Usefulness–Risk Paradox: AI Writing Tool Adoption

This post explains how the PLOS One study by Li, Liu, and Zou (published August 7, 2026) documents a usefulness–risk paradox in AI writing tool adoption and what that means for AI-enabled qualitative research. The primary keyword is AI writing tool adoption and this article is written for qualitative researchers, UX teams, and educators who need concrete guidance for analyzing open-ended survey responses, interview transcripts, and usage reports from students. According to the PLOS One study, 518 valid questionnaires were collected between January 25 and January 30, 2026, and results were analyzed by structural equation modeling to test an extended Technology Acceptance Model.

Key Takeaways

According to the PLOS One study by Li et al. (published August 7, 2026) PLOS One, Chinese undergraduates reported simultaneous perceptions of usefulness and risk when using AI writing tools, a pattern the authors call the usefulness–risk paradox.

518 valid responses were retained after data cleaning from an online survey run January 25–30, 2026, as reported in PLOS One.
The PLOS One analysis found a strong AI self-efficacy effect on ease of use (ASE → PEOU β = 0.792, p < 0.001) and a positive association between perceived usefulness and risk (PU → OPR β = 0.155, p = 0.030) in August 2026.
The PLOS One authors summarize that "functional value and risk awareness may coexist rather than offset one another in L2 writing contexts, " and they caution that "the findings should be interpreted as associations" because the data are cross-sectional.
Researchers should expect engagement to correlate with both perceived gains and perceived harms: the PLOS One model reported ATT → BE β = 0.804 (p < 0.001), BE → WPOS β = 0.679 (p < 0.001) and BE → WNEG β = 0.392 (p < 0.001).

What happened and how the study measured it

What happened: the PLOS One study surveyed undergraduate students with prior experience using AI writing tools and tested an extended Technology Acceptance Model to link AI self-efficacy, perceived ease of use, perceived usefulness, multidimensional perceived risk, attitude, behavioral engagement, and perceived writing outcomes.

According to PLOS One, the authors distributed the online questionnaire via Wenjuanxing from January 25–30, 2026 and retained 518 valid responses after excluding patterned answers and ineligible respondents.

According to PLOS One, all constructs were measured on 7-point Likert scales and analyzed with confirmatory factor analysis and structural equation modeling using WLSMV estimation in Mplus 8.3 to respect the ordinal nature of the data.

According to PLOS One, key model results included ASE → PEOU (β = 0.792, p < 0.001), PEOU → PU (β = 0.584, p < 0.001), PEOU → OPR (β = −0.465, p < 0.001), and PU → OPR (β = 0.155, p = 0.030), while OPR → ATT was non-significant (β = 0.015, p = 0.652).

Findings snapshot

Date	Metric	Value	Implication
Jan 25–30, 2026	Survey responses collected	518 valid questionnaires	Sizable convenience sample for associative analysis
Aug 7, 2026	Publication	PLOS One article (Li et al., 2026)	Peer-reviewed open access evidence for the usefulness–risk paradox
Sample composition (reported in PLOS One)	Use frequency	48.8% occasional, 45.4% frequent, 5.8% near-constant	Most respondents had some ongoing experience with AI writing tools
Model coefficients (reported in PLOS One)	Key paths	ASE→PEOU β=0.792; PEOU→PU β=0.584; PU→OPR β=0.155	Quantifies perceptions linking competence, utility, and risk awareness
Behavioral outcomes (reported in PLOS One)	Engagement→Outcomes	BE→WPOS β=0.679; BE→WNEG β=0.392	Engagement associated with both perceived benefits and harms

Implications for qualitative researchers studying AI writing tool adoption

Answer: qualitative researchers should design instruments to capture dual-awareness (both functional appreciation and risk consciousness) because the PLOS One study shows these attitudes can coexist.

According to PLOS One, perceived usefulness can correlate with higher risk awareness (PU → OPR β = 0.155, p = 0.030), so qualitative coding schemes should include codes for positive affordances (idea generation, language scaffolding, efficiency) and for concerns (academic integrity, dependence, privacy).

According to PLOS One, behavioral engagement related strongly to both perceived positive and negative outcomes, so mixed-method designs that combine open-ended responses with frequency measures will let researchers trace how specific engagement patterns map to perceived gains and harms.

Method practicalities: capture prompt examples, student rationales for accepting or rejecting AI outputs, and contextual rules instructors set. Use time-stamped diaries or recent-task interviews to reduce recall bias that the PLOS One authors note is a limitation of cross-sectional self-report data.

How Evidano Helps

Problem: Large open-ended datasets are slow to synthesize → Solution: AI-assisted thematic analysis

Evidano is an AI-powered qualitative data analysis platform that helps researchers analyze interviews, open-ended surveys, and documents.

Evidano can ingest 518 transcripts or responses like the PLOS One dataset, auto-generate thematic codes, and surface frequency and co-occurrence patterns that let researchers compare utility-related and risk-related themes at scale; see Evidano features.

