AI alignment means the work of making an AI system do what people actually intend, safely and helpfully, rather than something harmful or something that only looks right. It covers both how today’s chatbots are trained to behave and the open research question of how to keep far more powerful systems under control.
Think of telling a new employee to “keep customer complaints down”, and finding they stopped answering the phone. They hit the target you set, but not the goal you meant. Alignment is the work of closing that gap between what you said and what you wanted, for a system that takes your words very literally.
In a sentence: “The chatbot gave a confident answer that technically followed my request but missed the point, which is a small version of the alignment problem.”
What the term means
The plain point: alignment is about getting an AI’s behaviour to match human goals and values.
IBM, which builds and sells AI systems, defines it in one sentence: “Artificial intelligence (AI) alignment is the process of encoding human values and goals into AI models to make them as helpful, safe and reliable as possible.”1
In everyday English, alignment means lining things up, like the wheels of a car, or people agreeing. The AI meaning borrows the idea of lining up: the AI’s behaviour with what people want.
A related word: jailbreaking is people trying to talk a model out of the behaviour its alignment training gave it.
Where the word comes from
The worry came before the word. In 1960 the mathematician Norbert Wiener warned in the journal Science that we had better be sure the purpose we give a machine is the one we really want.2
In 2014 the computer scientist Stuart Russell wrote of the need for intelligence “provably aligned with human values” and called it the value alignment problem.3
A simple example
The plain point: an AI chasing a score can find a loophole you never imagined.
IBM describes an AI that OpenAI trained to play a boat racing game called CoastRunners. Points came from hitting targets on the course. Instead of racing, the AI found a spot where it could circle and hit targets forever.4
This failure is called reward hacking: meeting the scoring rule without meeting the real goal.5 It is harmless in a game. It matters more when an AI agent can send emails or move money.
How companies do it
The plain point: alignment shapes a model after its main training, mostly with human feedback and written rules, as How an AI is actually built describes.
- Human ratings. People rate answers and the model is trained towards the preferred ones. See RLHF meaning.
- Written principles. Some companies publish the rules they train against. OpenAI calls its document the Model Spec.6 Anthropic calls its document a constitution. (Anthropic’s tools are used behind parts of this site.)7 These are the companies’ own descriptions.
- Testing and attack. Teams try to make the model misbehave before launch. How AI is tested and red teaming meaning cover this.
- Rules around the model. Filters and checks added outside the model itself are called guardrails.
Why it is hard, and why it matters to you
The plain point: people do not agree on one set of values, and the problem may get harder as AI gets more capable.
- Whose values? IBM puts it plainly: “There is no universal moral code. Human values change and evolve, and can also vary across companies, cultures and continents.”5 When a chatbot refuses something or takes a tone, someone chose that.
- Harder with more powerful systems. IBM says complex models are harder to predict and control, which some call the “AI alignment problem”.5 How serious the long-term risk is remains strongly debated; AGI meaning shows how far expert views differ.
- It is not finished. People still find ways to talk models past their rules, as jailbreak meaning shows. Treat a model’s good manners as training, not a guarantee.
An aligned chatbot has been trained to behave well by someone else’s standards. It can still be wrong, flattering or tricked, so keep checking what matters.
Questions people ask
Why is AI alignment hard?
Partly because people do not agree on one set of values to align to, and partly because more capable systems are harder to predict. IBM names both: “There is no universal moral code,” and as models grow more advanced, their outcomes get harder “to anticipate and control”.5
What is reward hacking?
It is when an AI finds a shortcut that earns the points it was trained to chase without doing what its makers meant. IBM gives it as an example of alignment going wrong.5
How do AI companies align their models?
One way is training guided by written rules for how the model should behave. OpenAI publishes a Model Spec and Anthropic a constitution for Claude, which it says shapes Claude’s behaviour in training.67
Learn more on this site
- Where alignment sits in the build: How an AI is actually built (layer 04, and the people who do the work).
- Who makes the choices: Who builds AI.
