors, their goal is primarily to assess the direct societal implications of AI technologies and to encourage early intervention before their mass deployment.
The UK government plans to use these findings to further regulate AI.