Evidano’s thematic outputs are exportable and auditable, which helps researchers transparently connect qualitative themes to the PLOS One-style SEM findings.

Problem: Segment-level differences (e.g., major, use frequency) are hidden → Solution: cross-segment analysis

Evidano supports cross-segment analysis so you can compare themes and sentiment for subgroups reported in the PLOS One sample, such as STEM vs humanities students or occasional vs frequent users.

Evidano’s cross-tab and visualization tools help evidence whether the usefulness–risk paradox appears across demographics or is concentrated in particular subgroups.

Problem: Non-English responses and noisy transcripts → Solution: transcription and translation pipelines

Evidano provides transcription and translation features for multilingual datasets and supports custom dictionaries and PII redaction that are useful when students paste drafts or prompts into tools, a privacy concern highlighted by PLOS One.

Using these pipelines lets qualitative teams align multilingual responses for consistent coding and comparison.

Problem: Stakeholders want concise quotes and model validation → Solution: AI chat and audit-ready exports

Evidano’s AI chat over your documents enables rapid evidence summaries and extracts quotations tied to codes, which helps justify statements like the PLOS One quote that "functional value and risk awareness may coexist rather than offset one another in L2 writing contexts."

Evidano provides exportable codebooks, frequency tables, and visualizations for inclusion in reports and publications.

FAQ: AI writing tool adoption

What is the usefulness–risk paradox in AI writing tool adoption?

Answer: the usefulness–risk paradox is the empirical observation that users can perceive both high usefulness and high risk for the same AI writing tools.

According to PLOS One, the authors describe this paradox by showing a positive association between perceived usefulness and overall perceived risk and by noting that "functional value and risk awareness may coexist rather than offset one another in L2 writing contexts."

Can the PLOS One findings be interpreted as causal evidence that AI improves or harms writing?

Answer: no, the PLOS One authors explicitly state the findings are associative because the study uses cross-sectional self-report data.

According to PLOS One, the authors caution that the design cannot establish objective writing gains or causal effects and that outcomes are students’ perceived experiences rather than independently assessed writing scores.

Which quantitative and qualitative measures from the PLOS One study are most useful to replicate?

Answer: combine multi-item Likert scales for ASE, PEOU, PU, OPR, ATT, BI/AU and open-ended prompts asking students for examples of AI outputs they used and concerns.

According to PLOS One, the study used validated multi-item scales on 7-point Likert scales, collected demographic and use-frequency data (48.8% occasional, 45.4% frequent, 5.8% near-constant), and supplemented with open responses for balanced interpretation.

How can qualitative teams use AI-enabled tools to analyze similar datasets efficiently?

Answer: use AI-assisted ingestion, thematic coding, and cross-segment frequency analysis to accelerate pattern detection while keeping humans in the loop for validation.

Evidano’s workflow supports that approach: ingest transcripts or survey text, generate candidate codes, review and refine codes with subject-matter experts, then produce code frequency and co-occurrence visualizations for reporting.

Conclusion & Next Steps

The PLOS One study (Li et al., published August 7, 2026) shows that Chinese undergraduates can simultaneously perceive AI writing tools as useful and risky, a pattern with direct implications for qualitative study design and institutional policy.

According to PLOS One, the study’s key numeric results include a 518-person sample collected January 25–30, 2026 and model paths like ASE→PEOU β = 0.792 (p < 0.001) and PU→OPR β = 0.155 (p = 0.030), which you can use as benchmarks when planning mixed-method replication.

If you need an AI-enabled qualitative workflow to replicate, extend, or audit analyses like the PLOS One study, consider tooling that supports thematic coding at scale, cross-segment comparisons, and secure transcription and translation.

To test this approach on your own L2 writing or educational datasets, Try Evidano for free.

Topics

AI writing tool adoption
usefulness-risk paradox
qualitative analysis AI tools
AI in L2 writing

Keep reading

Browse all articles

Usefulness–Risk Paradox: AI Writing Tool Adoption

Key Takeaways

What happened and how the study measured it

Findings snapshot

Implications for qualitative researchers studying AI writing tool adoption

How Evidano Helps

Problem: Large open-ended datasets are slow to synthesize → Solution: AI-assisted thematic analysis

Problem: Segment-level differences (e.g., major, use frequency) are hidden → Solution: cross-segment analysis

Problem: Non-English responses and noisy transcripts → Solution: transcription and translation pipelines

Problem: Stakeholders want concise quotes and model validation → Solution: AI chat and audit-ready exports

FAQ: AI writing tool adoption

What is the usefulness–risk paradox in AI writing tool adoption?

Can the PLOS One findings be interpreted as causal evidence that AI improves or harms writing?

Which quantitative and qualitative measures from the PLOS One study are most useful to replicate?

How can qualitative teams use AI-enabled tools to analyze similar datasets efficiently?

Conclusion & Next Steps

Topics

Keep reading

Company

About

Preview Evidano

Account

Newsletter