[2505.01877] Humans can learn to detect AI-generated texts, or at least learn when they can’t

0 2 minutes read

250501877 Humans can learn to detect AI generated texts or at.png

[Submitted on 3 May 2025 (v1), last revised 6 May 2025 (this version, v2)]

PDF view of the paper entitled Humans can learn to discover texts created from artificial intelligence, or at least learn when they cannot, by Ji \ v {r} \ ‘I Mili \ v {C} KA and 3 other authors

PDF HTML (experimental) view

a summary:This study examines whether individuals can learn to discriminate accurately between the texts written on humans and AI when providing them with immediate comments, and if they can use these comments to re -calibrate their self -efficacy. We also explore the specific criteria that individuals depend on when making these decisions, focusing on the text style and the perceived reading.

We used GPT-4O to generate several hundreds of texts across different types and text species similar to Koditex, a multi-recorded collection of text written on humans. Then we presented random textual pairs to 255 Czech original speakers who identified the text written on the human being created by artificial intelligence. The participants were randomly appointed to two conditions: one receives immediate reactions after each experience, and the other does not receive comments until the experiment is completed. We recorded accuracy in identifying identity, confidence levels, response times, and provisions on the ability to read text along with demographic data and the participation of participants with artificial intelligence techniques before the experiment.

Participants receiving immediate reactions showed a significant improvement in accuracy and calibration of confidence. Participants initially made incorrect assumptions about the features of the text from artificial intelligence, including expectations about stylistic hardness and ease of reading. It is worth noting that the participants have made most of the mistakes specifically when feeling confident – a problem that has been largely solved between the counted feeding group.

The ability to distinguish between human texts and the creation of artificial intelligence can be learned effectively through targeted training with explicit reactions, which helps to correct misconceptions about AI’s stylistic features and reading capacity, as well as other potential variables that have not been explored, while facilitating more self -formation. This result may be of particular importance in educational contexts.