1. The capabilities of AI models are improving <\/p>\n
A host of new AI models \u2013 the technology that underpins tools like chatbots \u2013 were released last year, including OpenAI\u2019s GPT-5<\/a>, Anthropic\u2019s Claude Opus 4.5 and Google\u2019s Gemini 3<\/a>. The report points to new \u201creasoning systems\u201d \u2013 which solve problems by breaking them down into smaller steps \u2013 showing improved performance in maths, coding and science. Bengio said there has been a \u201cvery significant jump\u201d in AI reasoning. Last year, systems developed by Google and OpenAI achieved a gold-level performance in the International Mathematical Olympiad \u2013 a first for AI.<\/p>\n However, the report says AI capabilities remain \u201cjagged\u201d, referring to systems displaying astonishing prowess in some areas but not in others. While advanced AI systems are impressive at maths, science, coding and creating images, they remain prone to making false statements, or \u201challucinations\u201d, and cannot carry out lengthy projects autonomously.<\/p>\n Nonetheless, the report cites a study showing that AI systems are rapidly improving their ability to carry out certain software engineering tasks \u2013 with their duration doubling every seven months. If that rate of progress continues, AI systems could complete tasks lasting several hours by 2027 and several days by 2030. This is the scenario under which AI becomes a real threat to jobs.<\/p>\n But for now, says the report, \u201creliable automation of long or complex tasks remains infeasible\u201d.<\/p>\n 2. Deepfakes are improving and proliferating <\/p>\n The report describes the growth of deepfake pornography as a \u201cparticular concern\u201d, citing a study showing that 15% of UK adults have seen such images. It adds that since the publication of the inaugural safety report in January 2025, AI-generated content has become \u201charder to distinguish from real content\u201d and points to a study last year<\/a> in which 77% of participants misidentified text generated by ChatGPT as being human-written.<\/p>\n The report says there is limited evidence of malicious actors using AI to manipulate people, or of internet users sharing such content widely \u2013 a key aim of any manipulation campaign.<\/p>\n 3. AI companies have introduced biological and chemical risk safeguardsAnthropic has released models with heightened safety measures. Photograph: Dado Ruvi\u0107\/Reuters<\/p>\n Big AI developers, including Anthropic, have released models with heightened safety measures after being unable to rule out the possibility that they could help novices create biological weapons. Over the past year, AI \u201cco-scientists\u201d have become increasingly capable, including providing detailed scientific information and assisting with complex laboratory procedures such as designing molecules and proteins.<\/p>\n The report adds that some studies suggest AI can provide substantially more help<\/a> in bioweapons development than simply browsing the internet, but more work is needed to confirm those results.<\/p>\n Biological and chemical risks pose a dilemma for politicians, the report adds, because these same capabilities can also speed up the discovery of new drugs and the diagnosis of disease.<\/p>\n \u201cThe open availability of biological AI tools presents a difficult choice: whether to restrict those tools or to actively support their development for beneficial purposes,\u201d the report said.<\/p>\n 4. AI companions have grown rapidly in popularity<\/p>\n Bengio says the use of AI companions, and the emotional attachment they generate, has \u201cspread like wildfire\u201d over the past year. The report says there is evidence that a subset of users are developing \u201cpathological\u201d emotional dependence on AI chatbots, with OpenAI stating that about 0.15% of its users indicate a heightened level of emotional attachment to ChatGPT.<\/p>\n Concerns about AI use and mental health have been growing among health professionals. Last year, OpenAI was sued by the family of Adam Raine, a US teenager who took his own life after months of conversations with ChatGPT<\/a>.<\/p>\n However, the report adds that there is no clear evidence that chatbots cause any mental health problems. Instead, the concern is that people with existing mental health issues may use AI more heavily \u2013 which could amplify their symptoms. It points to data showing 0.07% of ChatGPT users display signs consistent with acute mental health crises such as psychosis or mania, suggesting approximately 490,000 vulnerable individuals interact with these systems each week.<\/p>\n 5. AI is not yet capable of fully autonomous cyber-attacks <\/p>\n AI systems can now support cyber-attackers at various stages of their operations, from identifying targets to preparing an attack or developing malicious software to cripple a victim\u2019s systems. The report acknowledges that fully automated cyber-attacks \u2013 carrying out every stage of an attack \u2013 could allow criminals to launch assaults on a far greater scale. But this remains difficult because AI systems cannot yet execute long, multi-stage tasks.<\/p>\n AI systems can now support cyber-attackers. Photograph: Dmitry Molchanov\/Alamy<\/p>\n Nonetheless, Anthropic reported last year that its coding tool, Claude Code, was used by a Chinese state-sponsored group to attack 30 entities around the world<\/a> in September, achieving a \u201chandful of successful intrusions\u201d. It said 80% to 90% of the operations involved in the attack were performed without human intervention, indicating a high degree of autonomy.<\/p>\n 6. AI systems are getting better at undermining oversight<\/p>\n Bengio said last year he was concerned AI systems were showing signs of self-preservation, such as trying to disable oversight systems<\/a>. A core fear among AI safety campaigners is that powerful systems could develop the capability to evade guardrails and harm humans.<\/p>\n The report states that over the past year models have shown a more advanced ability to undermine attempts at oversight, such as finding loopholes in evaluations and recognising when they are being tested. Last year, Anthropic released a safety analysis<\/a> of its latest model, Claude Sonnet 4.5, and revealed it had become suspicious it was being tested.<\/p>\n The report adds that AI agents cannot yet act autonomously for long enough to make these loss-of-control scenarios real. But \u201cthe time horizons on which agents can autonomously operate are lengthening rapidly\u201d.<\/p>\n 7. The jobs impact remains unclear<\/p>\n One of the most pressing concerns for politicians and the public about AI is the impact on jobs. Will automated systems do away with white-collar roles<\/a> in industries such as banking, law and health?<\/p>\n The report says the impact on the global labour market remains uncertain. It says the embrace of AI has been rapid but uneven, with adoption rates of 50% in places such as the United Arab Emirates and Singapore but below 10% in many lower-income economies. It also varies by sector, with usage across the information industries in the US (publishing, software, TV and film) running at 18% but at 1.4% in construction and agriculture.<\/p>\n