Can AI apply AI ethics to AI?


 

The challenge is to convert public information about AI-intensive technologies and companies into numerical representations of their human impact. It is to design algorithms and data to understand professional and social media with the mind of an AI ethicist.

 

 

The ESG example / Frictionless AI ethics


 

Natural language processing is already being used to ethically evaluate companies for Environmental, Social, and Governance investing. ESG investors use machine learning to scour public records, news and social media, and then rate companies in nonfinancial terms. Typically, the metrics align with the United Nation’s Sustainable Development Goals. The companies studied largely belong to the industrial economy.

The humanist measurement of AI-intensive companies adapts the ML/NLP work, but applies it with different metrics. Instead of correlating with the UN Sustainable Development Goals, the measurements are derived from artificial intelligence ethics. Instead of industrial categories (e.g. carbon production, customer health and safety, wages and working conditions), the natural language processing analyzes and measures AI ethics: Uses of personally identifying information, demographics of training data employed for machine learning , mechanisms for user appeals of algorithmic decisions....

Frictionless AI

Frictionless AI ethics adapts ML/NLP techniques developed by ESG investors. It translates them for contemporary AI ethics, and applies them to today's AI-intensive companies. The result is a profile of a technology’s human impact.

 

 

Examples of ML/NLP applied to ethically evaluate industrial economy companies

 

 

Elements of AI Human Impact


 

The human – as opposed to financial – impact of an AI-intensive company is measured along 6 values developed in recent AI ethics, and derived from the European Commission’s Ethics Guidelines for Trustworthy AI. They are:

Self-determination

  • Does the AI conform to my projects, or nudge and manipulate?

Privacy

  • Do I control access to my own personal information, or is it hoarded and exploited for uses beyond my authority?

Fairness

  • Are all treated equally by the algorithms?

Society

  • Is overall social wellbeing promoted?

Performance

  • Is the AI accurate, efficient, reliable? Does it provide convenience and facilitate pleasure?

Accountability

  • Can AI decisions be explained? Can they be contested? Is someone responsible when the AI goes wrong?

 

Deep dive: Origin and meaning of the AIHI values  

 

Some examples of applied ethics in AI  

 

 

Keywords AI Ethics


 

Self-determination

Does the AI conform to my projects, or nudge and manipulate?

Keywords +
Autonomy, agency, human agency, human oversight, enable, empower, create, self-determine, freedom, informed

Keywords -
Nudge, coerce, deceive, addict, addiction, gamification, manipulate, exploit, exploitation, commodify, dopamine, filter bubble, echo chamber, dark pattern, behaviorism

 

 

Privacy

Do I control access to my own personal information, or is it exploited for uses beyond my authority?

Keywords +
Privacy, anonymization, pseudonymization, cyber-security, data-security, privacy-by-design, data minimization, encryption, logging data access, oversight mechanisms for data processing, Data Protection Impact Assessment (DPIA), Data Protection Officer (DPO), right to withdraw consent, right to object, right to be forgotten, data protection, consent, privacy policy, The American Institute of Certified Public Accountants’ (AICPA) Service Organization Controls (SOC) for Cybersecurity, ISACA’s COBIT 5, ISO/IEC 27000-series, National Institute of Standards and Technology’s (NIST) Framework for Improving Critical Infrastructure Cybersecurity

Keywords -
Data breach, hacked, ransomware,

 

Fairness

Are all treated equally?

Note
This is not a governance question about the workplace, but a question about how the technology works. Still, the governance/workplace ESG rating may serve as a proxy for the technology’s user-facing fairness. For example, an e-insurance or e-microlending company with a strong/weak workplace equal opportunity may be expected to share that rating with their own products.

Keywords +
Fair, equality, accessibility, universal design, stakeholder participation, principle of proportionality

Keywords -
Unfair, bias, biased training data, incompleteness

 

Society

Is social wellbeing promoted by the AI?

Note
This category already appears in multiple ESG ratings, so results may be simply transferred. Typically, the category aligns with the United Nations Sustainable Development Goals.

Keywords + (All UN SDGs)

  • No Poverty
  • Zero Hunger
  • Good Health and Well-being
  • Quality Education
  • Gender Equality
  • Clean Water and Sanitation
  • Affordable and Clean Energy
  • Decent Work and Economic Growth
  • Industry, Innovation and Infrastructure
  • Reducing Inequality
  • Sustainable Cities and Communities
  • Responsible Consumption and Production
  • Climate Action
  • Life Below Water
  • Life On Land

 

Performance

Is the AI accurate, efficient, reliable? Does it provide convenience and facilitate pleasure?

Note
This category could be approximated with a straight sentiment index surrounding the AI (e.g. “Tesla driverless car”). Especially useful would be sentiment in professional publications (e.g. What do doctors write about an AI analysis of cardiovascular health as is produced by the startup Cardisio).

Keywords +
Accuracy,  correct judgments, precision, correctly classify, proper categories, correct predictions, correct recommendations, correct decisions, reliability and reproducibility, reproducible, reliable, robust, safe, resilient to attack, security

Keywords -
Inaccurate, imprecise, error, unsafe, data poisoning. adversarial attacks. potential abuse of the system by malicious actors

 

Accountability

Can AI decisions be explained? Can they be contested? Is someone responsible when the AI goes wrong?

Keywords +
Explainability, explicability, transparent, auditable, auditability traceable, traceability, logging, documentation, contestability, redress, redress-by-design, glass box, simple classification, regression, explainable boosting, decision-tree, Lime, Local Interpretable Model-agnostic Explanations), SHAP, SHapley Additive exPlanations

Keywords -
Blackbox

 

 

 

Resources


 

Here