Source-linked AI summary
The Value, Benefits, and Concerns of Generative AI-Powered Assistance in Writing
Zhuoyan Li, Chen Liang, Jing Peng, Ming Yin
TL;DR
The paper asks how much people value generative AI writing assistance and how it changes writing experience and performance. A randomized experiment compares assistance types and writing modes; people value assistance, especially content generation for creative tasks, while reporting productivity and confidence benefits alongside accountability and diversity risks.
Problem
The benefits and risks of AI writing assistance leave open whether people value it financially and how assistance type and writing task shape its effects on writing experience and performance.
Source: §1 · pp. 1–2Method
A randomized experiment assessed financial value and writing outcomes across independent writing, ChatGPT editing, and collaborative AI content generation.
Source: §1 · p. 2Results
Participants were willing to forgo $0.85 for AI-primary content generation, and assistance increased productivity and confidence while direct generation also reduced accountability and writing diversity.
Source: §1 · p. 2, §6 · p. 16Takeaways & Limitations
Writing-assistant design could adapt assistance to user characteristics and task context while preserving human control over generated content and the final writing product.
Source: §6 · p. 16The study examined only argumentative essays and creative stories and two assistance types, so its conclusions may not generalize to other writing tasks or more interactive assistance.
Source: §5.5 · p. 15
Abstract
from arXiv · showhide
Recent advances in generative AI technologies like large language models raise both excitement and concerns about the future of human-AI co-creation in writing. To unpack people's attitude towards and experience with generative AI-powered writing assistants, in this paper, we conduct an experiment to understand whether and how much value people attach to AI assistance, and how the incorporation of AI assistance in writing workflows changes people's writing perceptions and performance. Our results suggest that people are willing to forgo financial payments to receive writing assistance from AI, especially if AI can provide direct content generation assistance and the writing task is highly creative. Generative AI-powered assistance is found to offer benefits in increasing people's productivity and confidence in writing. However, direct content generation assistance offered by AI also comes with risks, including decreasing people's sense of accountability and diversity in writing. We conclude by discussing the implications of our findings.
1 INTRODUCTION
The paper tests how much people value AI writing assistance and how it affects their writing experience and performance. It finds that value is higher for content generation and creative tasks, while AI assistance brings productivity and confidence benefits alongside costs to accountability and writing diversity.
Introduction
- The study uses a randomized experiment to assess financial value and changes in writing experience and performance.It compares independent writing with editing assistance and collaborative content generation.Source: §1 · p. 2
- Participants were willing to forgo $0.85 for ChatGPT content generation in AI-primary mode, equal to 28.3% of the $3 they would receive for writing independently.Its value was particularly salient when the task demanded more creativity.Source: §1 · p. 2
- Both editing and content-generation assistance improved productivity and writing confidence, especially for argumentative essays.Reported productivity benefits included reduced writing time and improved grammar.Source: §1 · p. 2
- AI-primary assistance was associated with lower writing satisfaction and self-expression, less willingness to take responsibility for issues, and reduced diversity across writings.The reported costs included concerns about misinformation and diminished ability to express oneself through writing.Source: §1 · p. 2
- People with higher confidence in their own writing abilities generally attached lower value to AI assistance.Source: §1 · p. 2
2 RELATED WORK
Prior work examines how writing assistants interact with writers and how LLMs can support varied writing tasks. This study also draws on willingness-to-pay methods to quantify the value people attach to assistance.
2.1 Interaction between humans and writing assistants
- LLM writing assistance extends beyond word and grammar suggestions to ideation, advanced revisions, and support for diverse writing tasks.Earlier tools included word-level and sentence-level suggestions, which could sometimes clash with users’ typing habits and adversely affect their experience.Source: §2.1 · p. 2
- Studies have also investigated risks of LLM assistance, including shifts in text sentiment, writers’ opinions, and perceived trustworthiness of AI-generated profiles.Source: §2.1 · pp. 2–3
2.2 Value estimation via “willingness to pay”
- Discrete-choice experiments estimate willingness to pay by asking people to choose between options that differ in an amenity and its wage.Prior studies applied this approach to work flexibility, workplace composition, monitoring, and online labor-platform reviews.Source: §2.2 · p. 3
3 STUDY DESIGN
The study uses a randomized human-subject experiment to measure the value of AI writing assistance and assess its effects on writing experience and performance. Participants choose between independent and AI-assisted paid writing jobs, complete a writing task, and report perceptions and performance measures.
