What is Artificial Intelligence Fairness (AI Fairness) in one article

Artificial intelligence fairness: definitions and core connotations

AI fairness is the interdisciplinary field of ensuring that AI systems treat all individuals and groups of people in a fair and unbiased manner throughout their design, development, deployment, and operation lifecycle. The core goal is to prevent AI systems from producing discriminatory or biased decisions and outputs for specific groups of people based on sensitive characteristics such as race, gender, age, religious beliefs, socioeconomic status, and so on. Not only a technical indicator, but also a social commitment and ethical requirement.

Imagine an AI tool for screening resumes whose training data comes primarily from the historical hiring records of one gender, it is likely to inadvertently learn to devalue the resumes of job seekers of the other gender, leading to unfair hiring outcomes. AI fairness is about identifying, quantifying, and eliminating such problems.

AI fairness requires us to go beyond code and algorithms and take a deeper look at the historical biases that may be embedded in the data itself and how algorithmic models may amplify these biases. AI fairness seeks to establish a technological vision of AI systems that serve the diversity of society and promote equal opportunity, not solidify or exacerbate existing social inequalities. Understanding AI fairness is a critical first step in understanding how to responsibly create and utilize future technologies.

Artificial intelligence fairness: core concepts and multidimensional understanding

AI fairness is very rich and can be understood in terms of multiple interrelated and slightly differentiated dimensions that together form a three-dimensional framework for assessing the fairness of AI systems.

• Group Fairness:is the most intuitive concept of fairness. Group fairness requires that AI systems treat different protected groups (e.g., men vs. women, different racial groups) equally. Statistically, this is reflected in a balanced distribution of key metrics across groups, such as ensuring that loan approval rates, facial recognition accuracy, or error rates in crime risk prediction remain roughly the same across groups.
• Individual Fairness:It is emphasized that similar individuals should be treated similarly in the eyes of AI. Individual fairness is concerned with micro-level justice. For example, two job applicants with almost identical qualifications, experience, and backgrounds should be rated similarly in an AI hiring system, regardless of gender or race. Individual fairness requires that algorithms learn to ignore sensitive features that are not relevant.
• Anti-Classification (Anti-Classification):is a stricter standard that requires AI models to completely refrain from using sensitive features (e.g., race, gender) in the decision-making process. The idea is to cut off the possibility of discrimination at the source and ensure that decisions are completely unrelated to these characteristics.
• Equalized Opportunity:It is a more refined and practical concept of equity. Equalizing opportunities does not require identical outcomes, but rather equal "chances". Taking recruitment as an example, it is required that among all the truly qualified candidates, the proportion of those successfully selected by the AI is equal among different groups. This means that the model's ability to recognize "real talent" is equally accurate across groups.
• Counterfactual Fairness:is a more cutting-edge concept, anti-causal discrimination draws on causal reasoning to think about fairness. The question is: if a sensitive characteristic (e.g., gender) of an individual changes, but everything else remains the same, does the AI's decision change? If the decision changes with it, then there is causal discrimination based on that characteristic. This way of thinking helps to reveal more hidden biases.
• Procedural Fairness:Focus on the fairness of the decision-making process, not just the outcome. Procedural fairness requires that the decision-making process of an AI system be transparent, explainable, and open to challenge and appeal. Even if the results appear statistically fair, it is difficult to trust the fairness of the process if it is an unintelligible "black box".

Artificial intelligence fairness: urgent social importance

The promotion and realization of fairness in AI is far from being a technician's paper exercise, but has far-reaching and urgent social significance that relates to the roots of technological development - trust and justice.

