The limits and biases of AI
Where an AI goes wrong and why: invented statements, data biases (historical, representation, measurement, selection), drift over time, opacity, adversarial examples, energy cost; the safeguards (human review, per-group measurement, monitoring in service) and the four risk levels of the European AI regulation.
- 26 cards
- 6 topics
- Free
The outline, topic by topic
Each topic is weighted the way the exam weighs it.
- Concepts
- Questions
- Sort and link
- The numbers
- Scenarios
- Understand
A few cards from the deck
Algorithmic bias: what is it?
Show the answer
A systematic and unfair gap in the results of an automated system, which disadvantages certain groups of people.
It comes mostly from the data (an unjust past, an underrepresented group, a badly chosen measure, an unrepresentative sample) and from what the system is asked to optimise, rarely from a programmer's intent.
In which case is human review of a decision proposed by a model a real safeguard?
- The person has the time, the full case file and the power to overrule the model
- The person approves dozens of proposals a minute with a click, without reading them
- The person only sees the final decision, without the case elements that led to it
- The person steps in after the decision, only to announce it to the person concerned
Show the answer
The person has the time, the full case file and the power to overrule the model
A check without time, information or the power to refuse is not a check: people defer to the machine out of habit. The check must also come before the decision takes effect.
Data that reflect unjust past decisions produce a … bias; if a group is almost absent from the examples, it is a … bias.
Show the answer
Data that reflect unjust past decisions produce a historical bias; if a group is almost absent from the examples, it is a representation bias.
In both cases the model learns its data faithfully: it reproduces a past we did not want to reproduce, or it makes more mistakes on the group it has seen little.
Data drift: what is it?
Show the answer
The change over time in what the system receives in service, which moves away from what it saw in training and degrades its results without a single line of code changing.
Behaviours, prices, fraudsters, language all change. It is detected by monitoring performance in service; the remedy is to retrain on recent data.