They then presented "Argumentation for Explainable AI (XAI", which have 2 types (intrinstic and post-hoc), with 2 examples regarding bias detection and Dutch Police to prove the feasibility.

More lablings and situations were introduced, such as self-attack statement, and other kinds of labelings

The process of formalising argumentaion was introduced, including the definition of abstract argumentation, and how statements with 3 labels (IN, OUT, UNDEC) computationally form the argumentation, explained based on statements as nodes and relations as graphs

They then introduced two terms: 1) Definition of argumentation, and 2) Internal reasoning, which lead to the talk of the commonsense reasoning: defeasible

The tutorial start with two small interactions with audience: audience are asked to talk about their definition of "argumentation", and there is a argumentation topic "Should human cloning be illegal?" selected by spinning wheel. Interesting!

Finally the tutorial presented several scenarios for audience to interact with, such as hiring and loan application.

Amazing!

#FACCT2025

Discussions:

How should decision-makers address multiplicity?

Domain-Specifc Considerations:
Factual vs Normative Decisions, Single-Shot vs Multi-Shot Settings, Low vs High Aleatoric Uncertainty, Comprative & Individual Decisions

implications of multiplicity for explainability: explainability concern (multiple explanations, interpretability) => multiplicity intervention (combining models, secondary criteria to choose models

Implications of multiplicity for fairness: different concerns (conflicting outcomes, outcome homogenization, group fairness) => multiplicity intervention (combining models, randomizing among models, secondary criteria to choose models)

They used a vivid example regarding job application to explain the harmful outcome of homogenization, thus introducing ways to reduce the homogenization without reducing accuracy.

Another presenter introduced "Predictive Multiplicity" as the ability of a prediction problem to admit competing models that assign conflicting predictions, including for Binary Classification, for Probabilistic Classifcation, Allocation Multiplicity, Procedural Multiplicity, etc.

Rashomon Effect: A combination of a difference of perspective and equally plausible accounts, with the absense of evidence to elevate one above others.

In the context of AI, it often refers to a multitude of different descriptions in a class of funtions giving about the same minimum error rate.

USING the register - litigation, as well as advocacy and public affairs.

#FACCT2025

After methodologies, they presented different categories of information for the register. cmsassets.ombo.nsw.gov.au/assels/Ropor...

Methodologies:

1. Surveys

2. Publicly available reports and websites

3. Freedom of Information requests

4. Case studies

ADM system: a fully or partially automated technical system, used by state government organisation (state government department or agency, or local council), in administrative decision-making, and that affects people.

Few definitions:

Automatic decisions: one in which an automated system performs at least part of the decision-making process.
Automated systems can enter into government decision-making in a range of different ways;
• Decision support
• Streaming or triage
• Fully (or solely) automated

3. To set an example of what good transparency looks like when it comes to government use of ADM, and improve compliance with governments own transparency register.

Why a transparency register?

1. To better understand the overall state of government automation and the key risks and challenges.

2. To make real a right that PLP belleve people should have: a right to easy-to-understand information about decisions affecting them.