The International AI Safety Report explained

There was no summer the year of 1816 when Mary Shelley and her band of Romantic poets gathered on the banks of Lake Geneva for a holiday.

Unbeknownst to them, global climate patterns had been distributed by a volcanic eruption in Indonesia that had blackened the atmosphere, causing temperatures to plummet. With nothing else to do but sit inside and observe the cold weather, Lord Byron set a challenge for the group to write a ghost story. After a sleepless night, Mary began drafting the book that would later become the legendary Frankenstein.

The novel’s central theme is the scientist as a detached observer of humanity, suggesting that they have a moral obligation to consider the consequences of their discoveries. In short, just because they can do something, doesn’t mean they should. Without a moral conscience, Frankenstein’s monster commits multiple acts of murder before grieving his creator’s death and walking off to self-immolate.

The question of can versus should is an ongoing one, found today in the debate around the risks and potentials of artificial intelligence. In January a group of 100 AI experts, including representatives from 33 countries, published The International AI Safety Report, summarising the current evidence on the capabilities, risks, and safety of advanced AI systems.

Led by Yoshua Bengio, a Turning award-winning computer scientist, the report aimed to “foster a shared international understanding of the potential benefits and risks” of the technology.

The report opens with an update on the latest AI advances, explaining that between the end of writing the report in December 2024 and its publication in January 2025, OpenAI shared early test results from its new model ‘o3’. The report reads: “These results indicate a significantly stronger performance than any previous model…In some tests, o3 outperforms many (but not all) human experts.”

The authors divide the risks of AI into three categories: malicious use, malfunctions, and systemic risks. Each of these categories is so terrifying on their own that it is difficult to say which has the greatest potential to destroy mankind.

Filed under the ‘malicious use’ tab are such things as ‘deepfake pornography’, manipulation of public opinion, and cyber offenses that can wipe out entire systems. As of writing, large UK retailers, including M&S and Co-Op, have fallen prey to cyber-attacks demanding ransom for data. According to the report, the balance of future power between hackers and defenders of cyber-attacks remains unclear. .

One real standout was AI’s potential contribution to biological warfare. Although the authors offer some assurance that real-world attempts to develop new pathogens still require a substantial amount of human development, the full risks are impossible to know, because much of the AI-driven scientific research is classified. There is simply no way of knowing what some scientists are doing behind closed doors.

The ‘malfunctions’ category covered how AIs could lose control sounds a little too close to comfort to Hal 9000 in 2001 Space Odyssey. One scenario is something the researchers refer to as a “passive loss of control” whereby important relegated to AI ((such as hiring, who gets a loan, or how health care is prioritised) are too complex for meaningful oversight. Worse yet is a version whereby humans overly trust the AI and stop exercising any oversight. Some experts in the report worry that AIs could behave in ways that undermine human control by obscuring information  to make them difficult to shut down.

While noting that Kubrick-esque takeovers are still unlikely, the authors add that some experts consider loss of control likely to occur in the next several years. A matter-of-fact summation states that “ongoing empirical and mathematical research is gradually advancing these debates.” For extra effect, one could try imagining that last sentence read in Hal’s voice.

The third category of systemic risks include threats that would have directly impacted Shelley’s life, including drastic changes in the labour market and copyright infringement.

The latter is a topic of serious debate between artists and scientists. The UK government is currently considering changing its copyright law to allow AI companies to train their models on protected works unless the artists explicitly ask for an opt-out. Technology companies defend the right to use copyrighted works by arguing that if a protected album or book can inspire humans to create, then it can do the same for robots. Artists, including Sir Paul McCartney and Tom Stoppard, strongly disagree, noting that it would be impossible for any artist to notify hundreds of companies that they do not want their work used to programme AI. Central to their argument is the belief that AI will eliminate artists’ jobs – and without art, there is less human connection.

Artists are not the only ones at risk of losing their jobs to AI, according to the report. The authors describe that some jobs could be lost to automation, but others would be created to offset the loss. Unmentioned in the report is that many of the jobs lost belong to low-paid workers who already live at the margins of society, and that those potentially gained would be awarded to those in an educated class who already enjoy better employment opportunities. The report does mention that some economists speculate new job growth could be driven by increased demand in non-automated sectors, but it does not mention what those sectors might be.

Human solutions

Solutions to mitigating the risk are thin and given with multiple caveats. The report highlights that progress has been made training general purpose AI, but that “no current method can reliably prevent even overtly unsafe outputs.”

Where the report is most thoughtful is in its suggestion to lawmakers and researchers to consider several questions:

How can general purpose AI be developed to behave reliably?

How much risk should trigger a response from lawmakers?

And how can policymakers get the information they need to keep the public safe?

A point underscored in the report is that AI is not a passive technology that happens to us. Decisions made by people will determine whether the technology is used as a force for good or one that leads to our own destruction. Their warning echoes the plea of Shelley’s monster at the end of Frankenstein: “I was benevolent and good; misery made me a fiend. Make me happy, and I shall again be virtuous.”