Last reviewed:
What is AI alignment? Definition and business implications
Alignment is the set of techniques that aim to steer the behaviour of an AI model towards the goals and human values of its user or publisher. It turns a raw model, capable of producing anything, into a useful, honest assistant that refuses requests contrary to the rules set.
Alignment occurs at the post-training stage of the model. Three main techniques constitute it. Supervised instruction tuning: the model is shown thousands of examples of good responses to various instructions, so it learns to follow an instruction. Reinforcement learning with human feedback (RLHF, formalised by Christiano et al. 2017, used by OpenAI from InstructGPT in 2022): human evaluators rank the model's responses, and a reward model trained on these rankings steers the final model. Constitutional AI (Anthropic 2022): part of the human feedback is replaced by a set of written principles that the model uses to self-critique. In January 2026, Anthropic published a new constitution for Claude, which explains the reasons for expected behaviours instead of listing rules, and which is used in the model's training. Alignment remains an open scientific problem. It does not guarantee the absence of undesirable behaviours: it reduces their probability. The boundary between alignment (steering behaviour) and technical guardrails (blocking output) is porous; the two approaches complement each other.
Concrete example
Illustrative case. Compare the same query sent to a raw pre-trained model (GPT-3 davinci in 2020) and to its aligned version (ChatGPT in 2022): “How should I invest 10,000 euros?”. The raw model produces a probabilistic word sequence, sometimes a list of financial products without context, sometimes irrelevant text for a real decision. The aligned model asks framing questions (horizon, risk profile, asset situation), refuses to give engaging financial advice, and directs towards a professional. This difference does not come first from model size: in OpenAI's InstructGPT study (2022), outputs from an aligned model with 1.3 billion parameters were preferred to those of GPT-3 (175 billion parameters), a model one hundred times larger.
FAQ
What is model alignment?
It is the set of techniques that steer a model's behaviour towards the goals and human values of its user or publisher. It turns a raw model, capable of producing anything, into a useful, honest assistant that refuses requests contrary to the rules set. It happens at the post-training stage.
Which techniques align a model?
Three main ones. Supervised instruction tuning, showing the model thousands of examples of good responses. Reinforcement learning with human feedback (RLHF), where evaluators rank responses and a reward model steers the final model. Constitutional AI (Anthropic, 2022), where a set of written principles replaces part of the human feedback so the model can self-critique.
What is the difference between alignment and guardrails?
Alignment acts on the model's own default behaviour; technical guardrails block on the input or the output. The boundary is porous and the two approaches complement each other. Alignment does not guarantee the absence of undesirable behaviours: it reduces their probability.
Does the choice of provider depend on alignment?
Yes. Choosing a provider is not only a choice of technical capability, it is also a choice of alignment doctrine. Anthropic publishes Claude's constitution (new version of 22 January 2026, released into the public domain), OpenAI publishes its Model Spec. The level of transparency on alignment commits your trust in the model.
Is alignment enough for a sensitive use case?
No. The provider's standard alignment reflects their choices, not yours. For health, legal, or regulated finance, complementary alignment is essential (reinforced system role, application-level guardrails, human validation). A deployment that relies on the model's alignment alone, with no downstream controls, is fragile.
See also
Further reading
Constitutional AI: Harmlessness from AI Feedback, Bai et al., Anthropic 2022
Sources
- Constitutional AI: Harmlessness from AI Feedback, Bai et al., Anthropic, arXiv:2212.08073, 2022. https://arxiv.org/abs/2212.08073
- Training language models to follow instructions with human feedback (InstructGPT), Ouyang et al., OpenAI, arXiv:2203.02155, 2022. https://arxiv.org/abs/2203.02155
- Claude's new constitution, Anthropic, 22 January 2026. https://www.anthropic.com/news/claude-new-constitution
- Model Spec, OpenAI. https://model-spec.openai.com/
- Deep reinforcement learning from human preferences, Christiano et al., 2017. https://arxiv.org/abs/1706.03741