AI Alignment
The problem of making AI systems actually do what people intend — reliably pursuing the goals we want, not just the ones we accidentally specified.
Reviewed July 10, 2026Stable
When not to use it
- As a synonym for safety. Alignment is about systems pursuing intended goals; safety also covers misuse, reliability, security, and impact. Collapsing them hides real problems.
- As a reason to defer near-term duties. Long-term alignment debates don't excuse an unmonitored model making decisions about people today.
- As a marketing claim. "Aligned" is not a binary property a product can possess, and treating it as one is how the term gets emptied.
Reach for something else instead
- Evaluation and red-teaming — concrete, measurable, and what most teams actually need before they need alignment theory.
- Access control and scope limits. The strongest safety measure is usually not letting the system do the dangerous thing at all.
- Human oversight on consequential decisions, designed in rather than promised.
Read more on the blog
- AI alignment and safety, without the hype or dismissalAI safety is discussed either as impending doom or as overblown hype, and neither framing helps you understand it. The actual landscape: what alignment means, the concrete risks experts agree on, the speculative ones they don't, and why serious people land in very different places.
- Superintelligence: the empirical record is zeroSixty years after the intelligence explosion was described, no system has demonstrated sustained open-ended self-improvement. The public forecasts come from five people with the same financial interest.
- What is prompt injection, and why is it unsolved?Prompt injection is the number one security risk for AI applications, and researchers treat it as unsolved: not a bug waiting for a patch, but an architectural flaw in how language models work. Here is why a model cannot reliably tell instructions from data, why that makes AI agents dangerous, and what actually reduces the risk.
- Mechanistic interpretability: opening the AI black boxWe built AI systems that work without fully understanding how they work. We have every number inside them, yet the numbers do not obviously mean anything. Mechanistic interpretability is the effort to reverse-engineer that black box, and it has started to succeed, revealing why you cannot just read a neuron and how researchers now extract readable concepts from the tangle.
Further reading
- Amodei et al. (2016), Concrete Problems in AI Safety — still the clearest framing of the near-term technical issues.
- Christiano et al. (2017), Deep Reinforcement Learning from Human Preferences — the technique behind RLHF.
- Bai et al. (2022), Constitutional AI — one approach to supervision that doesn't scale with human labellers.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Assuming a model that behaves well in testing is aligned. It's evidence about the test, not the system.
- Confusing refusing to say things with being aligned. A model can be harmless and still pursue the wrong objective.
- Treating this as purely technical. What "intended behaviour" means is a question about people, and it doesn't have a purely engineering answer.
At a glance
FieldSafety & Ethics
Core ideaAI doing what we intend, not just what we said
Failure modereward hacking
Current methodsRLHF, red-teaming, oversight
DifficultyIntermediate → Advanced
Where this sits
30 concepts come first. Understanding it opens up 8 more.
14Levelsteps in
30Needs firstconcepts
8Opens up3% of 310
2Areascrosses fields
Computed from the prerequisite graph, not assigned. How this works