AI Output Error Likelihood Estimation Using Knowledge Neurons

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing AI systems do not effectively inform users about the likelihood of errors in their output data, making it difficult for users to recognize the degree of possibility that an element in the output includes an error.

Innovation Solution

An information processing system that includes a processor to acquire input data, generate output data using AI, and provide error possibility information indicating the likelihood of errors in the output data, utilizing a knowledge neuron analysis to identify and quantify the contribution of neurons in a large language model (LLM) to word predictions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If AI generates output data in response to input data, then productivity and automation are improved, but the reliability deteriorates due to undetectable errors in the output data

Engineering Contradiction:

ImproveAI output generation speedVSAvoiderror possibility in output data

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The system implements a feedback mechanism where the AI model's own predictions are fed back into the model to generate error possibility information. The processor acquires output data generated by AI, then uses this output as input to the same AI model to predict whether errors exist in the generated content, creating a self-validation feedback loop that maintains reliability while preserving productivity

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The patent introduces an intermediary validation process between AI output generation and final user delivery. The AI model acts as an intermediary to evaluate its own output by predicting error possibilities, thereby mediating between the raw AI generation and the final reliable output presented to users

Inventive Principle:

Principle #24Intermediary (Mediator)

2Reliability

If error possibility information is provided for each element in output data, then reliability and transparency are improved, but device complexity increases

Engineering Contradiction:

Improveerror possibility indicationVSAvoidsystem complexity for error analysis

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The AI model performs self-service validation by predicting errors in its own output without requiring external validation systems. The same AI model that generates output data also analyzes its own output to predict error possibilities, eliminating the need for separate complex validation hardware or software systems

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The AI model is designed to perform multiple functions: it both generates output data and validates its own output by predicting errors. This multi-functionality reduces device complexity by using a single system component for both creation and validation tasks rather than requiring separate specialized systems

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentUS20260065040A1Information processing system, non-transitory computer readable medium, and information processing method

Publication Date: 2026.03.05 FUJIFILM BUSINESS INNOVATION CORP

AI summary

An information processing system includes a processor configured to: acquire input data; acquire output data generated by artificial intelligence in response to the input data; and output the output data and error possibility information indicating a degree of possibility that an element of multiple elements in the output data includes an error by the artificial intelligence.