• Prevent the automation and amplification of historical bias:There are historical inequalities in many areas of society, and these inequalities are recorded in the form of data. If AI learns this data uncritically, it will automatically inherit and amplify the biases, allowing discrimination to repeat itself in a more efficient and insidious way, creating a vicious cycle of "garbage in, garbage out".
• Upholding fundamental human rights and social justice:Fair treatment is a fundamental right of every human being. The use of biased AI systems in areas such as health care, justice, credit, employment and other major life decisions can directly harm the vital interests of specific groups, deprive them of the opportunities they deserve, and exacerbate social cleavages and antagonisms.
• Building public trust in AI:Trust is the cornerstone for technology to be widely accepted by society. If AI systems are repeatedly exposed as discriminatory, the public will become fearful and resistant to them, ultimately hindering innovation and application in the field as a whole. Fairness is necessary to win public trust.
• Enhance business benefits and brand value:A fair AI system helps organizations make better, more comprehensive decisions, tap into a broader talent pool and customer base, and avoid brand image damage from discriminatory scandals. Practicing fair ethics is a sustainable business strategy in the long run.
• Promoting Technology for Good (TFG):Artificial intelligence has been touted as a central driver of the fourth industrial revolution. Ensuring its fairness ensures that the results of this revolution are inclusive and can truly be used to bridge social divides and enhance human well-being, rather than creating new privileged classes and oppressed groups.

Artificial intelligence fairness: a serious challenge in reality

Translating the ideal of fairness into reality faces a complex and intertwined set of challenges, many of which are rooted in the gray areas at the intersection of technology and society.

• Plurality and Conflict in the Definition of Fairness:There is no single "correct" mathematical definition of fairness. As mentioned earlier, the criteria of group fairness versus individual fairness, and anti-classification versus equalization of opportunity, are often in conflict with each other. Satisfying one criterion may necessarily violate another, forcing developers to make difficult value tradeoffs based on specific scenarios.
• Historical bias traps in data:Data is the food of AI, but historical data often reflects historical injustices. For example, the tech industry has been dominated by men in the past, resulting in resume data that contains far more men than women. Train a model on this data and it will assume that men are more likely to be "good programmers" and will score women's resumes lower. Cleaning and correcting for bias in the data is a huge challenge.
• The amplification effect of the algorithm itself:Even if the bias in the data is not obvious, complex algorithmic models may detect and amplify certain spurious patterns associated with sensitive features (Proxy Features) in the process of learning and generalization, which can produce unexpected discriminatory results.
• Fairness versus performance trade-off (Trade-off):In many cases, forcing a model to be constrained so that it meets some fairness metric may reduce its overall predictive accuracy to some extent. For example, criteria may have to be relaxed in order to increase the rate of minority loan approvals, thereby increasing the overall risk of loan defaults. How to balance "equity" and "efficiency" is a classic decision-making dilemma.
• Intersectionality (intersectionality) of multiple sensitive attributes:In reality, a person may belong to multiple disadvantaged groups at the same time (e.g., a low-income African-American woman). Prejudice has compounding effects at the "intersection" of these identities, which become more complex and difficult to measure. Equity measures optimized for a single group may not address intersectional discrimination.
• Interpretive and accountability deficits:Many state-of-the-art AI models (e.g., deep neural networks) are complex "black boxes" whose internal decision logic is difficult to interpret. When unfair results occur, it is difficult to trace the root cause of the problem back to the data, the algorithm, or some point in the deployment of the model, making accountability and fixes extremely difficult.

Artificial Intelligence Fairness: implementation methods and technical practices

Despite the magnitude of the challenge, researchers and engineers have developed a range of technical means to inject fairness considerations throughout the machine learning pipeline (Pipeline).

• Pre-processing method (Pre-processing):Processing the data before it enters the model. This includes: identifying and correcting for bias in the training data; resampling data from underrepresented populations; or stripping data of associations with sensitive features while retaining as much other useful information as possible through data transformation techniques.
• In-processing:This is a direct integration of fairness constraints into the model training algorithm itself. When designing the objective function, the developer not only requires accurate prediction, but also adds fairness as one of the optimization goals, so that the model can actively "learn" fairness in the learning process.
• Post-processing method (Post-processing):After the model is trained, its output is adjusted. For example, different decision thresholds (Threshold) are set for different groups. In a lending scenario, the approval threshold can be appropriately lowered for historically disadvantaged groups to achieve equal opportunity. The advantage of this approach is that there is no need to retrain the model.
• Regular audits and continuous monitoring:Fairness is not a "one and done" setup. Once a model is live, an ongoing monitoring system needs to be put in place, with regular fairness audits of its outputs using the latest real-world data, to ensure that its performance has not degraded or developed new biases over time.
• Use the Fairness Toolkit and Framework:Several open source toolkits have emerged in the industry to assist in achieving fairness, such as IBM's AIF360, Microsoft's FairLearn, and Google's What-If Tool. These tools provide harmonized metrics and mitigation algorithms that significantly lower the barriers to practice.
• Promote diverse teamwork:Technical solutions cannot be separated from the human element. Assembling development teams that are more diverse in terms of gender, ethnicity, culture, and professional backgrounds can help identify potential bias blind spots at the beginning of the design process and prevent fairness issues at the source.

