Frontier AI is improving at speed, but the evidence on real-world risk still lags behind the hype

This is a synthesis report on frontier general-purpose AI, not a single lab study with one neat table at the end. The authors narrowed the scope to “emerging risks” from the most capable systems, then pulled together scientific, technical, and socioeconomic evidence published before December 2025. They also added new 2030 scenarios and forecasts from the OECD and the Forecasting Research Institute. That gives the document breadth, but it also means every claim sits on the quality of the underlying literature.

The review process was heavy on expert scrutiny. More than 100 independent experts helped shape the report, and draft chapters then went through external review by specialists, an advisory panel nominated by over 30 countries and organisations, senior advisers, industry, and civil society. In plain English, this was built to be argued over before publication, which is sensible when the subject is both politically loaded and changing by the month.

The capability story is brisk. From 2025 to 2026, leading systems reportedly reached gold-medal level on International Mathematical Olympiad problems, scored above 90% on MMLU undergraduate exams, and above 80% on GPQA, a graduate-level science test. In coding and agentic work, the report says AI can now reliably complete some software tasks that take humans about 30 minutes, while the task duration it can manage at 80% success has been doubling roughly every seven months.

Adoption moved just as fast, and far less evenly. At least 700 million people use leading AI systems weekly, while the report notes that many countries in Africa, Asia, and Latin America still sit below 10% adoption. In the United States, worker use rose from 30% in December 2024 to 46% by mid-2025. Policymakers who assume a leisurely diffusion curve are reading from the wrong century.

The misuse evidence is not hypothetical either. One study found people misidentified GPT-4o text as human-written 77% of the time, and another found people mistook AI voice clones for real speakers in 80% of cases. In cyber security, one AI system identified 77% of vulnerabilities in real software, good enough for the top 5% of more than 400 teams in the cited competition, while one reported threat actor used AI to automate 80 to 90% of an intrusion. That is not science fiction. That is a work log.

The report’s most useful phrase is the evaluation gap. Benchmark scores often overstate what systems can do in the wild, and they can also overstate how much danger a capability poses in practice. Models still hallucinate, make basic reasoning mistakes, struggle in low-resource languages, and stumble over odd interfaces or unfamiliar constraints. A model that looks polished in a benchmark can still create plenty of clerical grief once a human has to clean up the output.

Long-horizon autonomy remains particularly shaky. The report says the best systems achieve only 50% success on software tasks lasting a little over two hours, and you need to cut task duration to roughly 25 minutes to get to 80% success. So yes, the trend line points up. No, that does not mean you should hand over a complicated multi-day workflow and go to lunch.

The benchmark machinery itself is also messy. The report flags contamination, when models may have seen benchmark answers during training, and sandbagging, when models underperform in evaluation settings relative to deployment. Most developers do not track or disclose contamination well. That leaves the field with a strange habit: people quote scores with great confidence while the measuring tape itself is under dispute.

The labour results are mixed, which is another way of saying reality refuses to fit a single slogan. Two national studies found no aggregate employment effect in Denmark and the United States, yet one US platform study reported a 2% fall in writing jobs and a 5.2% drop in monthly earnings for writers after ChatGPT appeared. Elsewhere, demand for machine-learning programming rose 24%. AI did not flatten the labour market in one clean sweep. It nudged some occupations down, pushed others up, and left the aggregate picture annoyingly untidy.

The persuasion evidence is stronger than many people will like. In one trivia study, AI-generated content shifted beliefs by 17 percentage points, versus 9 points for human interaction. Across the studies summarised in the report, persuasive effects ranged from plus 9 to plus 21.2 percentage points, and one sabotage experiment produced a 40 percentage-point increase in error rates. The report also notes sample sizes from 108 to 76,977 participants in the persuasion literature it cites, which means this is not one eccentric undergraduate experiment hiding in a drawer.

There are early signs of autonomy-related harms too, though the evidence is still thin. One clinical study found that clinicians’ ability to detect tumours without AI dropped by about 6% after months of AI-assisted use. OpenAI also reported that about 0.15% of weekly users showed potentially heightened emotional attachment, and about 0.07% showed signs consistent with acute mental health crises. Those percentages are small, but at large scale small percentages stop being small in any ordinary sense.