[Japan's First (*1)] Patenting a System to "Prevent AI from Learning Lies". Does the input data have integrity? "Multi-AI" pre-screens and blocks fake data. Solving the "flaw at the entrance" ignored by AI developers worldwide with a patent.
Cycal Trust Co., Ltd. has received a patent allowance for a system where a 'Multi-AI' council pre-verifies the authenticity of training data to block malicious inputs, solving fundamental flaws in AI development.
📋 Article Processing Timeline
- 📰 Published: April 22, 2026 at 04:11
- 🔍 Collected: April 23, 2026 at 01:01 (20h 50m after Published)
- 🤖 AI Analyzed: April 23, 2026 at 02:37 (1h 35m after Collected)
Cycal Trust Co., Ltd. (Headquarters: Shibuya-ku, Tokyo, CEO: Tsuyoshi Sue, hereinafter "Cycal Trust") is pleased to announce that it has received a Notice of Allowance for its patent application "Patent Application No. 2025-049833" (Patent number pending). The core of this patent is "a mechanism where a council of 'AIs operated by multiple distinct entities (hereinafter Multi-AI)' verifies whether the dataset (training data) set into an AI is correct, authentic, and free of malicious information before it is recorded," and it becomes a new addition to Cycal Trust's patent portfolio.
Including this case, Cycal Trust holds multiple patents related to guaranteeing the authenticity of AI input data (and continues to actively file more), and will proceed with implementing this technology as a solution combined with Distributed Ledger Technology (DLT) in parallel with international standardization activities in "ISO/TC307 (Blockchain and distributed ledger technologies)."
(*1) Based on our research targeting patent documents, academic literature, and published materials by companies and research institutions published in Japan as of April 21, 2026.
Chapter 1: What is the structural issue shared by AI worldwide?
(1)
Generative AI, conversational AI, autonomous driving AI, medical diagnostic AI, and other financial trading AIs—these may seem to belong to different technological domains at first glance, but they actually share a common structural issue.
That issue is "the absence of a mechanism to systematically verify the authenticity of the diverse data recorded and learned by AI."
(2)
Let's look at specific examples.
1. Autonomous Driving Sector
In the autonomous driving sector, if sensor data is spoofed, the AI might avoid non-existent obstacles while overlooking actual pedestrians. However, at present, a standard, mechanical mechanism to verify the authenticity of data passed from sensors to the AI prior to recording has not been established.
2. Medical AI Sector
In the medical AI sector, if electronic medical records are tampered with, the very foundation of the diagnosis collapses. Nevertheless, a practical mechanism to verify authenticity at the stage before data is input into the AI has yet to become an industry standard.
3. Semiconductor Supply Chain Sector
In the semiconductor supply chain, counterfeit semiconductors circulating with forged electronic Certificates of Authenticity ("eCOA") pose a significant risk to economic security. The technology to determine the authenticity of the data itself before recording quality data onto a blockchain has not been standardized.
4. Datasets for Generative AI / AI Agents (Training Data)
Regarding datasets (training data) for generative AI and AI agents, "data poisoning attacks," where malicious third parties inject fraudulent data into them, are increasing internationally. However, to Cycal Trust's knowledge, there are no established precedent cases globally for general-purpose technology that automatically determines the authenticity of datasets (training data) prior to input.
5. Summary
These are all variations of the same issue. While global R&D investments are concentrated on "making AI outputs correct," industrial efforts toward "confirming whether AI inputs are correct" have been extremely limited until now.
Chapter 2: A Paradigm Shift - Not "purifying the water" but "keeping dirty water out"
Until now, the primary focus of international R&D has been "improving AI output quality." Suppressing hallucinations (false AI outputs), correcting biases, and advancing output filters—all these are approaches akin to applying water purification treatment after the "polluted water" has already flowed through.
The design philosophy adopted by Cycal Trust's patent group is the exact opposite. That is, the approach is not to "purify polluted water," but rather "not let polluted water flow into the pipe in the first place." Prior to data recording and learning, we have patented the "entrance gate" itself, where a "Multi-AI" acts as a council to verify the authenticity of the data and allows only passed data to go through.
