AI risk
AI risk is the chance that a flaw, misleading operation, or malicious use of an artificial intelligence system causes harm to people, companies, or society.
In plain terms
AI risk is not a single danger but an umbrella term: everything that can go wrong when using artificial intelligence. This includes confidently stated incorrect answers, discrimination, data leaks, fraud, and cases where an autonomous agent acts without permission.
Think of driving. A car is useful, but everyone knows accidents can happen, which is why there are brakes, seat belts, traffic laws, and technical inspections. AI risk should be considered similarly: we do not ask whether to use it, but what can go wrong, how much damage could result, and how to prevent it.
Risk consists of two things: how likely the trouble is, and how large the damage is if it occurs. A bad movie recommendation is annoying, but a bad credit assessment or medical advice can have serious consequences. Therefore, the same technology is hardly risky in one situation but very risky in another.
This differs from traditional software risks: AI behavior cannot be written and verified line by line in advance. It may answer the same question differently, and its errors often sound convincing.
How it works
- Identification: they list what the system is used for, who it affects, and what can go wrong, from wrong answers to misuse.
- Assessment: for every risk, they estimate how likely it is and how much damage it would cause, then rank them based on this.
- Classification: the system receives a risk level based on its area of use. For example, EU regulation distinguishes between prohibited, high-risk, limited-risk, and minimal-risk applications, while imposing specific rules on large, general-purpose AI models.
- Mitigation: protective measures are built in, such as human approval for critical decisions, testing, restricted access rights for agents, and logging.
- Monitoring: the system is tracked in live operation because risks change when new data, new use cases, or new models arrive.
An everyday example
You ask a chatbot whether you can take two medications together and receive a definite but incorrect answer. If you believe it without verification, you bear the harm. This is the everyday form of AI risk, which is why it is worth checking reliable sources for health, legal, and financial questions.
Who is affected
A common misconception
Limits and risks
- Risk cannot be reduced to zero, only to an acceptable level, so the promise of 'safe AI' should always be treated with caution.
- Estimates are uncertain: experts do not agree on which risks are most severe or how likely they are.
- Debate about long-term, dramatic risks easily distracts from measurable harms already present today, and vice versa.
- The more autonomy and access an agent gets, the greater the risk, because its error is no longer just bad text, but actual action: a transfer, a deletion, a sent email.
Similar, but not the same
- AI safety – AI risk is the potential harm itself, while AI safety is the field and practice aimed at preventing or mitigating that harm.
- Existential risk – This is a broader concept: it refers to any threat to humanity's survival, such as nuclear war or a global pandemic. The AI-derived version is the most extreme slice of AI risk.
- Cybersecurity risk – Cybersecurity focuses on attacks against systems, while AI risk also covers cases where the system makes mistakes or causes damage on its own, without an attacker.
Related concepts
Technological risk, Risk management, Responsible AI
AI safety, AI regulation, High-risk AI system, Human oversight, Goal alignment
A brief history
Warnings about the dangers of thinking machines date back to the early days of computing, for example from Alan Turing in the early 1950s, Norbert Wiener in 1960, and I. J. Good in 1965. For a long time, however, the topic remained the concern of narrow research circles. It reached a wide public through Nick Bostrom's 2014 book Superintelligence and the late-2022 surge of generative AI. NIST released its AI Risk Management Framework in January 2023, and the EU's risk-based AI Act entered into force in August 2024.
Where things stand
By the end of 2026, AI risk has moved from professional debates into everyday corporate management and law. The obligations under the EU AI Act are becoming applicable in stages: prohibitions took effect in February 2025, and rules for general-purpose models have been in force since August 2025. The implementation of rules for high-risk systems is gradual, and the EU has placed the extension of certain deadlines on its agenda. More organizations are regularly assessing the risks of their AI systems. Most new questions arise from the spread of autonomous agents.
In the news
- ECB may review banks' data center lending over AI riskNov 6, 2026 · Money
- Philippine central bank: central banks must weigh AI risks and opportunitiesNov 2, 2026 · Money
- Miklós Szócska: erosion of doctors' diagnostic skills is an AI riskNov 2, 2026 · Health
- Anthropic says no single insurer may have capacity to cover a catastrophic AI incidentOct 26, 2026 · Money
- Bank of Japan, MAS and China's central bank focus on AI risks and strategyOct 20, 2026 · Money
- Generative AI a systemic risk to legal sector, and insurers struggle to price itOct 20, 2026 · Jog
- Bank regulators in the US and Europe tighten scrutiny of AI risksOct 13, 2026 · Money
- Grid security debate: Trump order and AI risk focus leave cyber problems unaddressedOct 12, 2026 · Ipar
Also known as
AI risk · AI risk · AI risks · AI risk · risks of artificial intelligence · AI risk · AI risks · AI hazard
Sources
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.