Home
16+
DOI: 10.18413/2313-8912-2026-12-3-0-1

When Human Writing Looks Artificial: Perceived AI-Likeness as a Property of Text

Tatiana A. Litvinova (Voronezh State Pedagogical University, Russia)
Olga V. Dekhnich (Belgorod State National Research University, Russia)

A piece of writing can look as if it were produced by AI even when it wasn’t, and conversely, text generated by AI can still be mistaken for human work. We investigated this mismatch between the true source of 28 Russian texts (14 written by humans and 14 generated by AI) and their perceived AI-likeness. The texts were assessed by professional linguists under three different evaluation setups. For each text, we extracted 46 embedding-based semantic alignment features and 12 stylometric measures. In the third evaluation regime (W3), we also used a balanced subset of 10 texts and collected ratings along 13 perceptual dimensions before participants made their source judgments.

Across both semantic and stylometric space, human and AI texts showed a clear separation. Unsupervised components related to origin produced large Human-AI differences (Hedges’ |g| ≈ 1.4), and nested leave-one-text-out ridge models achieved AUCs between .832 and .878. However, the objective dimensions that distinguished the actual origin were not sufficient on their own to explain AI attribution once the origin was controlled for. The importance of cues also varied across evaluation regimes. In particular, seven features showed FDR-significant regime × feature interactions between W2 and W3, all with reversed slopes; mean word length also shifted direction between W1 and W2. In W3, the perceptual ratings were largely driven by a dimension that contrasted authorial-experiential presence with schematic generality. Shifting toward the authorial-experiential end substantially lowered AI-oriented attribution on the original five-category scale (OR = .228, 95% CI [.145, .359], p < .001).

So, perceived AI-likeness cannot be viewed as a straightforward reflection of either actual authorship or the linguistic markers tied to origin. In these results, it functioned more like a perceptual property of the text: it depended on whether readers felt there was a situated author behind what they read. Meanwhile, the linguistic cues that supported AI attribution varied depending on the evaluation regime. This distinction is important because perceived AI authorship can affect trust, credibility, and later assessments even when the source judgment turns out to be incorrect.

Keywords: AI-Generated Text, Perceived AI-Likeness, AI Attribution, Human Text Perception, Authorial Presence, Semantic Features, Stylometry, Source Attribution.

Figures

Number of views: 205 (view statistics)
Количество скачиваний: 294
Full text (HTML)Full text (PDF)Скачать XMLTo articles list
Information for citation:

Litvinova, Т. А., Dekhnich, O. V. (2026). When human writing looks artificial: perceived AI-likeness as a property of text, Research Result. Theoretical and Applied Linguistics, 12 (3), 3–31.

  • User comments
  • Reference lists
  • Thanks

While nobody left any comments to this publication.
You can be first.

Leave comment: