Building Fairer AI Systems with Zero-Knowledge Proofs: Still a Fantasy?
Is AI Fair?
If we hadn’t asked this question today, most ordinary users likely wouldn’t have even considered it, subconsciously believing that AI is fair. After all, in the minds of many, AI doesn’t have “emotions,” so naturally it wouldn’t show favoritism.
However, let’s not forget: today’s AI operates based on large-scale data-driven learning models. In real-world applications, these models are influenced by the biases present in their training data, which can lead to unfair outcomes. To address this issue, researchers are now exploring the use of Zero-Knowledge Proofs (ZKPs) to verify the fairness of AI systems — while still protecting sensitive model information.
A September 2024 study from Imperial College London indicated that ZKPs can help companies verify whether their machine learning (ML) models treat all demographic groups equally, while preserving the privacy of the model’s structure and user data.
Bias and Unfairness in AI Learning
Artificial Intelligence (AI) systems may introduce or amplify human social biases during learning and decision-making. These biases stem mainly from imbalanced training data, flaws in algorithm design, and developers’ unconscious biases.
For example, Amazon once developed an AI recruitment tool to automatically screen resumes. However, because its training data was based on male-dominated job applications from the previous decade, the model favored male candidates and even penalized resumes containing the word “women’s” or those from women’s colleges. Eventually, Amazon abandoned the project after failing to eliminate gender bias.
Similar issues have emerged in the healthcare sector. Studies found that certain AI algorithms predicting medical needs perform inconsistently across races, leading to some patients being underserved. In some cases, AI models analyzing medical images could accurately identify a patient’s race — but researchers couldn’t explain how — which has raised concerns that AI may exacerbate racial disparities in healthcare decisions.
These examples show that AI systems are not inherently fair. Their decisions can be affected by the biases of both their training data and their designers. Therefore, rigorous audits and ongoing oversight are crucial to ensure the fairness and transparency of AI systems.
Sources of Bias in AI Learning
Dataset Bias: AI models rely on vast amounts of data. If certain demographic groups are underrepresented in the training data, the model may fail to accurately predict behaviors or characteristics for these groups.
Algorithmic Bias: A model’s design or optimization targets may unintentionally sacrifice fairness for some groups, such as optimizing overall accuracy at the expense of minority populations.
Human Bias: Conscious or unconscious biases from developers or users can be embedded in model design, data selection, and result interpretation, leading to unfair outcomes.
How Zero-Knowledge Proofs Enhance AI Fairness
AI is now widely used across industries like recruitment, healthcare, and finance. But we all know AI is not born fair — it learns from data, and that data can carry bias. So the big question is: how do we ensure AI decisions aren’t discriminatory? This is where Zero-Knowledge Proofs come in.
Their magic lies in the ability to prove an AI system made a fair decision — without revealing the model’s inner workings or exposing user data. For example, if an AI system decides whether to approve a loan, it can use ZKPs to prove that its scoring wasn’t negatively affected by the applicant’s gender or race — without disclosing how it calculated the score. This is a huge breakthrough for maintaining privacy while making AI decisions auditable and transparent.
More importantly, this isn’t just theoretical anymore. Several real-world systems already implement this idea. Projects like FairProof, OATH, and FairZK can now generate these “fairness proofs” for actual models. They can evaluate individual-level fairness, and even scale to massive neural networks. In other words, even models with tens of millions of parameters can now be verified for fairness via ZKPs.
This opens a whole new avenue for trusted AI, especially in contexts like government regulation, financial auditing, and medical screening, where ZKPs could soon become a standard compliance requirement.
- FairProof: Verifying Individual-Level Fairness
Researchers introduced a system called FairProof, which combines fairness certification algorithms with cryptographic protocols to publicly verify a model’s fairness while preserving model secrecy.
FairProof uses local Individual Fairness (IF) metrics and applies ZKP techniques to issue personalized fairness certificates to users — assuring them that the model’s decisions were fair to them on an individual level.
- OATH: An End-to-End Fairness Verification Framework
Another example is the OATH framework, the first efficient and flexible end-to-end ZKP system for machine learning fairness. OATH supports various scoring-based classifiers and offers a robust security model that preserves fairness and confidentiality during training, inference, and auditing.
Compared to previous systems, OATH achieves a 1343x speedup in generating ZKPs for neural networks and can scale to models with tens of millions of parameters.
- FairZK: A Scalable Fairness Verification System
FairZK proposes a scalable method for verifying machine learning model fairness using ZKPs, while preserving the model’s secrecy. It utilizes aggregate information from model parameters and inputs, without relying on any specific dataset — greatly improving verification efficiency.
Experiments show that FairZK can generate fairness proofs for a 47 million parameter model in just 343 seconds, outperforming existing solutions by orders of magnitude.
Conclusion
As AI is increasingly applied in finance, healthcare, and justice, ensuring the fairness and transparency of its decision-making processes has become critically important.
Zero-Knowledge Proofs offer a promising path to verify fairness while preserving confidentiality. In the future, as computational power grows and cryptographic technology matures, the use of ZKPs for AI fairness verification will become more practical and widespread.
At the same time, establishing standardized fairness benchmarks and strengthening interdisciplinary collaboration will accelerate the real-world adoption of ZKPs in AI systems — helping us build a more just and trustworthy AI-driven society.
