Writing with Ghosts: Creativity and Authenticity in AI-Assisted Screenwriting

State of Research

Screenwriting is characterized by rigid industry conventions, such as strict formatting rules, established dramatic structures, and specific storytelling expectations that require clarity, precision, and narrative coherence [9]Green (2023) - The Artist's Code: Technology and Optimization of Creativity. In contrast to more freely associative or experimental writing, professional screenwriting emphasizes functional use of language, structured narrative arcs, and efficient storytelling to visually convey meaning and emotion through subsequent film production [10]Pasolini (1986) - The Screenplay as a Structure That Wants to Be Another Structure, [11]Price (2012) - The screenplay: an accelerated critical history. These structured requirements align closely with the strengths of large language models (LLMs), making screenwriting a particularly insightful domain for studying human-AI collaborative practices.

Specialized screenwriting software has been used by professional screenwriters since the 1980s to ensure adherence to industry-standard formatting conventions and to track narrative structure and dramatic arcs [12]Plagens (2025) - Reinventing the typewriter: history of screenwriting software. The use of generative AI for screenwriting is a more recent phenomenon. Early experiments, such as the 2016 short film Sunspring (written by an LSTM neural network), demonstrated that machines could generate screenplay text, albeit with surreal and incoherent results [13]Cohn (2021) - The scientist of the holy ghost: 'sunspring' and reading nonsense. Modern LLMs [14]Vaswani et al. (2017) - Attention is all you need such as OpenAI's GPT-3 [15]Brown et al. (2020) - Language models are few-shot learners (GPT-3), GPT-4 [16]OpenAI et al. (2023) - GPT-4 Technical Report, or Anthropic's Claude [17]Anthropic (2025) - The Claude 3 model family have significantly improved language coherence, sparking interest in whether they can produce complete scripts.

A variety of research projects on dramatic writing with LLMs have since been conducted, but these studies mostly lack the input from or participation of professional writers, relying instead on crowdworkers recruited from Amazon Mechanical Turk [18]Lee et al. (2022) - CoAuthor: human-AI collaborative writing dataset, graduate students with creative writing aspirations [19]Yang et al. (2022) - AI as an active writer in collaborative fiction writing, screenwriting students [20]Cake (2025) - AI as a collaborative tool for script development, or a mix of nonprofessionals and professionals [21]Chung et al. (2022) - TaleBrush: sketching stories with GPT models, [22]Tang et al. (2025) - Understanding screenwriters' practices in human-AI co-creation. This reliance on non-experts, who often prefer novice work to professional output [23]Pawar et al. (2024) - AI and critics: exploratory study on creative content, is problematic because it overlooks the tacit expertise [24]Polanyi (1967) - The tacit dimension, [25]Schön (1983) - The reflective practitioner that experienced screenwriters possess regarding narrative structure, characterization, dialogue authenticity, and subtle dramatic elements. In addition, properly evaluating a screenplay requires mentally translating written descriptions into their intended audiovisual form, a specialized skill [26]Walter (2010) - Essentials of screenwriting not typically available to lay readers. Consequently, evaluations risk overlooking critical shortcomings of LLM tools, resulting in AI-generated screenplays that appear superficially coherent, but ultimately fall short of artistic standards and professional usability.

Breaking from this pattern, a limited number of studies have exclusively involved professional writers, including [27]Mirowski et al. (2022) - Co-writing screenplays with language models, which engages professional playwrights to evaluate a large language model-based co-writing system, [28]Gero (2023) - AI and the writer: how language models support creative writers, which develops AI-based writing aids that are tested on professional poets, and [29]Ippolito et al. (2022) - Creative writing with AI-powered writing assistant, which focuses exclusively on published fiction writers. However, to the best of our knowledge, no study has been published that focuses on the use of LLMs by professional screenwriters.

In addition, existing studies consistently take a top-down approach: researchers first define workflows, frameworks, and dedicated software tools, and then proceed to test these implementations on writers. While helpful in the early stages of generative AI screenwriting—when AI capabilities were still limited and had to be augmented by customized programming—this approach overlooks lessons learned directly from users' own independent and spontaneous adoption practices. Given the striking advances in modern generative AI systems, creative professionals across many fields now regularly incorporate these technologies into their workflows in unique and unexpected ways [30]Mathur (2025) - AI as a creative collaborator in art. In response to this widespread individual experimentation and adaptation, our novel research design proposes guided auto-ethnography [31]Markham et al. (2024) - Aggregation and guided autoethnography as a method that enables creative practitioners to deeply reflect on and analyze their own lived experiences. This bottom-up approach not only captures nuanced practices that traditional observational methods may miss but also foregrounds the reflective insights of practitioners directly engaged in everyday creative production.

Finally, previous research has focused almost exclusively on the English-speaking world (with [22]Tang et al. (2025) - Understanding screenwriters' practices in human-AI co-creation as a notable exception). The proposed research would be the first publication to examine the use of LLMs by professional German-language screenwriters, gaining additional relevance due to well-documented disparities in large language model performance between English and other languages [32]Zhang et al. (2023) - Don't trust ChatGPT when your question is not in English. Expanding academic research beyond the English-oriented mainstream provides an opportunity to clarify these limitations, raise awareness among software developers, and ultimately spur progress toward more effective multilingual AI.

