Voices of Freelance Professional Writers on AI: Limitations, Expectations, and Fears

arXiv:2504.05008 · cs.CL, cs.CY, cs.HC · Submitted 2025-04-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Voices of Freelance Professional Writers on AI".

Jane: The rapid development of AI-driven tools, particularly large language models (LLMs), is reshaping professional writing,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: Now that we know who wrote this—Anastasiia Ivanova from LMU Munich along with Fedorova, Tilga, and Artemova Toloka—we can dig into why they focused on these specific limitations and expectations. It’s about moving past just the excitement of the technology to the actual hurdles writers face.

Jane: Right, Tom. The title itself sets a tone because it immediately tells us this isn't a purely positive look; it’s about what’s holding people back, what they worry about regarding their craft, and what they actually expect from future AI tools.

Lu: What really stands out is the focus on voice and creativity concerns mentioned in the related work sections—it shows that writers aren't just worried about speed; they are deeply concerned about losing their unique way of expression when using these large models.

Meng: And thinking about the authors, seeing them from a European academic perspective gives us an interesting lens on how different cultural backgrounds might influence trust in these new systems, which is a very practical concern for developers to address.

Lalam: I think the implication here is that we need to build tools that respect the writer's identity, not just generate text; it’s about preserving that personal touch while using the AI as a collaborator.

The paper's summary: Tom: So, what does the actual research tell us about how writers are using these LLMs? It shows a pretty clear split: some people are using them heavily while others actively avoid them entirely because of certain issues they encounter.

Jane: The study found that there's this polarization among writers; you have people who lean into AI for efficiency, and then there's the group that is more cautious, maybe questioning the quality or worrying about accuracy.

Lu: The summary highlights how AI is being used for things like translation support and brainstorming, but it also points out that there's a significant struggle with maintaining language fidelity when writing outside of English.

Meng: That struggle with accuracy in non-English contexts is a major practical concern because if the AI messes up the nuance, it can seriously damage the final professional output and trust in its use.

Lalam: It also shows that while some writers feel they have to heavily edit everything, about half of them report that they don't see much change in their writing style even when using AI frequently.

The paper's improvements: Tom: Moving into suggestions for improvement, this research doesn't just stop at reporting problems; it actually proposes specific areas where developers need to step up their game to make these tools better for the user.

Jane: They point toward a few key fixes: first, they emphasize the need for much stronger fact-checking mechanisms and source attribution because misinformation is a real fear.

Lu: I agree with that, but they also strongly suggest enhancing cultural context awareness so the AI can flag or suggest edits when it detects potentially inappropriate references based on the writer's cultural background.

Meng: From an implementation view, this means building in explicit controls for style consistency so users can decide exactly how much the AI should change their tone versus just polishing grammar. That level of user control is what we need to see implemented practically.

Lalam: And I think the emphasis on personalization is crucial; if the AI can learn a specific writer's preferred vocabulary and rhythm, it moves from being a general tool to being a truly tailored assistant.

Conclusion: Tom: So, wrapping up on "Voices of Freelance Professional Writers on AI: Limitations, Expectations, and Fears," we see that while AI adoption is widespread—with seventy-three percent of respondents using these tools—the real discussion centers on language support, the risk of factual errors, and maintaining a unique writing voice.

Jane: It boils down to this: writers are excited about the potential for speed and creativity, but they are rightly concerned about cultural representation and authorship. The path forward seems to involve systems that offer greater transparency and more nuanced handling of different linguistic needs.

Lu: I just think it opens up so much creative possibility for how we design these assistants to be more attuned to human nuance rather than just focusing on output quantity.

Meng: From my side, the implication is that the next generation of AI tools needs to prioritize reliability and user-defined control over aggressive feature deployment.

Lalam: I'm really hopeful that these findings push us toward creating an environment where AI truly functions as a supportive collaborator for every writer, not just a fast typewriter.

Tom: That’s all the time we have for this discussion on the paper "Voices of Freelance Professional Writers on AI: Limitations, Expectations, and Fears." We'll be back next week with another look at some fresh research.

Anastasiia Ivanova, Natalia Fedorova, Sergei Tilga, Ekaterina Artemova

LMU Munich

cs.CL, cs.CY, cs.HC

Submitted: 2025-04-07

Updated: 2026-09-29

Code: https://github.com/Toloka/surveying-prof-writers-on-ai

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 70/100

The gist: The rapid development of AI-driven tools, particularly large language models (LLMs), is reshaping professional writing, and this research investigates how freelance professional writers perceive and

Key concepts

LLM Adoption
This refers to how freelance writers integrate large language models into their professional writing workflows. The study found that adoption is polarized: some writers rely heavily on AI for efficiency while others avoid it entirely, depending on their specific needs and tolerance for potential inaccuracies.
Domain and Style Adaptation
This examines whether AI can effectively handle different writing styles and subject matter across various languages. Writers noted that while AI helps with translation, adapting its output to maintain a specific cultural style or domain knowledge remains a significant challenge.
Authorship Rights and Ethics
This addresses the writers' concerns about who owns the final text when AI is involved. Most writers see AI as a tool, claiming full authorship, but there is some acknowledgment that human expertise and ethical considerations regarding originality still require careful navigation.

Terminology

Summary

The rapid development of AI-driven tools, particularly large language models (LLMs), is reshaping professional writing, and this research investigates how freelance professional writers perceive and utilize these technologies across various linguistic and ethical dimensions.

The findings demonstrate important insights reflecting upon the importance of: LLMs adoption for non-English speakers; the degree of misinformation, domain and style adaptation; usability and key features of LLMs.

How it works

The study conducted a large-scale questionnaire (N = 301 from 49 countries) and an interactive survey (N = 36) targeting freelance professional writers regularly using AI. The research examined LLM-assisted writing practices across over 25 languages, ethical concerns, and user expectations. The primary research questions addressed included: RQ1 How do freelance professional writers use AI chatbots? RQ2 How do AI chatbots impact writing experience? RQ3 How useful are they in languages other than English? RQ4 Do they represent and respect cultural knowledge from different user backgrounds well? RQ5 What are the main limitations and what improvements are expected in the future? And RQ6 What are the authors’ expectations and fears regarding their job security, text authorship rights and ethics in writing?

How it works

The methodology involved participant selection from Anonymized, a crowdsourcing platform for data labeling. Participants were verified freelance writers with professional backgrounds in relevant fields and self-disclosed proficiency in multiple languages. The questionnaire consisted of 64 questions organized into seven sections, covering demographics, AI chatbot usage patterns (including frequency and preferred chatbots), impact on writing tasks (advantages and disadvantages), experience with non-English language use, perception of cultural representation, limitations and expectations for future tools, and AI impact and ethical issues.

How it works

Data preprocessing involved encoding variables as ordinal or one-hot vectors based on the nature of the answer options. Factor analysis was performed on non-demographic and non-target variables to capture underlying attitude patterns, resulting in seven factors used for subsequent regression analyses to address research questions. Statistical methods included Spearman’s rank correlation coefficient for ordinal variables and chi-square tests for binary variables, with point-biserial correlation coefficients used to measure the strength of the relationship between binary and ordinal variables.

How it works

The results revealed that 73% of respondents use AI chatbots, with a polarization where writers tend to either avoid them or rely on them heavily. Factor analysis identified seven user profiles, including Users questioning AI-generated text quality and using other tools, Efficiency-focused users, and Heavy AI chatbot users. Efficiency-focused users particularly valued speed and cost reduction, while heavy AI chatbot users showed a correlation between positive impressions and willingness to pay.

How it works

Regarding the impact on writing practice, the main advantages cited were increased speed, quick responses, enhanced creativity, and translation support. Primary disadvantages included unnatural language, factual inaccuracies, and fabricated references. Furthermore, about half reported no writing style change despite frequent usage. The interactive survey showed that 58% of submissions required only minor or no edits to achieve satisfactory outputs.

How it works

Concerning non-English languages (RQ3), nearly half of multilingual respondents interact with AI only in English, and some lack language support. While most agree AI generates more accurate outputs in English, writers struggle with grammar and idioms in other languages. Translation dominates for non-English tasks, followed by summarization and brainstorming. However, 22% reported similar performance across languages for French, German, Spanish, Russian, Arabic, and Swahili.

How it works

On cultural knowledge (RQ4), over half of respondents consider cultural representation extremely important, though perceptions varied by region. No correlation existed between these preferences and geographic or language backgrounds. Writers face limitations such as loss of personal voice and lack of genuine creativity. Expectations for the future included a strong demand for fact-checking (58% of respondents), followed by up-to-date knowledge, personalization, and better handling of cultural content.

How it works

Regarding ethics and job security (RQ6), 43% view AI as a potential threat but believe human expertise remains essential. Nearly 50% claim full authorship of AI-generated text, viewing AI as a tool, while 18% support user ownership without full authorship. Writers who view themselves as collaborators with AI showed intermediate levels of both disclosure willingness and threat perception.

How it works

The interactive task results indicated that two dominant prompting strategies were used: (1) specifying tone and audience with targeted revision instructions, and (2) requesting multiple variations and synthesizing them. The average similarity score between AI-generated drafts and final edited versions was 0.84, suggesting most outputs require only polishing. Qualitative feedback highlighted that editing efforts focused on "improving naturalness, correcting grammar and style, and refining tone.

Improvements for AI systems

Based on the research presented in Voices of Freelance Professional Writers on AI: Limitations, Expectations, and Fears, here are specific improvements that can be implemented in AI systems to better serve freelance professional writers:

The following improvements focus on addressing the identified limitations across linguistic diversity, cultural nuance, voice preservation, and ethical trust.

  1. Acknowledge and Support Language-Specific Proficiency Gaps:

  2. Implement Dynamic Linguistic Adaptation for Non-English Users:

  3. Enhance Cultural Context Awareness in Generative Outputs:

  4. Develop Voice Preservation and Style Consistency Modules:

  5. Improve Factuality Verification and Source Attribution Mechanisms:

  6. Integrate User Preference Profiling for Personalized Interaction:

Specific capabilities of the improved AI system based on these improvements:

  1. Acknowledge and Support Language-Specific Proficiency Gaps:

  2. Implement Dynamic Linguistic Adaptation for Non-English Users: The system should move beyond simple translation or grammatical correction in non-English languages. It must be trained to recognize and respect regional dialects (e.g., Egyptian Arabic, Nigerian Pidgin English) and idioms, as current models struggle with these nuances (as noted in the interactive task results).

  3. Enhance Cultural Context Awareness in Generative Outputs: The AI should incorporate a mechanism that cross-references generated content against cultural knowledge databases to proactively flag or suggest modifications for culturally inappropriate references, festive contexts, or social norms. This directly addresses the feedback regarding lack of cultural context and cultural inaccuracies.

  4. Develop Voice Preservation and Style Consistency Modules: The system needs features that allow writers to explicitly define and maintain a unique voice. This could involve:

  5. Integrate User Preference Profiling for Personalized Interaction: The AI should learn the user's preferred tone, vocabulary complexity, sentence structure patterns, and rhetorical style over time (as suggested by the need for personalization). It must offer fine-grained control over style influence—allowing users to toggle between minimal stylistic change and complete voice transformation.

  6. Improve Factuality Verification and Source Attribution Mechanisms: The system should integrate robust tools that allow writers to easily verify claims. This includes:

  7. Fact-checking capabilities, allowing the AI to cite reliable sources (as a top expectation), provide immediate flags for fabricated references, and offer confidence scores for generated statements.

  8. Integrate User Preference Profiling for Personalized Interaction: The system should be designed as an assistant only, prioritizing collaboration over replacement, and explicitly delineate where human expertise remains essential (addressing the AI as assistant only sentiment).

  9. Develop Voice Preservation and Style Consistency Modules: The AI should provide iterative suggestions focused on improving naturalness and reducing redundancy, specifically by offering multiple rewording options that align with the user's desired stylistic parameters.

Abstract

The rapid development of AI-driven tools, particularly large language models (LLMs), is reshaping professional writing. Still, key aspects of their adoption such as language support, ethics, and long-term impact on writers' voice and creativity remain underexplored. In this work, we carried out a questionnaire (N = 301) and an interactive task (N = 36) targeting freelance professional writers regularly using AI. We examined AI-assisted writing practices across 25+ languages, ethical concerns, and user expectations. Our findings reveal that adoption is shaped more by peer influence and professional outlook than demographics, that multilingual writers face both performance and perception barriers to equitable AI use, and that job security concerns are widespread and adoption-independent.

Sources

Related papers