Ethics in AI

Week 7 - 12

Datafication

Abstraction of complex human systems into simplified and fallible measurements (data) to be legible and manipulable by technological system

Datafication Traits

• Key enabler automation/scalability

• Can destroy key information before it reaches a system and obscure it’s absence

• Can often shape systems to align with datafied abstraction (often to ill effect)

• Enabler of disparate impacts

• punish “failure to conform” to norm established by abstraction

Disparate Impacts

systems can wrongly neglect or abuse differences in measurable features of disadvantaged groups and wrongly punish “failure to conform” to norm established by abstraction

Automation

transformative changes to the world can happen automatically, either replacing human processes or introducing entirely new phenomena

Automation Traits

• Key enabler of scalability

• May increase or decrease opacity

• Can remove human oversight/insight from processes

• Often required for safety and success of systems in unanticipated ways

• Can lead to transformative disruptions of ecosystems, society, economy, culture

• Inexorably optimises for simplified/datified goals which might differ from social welfare or implicit goals of human predecessor

Scalability

Implementation of system on a large scale, which can aplify harms

Scalability Traits

• Problems can affect large populations before they are detected, or safety mechanisms can be established

• Feedback from effects becomes further detached/invisible

• The “work” required for scaling previously may have been fundamental to the success of scaled-up systems

• Disparate impacts can be amplified when system is widespread, e.g. credit scoring

• Massive impacts can be caused with little thought/effort

• Individuals and firms who own/create systems can quickly and undemocratically have immense power over people’s lives

Transformative Effects

Rapid large-scale automation can profoundly change social, cultural, ecological, and economic systems

• Interacting with technology as an intermediary or instead of others shapes how we relate to one another

• Automation can lead to deskilling where it replaces human labour, losing skills/knowledge of social and cultural value

• Rapid and profound impacts on labour market shift power towards elites able to directly profit from technology– faster= bigger whiplash

• Externalised costs/tragedy of commons associated with tech, e.g. impacts on environment and data privacy

Problems with Inscrutability/Opacity

• Designers and regulators may not understand well enough to identify fix problems

• Those impacted may not understand well enough to contest outcomes or consent to use

Inscrutability/Opacity

AI systems are often designed or used in ways that prevent stakeholders from understanding how they function

Mechanisms of Harm

• Datafication

• Automation

• Scalability

• Transformative Effects

• Inscrutability

When someone is accountable

If

• They are morally blameworthy or praiseworthy

• They are exposed or subject to to punishment or liable for restitution

• They can give an account or explanation for what they have done and its correctness (or what they are doing to correct it)

• They are responsible for ensuring outcomes or rectifying errors

Barriers to Accountability

1. The problem of many hands

2. Attitudes towards bugs

3. The computer as scapegoat

4. Ownership without liability

The problem of many hands

Barrier to accountability: “Where a mishap is the work of “many hands,” it may not be obvious who is to blame because frequently its most salient and immediate causal antecedents do not converge with its locus of decision-making

Attitudes towards bugs

“The view of bugs as inevitable hazards of programming implies that while harms and inconveniences caused by bugs are regrettable, they cannot–except in cases of obvious sloppiness–be helped.

In turn, this suggests that it is unreasonable to hold programmers, system engineers, and designers, to blame for imperfections in their systems.”

Computer as scapegoat

- Technology may appear performing roles that would historically be associated with the “accountable party”

- Presents target for others to evade responsibility, but cannot perform social activities of accountability

Ownership without liability

- Law struggling to adapt to changes discussed

- Result is decoupling of positive aspects of accountability (e.g. praiseworthiness and profitability) with negative (blameworthiness and liability)

Solutions and mitigations of accountability

• Explicit standard of care

• Distinguish accountability from liability

• Strict liability and producer responsibility

Justification of Opacity due to Corporate or State Secrecy

When/why might it be justified?

● Security/privacy risks

● IP protection

Cost balanced against for Opacity due to Corporate or State Secrecy

consumer/citizen rights to information/consent

Opacity due to technical illiteracy

Even with “full transparency”, I might not know what I’m looking at

Opacity due to fundamental characteristics of ML

Structure has multiple layers and dependencies which may not be easy for individuals to easily decipher

Creel’s 3 Levels

• Functional Transparency

• Structural Transparency

• Implementation Transparency

Functional Transparency

awareness of high level algorithm system is implementing

Structural Transparency

how that algorithm is realized in code

Implementation Transparency

how code runs on hardware + broader implementation details

Transparency vs. Accuracy

• Procedural fairness sometimes trumps “accuracy”

• “Accuracy” may not be a first order value

Types of Trust

Intrinsic, Extrinsic, Relationship, Values

Intrinsic Trust

Based on understanding of how decisions are made

Extrinsic Trust

Based on track record or testimony of other experts

Relationship Trust

personal experience with marker demonstrates trustworthiness

Value Trust

belief that individual is motivated your best interests & shares your values

Key issues in trusting AI

1. How vulnerable are you to its performance?

2. How clearly circumscribed is its domain of presumed competence?

3. How necessary is exercising discretion in order to meet expectations?

Problem of Induction

How informative is past performance on future trustworthiness?

How user gains intrinsic trust

reasoning process aligns with human reasoning → user observes reasoning process

How user gains extrinsic trust

external symptoms of model behavior are trustworthy → user observes model behavior

Explanation

answer to why question: how come? and/or what for?

• cognitive and social process

Explanation cognitive process

explanee coming to understand the answer to a why question by filling in gaps and considering counterfactuals

Explanation social process

explainer transferring knowledge and establishing mutual understanding with an explainee

Good explanation

selective, salient, and simple. They respond to what we don’t know, what we care about, what surprises us.

Contrastive explanations

explain why one outcome happened instead of another outcome. Rather than simply answering "Why did X happen?", they answer: "Why did X happen instead of Y?"

Causal Attribution

Process of identifying the cause or causes of an event, behavior, or outcome. What caused this event?

xAI Approach Properties

• Global vs. Local

• Intrinsic vs. post-hoc

• Model Agnostic vs. Model Specific

xAI Approach Foundational Models

• Attribution-based/ Saliency maps

• Example-based

• Extract rules

• Surrogate models

• Adversarial learning

xAI Global vs. Local

Understand the model's overall behavior vs. the local explanation, which focuses on a single prediction.

xAI Intrinsic vs. Post-hoc

xAI Property: Can easily follow the model logic and understand why a decision was made transparently vs. an explanation generated after a model has already made its prediction.

Model-Agnostic vs. Model-specific

xAI Property: Uses only inputs and outputs for explanation vs. inner workings of model used for xAI method

xAI attribution/saliency maps

xAI Foundational Model: show which parts of the input influenced a model's prediction

xAI example based

xAI Foundational Model: explains a prediction by showing similar examples from the data that influenced or justified the decision

xAI rule based

xAI Foundational Model: model’s decisions are explained using explicit “if–then” rules that describe how inputs lead to outputs.

xAI Surrogate models

xAI Foundational Model: Approximate global or local behavior with an explainable model to produce an explanation or replace an unexplainable system, e.g. linear approximation

xAI Adversarial learning

xAI Foundational Model: Clarify and expand conditions of applicability by trying to “trick” systems

Accessibility

technology is accessible if it can be used as effectively by people with disabilities as by those without

Universal Usability

“design for all” approach which is about making a product as accessible as possible to as wide a group of people as possible. The term originated from architecture

Equity

the quality of being fair and just, especially in a way that takes account of and seeks to address existing inequalities

Data-Driven Bias

training datasets

Algorithmic Bias

assumptions/simplifications (e.g. problematic feature selection); or just plain errors

Prejudice Bias

training data labelling/preprocessing, the design of the algorithm (including, e.g., non-diverse teams), or the operational environment.

Ways to get rid of bias in tech

• Debiasing, diversifying the datasets.

• Remove/mask group information such as gender (and other sensitive attributes) during the data processing.

• Reduce model bias

Ways to get rid of bias in non-tech

• Diverse development teams

• Bias Audits

• Regulation

Bias Audits

Regular audits of AI systems to check for biases in data, algorithms, and outcomes.

Factors in the digital divide

1. Access

2. Skills and Usage Patterns

3. Disparities in returns

Relationship between the spaces and biases

• Potential Space (PS): potential at birth

• Construct Space (CS): life experiences makes us realize abilities to potentially different degrees

• Observed Space (OS): realized abilities measured here

• Decision Space (DS): OS used on basis of predictions here

Original Position (OP)

John Rawls’ idea of justice is that the fairest rules are made when people decide them without knowing anything about their own identity or situation.

Put in the shoes of others

Data Governance

Represents the program used by an organization to manage the organizational bodies, policies, principles, and quality that will ensure access to accurate and risk-free data and information

Asymmetrical Relationship With Tech

Tech companies benefit off of us more than we do them

GDPR

It is the European Union’s (EU) comprehensive data privacy and security law. Enacted in 2018, it establishes strict rules on how organizations globally must collect, handle, and protect the personal data of individuals located within the EU and the European Economic Area.  

Any “organization that processes the personal data of people in the EU must comply with the GDPR.

GDPR Data Protection Principles

• Lawfulness, fairness and transparency

• Purpose limitation

• Data minimization

• Accuracy

• Storage limitation

• Integrity and confidentiality

• Accountability

DFPR Lawfulness, fairness and trainsparency

Processing must be lawful, fair and transparent to the data subject

GDPR Purpose Limitation

You must process data for the legitimate purpose specified especually to the data subject when you collected it

GDPR Data Minimization

You should collect and process only as much data as absolutely necesssary fort he purposes specified

Accuracy

You must keep personal data accurate and up to date

Storage Limitation

You may only store personally identifying data for as long as necesary for the specified purpose

Integrity and confidentiality

Processing must be done in such a way as to ensure appropriate security, integrity, and confidentiality

Accountability

The data controller is responsible for being able to demonstrate GDPR compliance with all of these principles

Data Commons

System where participants contribute voluntairly pooling their data for the benefit of the wider community. Data is usually shared with the public under an open license.

Data Cooperatives (DCs)

Generally understood as member-owned organizational structures. DCs are deemed a specific subcategory of platform cooperatives that primarily deal with democratic data management. Stewardship of data for the membership, commons focus “primarily for the benefit of the community at large.