Ai Ethics And Fairness

AI ethics and fairness is the part of artificial intelligence that asks not just whether a system works but whether it works justly: whether its benefits and burdens fall fairly, whether it treats people with respect, and whether anyone can be held to account when it goes wrong.

12 frameworks, from 1970 to today · Artificial Intelligence · Computer Science

still a working view ended · each leads to its entry below

Overview

AI ethics and fairness is the part of artificial intelligence that asks not just whether a system works but whether it works justly: whether its benefits and burdens fall fairly, whether it treats people with respect, and whether anyone can be held to account when it goes wrong. The stakes are concrete. A hiring model that quietly down-ranks women's résumés, a medical triage tool that performs worse on darker-skinned patients, a predictive-policing system that sends patrols to poor neighbourhoods—these are not hypotheticals but documented failures. The field's history is the story of how researchers, engineers, philosophers, and eventually regulators came to see such failures as a technical and moral problem with its own methods, and of the competing ideas about what "fair" even means.

From Computer Ethics to Machine Morality

The oldest strand is computer ethics, which predates modern machine learning by decades1. Norbert Wiener's The Human Use of Human Beings (1950) warned that automating decisions would concentrate power and demanded that engineers consider consequences. In the 1970s and 1980s, Walter Maner and Deborah Johnson built a teaching tradition arguing that computing professionals face ethical dilemmas—privacy, accuracy, access—that general ethics did not address. This was a humanistic, professional-responsibility field: it asked what people using computers ought to do12.

A more speculative branch asked what machines ought to do32. In 1967 the philosopher Philippa Foot published her analysis of the "trolley problem," a thought experiment about choosing whom to harm. Decades later it became a staple of debates about self-driving cars, though its use there was often criticized as a distraction from the real statistical trade-offs of road safety. The idea of an explicit "machine ethics" took shape in the 2000s, with work by James Moor and others on whether autonomous systems could or should be given moral reasoning. This line remained largely philosophical; its lasting contribution was to make the question of whose values get encoded unavoidable.

The Fairness Problem Becomes Technical

The decisive shift came when machine learning spread into decisions about people. Through the 2000s, credit scoring, hiring, and criminal-justice risk assessment began using statistical models trained on historical data. Because those models learn from the past, they reproduce its patterns—including discrimination3. A landmark moment was the 2016 ProPublica investigation of the COMPAS recidivism tool, which found that its error rates differed by race: Black defendants were more often falsely flagged as high-risk, white defendants more often falsely cleared. The company disputed the analysis, and the dispute itself was instructive—it showed that "fair" could mean several incompatible things at once.

This crystallized into algorithmic fairness formalization4. Researchers proposed precise mathematical definitions. Demographic parity requires equal positive rates across groups. Equalized odds requires equal true- and false-positive rates. Individual fairness requires that similar people be treated similarly. In a widely cited 2016 paper, Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan—and independently Alexandra Chouldechova—proved an impossibility: except in degenerate cases, a classifier cannot satisfy several of these definitions simultaneously when base rates differ between groups. The result reframed the debate. Fairness was not a bug to be patched but a choice among values that had to be made explicitly.

Alongside this, bias measurement and mitigation developed as an engineering practice5. Researchers built datasets and metrics to detect disparate impact, and methods to correct it: reweighting training data, removing sensitive attributes, or adjusting predictions after the fact. A recurring lesson was that simply deleting race or gender from the inputs does not remove bias, because correlated features (zip code, name, language) act as proxies3. The field learned to measure outcomes, not intentions.

Privacy, Explanation, and the Turn to Principles

Two neighbouring technical concerns matured in the same period. Differential privacy, introduced by Cynthia Dwork and colleagues in 2006, gave a rigorous guarantee: a computation is differentially private if its output changes only negligibly whether or not any single person's data is included. This let organizations learn from populations while bounding what could be inferred about individuals—a fairness-adjacent promise about who bears the risk of data use6.

