Responsible AI, Safety and Security
Making AI systems fair, understandable, secure against attacks and aligned with what people actually want.
AlsoAI safety · AI ethics · trustworthy AI · AI security
A powerful tool needs safety rules. This part of AI asks: is it fair to everyone, can we tell why it decided something, can bad actors trick it, and does it do what we really meant?
Where it fits
Not shown: 6 more kinds and parts. See the full list below.
+All relationships as a list
Overview
As AI systems make or support real decisions, how they fail starts to matter as much as how well they work. This domain covers four overlapping concerns:
- Fairness: models learn from historical data, including its biases, and can repeat them at scale.
- Understanding: explainability methods show which inputs drove a decision, from simple feature importance to research into the inner workings of neural networks.
- Security: models can be attacked. Adversarial examples fool image classifiers with invisible changes; prompt injection smuggles instructions into language models.
- Alignment: making sure capable systems pursue the goals we intend, not a literal or gamed version of them (see alignment and reward hacking).
Everyday use cases
- Case 01
Fair lending and hiring
Checking that models do not systematically disadvantage groups of people.
- Case 02
Safe assistants
Making chatbots refuse harmful requests and resist manipulation.
- Case 03
Explaining decisions
Telling someone why a loan was denied, as some laws require.