SEOUL, Sept. 4, 2026 — A Korean developer has developed a new technology designed to systematically verify and block hallucinations, one of the most persistent challenges in generative artificial intelligence, and has completed a patent application in South Korea.

The technology goes beyond simply searching again to confirm an AI-generated answer. It examines whether evidence supporting a significant AI-generated claim actually exists, whether the evidence genuinely supports the claim, whether the figures and logic are accurate, whether important conditions or exceptions have been omitted, and whether contradictory evidence exists. Content that fails these checks is restricted from entering the final output through a process known as the “Verification Gate.”

A Korean patent application for the technology was completed on Sept. 4, 2026, under application No. 10-2026-0168063.

AI’s persistent problem of being convincingly wrong

Generative AI services such as ChatGPT are rapidly becoming part of the global infrastructure for information and professional work. OpenAI says ChatGPT is now available to more than 1 billion weekly active users. Together with Google’s Gemini, Anthropic’s Claude and other services, generative AI is already having a substantial impact on how individuals and businesses work.

Despite these advances, hallucinations—instances in which AI systems generate plausible but nonexistent facts or sources—remain a significant barrier to reliability.

Researchers do not generally expect the problem to disappear as readily as an ordinary software defect. Research published in Nature portfolio journals has found that hallucinations and material omissions can persist in current language models, with potentially serious consequences when unverified outputs are used in high-stakes fields.

The technology developed in Korea reframes the challenge by focusing on whether an AI-generated error can be prevented from entering a final output as though it were a verified fact.

Jung Yeon-Tae, President of the Korea Information Service of the Construction Industry, explained the thinking behind the Verification Gate.

“The goal is not to ensure that AI never makes mistakes,” Jung said. “It is to embed systematic verification into the generation process so that unverified information cannot enter the final output as fact.”

Internal blind test detected all planted hallucination cases

During development, the system underwent an internal blind test using a fictional complex decision-making report combining medical, legal, industrial safety, technology and investment issues.

Different types of verification traps were planted in advance, including fabricated papers and sources, numerical misinterpretations, exaggerated causal relationships, unsupported extensions of scope, overstated safety claims, and omissions of material conditions and risks.

The system was instructed to independently verify the report without being given the correct answers. Its findings were then compared with a reference answer sheet prepared beforehand.

Of the 40 combined verification traps, the system fully detected 39 and partially detected one. No item was completely missed.

When half credit was assigned to the partially detected case, the system received an overall score of 98.75%, or approximately 98.8%.

When only the deliberately planted hallucination cases were assessed, separate from the broader range of errors, the system detected and blocked every prepared case.

The hallucination detection and defence rate was therefore 100% on the internal blind evaluation set, with no complete misses observed.

This does not conflict with the broader score of 98.8%. The overall assessment covered not only hallucinations but also numerical, logical, causal, professional-judgment and omission-related errors.

The 100% figure, however, represents the result observed on this particular blind evaluation set. It does not constitute a universal guarantee that the system can eliminate every hallucination in all circumstances.

Moving beyond post-generation checks of ChatGPT, Gemini and Claude

Another notable aspect of the technology is that it is not limited to post-generation verification.

A report or research paper produced by an existing generative AI system can be submitted to the Verification Gate after completion. The system has also been built to support the preparation of reports, academic papers, legal materials and investment analyses within the verification framework from the outset.

Under this approach, significant claims and supporting evidence become subject to verification during the generation process.

If the system identifies an unverified source, an inaccurate figure or a conclusion stronger than its supporting evidence, the material is not allowed to pass directly into the final document. It is revised or restricted and then reviewed again. The design also treats a revised document as a new output that must undergo another verification cycle to determine whether the revisions have introduced additional hallucinations.

The development direction represents a shift from having AI draft a document before a person searches for errors to integrating verification into the AI-assisted drafting process itself.

If the results can be reproduced through large-scale external testing, the approach could change how AI is used to prepare professional documents in fields where reliable sources and factual accuracy are critical, including academic research, legal and contractual reviews, investment analysis and public policy.

Verification extends to important omissions

The system’s scope is not limited to statements that are demonstrably false.

A legal explanation may omit a significant exception. An investment proposal may present only the potential benefits while excluding downside risks. A research summary may leave out adverse effects and methodological limitations. In such cases, individual sentences may be factually correct, yet the document as a whole can appear substantially more certain or safer than the underlying evidence warrants.

The system is therefore designed to treat significant omissions involving conditions, exceptions, risks, alternatives and contradictory evidence as matters requiring verification.

Its procedures also include assuming that the current conclusion may be wrong and re-examining conflicting evidence, alternative explanations, hidden assumptions and the worst reasonable scenario.

Targeting verification of professional documents hundreds of pages long

The GPT-based system is intended to verify not only short questions and answers but also lengthy reports, research papers, legal opinions, contracts and investment materials.

For longer documents, the system does not rely solely on summarising the entire text. It tracks important conclusions and their supporting evidence, examines whether related claims and figures dispersed across different sections are consistent, and reassesses how revised information affects other conclusions.

This approach may be particularly relevant to corporate due diligence materials, legal opinions and research documents running to hundreds of pages, where an error in one section can alter a final conclusion presented much later in the document.

Patent filed in Korea, with independent global testing ahead

Generative AI hallucinations are not a problem confined to a single company.

They present a shared reliability challenge for companies and developers building services on language models such as ChatGPT, Gemini and Claude, as well as for users worldwide who apply these systems to professional work and research.

The Korean development is noteworthy because it implements a different approach from model-level competition aimed at eliminating hallucinations altogether. It seeks to prevent unverified errors from entering final outputs and has now reached the patent-application stage as an operational GPT-based system.

The next test will be larger and a fully independent evaluation.

Because the current findings were produced through an internal blind test conducted during development, the system will need to be assessed repeatedly using new evaluation materials prepared by researchers, companies and testing organisations independent of the developer.

If the observed hallucination-detection performance can be reproduced under those conditions, a longstanding competitive benchmark in generative AI could begin to change.

The focus may expand beyond which AI produces the most convincing answers to which system most reliably prevents unverified answers from reaching the final output.

The global AI industry will be watching to see whether this technology, developed in Korea, can help reshape how generative AI reliability is evaluated and enforced.

Patent application: Republic of Korea Patent Application No. 10-2026-0168063, filed Sept. 4, 2026
Technology: OXMV

Korean version available here

Yeon Choul-woong More by this reporter