Generative AI Bias Evaluation Using Masked Statement Reconstruction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Generative AI models can produce biased or unfair outputs due to training on incorrect or biased information, particularly in fields requiring fair decision-making, such as gender equality and climate action, leading to skewed analysis and outcomes.

Innovation Solution

A method involving a primary generative AI model generating question prompts and statements, masking key terms, and using a secondary generative AI model to unmask these terms, allowing evaluation of bias through comparison with a statement set, followed by retraining to correct biases.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If generative AI models are trained on existing knowledge bases, then the models can perform analytical and creative tasks, but the models may inherit biases and produce unfair outputs

Engineering Contradiction:
ImproveAI task performanceVSAvoidoutput fairness
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The system performs preliminary bias detection and evaluation before the AI model produces final outputs. By evaluating the knowledge base and training data beforehand, and continuously monitoring outputs during operation, the system can identify and mitigate biases before they affect decision-making processes.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system implements continuous feedback loops where AI outputs are evaluated for bias, and this evaluation information is fed back to adjust the knowledge base, training data, or model parameters. This creates a self-correcting mechanism that progressively reduces bias while maintaining productivity.

Inventive Principle:
Principle #23Feedback

2Adaptability or versatility

If AI models are trained to understand complex fields, then the models can perform specialized analysis tasks, but the models may amplify existing biases in those fields

Engineering Contradiction:
Improvefield-specific analysis capabilityVSAvoidbias amplification
Core Design Contradiction:
Adaptability or versatilityVSObject-generated harmful factors

Solution Approach 1:

The system applies different evaluation criteria and bias detection methods tailored to specific fields and contexts. Rather than using a uniform approach, it adapts the evaluation framework to the particular domain (e.g., healthcare, law, finance), recognizing that different fields have different types of biases and fairness requirements.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The system introduces an intermediary evaluation layer between the AI model and the knowledge base/training data. This intermediary component (the bias evaluation system) mediates by filtering, adjusting, or flagging biased content before it influences model training or output generation.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Adaptability or versatility

If generative AI models generate diverse outputs, then the models can address multiple scenarios, but the models may produce inconsistent fairness across different outputs

Engineering Contradiction:
Improveoutput diversityVSAvoidfairness consistency
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The system implements a universal bias evaluation framework that can assess multiple types of outputs across different fields and contexts. This multi-functional evaluation system maintains consistent fairness criteria while adapting to diverse output types, ensuring that the same fairness standards apply regardless of the specific scenario or domain.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4579528A1Examining generative artificial intelligence for bias
Publication Date: 2025.07.02 FUJITSU LTD
  • EP4579528A1 patent drawingFigure 1
  • EP4579528A1 patent drawingFigure 2A
  • EP4579528A1 patent drawingFigure 2B

AI summary

A method may include obtaining a topic and an artificial intelligence (AI) role relating to the topic in which the topic relates to a field of study and the AI role represents an occupational role in the field of study. The method may include generating, by a first generative AI model, a question prompt based on the topic and AI role. The method may include generating, by the first generative AI model, one or more statements as a statement set corresponding to the question prompt and masking key terms included in the statements in which each statement includes at least one respective key term. The method may include determining, by a second generative AI model, unmasked statements corresponding to the masked statements. The method may include evaluating performance of the second generative AI model by comparing the unmasked statements to corresponding statements of the statement set.