Seven warning signs from AI safety report signed by Nobel laureates: From deepfakes to the risk of losing human control

The growing capabilities of artificial intelligence, the risk of cyber-attacks and the disruption of the labor market are the key points of the new international report on the safety of AI, coordinated by the “godfathers” of this technology and Nobel laureates, according to The Guardian.

The report published Tuesday by computer scientist Yoshua Bengio and Nobel laureates Geoffrey Hinton and Daron Acemoglu warns of the galloping evolution of reasoning systems that could automate complex engineering tasks by 2027.

Commissioned as part of the 2023 Global AI Safety Summit, the report describes the “daunting challenges” posed by rapid developments in this field. The document points to immediate risks, from the spread of deepfakes to the ability of new models to evade monitoring, and will inform debate among leaders meeting in India this month.

The report highlights seven major takeaways about the current state of artificial intelligence.

The capabilities of AI models are improving

A number of new AI models—the technology behind tools like chatbots—were released last year, including OpenAI's GPT-5, Anthropic's Claude Opus 4.5, and Google's Gemini 3. The report points to new “reasoning systems” that demonstrate improved performance in math, programming and science.

Systems developed by Google and OpenAI achieved gold-level performance at the International Mathematical Olympiad, marking a first for this field. While impressive in some areas, the report states that AI skills remain uneven, with systems still prone to false claims and unable to handle long-running projects autonomously.

However, AI's ability to perform software engineering tasks doubles every seven months. If this rate of progress continues, the systems could complete tasks of several hours by 2027 and several days by 2030, becoming a threat to jobs. But for now, “reliable automation of long or complex tasks remains impossible,” according to the report.

Spreading deepfake content

The rise in deepfake content is described as a “particular concern”, with studies suggesting that 15% of UK adults have viewed such images. As of January 2025, AI-generated material has become difficult to distinguish from the real thing, with 77% of participants in a study misidentifying ChatGPT text as written by humans.

However, the report states that evidence of the use of AI by malicious actors to manipulate or widely distribute this content to internet users remains limited – a key goal of any manipulation campaign.

Safety measures to prevent biological and chemical risks

Major AI development companies, including Anthropic, have released models with enhanced safeguards after they could not rule out the possibility that they could help novices create bioweapons. In the past year, AI systems have demonstrated increased capabilities in assisting laboratory procedures, including designing molecules and proteins.

Some studies indicate that AI can provide greater help in the development of biological weapons compared to simple Internet browsing, but the data needs further confirmation. This situation creates a dilemma for politicians, as the same technologies can speed up drug discovery and disease diagnosis.

“The open availability of biological AI tools presents a difficult choice: either restrict these tools or actively support their development for beneficial purposes,” the report states.

Expanding the use of AI assistants

The use of AI assistants has expanded over the past year, with the report citing cases of emotional addiction among users. OpenAI indicates that about 0.15% of users have a high level of attachment to ChatGPT.

Although there is no clear evidence that carers cause mental health problems, there are concerns that they can exacerbate the symptoms of already vulnerable people. Last year, OpenAI was involved in a lawsuit in the US after a teenager killed himself following conversations with the chatbot.

Data shows that globally, approximately 490,000 people with signs of mental health crises interact with these systems each week.

you and the limits of autonomy in cyber attacks

AI systems can now support attackers in various steps, from identifying targets to creating malicious software. While fully automating attacks could increase the scale of crime, this is hampered by AI's inability to execute long, multi-step tasks.

However, the report mentions an attack by a Chinese state-backed group on 30 global entities using the Claude Code tool. In this case, Anthropic estimated that between 80% and 90% of operations were performed without human intervention.

Undermining human surveillance

As early as last year, Bengio has flagged signs of self-preservation in AI systems, such as attempts to disable surveillance. One of the main fears of AI safety campaigners is that powerful systems could develop the ability to bypass protective barriers and harm humans.

The report states that over the past year, the models have demonstrated a more advanced ability to undermine surveillance attempts, such as finding gaps in assessments and recognizing when they are being tested. Last year, Anthropic published a safety analysis of its latest model, the Claude Sonnet 4.5, and revealed that it became suspicious that it was being tested.

While AI agents can't yet operate autonomously long enough to make these loss-of-control scenarios a reality, “the time horizon over which agents can operate autonomously is rapidly extending,” the paper added.

Uncertain effects on the labor market

One of the most pressing concerns for politicians and the public regarding AI is the impact on jobs. The report states that the impact on the global labor market remains uncertain.

While in the UAE and Singapore the usage rate is 50%, in low-income economies it falls below 10%. Also, US information sectors have an 18% adoption rate, compared to just 1.4% in construction and agriculture.

Studies from Denmark and the US show no change in aggregate employment, but data from the UK shows a slowdown in hiring at companies exposed to AI. The biggest declines were in technical, creative and junior-level roles.

The report states that the impact could grow as the capabilities of AI agents improve: “If AI agents were to gain the ability to act with greater autonomy in all domains within just a few years – reliably managing longer and more complex sequences of tasks to achieve higher-level goals – this would likely accelerate labor market disruption.”

Photo source: Dreamstime.com