Chapter 3: Technical Structure of This Patent - What is council verification by "Multi-AI"?
The mechanism established by this patent consists of the following four stages:
Stage 1: Data Acquisition
Stage 2: Independent verification by "Multi-AI" (Solving the Single Point of Failure problem)
Stage 3: Weighted verification by a council system
Stage 4: Pass/Fail judgment
Chapter 4: Why has no one been able to solve this problem until now?
The background as to why technological development in this area has not progressed until now can be explained structurally.
First, investments in the AI industry have concentrated on the "model side." Larger models, faster inference, lower hallucination rates—these compe
Including this case, Cycal Trust holds multiple patents related to guaranteeing the authenticity of AI input data (and continues to actively file more), and will proceed with implementing this technology as a solution combined with Distributed Ledger Technology (DLT) in parallel with international standardization activities in "ISO/TC307 (Blockchain and distributed ledger technologies)."
(*1) Based on our research targeting patent documents, academic literature, and published materials by companies and research institutions published in Japan as of April 21, 2026.
Chapter 1: What is the structural issue shared by AI worldwide?
(1)
Generative AI, conversational AI, autonomous driving AI, medical diagnostic AI, and other financial trading AIs—these may seem to belong to different technological domains at first glance, but they actually share a common structural issue.
That issue is "the absence of a mechanism to systematically verify the authenticity of the diverse data recorded and learned by AI."
(2)
Let's look at specific examples.
1. Autonomous Driving Sector
In the autonomous driving sector, if sensor data is spoofed, the AI might avoid non-existent obstacles while overlooking actual pedestrians. However, at present, a standard, mechanical mechanism to verify the authenticity of data passed from sensors to the AI prior to recording has not been established.
2. Medical AI Sector
In the medical AI sector, if electronic medical records are tampered with, the very foundation of the diagnosis collapses. Nevertheless, a practical mechanism to verify authenticity at the stage before data is input into the AI has yet to become an industry standard.
3. Semiconductor Supply Chain Sector
In the semiconductor supply chain, counterfeit semiconductors circulating with forged electronic Certificates of Authenticity ("eCOA") pose a significant risk to economic security. The technology to determine the authenticity of the data itself before recording quality data onto a blockchain has not been standardized.
4. Datasets for Generative AI / AI Agents (Training Data)
Regarding datasets (training data) for generative AI and AI agents, "data poisoning attacks," where malicious third parties inject fraudulent data into them, are increasing internationally. However, to Cycal Trust's knowledge, there are no established precedent cases globally for general-purpose technology that automatically determines the authenticity of datasets (training data) prior to input.
5. Summary
These are all variations of the same issue. While global R&D investments are concentrated on "making AI outputs correct," industrial efforts toward "confirming whether AI inputs are correct" have been extremely limited until now.
Chapter 2: A Paradigm Shift - Not "purifying the water" but "keeping dirty water out"
Until now, the primary focus of international R&D has been "improving AI output quality." Suppressing hallucinations (false AI outputs), correcting biases, and advancing output filters—all these are approaches akin to applying water purification treatment after the "polluted water" has already flowed through.
The design philosophy adopted by Cycal Trust's patent group is the exact opposite. That is, the approach is not to "purify polluted water," but rather "not let polluted water flow into the pipe in the first place." Prior to data recording and learning, we have patented the "entrance gate" itself, where a "Multi-AI" acts as a council to verify the authenticity of the data and allows only passed data to go through.
Chapter 3: Technical Structure of This Patent - What is council verification by "Multi-AI"?
The mechanism established by this patent consists of the following four stages:
Stage 1: Data Acquisition
Stage 2: Independent verification by "Multi-AI" (Solving the Single Point of Failure problem)
Stage 3: Weighted verification by a council system
Stage 4: Pass/Fail judgment
Chapter 4: Why has no one been able to solve this problem until now?
The background as to why technological development in this area has not progressed until now can be explained structurally.
First, investments in the AI industry have concentrated on the "model side." Larger models, faster inference, lower hallucination rates—these compe