Users favored GPT-5 definitions in direct comparisons

Users preferred definitions generated by GPT-5 in direct comparisons with dictionary entries, but AI suggestions still need factual and linguistic checks, Maria Tuulik, a senior researcher at the Institute of the Estonian Language, writes in ERR’s “Keeleminutid” segment.

The institute surveyed 186 people last autumn to find out whether AI could help write dictionary definitions and what risks its use might bring. Participants compared AI-generated definitions with entries from the institute’s Unified Dictionary. In a blind test, they were not told which definitions had been written by people. The researchers had first found that GPT-5 was best at imitating the dictionary’s style.

One concern emerged while the test was being prepared: AI occasionally suggested meanings that were not supported by actual use in language data, particularly for rare words. For example, GPT-5 defined the Estonian word “sabajas” as “a person who tends to follow someone constantly and attach themselves to others.” The dictionary defines it as “resembling a tail in shape or movement.” Nine participants considered the two definitions equivalent, and seven preferred the AI version. Tuulik said this showed that users may be especially vulnerable when encountering an unfamiliar word.

In the first part of the test, participants rated definitions on a scale, and the scores for dictionary and AI entries were similar. In a second task, they saw two definitions side by side and chose a preference; overall, they chose GPT-5’s more often. Comments suggested that AI definitions were liked for being shorter and simpler. For “mjäu,” for instance, participants preferred “a cat’s vocalisation” to “imitates a cat’s vocalisation.” Dictionary definitions were favoured when precision mattered, especially for more complex and specialist terms.

Both kinds of definitions were generally rated highly, with participants usually saying they described a word’s meaning “very well” or “rather well.” Tuulik said AI could already help speed up dictionary work, but its suggestions still require factual and linguistic review.