STEPI analysis of 120 AI fiscal projects finds expert consensus in only 21.7% of cases

An AI-generated image [123RF]
An AI-generated image [123RF]

Amid a surge in government fiscal projects tied to AI, experts are calling for clearer standards to distinguish genuine research and development from simple technology adoption.

The Science and Technology Policy Institute, known as STEPI, said Tuesday that criteria for classifying AI fiscal projects as R&D need to be overhauled, in a report published in the 369th issue of its STEPI Insight series titled "Issues and Implications of R&D Classification Standards for AI Fiscal Projects."

Current government budget guidelines contain no separate criteria for determining R&D status that account for the distinct characteristics of AI and software projects. As AI spreads across industries, the institute warned, projects that merely introduce or apply existing AI technology risk being counted as R&D.

The stakes are significant: how a project is classified determines whether it qualifies for government funding and tax support. If non-research activities are labeled R&D, limited research funds could flow into work that does not advance knowledge. Conversely, genuine research projects that resolve technological uncertainty could lose investment incentives if misclassified as non-R&D.

Researchers analyzed 120 AI fiscal projects in the information and communications sector, using generative AI tools including ChatGPT and Claude to assess whether each project's R&D classification was appropriate.

When the OECD's 2025 supplementary guidelines on software development were factored in, the number of projects that both AI models agreed qualified as R&D fell from 49 to 44.

Of the 13 projects that received a different verdict from the current classification, 11 were cases where at least one AI model judged an existing R&D project to be non-R&D. Those cases were concentrated mainly in research infrastructure development, technology demonstration and talent-training projects.

The researchers cautioned, however, that these findings should not be read as direct evidence of misclassification, but rather as signals warranting further review.

When 10 experts were asked to evaluate 46 projects pre-screened by generative AI, unanimous agreement was reached on only 10 projects — just 21.7 percent of the total.

Expert opinions diverged on the remaining 36 projects. Classification confusion was particularly pronounced in projects that blended R&D and non-R&D activities, such as testbed and research infrastructure development, talent training and voucher support programs.

Researchers said whether an AI project qualifies as R&D should be determined by the actual activities carried out, not by the project's name or its use of AI technology.

To that end, the report recommended incorporating the OECD Frascati Manual's 2025 supplementary guidelines on software development into the government's budget preparation guidelines, and allocating costs by phase and activity for projects that mix R&D and non-R&D work.

The researchers also proposed establishing shared classification standards and a post-hoc eligibility review system that relevant agencies can apply consistently, as well as using generative AI as a first-pass screening tool for R&D projects.

"R&D status should not be determined solely by whether AI is used or by a project's name," the researchers said. "The core criterion must be whether the project systematically resolves scientific or technological uncertainty and generates new knowledge or technological advances."


nbgkoo@heraldcorp.com