Risks & safety
AI Alignment
Also known as: alignment problem
The challenge of making AI systems reliably pursue what people actually intend, and behave safely and honestly while doing it.
Alignment is about making sure an AI does what we actually mean, not just what we literally said, and does it safely and honestly. Computers follow instructions very literally, so an AI told to maximise something can find clever but unwanted shortcuts. Alignment is the ongoing work of getting powerful systems to share our intentions and values, and to be helpful, harmless and truthful.
The alignment problem is the challenge of ensuring an AI system's behaviour matches human intentions and values. It is hard for a few reasons: our goals are hard to specify completely, systems can satisfy the letter of an objective while missing its spirit ("specification gaming"), and a capable system optimising a flawed goal can produce harmful results.
How it is approached today
For language models, alignment is done partly through training: supervised fine-tuning on good examples, then learning from human preferences (RLHF) to make outputs helpful, honest and harmless. It also involves testing systems adversarially ("red-teaming") to find failures before deployment.
Why it matters more over time
As systems become more capable and more autonomous (agents), the cost of misalignment grows, which is why alignment sits at the heart of AI safety.
Alignment concerns whether an AI system reliably optimises for the objectives its designers and users actually intend, spanning outer alignment (specifying the right objective) and inner alignment (ensuring the learned policy internalises it rather than a proxy). Classic failure modes, reward hacking, specification gaming, Goodhart's law, and instrumental convergence, motivate the concern that competent optimisers can pursue misspecified goals in undesired ways (Christian 2020).
Current practice for LLMs centres on preference-based fine-tuning (RLHF and direct alternatives such as DPO), constitutional/AI-feedback methods, and evaluation via red-teaming and behavioural benchmarks (Ouyang et al. 2022). Open research problems include scalable oversight (supervising systems that may exceed human ability to evaluate), faithfulness and honesty (models that do not misrepresent their reasoning or knowledge, cf. hallucination), robustness to distribution shift and adversarial input, and interpretability as a tool for verifying internal objectives. Alignment is the technical core of the broader AI safety agenda.