1/21
Looks like no tags are added yet.
AI governance
How we steer AI
EU Model
Risk based regulation. Tiered system from high risk, transparency risk, general purpose risk, and minimal risk. Has red lines and prohibited AI. Its goal is to regulate before harm can occur. It can be hard to make changes to official regulations in the quickly changing AI space.
US Model
No comprehensive AI law. The US relies on executive action like executive orders and state laws after harm occurs. This could be beneficial in a time where AI development is moving very quickly, the US model is very flexible.
Rational choice
Actors have a purpose and pick the option that best suits their goals using cost benefit analysis. It assumes they fully understand the problem, weigh the alternatives, and evaluate the consequences.
Game theory
Actors maximize their own preferences based on what others will do. This is called strategic interdependence.
Prisoner’s dilemma
A paradox where two individuals acting in their own self interest do not produce the optimal outcome. Cooperation cannot immediately emerge; it requires repeated interaction. This is important for AI because frequent international discussions like summits would improve cooperation.
Bounded rationality
Decisions are limited by information, cognitive capacity, and time. Not the same as irrational. The goal is rationality; it just isn't always achieved.
Satisficing
Stopping the search once an option clears an accessibility threshold. This doesn’t take all the consequences into consideration.
Good AI
No single definition. It is a bundle of different types of goods like responsibility, fairness, safety, etc… Always ask good as what, good for what, and good for whom? All the different goods do not immediately fit together; there might be some trade-offs required. Could be very accurate but violates privacy, could be very efficient but biased.
4Ps
Must ask Purpose (goal), Protection (what must be protected), People (who carries the benefit and the cost), Power (who controls it and who answers for it).
Institutionalism
The concept that institutions shape decisions. Rules define authority, roles define what officials are expected to do, and routines make some responses feel normal.
Path Dependence
Early choices create increasing returns, so continuing gets easier as switching becomes costlier. For example in our simulations, piloting something sets new norms and it will be hard to turn back from this once it gets set up.
Principal Agent Theory
Principal delegates the task. Agent completes the task. When there are several principals, none fully control the agent, and many issues can arise from this. This problem matters because policy drift, oversight gaps, misleading incentives, blurred accountability, can occur.
Prospect theory
People judge outcomes as gains or losses relative to a reference point.
Loss aversion
losses feel about twice as strong as equal gains.
Risk attitudes
people are risk-averse when facing gains and risk-seeking when facing
Framing
describing the same choice as a gain or a loss changes the decision.
Dual-Process Theory
There are two modes of thinking. System 1 is fast, intuitive, automatic, and also more likely to get things wrong. Kind of like a gut reaction. System 2 is slow, reflective, deliberative, and more reliable for complex decisions.
Identify based approaches
decisions follow from who the actor is. This is the logic of appropriateness.
Rule Based Programming
intelligence as symbols, logic, and explicit rules written by humans.
Machine Learning
The system learns patterns from training data instead of humans writing every rule.
Evidence informed policy making
The same evidence about AI does not produce the same decision because actors have different goals, incentives, institutional positions, values, and risks. Can explain why the EU and US differ.