STEPI analyzes 120 AI-funded projects... Findings vary depending on the criteria
All 10 experts agreed on only 21.7%... Assess whether projects achieve technological advances

Concerns have been raised that government-funded projects should not be classified as research and development (R&D) solely because they use artificial intelligence (AI). For projects that combine R&D with general support activities, such as building AI infrastructure or training talent, the key question is whether they actually create new technologies or knowledge. An analysis of 120 government AI-funded projects found that their R&D classification varied depending on the criteria applied, and experts’ assessments differed considerably.


On November 10, the Science and Technology Policy Institute (STEPI) announced the results of a generative AI analysis and expert survey covering 120 AI-funded projects in the information and communications sector. The findings were presented in a report titled “Issues and Implications in the R&D Classification Criteria for AI-Funded Projects.”

"Using AI Doesn't Make Everything R&D"... Government Should Revisit Classification Criteria for Fiscal Projects View original image

The government’s current detailed guidelines for formulating budget proposals do not include separate R&D classification criteria that reflect the specific characteristics of AI and software projects. The research team noted that this leaves room for projects that use existing technologies to build systems or train personnel to be included in R&D.


The research team used ChatGPT and Claude to assess whether the funded projects qualified as R&D. When only the existing Organisation for Economic Co-operation and Development (OECD) Frascati Manual was applied, both models classified 49 projects as R&D. When the 2025 supplementary guidelines were incorporated, that figure fell to 44.


The analysis using the supplementary guidelines identified 13 cases in which the classification differed from that of the current projects. In 11 of these cases, at least one AI model classified a project currently designated as R&D as non-R&D. The differences arose mainly in projects involving infrastructure development, demonstration, and talent training.


Experts also differed in their assessments. Of the 46 projects classified as R&D in the AI preliminary screening, only 10 (21.7%) received unanimous assessments from all 10 experts. Differences of opinion were especially pronounced for projects combining R&D and non-R&D activities, including testbeds, infrastructure development, talent training, and voucher support.


The research team noted, however, that these findings do not prove that current projects have been misclassified; rather, they signal the need to review the classification criteria.


"Classification Should Be Based on Actual Research Activities, Not Project Titles"


To improve the classification system for AI-funded projects, the research team proposed incorporating the supplementary guidelines to the OECD Frascati Manual into the government’s detailed guidelines for formulating budget proposals. The key criterion, it said, should be whether a project resolves scientific or technological uncertainties and generates new knowledge or technological advances, rather than whether its title includes terms such as “AI adoption” or “AI talent training.”


The team stressed that for projects combining R&D and non-R&D activities, the costs actually allocated to research and development should be identified separately, and the agencies responsible for R&D statistics, budgets, and tax support should apply common assessment criteria. It also proposed checking R&D eligibility through sample reviews after projects have been designated.


Generative AI should be used as an auxiliary tool to identify projects that need further review, the team said, while final decisions should be made following expert review and an explanation process involving the relevant ministries.



“As AI spreads as a general-purpose technology, simply ‘using AI’ is no longer enough to determine whether something qualifies as R&D,” the research team said. “The assessment should focus not on the project title, but on whether it actually resolves scientific or technological uncertainties.”


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing