11_ethical considerations in AI

Question 1

A survey of 84 AI ethics guidelines identified 11 common principles:

Answer 1

• Transparency
• Trust
• Non-maleficence
• freedom and autonomy
• solidarity
• privacy
• justice and fairness
• sustainability
• responsibility
• beneficence
• dignity

Question 2

What is Transparency referring to?

Answer 2

ability to understand an interpret the internal workings and decision-making processes
(helps increase trust in the model and identify potential biases or errors in its reasoning)
–> software and mathematical level

–> publishing source code and training datasets, disclosure of model-relevant information

Question 3

What does explainability refer to?

Answer 3

ability of a model or algorithm to provide explanations for its predictions or decisions
–> technical level

Question 4

What does interpretability refer to?

Answer 4

to understand and explain the internal workings and decision-making processes
–> domain level

does the model correctly use characteristic features to make its decision?

Question 5

How can explainability be achieved?

Answer 5

eg through CAMs (Class Activation Maps)

heatmap that indicates areas that supposedly contributed to the decision-making process

Question 6

Why do people think AIs are black boxes?

Answer 6

parameters might not be accessible

interpreting how the parameters affect the decision-making process is difficult and expensive

Question 7

Why are AIs not black boxes?

Answer 7

by default, they are fully transparent

–> based on (rather simple) mathematical operations and their entire knowledge is stored in its parameters

Question 8

What does fairness refer to?

Answer 8

outputs and decisions of a machine learning model should not be biased or discriminatory towards certain groups of individuals

Question 9

When are AI model predictions biased?

Answer 9

• the data used for training does not generalize well
• the model is not suited for the data or the task

Question 10

What does non-maleficence refer to?

Answer 10

principle of avoiding harm or adverse effects on individuals or society as a result of the use of ML algorithms and techniques

–> you have to properly design a model’s use case

Question 11

What does privacy refer to?

Answer 11

protection of individuals’ personal information and data from unauthorized access, use or disclosure

–> is possible with federated learning techniques

Question 12

What are federated learning techniques?

Answer 12

Central server (for hosting the model) and decentral servers (containing the data)

–> advantage: the central server never sees the actual data, therefore conserving privacy

Question 13

What does trust refer to?

Answer 13

confidence and reliance that individuals and stakeholders have in the accuracy, reliability and fairness of ML algorithms and technologies

–> can be built by following the other AI principles

Question 14

What are the 5 ethical AI guidelines we learned in class?

Answer 14

1) transparency
2) explainability and interpretability
3) justice and fairness
4) non-maleficence
5) privacy

Question 15

What are the 5 ethical AI guidelines we learned in class?

Answer 15

1) transparency
2) explainability and interpretability
3) justice and fairness
4) non-maleficence
5) privacy