Authorship Attribution and Trust in AI-Generated Scientific Texts: An Experimental Study

In the paper

“Taken together with evidence on algorithm aversion and authorship norms (Magni et al., 2024;Draxler et al., 2024), our data suggest that framing AI involvement through human oversight is critical for preserving trust without affecting comprehension.”

What the sources say

• Supported

  The Abstract and General Discussion report that identifying a creator as AI rather than human sometimes lowers creativity ratings, supporting the cited background point about reactions to AI authorship. The Practical Implications section also suggests emphasizing human involvement to reduce bias. The sentence presents the trust-and-comprehension conclusion as the citing authors’ own inference from their data, not as a finding attributed to Magni et al. (LLM, conf=0.87)

  cites Magni et al., 2024 source
• Supported

  Section 6.4 discusses norms and expectations around declaring AI contributions and how those declarations may vary across cultures. That supports citing this paper as evidence on authorship norms. The sentence presents the claim about trust and comprehension as the citing authors’ conclusion from their own data, not as a finding attributed to Draxler et al. (LLM, conf=0.88)

  cites Draxler et al., 2024 source

In the paper

“Prior work shows that content labeled as machine-produced is often viewed more skeptically than identical human-attributed text.”

What the sources say

• Supported

  The Introduction reports that the same human-generated headlines were rated less accurate when labeled AI-generated than when labeled human-generated. The paper’s Abstract also reports lower perceived accuracy for AI-labeled headlines regardless of whether they were written by humans or AI. (LLM, conf=0.91)

  cites Altay & Gilardi, 2024 source
• Supported

  In §3.1 (page unavailable), the paper reports that participants rated the same GPT-written science journalism text as more credible when it was labeled human-authored rather than AI-authored (d = 0.36). This supports the claim’s comparison, though the experiment tested one article rather than how often the effect occurs across content generally. (LLM, conf=0.92)

  cites Lermann Henestrosa & Kimmerle, 2024 source
• Partly supported

  The Conclusions report that identical productions were sometimes rated less creative when attributed to AI, particularly paintings. But the Results for Study 3 found no significant difference in creativity ratings for written business ideas attributed to AI versus humans. The source therefore supports a label effect for some content, not the claim that machine-labeled text is often viewed more skeptically. (LLM, conf=0.88)

  cites Magni et al., 2024 sourceExperimental — accuracy not yet measured

In the paper

“Similar algorithm-aversion appears for creative outputs (Magni et al., 2024).”

What the sources say

• Supported

  The Conclusions (page not provided) report that evaluators sometimes rated a product as less creative when told AI rather than a human made it. The effect appeared for paintings, though not for advertisement posters or business ideas, so the claim is supported as an occurrence rather than a universal pattern. (LLM, conf=0.91)

  cites Magni et al., 2024 source

In the paper

“A philosophical critique adds that LLM text can appear truth-apt without truth-tracking intentions, inviting baseline skepticism toward AI authorship (Hicks et al., 2024).”

What the sources say

• Supported

  The Abstract says LLMs produce text that looks truth-apt without concern for truth. The section “Bullshit? hallucinations? confabulations? The need for new terminology” adds that any truth-tracking is indirect and incidental and that readers should be wary of relying on their outputs. (LLM, conf=0.93)

  cites Hicks et al., 2024 source

In the paper

“Our results align with policy arguments that transparent disclosure must be paired with clear human accountability to sustain trust in AI-assisted scientific communication (Gautam, 2025).”

What the sources say

• Partly supported

  The Abstract calls for standardized AI-use disclosure and warns that an uncredited AI co-author can erode the credibility of scientific writing. It supports the transparency-and-trust argument, but does not explicitly say disclosure must be paired with clear human accountability. (LLM, conf=0.82)

  cites Gautam, 2025 sourceExperimental — accuracy not yet measured

In the paper

“In news, labeling headlines as”

• Also stated in: § AUTHORSHIP ATTRIBUTION AND TRUST IN AI-GENERATED SCIENTIFIC TEXTS: AN EXPERIMENTAL STUDY · plain

  “AI-generated”

• Also stated in: § М.О. Матвеев, Е.А. Матвеева. Доверие к научным текстам, сгенерированным ИИ 255 · plain

  “reduces perceived accuracy and sharing intent (Altay & Gilardi, 2024)”

What the sources say

• Supported

  The Discussion states that labeling headlines as AI-generated reduced both perceived accuracy and participants’ intention to share them. It reports this effect for headlines regardless of whether they were human- or AI-generated and true or false. (LLM, conf=0.94)

  cites Altay & Gilardi, 2024 source

In the paper

“In contrast, comparisons between AI and AI+Human were significant (Trust: t(117.3)= 6.77, p < .001,g = 1.24;Intent: t(117.7)= 6.31, p < .001,g = 1.18).”

In the paper

“Mean comprehension scores were uniformly high: Control M = 4.1 (SD = 0.8), AI M = 4.0 (SD = 0.9), AI+Human M = 4.2 (SD = 0.7) out of 5 (see Table 1).”

