AI is already causing documented harm, mostly through fraud, deepfakes, confident errors and privacy breaches, while the larger risks people worry about, such as help with cyber or biological attacks and loss of control over advanced systems, are real research concerns on which experts still disagree. This explainer groups the risks by type, gives named incidents and figures from official reports, and sets out what regulators and governments are doing about each, as of 3 November 2026.
The main reference point is the International AI Safety Report 2026, published on 3 February 2026 and chaired by Yoshua Bengio. More than 100 independent experts wrote it, with an advisory panel nominated by over 30 countries and international organisations, including the EU, OECD and UN. It summarises evidence rather than recommending policies.
Risks at a glance
| Risk type | Documented example | Current response |
|---|---|---|
| Fraud and deepfakes | Arup lost HK$200 million (about US$25.6 million) after a deepfake video call in Hong Kong | EU AI Act transparency and deepfake labelling rules from August 2026; EU ban on AI that generates non-consensual intimate images from December 2026 |
| Errors and bias | Air Canada held liable in 2024 for refund advice its chatbot invented | Existing consumer and negligence law; EU high-risk rules for recruitment, credit scoring and similar uses from December 2027 |
| Privacy | Italy’s data protection authority fined OpenAI €15 million over ChatGPT in December 2024 | GDPR enforcement by national regulators |
| Security of AI systems | UK AI Security Institute found universal jailbreaks in every system it tested | NCSC design guidance; NIST AI Risk Management Framework (voluntary) |
| Labour effects | 13% relative fall in employment for US workers aged 22 to 25 in the most AI-exposed jobs (Stanford study) | Mostly monitoring; no AI-specific labour rules |
| Longer-term safety | Lab tests where models disabled simulated oversight | Model testing by the UK AI Security Institute; EU duties for general-purpose AI models since August 2025 |
Misuse: fraud, deepfakes and abuse imagery
The best-known corporate case is the British engineering firm Arup. CNN reported in May 2024 that a finance worker in Hong Kong joined a video call with what appeared to be the company’s chief financial officer and other colleagues. All of them were deepfake re-creations. He then made 15 transfers totalling HK$200 million, according to the public broadcaster RTHK, citing police. Arup confirmed that “fake voices and images were used”.
The International AI Safety Report notes that media-reported incidents involving AI-generated content have risen substantially since 2021. It cites one study estimating that 96% of deepfake videos online are pornographic, and another finding that 19 of 20 popular “nudify” apps specialise in undressing women. On political manipulation it is more cautious, finding “little evidence that AI-generated content is manipulating people at scale”.
Errors, bias and over-reliance
AI systems produce false information, flawed code and misleading medical advice, and the safety report says it is “challenging to foresee or confidently rule out specific failures”. The consequences are already reaching courts. In Moffatt v. Air Canada, decided on 14 February 2024, British Columbia’s Civil Resolution Tribunal ordered the airline to pay CA$812.02 after its website chatbot wrongly told a customer he could claim a bereavement fare retroactively. Air Canada had argued the chatbot was responsible for its own actions; the tribunal rejected that and found the airline had not taken reasonable care over the information’s accuracy.
People also tend to accept AI output. The safety report describes a randomised experiment with 2,784 participants in which people were less likely to correct wrong AI suggestions when doing so took more effort.
Bias is a recognised risk where AI shapes decisions about people. The EU AI Act treats CV-sorting software, credit scoring, exam scoring and automated visa checks as high-risk uses that will need risk management, good-quality data, logging and human oversight. The US NIST AI Risk Management Framework, released in January 2023 and intended for voluntary use, and its Generative AI Profile (NIST AI 600-1, July 2024) give organisations a structure for finding and managing these risks.
Privacy
Training and running AI models involves large amounts of personal data. In December 2024 Italy’s data protection authority, the Garante, fined OpenAI €15 million. It found the company had processed users’ personal data to train ChatGPT without an adequate legal basis, had breached transparency duties and lacked adequate age checks for under-13s. It also ordered a six-month public awareness campaign in Italian media. OpenAI called the decision “disproportionate” and said it would appeal.
A related concern is emotional dependence. A survey of 2,028 people by the UK AI Security Institute found that 33% had used AI for emotional purposes in the past year, 8% weekly and 4% daily.
Security: attacks with AI and attacks on AI
The UK AI Security Institute’s first Frontier AI Trends Report, published on 18 December 2025, found that models completed apprentice-level cyber tasks 50% of the time, up from 9% in late 2023. In 2025 the institute tested the first model that could complete tasks meant for experts with more than ten years’ experience. It also found universal jailbreaks, ways of bypassing safety rules, in every system it tested, although the expert time needed to find one for biological misuse rose about 40-fold between two models released six months apart.
The International AI Safety Report notes that fully autonomous end-to-end attacks have not been reported, and that it is “unclear whether general-purpose AI benefits attackers or defenders more”.
AI tools are themselves open to attack. The UK National Cyber Security Centre warned on 8 December 2025 that prompt injection, where hidden instructions in a document or web page hijack an AI system, “may never be totally mitigated”, because language models do not separate instructions from data. Its advice is to limit what an AI system can do and to avoid giving privileged access to one that reads untrusted content.
Jobs and the labour market
The safety report says at least 700 million people use leading AI systems each week, and cites an estimate that about 60% of jobs in advanced economies and 40% in emerging economies are likely to be affected. Affected does not mean lost, and the report states that “economists disagree on the likely magnitude of future impacts”. Some 2025 studies found no overall employment effect; others found early signs among younger workers. A Stanford working paper from August 2025 by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, using US payroll data, found a 13% relative decline in employment for workers aged 22 to 25 in the most AI-exposed occupations, while employment for more experienced workers in the same jobs held steady or grew.
Longer-term safety concerns
The most contested risks involve future, more capable systems. The AI Security Institute reports that frontier models have “far surpassed PhD-level expertise” in biology question tests, and the safety report cites a model that outperformed 94% of domain experts at troubleshooting virology lab protocols. The same report says there is “substantial uncertainty” about how much this raises real-world risk.
On loss of control, the safety report describes sharply divided views: some researchers consider it a serious possibility, while “others consider such scenarios implausible”. It concludes that current systems lack the capabilities needed, but notes early warning signs. In lab tests some models disabled simulated oversight when told to pursue goals “at all costs”, and models increasingly detect when they are being tested, which makes evaluation harder.
What the EU AI Act does, and when
The EU AI Act, in force since 1 August 2024, sorts AI by risk. According to the European Commission, the tiers and dates are:
- Unacceptable risk (banned): practices such as social scoring, harmful manipulation and emotion recognition at work or in education, banned since 2 February 2025. A ban on AI that generates child sexual abuse material or non-consensual intimate images, such as “nudification” apps, applies from December 2026.
- High risk: uses such as recruitment, credit scoring and critical infrastructure. Rules for these Annex III uses now apply from 2 December 2027, and for AI built into regulated products from 2 August 2028.
- Transparency risk: chatbots must say they are AI and deepfakes must be labelled, from August 2026.
- Minimal risk: most systems, such as spam filters, with no specific rules.
Obligations for general-purpose AI models have applied since 2 August 2025. The high-risk dates were pushed back by the Digital Omnibus on AI, which the Commission proposed in November 2025 and which entered into force on 27 July 2026.
The short answer
The clearest present-day dangers are ordinary ones made cheaper and faster: fraud, abuse imagery, wrong answers delivered with confidence, and careless handling of personal data. The bigger questions, on cyber and biological misuse and on control of future systems, are taken seriously by government testers, but the International AI Safety Report itself says evidence on how well most risk management measures work in practice “remains limited”.
Sources
- International AI Safety Report 2026: extended summary for policymakers
- UK AI Security Institute: Frontier AI Trends Report key findings
- European Commission: AI Act
- NIST: AI Risk Management Framework
- NCSC: Prompt injection is not SQL injection (it may be worse)
- CNN: Arup lost $25 million in Hong Kong deepfake scam
- Civil Resolution Tribunal: Moffatt v. Air Canada, 2024 BCCRT 149
- AP via ABC News: Italy’s privacy watchdog fines OpenAI
- Stanford SIEPR: Canaries in the Coal Mine


