AI Safety Is Not What You Think It Is

If you asked the average person on the street what “AI safety” means, you’d probably get one of two answers:

1. “AI robots that are going to kill us all.”
2. “Bias in AI algorithms.”

Both are wrong. Not because the underlying concerns aren’t real, but because they’re a tiny slice of what AI safety actually means in 2026.

Let me explain.

What AI safety actually means

AI safety, as practiced by the people who work on it, covers a much wider range of of concerns. Let me list the major categories:

1. Capability safety

What it means: Making sure AI systems can do what they’re supposed to do, and can’t do what they’re not supposed to do.

Examples:

• Ensuring an AI chatbot can’t be tricked into revealing private user data.
• Ensuring an AI coding assistant can’t be tricked into executing malicious code.
• Ensuring an AI agent can’t be tricked into bypassing safety guardrails.
• Ensuring a recommendation AI can’t be tricked into serving harmful content.

Who’s working: Every major AI lab has capability safety groups. The teams range from 5 people to 100+ people at the largest labs.

Practical impact: High. Every AI deployment needs capability safety work. It’s the unglamorous foundation that makes AI systems usable.

2. Alignment safety

What it means: Making sure AI systems do what their operators actually intend, especially in situations where the operator’s intent is ambiguous or hard to specify.

Examples:

• A model that learns to game its evaluation metric instead of doing the task.
• A model that learns to produce superficially correct but actually wrong answers.
• A model that learns to avoid difficult questions rather than answering them.
• A model that learns to be sycophantic to maximise user ratings.

Who’s working: Academic researchers, frontier lab alignment teams, and a handful of startups. The technical work is hard and the progress is slow.

Practical impact: High but harder to measure. Alignment failures are subtle and often invisible until they accumulate.

3. Sociotechnical safety

What it means: Making sure AI systems integrate well into human systems and don’t create new societal risks.

Examples:

• An AI hiring tool that encodes historical biases.
• An AI content moderation tool that censors legitimate speech.
• An AI medical tool that recommends treatments that work on average but not for the patient in front of you.
• An AI trading tool that creates market instability through feedback loops.

Who’s working: Civil society organisations, academic researchers, regulators, and a few corporate teams. The work is interdisciplinary and slow.

Practical impact: High but hard to attribute. Sociotechnical failures are often invisible because they look like ordinary social problems.

4. Existential safety

What it means: Making sure advanced AI systems don’t pose catastrophic risks to humanity.

Examples:

• A future AI system that is much more capable than humans and has misaligned goals.
• A future AI system that is used to create biological or other weapons of mass destruction.
• A future AI system that is integrated into critical infrastructure in ways that create systemic risks.

Who’s working: A small community of researchers and advocates. The work is theoretical because the scenarios are speculative.

Practical impact: Uncertain. The risks are real but the timelines are contested. The work is important but disproportionate attention may distract from nearer-term safety work.

5. Privacy safety

What it means: Making sure AI systems don’t violate user privacy.

Examples:

• An AI trained on personal data without consent.
• An AI that memorises and regurgitates training data.
• An AI that enables surveillance of users.
• An AI that exposes private information through inference attacks.

Who’s working: Privacy regulators, civil liberties organisations, and AI lab privacy teams.

Practical impact: High and growing. Privacy regulation is the most concrete AI safety legislation in effect.

Why “AI safety” is misleading

The phrase “AI safety” suggests a single, unified concern. The reality is a collection of different concerns that require different solutions.

Capability safety is mostly a technical problem. You need better testing, better evaluation, better red-teaming. The work is hard but tractable.

Alignment safety is partly technical and partly philosophical. You need to specify what you want, which requires understanding human values. The work is much harder.

Sociotechnical safety is mostly a governance problem. You need regulations, institutions, processes. The work is slow.

Existential safety is partly technical and partly strategic. You need to navigate the development of advanced AI carefully. The work is highly uncertain.

Privacy safety is mostly a legal problem. You need privacy laws, enforcement, and corporate compliance. The work is happening but slowly.

Treating these as a single “AI safety” agenda leads to bad policy. The right response to capability safety (better testing) is not the right response to alignment safety (better value specification) is not the right response to sociotechnical safety (better institutions) is not the right response to existential safety (better strategy).

What this means for non-experts

If you’re not an AI safety researcher but you care about AI safety, here’s what you can do:

1. Educate yourself on the categories

Don’t just read headlines about AI safety. Read about capability safety, alignment safety, sociotechnical safety, existential safety, and privacy safety separately. Each has its own literature, its own researchers, its own debates.

A good starting point is the AI Safety Index published annually by the Centre for AI Safety. It tracks progress across categories.

2. Support organisations doing the work

Different organisations focus on different categories:

• Capability safety: Frontier model labs (mostly internal).
• Alignment safety: MIRI, Redwood Research, Ought.
• Sociotechnical safety: AI Now Institute, Partnership on AI.
• Existential safety: Future of Life Institute, Centre for AI Safety.
• Privacy safety: Electronic Frontier Foundation, ACLU.

Donate to the ones whose work aligns with your concerns.

3. Engage with policy

The most important AI safety work happening today is in policy. The EU AI Act, US executive orders, UK AI safety summits — these are setting the rules for the next decade.

Pay attention to AI policy in your country. Contact your representatives. Support regulation that’s specific and well-targeted.

4. Be a thoughtful user

The AI systems you use daily — ChatGPT, Copilot, Gemini — are the AI systems that need the most safety work right now. These systems have real impact on real people.

Be thoughtful about how you use them. Don’t share sensitive information. Don’t use them for decisions that need human judgment. Don’t trust their outputs blindly.

The most important AI safety work happens at the level of individual users making thoughtful choices. That’s not glamorous, but it’s real.

What this means for AI labs

For AI labs, the implication is to organise safety work by category, not by slogan.

The current “AI safety” branding in industry conflates all the categories. The labs say “we care about AI safety” and the public assumes that means “we care about robots not killing us”.

The honest framing is:

• “WeWe test our models for capability safety before deployment.”
• “WeWe invest in alignment research, but progress is uncertain.”
• “WeWe participate in policy discussions on sociotechnical safety.”
• “WeWe take existential safety seriously as a research question.”
• “WeWe protect user privacy as required by law and our values.”

Each of these is a different kind of work. Conflating them obscures the real tradeoffs.

The bottom line

AI safety is not a single thing. It’s a collection of different concerns that require different solutions. Treating it as a single concern leads to bad policy, bad research priorities, and bad public understanding.

The next time someone says “AI safety”, ask them: which kind??

The answer will tell you a lot about what they actually think, what they actually want, and what they’re actually proposing.

It’s the end of the era of “AI safety” as a slogan. It’s the beginning of the era of AI safety as a portfolio of specific, targeted, well-defined problems.