Explainable AI addressed a different gap7. As deep neural networks became accurate but opaque, users and regulators demanded reasons. Early work explained models by approximation—LIME (2016) fit simple local models around individual predictions; SHAP (2017) borrowed game-theoretic Shapley values to attribute a prediction to its inputs. These tools made black boxes legible but also raised a caution: an explanation can be plausible without being faithful to what the model actually computed.

By the mid-2010s, the field's centre of gravity moved to principles and guidelines8. The Asilomar AI Principles (2017), the OECD AI Principles (2019), and countless corporate and national charters converged on a familiar list: fairness, transparency, accountability, privacy, human oversight, and beneficence. This was a genuine achievement of consensus—but critics noted that principles are cheap to sign and hard to operationalize, and that the same companies writing them were deploying the systems at issue.

Responsible AI, Participation, and Governance

The response to "principle-washing" was to build responsible AI as an institutional practice: ethics review boards, impact assessments, model cards and datasheets documenting datasets and systems, and dedicated teams inside firms. It borrowed from older traditions of technology assessment and safety engineering, and it made ethics a matter of process rather than pronouncement.

A parallel movement insisted that ethics be participatory and community-centered. Rather than experts deciding values in advance, this approach involves affected communities in design and evaluation9. Its clearest expression is participatory design and the "nothing about us without us" principle carried over from disability rights. Related work on data sovereignty—for example, Indigenous data-governance frameworks—argued that communities should control data about themselves. This strand pushed back against the assumption that fairness could be settled by a formula alone.

Two further practices rounded out the toolkit. Adversarial testing and red-teaming, adapted from security research, deliberately probes systems for harmful or biased behaviour before deployment—stress-testing a model with edge cases, hostile inputs, or demographic probes. And risk-based AI governance turned ethics into law-adjacent regulation310. The European Union's AI Act, proposed in 2021 and adopted in 2024, classifies systems by risk level and imposes heavier obligations on high-risk uses such as hiring, credit, and policing. This regulatory turn descends directly from the fairness and accountability debates: it tries to make the abstract principles enforceable.

Where the Field Stands

Today the main approaches coexist rather than compete. Formal fairness definitions and bias-mitigation methods supply the measurement; explainability supplies the audit trail; differential privacy bounds data harm; responsible-AI processes and red-teaming supply the operational discipline; participatory methods supply legitimacy; and risk-based regulation supplies teeth. The open questions are the ones the history kept surfacing: which conception of fairness should govern a given decision, and who gets to choose; whether explanations and audits can keep pace with ever-larger models; how to extend accountability across complex supply chains of data and pretrained components; and whether governance can act fast enough to matter. The field began by asking what computers ought to do and ended by asking who decides—and that question remains live.

References

  1. Computer ethics (introduction, History), Wikipedia
  2. Machine ethics, Wikipedia
  3. Ethics of artificial intelligence (Machine ethics, Racial and gender biases, Regulation), Wikipedia
  4. Fairness (machine learning), Wikipedia
  5. Algorithmic bias, Wikipedia
  6. Differential privacy, Wikipedia
  7. Explainable artificial intelligence, Wikipedia
  8. Regulation of artificial intelligence (introduction, Background), Wikipedia
  9. Participatory design, Wikipedia
  10. Artificial Intelligence Act, Wikipedia

The frameworks of Ai Ethics And Fairness

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1970-2005
  2. 2000-Present
  3. 2000-Present
  4. 2006-Present
  5. 2010-Present
  6. 2016-Present
  7. 2016-Present
  8. 2016-Present
  9. 2018-Present
  10. 2018-Present
  11. 2018-Present
  12. 2021-Present
A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Ai Ethics And Fairness is a planet in Nousfarer

Its city has a tower for each of these 12 frameworks. Land, walk in, and climb one concept at a time.

Land on Ai Ethics And Fairness

Related fields

More in Artificial Intelligence

See where Ai Ethics And Fairness sits on the Map of Human Knowledge