Sony Patent Targets AI Bias by Making It Visible and Correctable

By Patentlyze Team · Updated Jul 24, 2026

Sony is patenting a system that forces an AI to show its work, specifically so engineers can find out whether the model is making decisions based on things it shouldn't be, like race, age, or gender.

What Sony's AI bias-removal system actually does

Imagine a hiring tool powered by AI that learns to favor applicants from certain zip codes, not because zip codes matter, but because they correlate with race. The AI isn't obviously doing anything wrong, but the bias is baked in and nearly invisible.

Sony's patent describes a system to catch exactly that kind of hidden problem. It uses a technique called explainable AI (XAI), which forces an AI model to identify which pieces of data most influenced its decision. Sony's system then looks at whether those influential data points are secretly acting as stand-ins for protected characteristics like age or ethnicity.

Once the system maps out where the bias is coming from, it can automatically adjust the data or the model to reduce it. Think of it as a bias audit tool built directly into the AI pipeline, rather than something bolted on after the fact.

How the system traces bias back to specific data points

The patent describes an information processing device with two main components working together.

First, an acquisition unit scans input data for what the patent calls sensitive attributes (characteristics like age, gender, or race that shouldn't drive AI decisions) and non-sensitive attributes (everything else, like income, location, or purchase history). It builds a secondary prediction model that tries to guess the sensitive attribute using only the non-sensitive data. If that model succeeds, it means the non-sensitive data is carrying hidden information about the protected group.

Second, a contribution degree calculation unit uses XAI (explainable artificial intelligence, a class of methods that make AI reasoning transparent rather than opaque) to score how much each non-sensitive data point contributed to predicting the sensitive attribute. High-scoring items are flagged as proxies for bias.

Finally, a processing unit takes that ranked list and mitigates the bias, either by adjusting the training data or reweighting the model itself.

• Detects when non-sensitive data leaks protected-group information
• Scores each data field by how much it drives that leakage
• Applies corrections to data or model weights automatically

What this means for AI fairness in consumer products

AI fairness is increasingly a legal and regulatory concern, especially in Europe under the AI Act and in US employment and lending contexts. A tool that can automatically audit and correct bias rather than relying on manual review could become standard infrastructure for any company deploying AI in high-stakes decisions. Sony, which builds AI into cameras, sensors, and entertainment systems, would benefit from being able to certify that its models don't carry unintended demographic skew.

For you as a consumer, this kind of system matters most when AI is making decisions about you, whether that's a recommendation algorithm, a facial recognition gate, or a credit check. A built-in bias audit layer is a meaningful step toward accountability, though the effectiveness depends entirely on implementation.

Editorial commentary on a publicly published patent application. Not legal advice.