Artificial intelligence fairness: a wide range of application scenarios and examples

The issue of AI fairness is far from a theoretical pipe dream; it has literally appeared in all corners of our lives, and every case warns of its importance.

• Recruitment and Human Resource Management:Amazon developed an internal AI resume screening tool that learned to penalize candidates whose resumes included the word "female" (e.g., "captain of the women's chess club") because most of the training data came from male resumes, leading to systemic discrimination against This led to systematic discrimination against female applicants. The company eventually abandoned the program.
• Criminal Justice and Risk Assessment:Recidivism risk assessment systems such as COMPAS, used by some courts in the United States, have been investigated by media outlets such as ProPublica, which found that they systematically overestimate recidivism risk for black defendants while tending to underestimate it for white defendants, sparking a major global debate about algorithmic justice.
• Financial services and credit approval:AI credit models that are primarily trained using historical loan data, and that have historically "redlined" certain zip code areas (often minority neighborhoods) from being denied service by financial institutions, will learn to associate those zip codes with "high risk" and deny loan applications to residents of those areas, thus creating "redlining" in the digital age. The model then learns to associate those zip codes with "high risk" and rejects loan applications from residents of those areas, thus creating "redlining" in the digital age.
• Healthcare and Diagnostic Aids:If the image data used to train the skin cancer diagnostic AI is overwhelmingly light-skinned populations, the model's diagnostic accuracy for darker-skinned patients is significantly reduced, potentially leading to misdiagnosis or delays in treatment, resulting in serious health inequities.
• Face Recognition and Surveillance Technology:Several academic studies and the MIT Gender Shades project have shown that many commercial facial recognition systems have significantly higher error rates in recognizing dark-skinned women than light-skinned men. Misuse of this technology by law enforcement could lead to erroneous identifications and serious consequences.
• Content recommendation and information cocooning:Although social media and news push algorithms do not directly make "decisions," their recommendation mechanisms based on user engagement may prioritize biased, false, but eye-catching content, thus creating a cocoon of information that reinforces bias for groups with different political positions and cultural backgrounds, and affecting the formation of social consensus. The formation of social consensus will be affected.

Artificial intelligence fairness: an indispensable ethical and social dimension

AI fairness is essentially a Sociotechnical System issue, with ethical choices at its core that profoundly affect the structure of society.

• Value-Sensitive Design (VSD):Require technology designers to make human values (e.g., fairness, privacy, autonomy) a core design consideration at the earliest conceptual stage, rather than an afterthought. Technology is not value-neutral; it is embedded in the designer's choices and preferences from its inception.
• Algorithmic Accountability (AA):There must be a clear chain of accountability when AI systems make unfair decisions. It needs to be clear who is responsible - is it the developer, the deployer, the company or the regulator? Establishing accountability is a key mechanism to force relevant parties to focus on fairness.
• Digital Divide and Empowerment:Inequity is not only found within algorithms, but also in the access and use of technology. Ensuring that all groups, regardless of socio-economic background, have access to, understand and benefit from AI technologies is a topic of fairness at a more macro level.
• Public participation and democratic deliberation:The discussion of "what is fair" should not be confined to engineers and companies, but should include the participation of philosophers, sociologists, legal experts and members of the public in the communities that may be affected. This is a public issue that requires democratic deliberation by society as a whole.
• Technology for Good Business Ethics:Rather than just "do no evil" as a bottom line, companies should actively practice the positive ethics of "Tech for Good". Putting ethical principles such as fairness ahead of short-term profits is the only way for tech companies to win long-term trust and social license.
• Global and Cultural Perspectives:The definition of fairness is culturally relative. What is considered fair in one culture may not be accepted in another. AI systems developing global applications must take this diversity into account and avoid Techno-colonialism.