In the paper

“readers tend to rate human-written items as higher quality than machine-written ones (Lermann Henestrosa & Kimmerle, 2024).”

What the sources say

• Not checked yet

  Checked before — its earlier verdict is held back until a new check runs.

  cites Lermann Henestrosa & Kimmerle, 2024 source

In the paper

“Critically, the core text was identical for all participants.”

• Also stated in: § Materials · plain

  “Thus, the only difference between conditions was the small final phrase.”

In the paper

“Second, the sample was predominantly Russian undergraduate students (mean age ~22) recruited via university channels.”

In the paper

“The semantic differential scale we used is conceptually similar to established trust measures for online content (Brühlmann et al., 2017).”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Brühlmann et al., 2017

In the paper

“People anthropomorphize non-human agents and respond socially to computers (Epley et al., 2007;Nass & Moon, 2000).”

What the sources say

• ?Not checked — no full text of the sourcecites Epley et al., 2007 source
• ?Not checked — no full text of the sourcecites Nass & Moon, 2000 source

In the paper

“They were randomly assigned (approximately equal n per cell) to one of three between-subjects conditions: Control (no authorship label), AI (”

• Also stated in: § AUTHORSHIP ATTRIBUTION AND TRUST IN AI-GENERATED SCIENTIFIC TEXTS: AN EXPERIMENTAL STUDY · plain

  “AI-generated”

• Also stated in: § Method · plain

  “), or AI+Human (”

In the paper

“Although our study did not measure interpersonal trust, the overall negative reaction to raw AI content may reflect a broader tendency to distrust automated systems unless human agency is apparent.”

In the paper

“Importantly, the negative effect was fully counteracted by adding”

• Also stated in: § Discussion · hedged

  “The AI+Human label produced trust and intent ratings statistically indistinguishable from control.This suggests that highlighting human involvement in AIgenerated content restores confidence.”

In the paper

“KR-20 for the MCQ battery was .606(AI), .839(Control), and .846(AI+Human).”

In the paper

“Experiments by Sparks and Rapp (2011) showed minimal effects of source credibility on online processing and comprehension unless readers were prompted to consider it, suggesting that accurate AI text should yield comprehension comparable to human text.”

What the sources say

• ?Not checked — no full text of the sourcecites Sparks & Rapp, 2011 source

In the paper

“Educators and publishers should note that the simple label”

• Also stated in: § Discussion · plain

  “AIgenerated”

• Also stated in: § Conclusion · hedged

  “may inadvertently lower students' confidence in material, whereas coupling it with human involvement reassures readers.”

In the paper

“This resonates with broader adoption research: Đerić et al. (2025) reported that trust predicts intention to use generative AI tools in education.”

What the sources say

• Not checked yet

  Checked before — its earlier verdict is held back until a new check runs.

  cites Đerić et al., 2025 source

In the paper

“These effects are consistent with an AI aversion: merely labeling correct content as AI-generated greatly reduced trust and willingness to use or share it, even though accuracy was unchanged (Altay & Gilardi, 2024).”

What the sources say

• Not checked yetcites Altay & Gilardi, 2024 source

In the paper

“Independence was satisfied by design; unequal variances were accommodated by Welch's test; approximate normality of residuals was inspected via Q-Q plots and found acceptable.”

In the paper

“If AI authorship is always disclosed, purely AI-produced materials might be viewed skeptically.However, indicating that experts vetted or edited the content can maintain trust.”

In the paper

“These results replicate and extend the”

• Also stated in: § Discussion · plain

  “AI label aversion”

• Also stated in: § Discussion · plain

  “reported by Altay and Gilardi (2024) for news headlines, showing it also applies in educational/scientific contexts.”

What the sources say

• Not checked yetcites Altay & Gilardi, 2024 source

In the paper

“The five multiple-choice comprehension items were presented in a randomized order for each participant to reduce order effects.”

In the paper

“Trust calibration emphasizes aligning reliance with system capability (Lee & See, 2004), while automation bias predicts over-or under-reliance as a function of agency and accountability cues (Parasuraman & Riley, 1997;Skitka et al., 1999).”

What the sources say

• ?Not checked — no full text of the sourcecites Lee & See, 2004 source
• ?Not checked — no full text of the sourcecites Parasuraman & Riley, 1997 source
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Skitka et al., 1999

In the paper

“In sum, this controlled experiment showed that disclosing AI authorship of an informational text has no impact on comprehension but significantly undermines trust and intended use, consistent with recent findings in the media domain.”

In the paper

“For example, Dubrov (2023) found that higher involvement with digital technologies was associated with lower interpersonal trust in families (Ibid.).”

What the sources say

• Not checked yet

  Checked before — its earlier verdict is held back until a new check runs.

  cites Dubrov, 2023 source

In the paper

“We applied minimal data quality filters: responses faster than 30 seconds for the entire instrument were excluded as implausible.”

In the paper

“No additional attention checks were embedded.”

In the paper

“The 5-item MCQ produced high mean scores (ceiling tendency), which reduces sensitivity to small between-group differences.”

In the paper

“Cronbach's ␣ for the Quality/Trust composite was .90;␣ for the Intent composite was .82.”