Learn AI /Training and Model Adaptation /Alignment
Alignment
- Intermediate
- Training and Model Adaptation
- Policy officers
- Managers & front offices
Also called: AI alignment, aligned.
Back to the milestone you came from
In five seconds
What it is
Making an AI system's behaviour better match intended goals or preferences, often through methods applied after broad training.
In plain words
After a model has learned general patterns, developers can further adjust it so it follows instructions and preferred behaviour more reliably. That adjustment is often what people mean by alignment.
An everyday example
An experienced cook already knows how to prepare many dishes. In a new kitchen, the head chef provides sample plates, written standards and repeated choices between two versions of the same dish. The cook gradually learns that this kitchen prefers less salt, a particular presentation and no substitutions without asking. The cook has not relearned cooking from the beginning. Their existing skill has been adjusted towards the new kitchen’s preferences. They can still misread an order or make a poor judgement, so the final plate is still checked.
A student can reflect on reasons and goals in a human sense; a model does not do that in the same way. Also, examination feedback usually comes from one teacher and a clear marking scheme, while AI alignment can involve many raters, many criteria and statistical optimisation across huge numbers of examples. Finally, aligned behaviour on tests does not prove reliable behaviour in every new situation.
Remember thisAlignment usually means steering a broad model towards selected behaviour after pretraining. It improves conformity to chosen targets, but does not prove that the model is safe, correct or fully controlled.
This term is contested
The term is genuinely contested. In product practice it often means post-training to improve behaviour, while in research and policy it can mean the broader problem of making advanced AI systems reliably pursue intended goals.
Understand it
Understand it
Why it matters
If someone says a model is aligned, that sounds reassuring, but it may refer to very different things. For buyers, managers and policymakers, the key question is aligned to what, by which method, tested how, and with what trade-offs. A system tuned to be polite may still be inaccurate. A system tuned to refuse risky requests may also refuse legitimate ones. Clear understanding helps with procurement, risk assessment, governance, incident review and public communication. It also prevents a common mistake: treating a general claim about alignment as proof that a system is safe, fair or reliable in your own setting.
From broad training to more usable behaviour
Show how a broadly trained model is further steered using examples, feedback and tests, and why that is not the same as a simple output filter.
01
1. Broad model
Start with a pretrained model that has learned general language patterns from large text collections.
Draw a large rounded rectangle labelled 'Pretrained model'. Add faint background lines to suggest many training documents. Use neutral blue-grey.
02
2. Desired behaviour examples
Add examples of good task-following answers, plus rules or policies about what behaviour is wanted or unwanted.
Place a second box to the right labelled 'Examples, instructions, policies'. Show three small document icons feeding into it.
03
3. Feedback-based adjustment
Use human judgements or preference data to update the model so preferred responses become more likely in similar situations.
Connect to a third box labelled 'Alignment methods'. Inside, show two simple chips: 'instruction tuning' and 'human feedback'. Add a gear icon to suggest updating weights.
04
4. New version and checks
Evaluate the updated model on tests for helpfulness, policy-following and failure cases, then release it as a new version if results are acceptable.
Draw a final box labelled 'Aligned model version'. Beneath it, add a checklist icon and a smaller side box labelled 'optional guardrails outside model' connected with a dotted line.
A left-to-right step diagram with four boxes. Read it as a process: broad model first, then desired behaviour examples, then feedback-based adjustment, then testing and release as a new version. A small note under the final box says some controls may still sit outside the model as guardrails.
On hover or tap, each step reveals one sentence explaining what goes in, what changes, and what comes out.
Describe this without the diagram
The diagram is a four-step process read from left to right. Step 1 is a pretrained model that already learned broad language patterns. Step 2 adds desired behaviour examples, instructions and policies. Step 3 applies alignment methods, such as instruction tuning and feedback from human judgements, to make preferred responses more likely and less preferred responses less likely. Step 4 evaluates the updated model and releases it as a new version if it performs acceptably. A side note shows that some guardrails may still sit outside the model, meaning alignment is not the same thing as a runtime filter.
With reduced motion: A static four-box process with arrows from 'Pretrained model' to 'Examples, instructions, policies' to 'Alignment methods' to 'Aligned model version and checks', plus a dotted side note for optional external guardrails.
People use this word in more than one way, so it helps to be precise. In everyday product talk, it often means adjusting a model after its broad initial training so it behaves in ways people find more useful, safer or more consistent. In research and policy discussions, it can also mean the wider problem of making advanced AI systems reliably pursue intended goals.
For a language model, the input to alignment work usually includes example prompts and preferred answers, ratings from people, rules about what the system should avoid, and tests that check its behaviour. A model is first broadly trained on large amounts of text in pretraining. Then extra methods are used to steer it. These may include instruction tuning, preference learning, or reinforcement learning from human feedback, where human judgements are used as a training signal.
What happens during alignment is not magic and it does not guarantee perfect behaviour. Developers use training methods to change the model so certain kinds of responses become more likely and others less likely. That can improve helpfulness, reduce some harmful outputs and make answers follow a task more closely. But it can also introduce trade-offs, such as making the model more cautious, more verbose or less flexible.
The output of alignment work is usually a model version that better matches chosen instructions or preferences on the tests that were used. This differs from fine-tuning, which is a parameter-update method that can serve alignment or another adaptation goal. It also differs from a guardrail, which is a rule, check or control that limits outputs or actions before people rely on them. AI safety is broader and covers risks, governance and real-world impacts beyond this training stage.
How “An experienced cook joining a new kitchen” maps onto the real thing
| In the story | In the AI system |
|---|---|
| the student after studying many subjects already | a pretrained model with broad general patterns learned earlier |
| the teacher's marked feedback | human ratings or preference signals used to steer later training |
| practising the style expected by the examination | instruction tuning and related methods that make outputs follow desired formats and behaviours |
| better performance on similar practice papers | improved behaviour on evaluation tasks used during alignment |
A worked example
A translation service is adapting an existing multilingual model to produce first drafts that follow approved terminology, preserve qualifications and avoid adding content not present in the source.
Starting from: Curated source-and-draft examples, paired responses ranked by revisers, relevant drafting guidance and a held-out evaluation set containing texts not used during adaptation.
- The developers begin with an existing pretrained model rather than training a language model from scratch.
- They use supervised instruction tuning on curated examples so the model practises following terminology, formatting and fidelity instructions.
- They apply preference learning using pairs of candidate drafts, giving greater weight to versions that revisers judged accurate, restrained and consistent with the guidance.
- They test the adapted model on unseen texts, checking instruction-following, terminology, omissions, unsupported additions and cases where the source is ambiguous.
Result: The adapted model follows the selected drafting preferences more consistently than the base model. It still requires human revision, particularly where wording is politically sensitive, context-dependent or genuinely ambiguous.
In a United Nations workplace
At a shared services centre, a reviser is checking a multilingual drafting assistant before a side-event concept note goes for clearance. The base model produces fluent text, but it sometimes ignores approved terminology, adds unnecessary flourishes and changes carefully negotiated wording.
Alignment is the work done after broad training to make the model follow selected instructions and preferences more consistently. The reviser should ask what supervised examples, preference judgements and evaluation sets were used, and whether the resulting behaviour was tested on unseen texts rather than demonstrated on a few convenient prompts.
Before you try thisDo not enter an uncleared concept note, negotiating text or other non-public material into an external tool that has not been approved for that information. Alignment to house style does not make a model secure, confidential or reliable enough to replace revision and clearance.
Where you may hear it
- During a vendor demo to an ICT governance board
- In an evaluation readout before clearance of an AI pilot
- At an inter-agency working group reviewing generative AI safeguards
- In a procurement meeting examining evidence about model adaptation
Avoid confusion
Avoid confusion
Often confused with
AI safety
AI safety means designing, testing and using AI systems so they are less likely to cause harm, especially in situations where people could be affected.
Guardrail
A guardrail is a rule, check or control that limits unsafe or unacceptable AI outputs or actions before people rely on them.
Fine-tuning
Updating a pre-trained model's parameters with additional examples to improve performance on a narrower task, style or domain.
Common misconceptions
Common misconception: Alignment just means adding a safety filter on top of the model.
Correction: Sometimes a product includes both, but they are different layers. Alignment usually refers to changing the model's behaviour through post-training or feedback-based methods. A guardrail is often a separate rule, classifier or filter used at runtime. A model can be aligned, guarded, both or neither.
Common misconception: Once a model is aligned, it will reliably behave as intended in every situation.
Correction: Alignment improves behaviour on the data, criteria and tests used, but it does not remove all errors or surprises. New contexts, ambiguous prompts, conflicting goals and weak evaluations can still produce unwanted outputs. This is why organisations still need human oversight, testing and monitoring.
Common misconception: Alignment and AI safety mean the same thing.
Correction: They overlap, but they are not identical. Alignment often focuses on making system behaviour better match intended instructions or human preferences. AI safety is broader and can include misuse prevention, robustness, governance, deployment conditions, social impacts and long-term risks.
Quick check
Which statement best distinguishes alignment from a guardrail?
This is the key distinction. Alignment usually refers to post-training or related methods that steer the model itself. Guardrails are commonly external controls such as filters, classifiers or hard rules applied when the model is used. In real products, both may be present.
Go deeper
Go deeper
The simple version
After a model has learned general patterns, developers can further adjust it so it follows instructions and preferred behaviour more reliably. That adjustment is often what people mean by alignment.
The technical version
In current language-model practice, alignment often refers to post-training methods that optimise a pretrained model toward selected behavioural targets, such as instruction-following, policy compliance, response style or preference satisfaction. Typical inputs include supervised examples, pairwise preference data, constitutions or policy documents, and evaluation sets. Common mechanisms include supervised fine-tuning, instruction tuning, preference learning and reinforcement learning from human feedback. These methods alter model parameters, or in some systems combine parameter updates with external controls, so the distribution of likely outputs shifts toward behaviours that raters or designers prefer. In a broader research sense, 'alignment' can also refer to the harder problem of ensuring advanced systems robustly pursue intended objectives under distribution shift and imperfect specification. Because the term spans both practical post-training and deeper control questions, claims about alignment need context.
Reviewed external resources
EncyclopaediaAI alignment (opens in a new tab)Wikipedia
Wikipedia gives a broad overview of how the term is used and is suitable for readers who want orientation before technical detail.
Intermediatelink checked 2026-07-31
CourseLLM Course (opens in a new tab)Hugging Face
The Hugging Face LLM Course helps readers who want practical context on post-training methods that are often discussed under alignment.
Deeperlink checked 2026-07-31
CourseShort Courses (opens in a new tab)DeepLearning.AI
DeepLearning.AI short courses suit learners who prefer guided lessons and examples related to tuning and evaluating model behaviour.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Alignment (opens in a new tab)YouTube search
This YouTube search results page helps readers find different video explanations of alignment and suits people who learn well by comparison, but it is not one vetted source.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Alignment (opens in a new tab)Reddit search
This Reddit search results page shows community discussions and differing interpretations of alignment and suits readers who want to see common questions, but it is not one authoritative explanation.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Alignment usually means steering a broad model towards selected behaviour after pretraining. It improves conformity to chosen targets, but does not prove that the model is safe, correct or fully controlled.