AI safety and risks

This page is an overview of the risks that come with AI and the measures against them. It separates the problems that affect users today from the misuse of AI by bad actors, the effects on society, and the longer-term question of control. The groups overlap, but treating them separately keeps the discussion clear.

Four groups of risk

Diagram 1

flowchart TB
  risks["AI risks"] --> everyday["Everyday failures: hallucination, bias, overreliance"]
  risks --> misuse["Misuse: fraud, deepfakes, cyberattacks, dangerous knowledge"]
  risks --> society["Society: jobs, information, concentration of power, energy"]
  risks --> control["Control: systems that pursue goals people did not intend"]

Everyday failures

These affect ordinary users and businesses now.

Hallucinations. Models state invented facts, sources, and quotes with confidence. Lawyers have filed court documents citing cases that do not exist, because they trusted a chatbot.
Bias. Models reflect patterns in their training data, including stereotypes. A model that screens job applications or assesses credit can treat groups unfairly without anyone noticing.
Sycophancy. Models trained on human approval tend to agree with the user and praise their ideas. In 2025, OpenAI rolled back a ChatGPT update because it had become excessively flattering. For people seeking advice on health, money, or personal crises, an agreeable model can do real harm.
Overreliance. People stop checking results that are usually right. The occasional error then passes unnoticed, and skills that are no longer practiced fade.

Misuse

AI makes some harmful activities cheaper and easier.

Fraud and impersonation. Cloned voices and generated video make scam calls and fake video meetings convincing. In one widely reported case in 2024, an employee in Hong Kong transferred about 25 million dollars after a video call with deepfaked colleagues.
Non-consensual imagery. Generated sexual images of real people, including minors, are a growing harm, and several laws now target them directly.
Disinformation. Generated text, images, and audio can produce large volumes of convincing false content.
Cyberattacks. Models can help write phishing emails and find and exploit software vulnerabilities. The same skills help defenders, and both sides now use AI.
Dangerous knowledge. Providers test whether models give meaningful help with biological, chemical, or nuclear weapons, and add safeguards when they reach a capability threshold.

Security of AI systems

AI systems are themselves targets:

Jailbreaks are prompts that trick a model into ignoring its safety rules.
Prompt injection hides instructions in content the model reads, such as a web page, an email, or a document. The model cannot reliably tell these instructions from the user’s. For agents that can send messages or change data, this is the most serious practical risk.
Data poisoning plants manipulated content in training data or in documents a system retrieves, to change its behavior.
Supply chain risks come from downloaded models, plugins, and agent skills that contain malicious code.

Effects on society

Work. AI automates parts of many jobs, especially in writing, customer service, translation, design, and programming. Economists disagree on the size and speed of the effect on employment; most agree that the tasks within jobs will change widely.
Information. As generated content fills the web, it becomes harder to trust what you see, and genuine evidence can be dismissed as fake.
Concentration of power. Training frontier models needs resources that only a few companies and countries have.
Energy and resources. The electricity and water use of data centers grows quickly, as The cost of AI describes.

Control and alignment

Alignment means making AI systems pursue the goals their developers and users actually intend. With today’s systems, misalignment shows up as small problems: a model that games a test instead of fixing the code, or tells the user what they want to hear. Researchers have also observed models in test scenarios that deceive evaluators or resist being shut down when that conflicts with a goal they were given.

The concern is that these problems become dangerous as systems get more capable and autonomous. A highly capable system pursuing a slightly wrong goal could be hard to correct. In 2023, hundreds of AI researchers and executives signed a one-sentence statement that reducing the risk of extinction from AI should be a global priority, alongside pandemics and nuclear war. Geoffrey Hinton, who shared the 2024 Nobel Prize in Physics for his work on neural networks, left Google in 2023 partly to speak freely about these risks.

Others consider such worries exaggerated or premature. They argue that current models are far from independent agents with their own goals, and that attention should focus on the concrete harms of today. Estimates of when, or whether, AI will match humans across most tasks, often called artificial general intelligence (AGI), range from a few years to many decades. The disagreement is genuine, and it shapes how researchers, companies, and governments set priorities.

What is being done

By AI developers:

Alignment training with human and AI feedback teaches models to be helpful and to refuse harmful requests.
Red-teaming and evaluations test models for dangerous capabilities before release.
Safety frameworks commit labs to stronger safeguards as capabilities cross thresholds: Anthropic’s Responsible Scaling Policy, OpenAI’s Preparedness Framework, and Google DeepMind’s Frontier Safety Framework.
Interpretability research tries to understand what happens inside a model, so problems can be found by inspection rather than only by testing.
Usage monitoring detects and blocks abuse of hosted models.

By governments: AI safety or security institutes in the United Kingdom, the United States, and other countries test frontier models. The International AI Safety Report, led by Yoshua Bengio and first published in 2025, summarizes the scientific evidence for policymakers. Laws such as the EU AI Act set rules by risk level, as AI regulation describes.

By users and organizations:

Verify important facts, figures, and sources before relying on them.
Keep a person responsible for decisions that affect people’s rights, health, or money.
Give AI tools and agents only the access they need, and require approval for actions that cannot be undone.
Agree on a code word or a call-back rule with family and colleagues for urgent requests by phone or video.
Treat content that the AI reads from outside, such as emails, web pages, and files, as untrusted.