3 STUDY DESIGN
- A randomized human-subject experiment compares AI-assisted writing with independent writing to measure assistance value and effects on perceptions and performance.Source: §3 · p. 3
3.1 Writing Task
- Participants wrote a 200–250 word argumentative essay or creative story within 45 minutes.The two task types were selected to vary in structure, purpose, and creative demand; a pilot assessed task difficulty.Source: §3.1 · p. 3
3.2 Experimental Treatments
- The three writing modes were independent writing, ChatGPT editing and polishing, and collaborative writing with ChatGPT generating the initial draft.In AI-primary mode, participants could steer revisions and incorporate generated text into their final composition.Source: §3.2 · pp. 3–4
3.3 Experimental Procedure
- Participants chose between a $3 independent-writing offer and a variable-pay offer for human-primary or AI-primary writing; the variable amount was randomly drawn from a set spanning $1.50 to $4.50.Source: §3.2 · p. 4
3.4 Analysis Methods
- A Probit model estimated AI-mode choice by wage premium, and its 0.5 choice threshold yielded the estimated value of assistance.Regression analyses compared writing perceptions and performance across modes while accounting for covariates and correcting p-values for multiple comparisons.Source: §3.4 · pp. 5–6
4 RESULTS
Participants valued AI writing assistance financially, especially content generation, and its estimated value varied by writing task, writer confidence, and ChatGPT familiarity.
4.1 Estimating the value of AI assistance
- Participants were willing to forgo $0.85 for AI-primary content generation, compared with $0.10 for human-primary editing assistance.Source: §4.1 · pp. 6–7
- Content-generation assistance was valued more for creative stories than argumentative essays ($1.24 vs. $0.45), while editing-assistance values varied little by task.Source: §4.1 · p. 7
- Across both writing tasks, higher-confidence writers valued AI assistance less than lower-confidence writers.Source: Figure 2 · p. 7, §4.1 · p. 7
- Greater ChatGPT familiarity was associated with lower valuations of both assistance modes for argumentative essays, but higher valuations of creative-writing editing assistance.Source: §4.1 · pp. 7–8, Figure 3 · p. 8
4.2 Understanding the impacts of AI assistance in people’s perceptions of the writing experience
AI assistance changed participants’ writing perceptions in task- and assistance-specific ways: it often boosted future-writing confidence, while content generation could reduce ownership, enjoyment, and accountability.
4.2.1 Cognitive load.
- AI assistance significantly reduced cognitive load for argumentative essays, but not creative stories; editing-assisted story writers reported higher load than AI-primary story writers.For essays, independent writers’ load was higher than both Human-Primary and AI-Primary writers’.Source: §4.2.1 · p. 8, Figure 4 · p. 8
4.2.2 Perceptions about the writing process.
- For argumentative essays, AI content generation lowered enjoyment and ability to express creative goals; AI-primary participants also reported lower process satisfaction than human-primary participants.Editing assistance in Human-Primary often received the highest process ratings, while AI-Primary often received the lowest.Source: §4.2.2 · pp. 8–9, Figure 5 · p. 9
4.2.3 Perceptions of the final writing outcome.
- Content generation reduced perceived ownership across both task types and, for argumentative essays, also reduced pride and perceived uniqueness.The ownership difference between independent and AI-Primary writing was significant; the reported pride and uniqueness differences were also significant.Source: §4.2.3 · p. 9, Figure 6 · p. 9
4.2.4 Change of confidence in writing.
- After essay writing, Human-Primary generally increased confidence across future task types; AI-Primary significantly increased confidence for future creative and other tasks versus independent writing.After creative-story writing, AI-Primary increased confidence for future argumentative and other tasks, but not creative writing; Human-Primary showed no significant boost.Source: §4.2.4 · p. 10, Figure 7 · p. 10
4.2.5 Perceptions of accountability.
- AI-primary writers were less willing to take responsibility for several criticisms of argumentative essays and for deceptive content in creative stories than independent writers.For essays, the reported differences concerned deceptive content, privacy invasion, and bias or discrimination.Source: Figure 8 · p. 10, §4.2.5 · p. 11
4.3 Understanding the impacts of AI assistance on people’s writing performance
Writing performance was assessed with objective measures. AI content generation shortened argumentative-essay completion time and reduced same-topic writing diversity, while both forms of assistance reduced language errors and neither significantly changed coherence.
4.3.1 Completion time.
- AI-primary participants completed argumentative essays significantly faster than both independent and Human-Primary writers.Editing assistance did not significantly affect completion efficiency.Source: §4.3.1 · p. 11
4.3.2 Grammar and spelling.
- Both editing and content-generation assistance significantly reduced grammar and spelling mistakes compared with independent writing across the two task types.All independent-versus-AI-assisted comparisons were reported at p<0.01.Source: §4.3.2 · p. 11, Figure 10 · p. 11
4.3.3 Content coherence.
- AI assistance showed no significant effect on the coherence of participants’ final articles.Coherence was assessed using topic modeling and semantic similarity between highly scored words within topics.Source: §4.3.3 · p. 11, Figure 10 · p. 11
4.3.4 Content diversity.
- AI-primary writing had significantly lower same-topic diversity than both other modes for essays on both metrics; for stories, it was lower than independent writing on both metrics.For creative stories, the remote clique score was also significantly lower than Human-Primary; the Chamfer-distance comparison was not reported as significant.Source: §4.3.4 · p. 12, Figure 11 · p. 12
5 DISCUSSIONS
The discussion finds that writing-task differences shape both users’ valuation of AI assistance and its effects, while content generation’s higher perceived value may not reflect its broader undesirable impacts. It proposes context-sensitive assistance and cautions that the findings are bounded by the study’s tasks, setup, and measures.
5.1 The moderating role of writing task nature
- Participants valued content generation more for creative stories, while productivity benefits such as reduced cognitive load and completion time appeared mainly for argumentative essays.The authors conjecture that the differences reflect the tasks’ distinct nature and objectives.Source: §5.1 · p. 12
5.2 Potential mismatch between human
- People value content generation assistance more than editing, although it brings a larger set of undesirable impacts, suggesting a mismatch between perceived price and “true value.”The authors offer possible explanations, including productivity considerations and the experiment’s relatively low stakes.Source: §5.2 · pp. 12–14
5.3 Implications for designing AI writing assistance
- The authors propose dynamic pricing and assistance tailored to user characteristics and task needs, alongside controls that preserve user direction over generated content.Suggested features include on-demand suggestions, pause controls, and intensity sliders; they also propose exploring ways to reduce over-reliance and writing homogenization.Source: §5.3 · pp. 14–15
5.4 The potential influence of study setup
- Because the experiment used low-stakes, domain-knowledge-free tasks and paid online writing jobs, its findings may not transfer to high-stakes expert work or settings driven by intrinsic writing motivation.The authors note that Prolific participation may have amplified the emphasis on extrinsic motivation.Source: §5.4 · p. 15
5.5 Limitations and future work
- The two selected tasks may not represent other writing activities, and the study tested only editing/polishing and free-form generation through its chosen interfaces and ChatGPT model.The authors call for studies spanning other tasks, models, interfaces, and functionalities.Source: §5.5 · pp. 15–16
- The study did not exhaustively measure writing perceptions and performance or examine long-term effects on writing skills, habits, reliance, and valuation.The authors identify possible unmeasured risks such as biased or misleading generated content and call for longitudinal research.Source: §5.5 · p. 16
6 CONCLUSION
The study examines how people value AI writing assistance and how it affects writing perceptions and performance. It finds productivity and confidence benefits alongside risks from direct generation, and recommends adapting assistance to users and tasks while preserving human control.
6 CONCLUSION
- People were willing to forgo payment for AI writing assistance, especially direct content generation for highly creative tasks, which also carried risks to accountability and writing diversity.The study reports productivity and confidence benefits alongside those risks.Source: §6 · p. 16
- The authors recommend tailoring assistance to user characteristics and writing tasks, while preserving users’ control over generated content and the final product.They also call for research across more writing tasks, assistant types, and user populations.Source: §6 · p. 16
- Future work should test whether the findings generalize across a wider range of writing tasks, AI assistants, and user populations.Source: §6 · p. 16
A TASK INTERFACES
The appendix identifies the interfaces used in the two AI-assisted writing modes.
A TASK INTERFACES
- Figures A1 and A2 show the task interfaces used in the human-primary and AI-primary writing modes, respectively.Source: A TASK INTERFACES · p. 18
B DEMOGRAPHIC INFORMATION OF PARTICIPANTS
The appendix reports the study sample and treatment allocation.
B DEMOGRAPHIC INFORMATION OF PARTICIPANTS
- The study included 379 Prolific workers who passed the attention check.Source: B DEMOGRAPHIC INFORMATION OF… · p. 18
- Of these workers, 183 were assigned to the independent-versus-human-primary treatment and 196 to the independent-versus-AI-primary treatment.Source: B DEMOGRAPHIC INFORMATION OF… · p. 18
- Table A1 contains the participants’ full demographic information.Source: B DEMOGRAPHIC INFORMATION OF… · p. 18
C FULL REGRESSION RESULTS
The regression analyses cover writers’ experiences, perceptions of their final articles, responsibility, and writing performance. Writing mode is the primary predictor, with independent writing as the reference and additional covariates included.
C FULL REGRESSION RESULTS
- The 19 dependent variables span writing experience, final-article perceptions, responsibility, confidence, completion time, grammar and spelling errors, and coherence.They include mental load, satisfaction, enjoyment, ease, creative expression, ownership, pride, and uniqueness.Source: C FULL REGRESSION RESULTS · p. 18
- Writing mode is the primary independent variable, with independent writing used as the reference category.Source: C FULL REGRESSION RESULTS · p. 18
- The models also incorporate demographic characteristics, payment, writing confidence, and familiarity with and usage frequency of ChatGPT.Source: C FULL REGRESSION RESULTS · p. 18
- Complete regression results are provided separately for argumentative essays and creative stories.Source: C FULL REGRESSION RESULTS · p. 18
D.1 Detailed results on people’s perceptions of the final writing outcome
Content-generation assistance was associated with lower perceived ownership of the final article across writing tasks. In argumentative writing, it was also associated with lower pride and uniqueness ratings in specified comparisons.
D.1 Detailed results on people’s perceptions of the final writing outcome
- AI-primary participants reported significantly lower ownership than independent and human-primary participants for argumentative essays, and than independent writers for creative stories.Source: §D.1 · p. 18
- In argumentative writing, AI-primary participants reported lower pride than independent writers (adjusted-p=0.002) and human-primary writers (adjusted-p=0.007).Source: §D.1 · p. 19
- Participants in AI-primary mode reported a noticeable decrease relative to independent-writing participants (adjusted-p<0.001) and human-primary-writing participants (adjusted-p=0.005).Source: §D.1 · p. 19
D.2 Detailed results on people’s writing performance
Both human-primary editing assistance and AI-primary content-generation assistance were associated with fewer grammar and spelling errors than independent writing, for both argumentative essays and creative stories.
D.2 Detailed results on people’s writing performance
- Both AI-assisted modes had fewer grammar and spelling errors than independent writing in argumentative essays and creative stories.For argumentative essays, independent versus human-primary and independent versus AI-primary comparisons each had p=0.003; for creative stories, the respective values were p<0.001 and p=0.005.Source: §D.2 · pp. 19–20