Fairness in artificial intelligence: a progressive legal and policy framework

Lawmakers and regulators around the globe are moving quickly to try to draw red lines and build a governance framework for the fair development of AI.

• The EU's Artificial Intelligence Act (AI Act):This groundbreaking legislation takes a risk-based approach to regulation, categorizing AI systems into different risk levels. It prohibits the use of AI systems that pose an "unacceptable risk" to human safety, livelihoods and rights (e.g., social scoring) and imposes strict obligations on "high-risk" AI systems (used for recruitment, critical infrastructure, etc.), including data governance, transparency, human oversight and rigorous compliance assessments, including fairness requirements.
• Legislative and Executive Action in the United States:The U.S. currently has no comprehensive federal AI law, but promotes AI fairness through departmental regulations (e.g., Federal Trade Commission FTC enforcement), state legislation (e.g., Illinois' Artificial Intelligence Video Interviewing Act), and presidential executive orders (Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence) that emphasize the right to protect citizens from algorithmic discrimination.
• Algorithmic governance and regulation in China:The Regulations on the Administration of Algorithmic Recommendations for Internet Information Services and the Interim Measures for the Administration of Generative Artificial Intelligence Services issued by China's State Internet Information Office and other departments clearly require that providers of algorithmic recommendation services should adhere to the principles of fairness, equity and openness and transparency, prohibit price discrimination and other unreasonable treatment based on user characteristics, and establish and improve the mechanisms used to identify and correct bias and discrimination.
• Mandatory impact assessment:Similar to environmental impact assessments, many regulatory trends require fundamental rights impact assessments or algorithmic impact assessments for high-risk AI systems, forcing companies to systematically assess the potential impacts of their AI systems on fairness, privacy, and other rights, and to make the results publicly available.
• "Safe Harbor" and Sandbox Mechanisms:Some policy frameworks balance innovation and regulation by providing "safe harbors" for firms that proactively adopt compliance best practices and proactively conduct self-testing and evaluation, or by allowing them to test innovative products in a regulatory sandbox.

Artificial intelligence fairness: future research directions and trends

The field of AI fairness is still growing at a rapid pace, and researchers are exploring deeper and more effective solutions to future challenges.

• The deepening of artificial intelligence (XAI) can be explained:Promote models that are no longer "black boxes" but can provide clear and understandable rationales for decisions. When decisions are explainable, it is easier for auditors and companies to identify and correct specific paths that lead to unfairness.
• Frontier Explorations in Causal Fairness:The fairness framework based on Causal Inference (CI) is a current research hotspot. It tries to go beyond statistical correlation to understand the causal relationship between sensitive features and decision outcomes, so as to develop a more scientific and fundamental de-biasing strategy.
• Fairness Audit of Large Language Models (LLMs):With the popularity of generative AI such as ChatGPT, it has become a whole new challenge to audit social bias in these behemoths. Studying how to detect and mitigate gender stereotypes, cultural biases, and harmful content that LLMs may produce in their output is extremely urgent.
• Dynamics and long-term equity:Decisions made by AI systems change the user's future state, which in turn affects subsequent decision data. Studying the fairness impact of AI systems in long-term, multi-round interactions and how to avoid the "Matthew Effect" is a complex future direction.
• Equity in Federal Learning:In federated learning scenarios where data are not localized and joint training is performed by exchanging model parameters, it is a challenging technical problem to ensure the fairness of the final global model while protecting the privacy of each data source.
• Standardization and cross-cutting cooperation:Promote the standardization of AI fairness assessment indicators, processes and tools for comparison among different systems and industries. At the same time, strengthen in-depth cross-collaboration among computer science, law, ethics, sociology and other fields to jointly overcome this major challenge.