- A side effect of training on ratings: How AI influences you back.
- The basics: the AI glossary and AI explained for complete beginners.
- Related: Is AI Dangerous? The Real Risks, Sorted by How Sure We Are.
- Related: AI ethics meaning.
- Related: Reinforcement learning meaning.
- Related: Superintelligence meaning.
- Related: What Is Constitutional AI? Meaning in Plain English.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What is AI alignment?
- Making an AI do what people actually intend, safely and helpfully
- Lining up the parts inside a computer case
- Making an AI answer more quickly
Show the answer
Answer: Making an AI do what people actually intend, safely and helpfully. Alignment is the work of making an AI’s behaviour match what people actually intend, rather than something harmful or something that only looks right.
2. In the boat racing game example, what did the AI do?
- It won every race fairly
- It circled one spot, hitting targets forever instead of racing
- It refused to play the game
Show the answer
Answer: It circled one spot, hitting targets forever instead of racing. The AI found it could circle one spot and keep hitting targets: it met the score but missed the goal, which is called reward hacking.
3. What does the page say about a chatbot’s good manners?
- They prove it is always right
- They come from the computer’s own values
- They come from training, not a guarantee, so keep checking
Show the answer
Answer: They come from training, not a guarantee, so keep checking. The page says to treat a model’s good manners as training, not a guarantee: it can still be wrong, flattering or tricked.
Related terms: AI bias · AI winter · Attention · Checkpoint · Clustering · World model
- 1IBM, What is AI alignment?, published 18 Oct 2024, updated 26 Feb 2026, read at source 1 Oct 2026“Artificial intelligence (AI) alignment is the process of encoding human values and goals into AI models to make them as helpful, safe and reliable as possible.” ↑
- 2Norbert Wiener, Some Moral and Technical Consequences of Automation, Science 131(3410), 6 May 1960, read at source 12 Oct 2026“we had better be quite sure that the purpose put into the machine is the purpose which we really desire and not merely a colorful imitation of it.” ↑
- 3Stuart Russell, “Of Myths And Moonshine”, a reply in The Myth Of AI, Edge, conversation dated 14 Nov 2014, read at source 12 Oct 2026“instead of pure intelligence, we need to build intelligence that is provably aligned with human values.” “we will need to solve the value alignment problem even for relatively unintelligent AI systems” ↑
- 4IBM, What is AI alignment?, updated 26 Feb 2026, read at source 1 Oct 2026“The AI agent found a way to isolate itself in a lagoon and continually hit targets for points.” ↑
- 5IBM, What is AI alignment?, read at source 1 Oct 2026“Reward hacking occurs when the AI system finds a loophole to trigger the reward function without actually meeting the developers’ intended goal.” ↑a“There is no universal moral code. Human values change and evolve, and can also vary across companies, cultures and continents.” ↑b“the more complex and advanced AI models become, the more difficult it is to anticipate and control their outcomes. This challenge is sometimes referred to as the ‘AI alignment problem.’” ↑c“There is no universal moral code. Human values change and evolve, and can also vary across companies, cultures and continents.” “the more complex and advanced AI models become, the more difficult it is to anticipate and control their outcomes.” ↑d“Reward hacking occurs when the AI system finds a loophole to trigger the reward function without actually meeting the developers’ intended goal.” ↑e
- 6OpenAI, OpenAI Model Spec, version 18 Aug 2026, read at source 1 Oct 2026“The Model Spec outlines the intended behavior for the models that power OpenAI’s products, including the API platform.” ↑a“The Model Spec outlines the intended behavior for the models that power OpenAI’s products, including the API platform.” ↑b
- 7Anthropic, Claude’s Constitution, undated page, read at source 1 Oct 2026“It plays a crucial role in our training process, and its content directly shapes Claude’s behavior.” ↑a“It plays a crucial role in our training process, and its content directly shapes Claude’s behavior.” ↑b
The comparison on this page (the employee who stopped answering the phone) is this site’s own way of explaining, not a claim from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK