Fairness, Accountability, Principlism
Learning Outcomes Module 4
Copyright By PowCoder代写 加微信 powcoder
At the end of this module, you should be able to:
• Explain framework of Principlism in AI ethics
• Explain the concepts of fairness and accountability in relation to AI
• Intelligently apply the concepts of fairness and accountability to cases involving AI
What questions do you have on anything so far in the subject??
https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing 4
Correctional Offender Management Profiling for Alternative Sanctions Northpointe (now equivant – tagline: “Software for justice”) Recidivism algorithm
• Risk score for reoffending (‘recidivating’) after initial arrest
• Violent crimes, nonviolent crimes
• Blacks and whites involved
• Guides officers in determining bail
• May reduce rates of detention (allow bail)
• Factors may include: current charges, pending charges, prior arrest history, previous pretrial failure, residential stability, employment status, community ties, substance abuse, age
• NOT race – ‘protected attribute’ like e.g. gender
• Actual algorithms are trade secrets
• Used actual re-arrest rates to determine actual offending post-COMPAS
• Biased against blacks
• Black nonoffenders higher risk scores than white nonoffenders
• Disparate impact
ProPublica
• Equal accuracy (61%) for both whites and blacks – true
• But – the wrong predictions (39%) went wrong in different ways
• Not disparate treatment – COMPAS fair
• No explicit use of arbitrary differences (race)
• Yes, can be proxies for race (e.g. suburb)…but…
• Same accuracy for each race
• Reoffending rate for blacks and whites equal at each COMPAS scale (‘calibration’)
• E.g. At risk factor 3, blacks and whites had same rate of actual reoffending
Other replies
• Better than biased humans
• Historically: inconsistent, gut instinct
• One study (Dressel et al): COMPAS more accurate than individuals lacking criminal justice expertise and slightly less accurate than groups of individuals
• Decreased accuracy: more crime?
• Problem identified by ProPublica is less bad than e.g. poor ‘calibration’
(could violate discrimination laws)
• How do we decide what is fair?
Is algorithmic (un)fairness purely a technical problem?
Who else apart from ‘AI experts’ might be required to help solve it?
Principlism
• Distill theories into handy principles for AI ethics
• Midlevel principles: b/w theory and detailed rules
• Theory can guide these broad principles
• Derived from medical ethics principles
Principlism:
4 principles +1
1. Non-maleficence – do no harm
• Predict harm, avoid causing harm, minimize harm, short and long term
2. Beneficence – do good
• Anticipate good outcomes, short and long term
3. Respect autonomy – respect people’s values, choices, life plans
• Understand what others’ value, don’t override their choices, be honest
Principlism
4. Justice – fairness
• Distribute benefits and harms fairly, fair processes, don’t unfairly discriminate
4+1. Explicability – transparency and accountability (Floridi*)
• Complements the 4 principles
• Ensure those potentially impacted have sufficient understanding of the AI and that relevant people are held to account
Principles: need to be balanced against one another; all are ‘equal’
*Floridi, Luciano, et al. “AI4People—an ethical framework for a good AI society: opportunities, risks, principles, and recommendations.” Minds and Machines 28.4 (2018): 689-707.
Explicability
• ‘Democratizing Algorithmic Fairness’ (2020) by Pak-Hang Wong • Fairness
• Accountability
Shin, D., & Park, Y. J. (2019). Role of fairness, accountability, and transparency in algorithmic affordance. Computers in Human Behavior, 98, 277-284.
• • • • • • • •
Justice and Fairness
Justice-fairness often used interchangeably
High stakes decisions – AI Bail
Sentencing
Catching criminals
Job applications
Assigning grades
Censoring or generating misinformation Diagnosing illness
Justice/fairness
Broad definition: ‘Giving each their due’ or ‘what they are owed’. Treat similar cases similarly; ‘blind’ to arbitrary differences
Equal regardless of race, religion, class, sex, gender, sexual orientation, etc. ‘Equal treatment/consideration/respect’
• U – all similar interests are counted equal in calculus
• D – duty of justice
• Kant – Dignity equal; 2nd C.I.: always ends, never merely means
• always ends, never merely means
• VE – justice: praiseworthy trait
• CE – justice in care
Distributive justice
Resources, opportunities Need, merit, contracts etc.
Should everyone be given the same career opportunities regardless of talent? Income?
Should we help people suffer bad luck?
Should we favour people who are morally responsible rather than selfish?
Positive/reverse discrimination Affirmative action
More resources/opportunities/advantages to disadvantaged and historically oppressed groups
Procedural justice/fairness
• Fair procedure or process allocate benefits or harms
• E.g. only 10 spaces at uni and 15 equally qualified candidates
• Random, queue
• Social psychologists: many people care more about being treated fairly by
institutions than about actual outcomes
• Pure procedural justice: no question of need/merit etc. E.g. one lollipop and two best friends – who gets it?
Other types
Retributive justice
Impose penalty due to wrongdoing
Based on actual guilt and fair procedure (trial)
Reparative justice
Remediation for unfair treatment
Accountability
Assuming accountability (responsibility) E.g. holding myself to fair procedures
Being held accountable (responsible)
By external pressures and mechanisms
E.g. law, profession, codes of practice, colleagues, elections
Who is accountable? E.g. researchers, engineers, organizations (private and public), deployers, authorities
What mechanisms are the right and fair ones?
Algorithmic Fairness
ML algorithms can have embedded bias – unfair
E.g. Discriminate against groups unfairly e.g. race, gender
Either explicitly or by prox
Technical solutions to minimize unfairness E.g. change inputs
E.g. improve processing of dataset
E.g. change weighting of false –ves vs. +ves
Mathematical measures e.g. (Berk et al 2018) 1. overall accuracy equality
2. statistical parity
3. conditional procedure accuracy equality
4. conditional use accuracy equality 5. treatment equality
6. total fairness (1-5 achieved)
Ethical vs. technical problem: Whether mathematical fairness really is fair depends on the standard of fairness adopted. And this standard is disputed.
Hence: ethical question and debate
Also…Cannot always have perfectly fair algorithms due to:
The Impossibility Theorem: “mathematically impossible for an algorithm to simultaneously satisfy different popular fairness measures”
E.g. group parity that unfairly punishes group X who broke the law less often àsame treatment and disparate impact
The Inherent Tradeoff: between fairness and performance
E.g. Increased group fairnessàdecreased accuracy of recidivism prediction for bail
Decrease false positives (defendants falsely scored as high risk) but increase
false negatives (miss some high risk defendants) (social cost) 27
Ethical frameworks and fairness/justice
Utilitarianism: Fair/just = maximizes net wellbeing, even if some individual’s must be made worse off than others. Everyone’s similar interests are still considered equal
Kant’s deontology: Recognise human dignity and respect autonomy. Treat autonomous agents always as ends, never merely as means
Virtue ethics: Consider what a fair and just person would do
Ethics of care: Consider special relationships and roles and responsibilities
that flow from them. Consider impacts on the most vulnerable
All these frameworks recognize basic human equality
They may have different/similar views on algorithms that:
Reinforce existing disadvantage e.g. increasing policing for some groups
Overlook past oppression e.g. that affect algorithmic prediction that results
in biases against disadvantaged groups or don’t positively discriminate to
help disadvantaged groups
Pak-Hang Wong (reading)
• AI may necessarily create winners and losers – harms and benefits for different people.
• Determining what is fair in high stakes AI is not purely a technical task, but an ethical one
• But: how to ensure this determination is itself fair?
• Recall: no consensus about what is fair + perfect algorithmic fairness is
impossible
• Wong: procedural justice
• What mechanism is fairest for holding AI designers and owners accountable?
• E.g. panel of AI ethics experts?
• Political mechanism
Accountability for reasonableness (AFR)
• AFR developed from health ethics
• Wong: “ensure decisions are morally and politically acceptable to those affected by algorithms through inclusion and accommodation of their views and voices”
• AFR assumes no totally final and ‘right’ answer: answers emerge through open, democratic, good-faith dialogue and reason-giving involving stakeholders
• Not just developers and researchers determining what is fair AI
• Four conditions
Four conditions for AFR
1. Publicity condition: Decisions about algorithmic fairness and their rationales must be publicly accessible, transparent, and understandable to non-technical people.
2. Full Acceptability condition: Provide a reasonable explanation of the chosen fairness parameters i.e. give evidence, principles, reasons that fair-minded persons could accept – for all affected stakeholders, especially the vulnerable.
3. Revision and appeals condition: Have ongoing (not one-off) mechanisms for challenge and dispute resolution and revision of policies.
4. Regulative condition: Have ongoing public regulation of process to ensure conditions (1)–(3) are met.
Northpointe’s COMPAS – recidivism prediction AI
1. Publicity condition: Explain clearly what measures of fairness will be used to predict re-offending
2. Full Acceptability condition: Justify why the chosen parameters and impacts are relevant. Could those impacted accept these reasons? E.g. allowing ‘disparate impact’ on historically disadvantaged black people while ‘avoiding disparate treatment’? Using education level or being the victim of crime – that negatively affects certain racial groups more?
3. Revision and appeals condition: Mechanism for those impacted to contest those reasons e.g. vulnerable groups, representatives of broader society
4. Regulative condition: Media put spotlight on COMPAS, but no stronger regulation or enforcement
Northpointe did not follow the AFR approach and were not held accountable by public regulation
A final point: should AI be used at all for this purpose? 32
Other accountability mechanisms?
• Committees
• ‘Turing’ stamps – suitable body
• Open-source software – ‘crowd’
程序代写 CS代考 加微信: powcoder QQ: 1823890830 Email: powcoder@163.com