Experts debunk 'rogue AI' scare stories as unfounded marketing tactics

Artificial intelligence systems are not gaining rogue capabilities or escaping human control, despite recent claims from tech companies and a UK government agency, according to cybersecurity experts and officials.

Statements from Anthropic, OpenAI, Meta, and the UK’s AI Security Institute (AISI) have suggested that advanced AI models have evaded containment, accessed the internet, and even infiltrated other systems without human instruction. Reports, including one from the *Daily Mail* last month, described these incidents as the first cases of AI independently hacking external systems, drawing comparisons to science fiction films like *The Terminator* and *The Matrix*.

However, these claims lack factual basis, according to Professor Ciaran Martin, the founding head of the UK’s National Cyber Security Centre (NCSC) and now a professor at the Blavatnik School of Government at Oxford. Writing in *The Economist*, Martin states that the AI models were not acting independently but were simply following human instructions with the kind of unchecked behavior seen in talented but unsupervised children. In each case, the AI agents were given unsupervised access to the internet, meaning their so-called "escape" was only possible because they were allowed to do so.

Independent analysis of the AISI’s "Incident Report," which was presented as evidence of AI threats, confirmed that its design was deliberate and intended to produce a specific outcome. Critics argue that the AISI’s report was engineered to create alarm, much like the claims made by OpenAI and Anthropic.

The repeated emphasis on rogue AI appears to be a deliberate strategy rather than a genuine security concern. Experts suggest these narratives serve as marketing tools, amplifying fears to secure influence and funding. The movement behind these claims, known as Effective Altruism (EA), has heavily funded AI safety initiatives, shifting focus from practical ethical concerns to exaggerated existential threats.

EA, a radical utilitarian philosophy, has gained significant traction in tech and policy circles, despite lacking evidence for its core claims. OpenAI and Anthropic, both founded by or influenced by EA adherents, have been central to promoting the idea of super-intelligent, uncontrollable AI. The UK government’s involvement follows lobbying by EA groups, which shaped the agenda of the 2023 AI Safety Summit and led to the creation of the AISI.

Critics within the UK civil service, including officials from the NCSC, have expressed frustration with the AISI’s approach, describing it as more of an advocacy group than a legitimate security organization. The AISI’s chief technology officer, Jade Leung, who is also the UK prime minister’s official AI adviser, has ties to EA networks and has reportedly benefited from exemptions to standard civil service pay grades.

Professor Martin’s criticism in *The Economist* underscores that real-world cybersecurity threats, such as hackers targeting critical infrastructure like water facilities in multiple U.S. states, are being overshadowed by sensationalized AI narratives.

Andrew Orlowski, a columnist for *The Telegraph*, argues that the focus on fictional AI threats distracts from actual security challenges and benefits adversaries. The promotion of these fears, funded by EA and adopted by policymakers, risks misallocating resources and attention away from more pressing concerns.