The Risks of Generative AI Non-Verifiable Interpretation
Exploring Japanese youkai in English
Bibliographic Data
| ID | 16928477 |
|---|---|
| Authors | Minoru Moriguchi (Nagahama Institute of Bio-Science and Technology), Olivia Kennedy (0000-0003-0144-2516, Nagahama Institute of Bio-Science and Technology) |
| Year | 2025 |
| Volume | 12 |
| Issue | 1 |
| Pages | 39-56 |
| Publication date | 2025-08-01 |
| Peer Reviewed | Yes |
| Open Access | No |
| Type | ARTICLE |
| Venue | Lexicography (JOURNAL) |
| Journal identifiers | ISSN: 2197-4292 • E-ISSN: 2197-4306 |
| Publisher | University of Toronto Press Inc. (UTPress) (PUBLISHER) |
| DOI | 10.3138/lexi-2025-0005 |
| OpenAlex | W4412828788 |
| Language | EN |
| References cited | 9 |
Since Open AI's release of ChatGPT 3.5 in November 2022, generative AI has greatly impacted lexicography. During this time, Japanese popular culture including manga and anime has become increasingly popular throughout the world, in which one of the key themes is youkai. This paper explores how generative AI presently explains non-English-language concepts in English and how it can be used for the creation of dictionary entries, using youkai as an interesting contemporary example. We first define the Japanese term, youkai, based on existing literature. Next, we select 10 youkai, both well-known and less common, and examine how we can use generative AI to produce definitions. In order to also measure the progress of AI, we compare outputs generated in June and December 2024. Based on the generated results, we find that (1) definitions generated by AI are not linguistically problematic, (2) English and Japanese prompts generate slightly different definitions, (3) some definitions generated by AI include inferences which seem to be valid but cannot be verified in existing reference materials, which we term non-verifiable interpretation (NVI), (4) whether NVI appears in the definition is not necessarily related to the size of the large language data set, and (5) while the later outputs are better on the whole, they tend to include more NVI. We conclude that generative AI can be helpful in dictionary making, but lexicographers must work to combat not only hallucination, but also NVI
Generative grammar · Linguistics · Natural language processing · Programming language · Verifiable secret sharing · Computer Science · Natural Language Processing Techniques · Philosophy · Psychology · Text Readability and Simplification · Translation Studies and Practices · Artificial Intelligence
| Citation velocity | historical |
|---|---|
| Highly cited | No |