We are aware of recent examples where filmmakers have completely scripted and directed feature films using large language models, e.g. [33]Lis (2023) - Il diario di Sisifo. However, such cases are intentionally excluded from this research, as these experiments are primarily aimed at proving technical feasibility rather than achieving artistic excellence. Evaluations suggest that such fully AI-scripted films generally exhibit narrative and emotional mediocrity, demonstrating the current limitations rather than the creative potential of generative AI in professional artistic practice [34]Maher (2025) - The Last Screenwriter review. Since our research explicitly focuses on creative practice by experienced professionals, these cases fall outside our targeted research interest.

Research Plan

Goals

The primary goal of this research is to critically examine how professional screenwriters integrate large language models (LLMs) into their workflows, and how AI impacts their perceptions of creativity, originality, authenticity, and artistic identity.

Specifically, the study seeks to:

1. Understand how AI-assisted screenwriting shapes practitioners' views of their own craft and creative voice
2. Identify mental models and metaphors used by screenwriters to conceptualize their interactions with LLMs
3. Identify unmet creative and productivity needs among professionals and clarify desired AI features to fill these gaps
4. Document and analyze practitioners' strategies for balancing skepticism about authenticity, originality and ethical considerations with pragmatic AI adoption driven by industry pressures and productivity demands
5. Provide perspectives on how cultural contexts and industry structures in Germany, Austria, and Switzerland influence practitioners' approaches to AI adoption
6. Extract and share effective workflow techniques and creative strategies employed by professional screenwriters using LLMs to facilitate knowledge sharing and cross-disciplinary learning among practitioners in screenwriting and other artistic fields

Methods

For this study, we focus specifically on screenwriters who do not routinely direct their own scripts (non-directing screenwriters) to ensure clarity regarding AI's influence on writing practices, rather than on broader filmmaking processes. Unlike writer-directors—who have full creative authority to adapt scripts during the production phase—non-directing screenwriters must communicate their creative intent primarily through text alone. This makes their reliance on precise narrative articulation especially critical, highlighting the direct impact of AI tools on writing quality, narrative coherence, and creative authenticity without the confounding influences inherent in directing.

The research will focus on screenwriters working in Germany, Austria, and Switzerland, in order to gain insights into culturally similar but structurally distinct markets. Although these countries share linguistic characteristics and some cultural underpinnings, their film and television industries are structurally diverse, characterized by differences in market size, levels of public funding, production ecosystems, and professional practitioner networks.

Phase 1: Online Survey

The first phase of the data collection will be carried out in close cooperation with three professional screenwriters' associations: Deutscher Drehbuchverband e.V. (Germany), drehbuchVERBAND Austria and ARF/FDS Verband für Filmregie und Drehbuch (Switzerland). Through these partnerships, we will distribute a comprehensive, anonymous online survey to the members of the associations. The online questionnaire will target approximately 200–300 professional screenwriters from various sectors (theatrical feature films, TV movies and TV/streaming series) and will provide comprehensive statistics on screenwriters' frequency of use of LLMs, perceived benefits, concerns, and unmet creative and productivity needs.

Phase 2: Guided Auto-Ethnographic Study

From within the cohort of respondents, we will select a subset of four to six screenwriters who demonstrate a particular engagement with, or reflective interest in, the AI-assisted writing process. These selected practitioners will be invited to participate in a guided auto-ethnographic study. Over the course of four to six weeks, they will use a think-aloud protocol to verbalize their real-time thoughts, decisions, and immediate emotional responses as they write and interact with LLMs. Audio recordings of these think-aloud sessions will be transcribed using AI for further analysis. In addition, participants will engage in systematic reflexive journaling to document reflections, techniques, emotional reactions, and evolving attitudes before and after their writing sessions. Over the course of actual professional projects, these methods will capture a nuanced narrative of the screenwriters' lived experiences with LLMs in their workflow, crystallizing experiential dynamics not fully accessible through surveys or interviews alone. By treating the screenwriters as artistic co-investigators, the project will be able to capture rich qualitative data over time, not just in one-off trials. Recent human-computer interaction studies have noted the need for such longitudinal, practice-based insights, noting that few studies so far have examined the "long-term qualitative experience" of writers actively co-creating with AI.

Phase 3: Focus Group Discussion

In a final research step, the researchers will organize a structured focus group discussion (either in person or online, depending on logistical feasibility) with the auto-ethnographic subgroup participants and three additional screenwriters selected from the survey respondent cohort to mitigate biases introduced by the intensive auto-ethnographic process. This group will triangulate previously collected data and reflections, critically discuss and validate key findings, confirm thematic insights, clarify recurring tensions, and collaboratively envision potential solutions or developments in AI-driven creative tools.

This methodologically integrative approach leverages the complementary strengths of quantitative survey data (breadth and representativeness), auto-ethnographic reflections (nuanced phenomenological understanding), and confirmatory focus group discussions (collective validation).

Expected Results

The project will provide theoretical and practical insights into co-creativity, illuminating how human imaginative processes are augmented or perturbed by AI's contributions, and how writers negotiate creative control with a non-human collaborator. By documenting the push and pull between a screenwriter's creative intuition and the AI's suggestions, the project offers a new understanding of collaborative creativity that goes beyond existing models of computational creativity. This aligns with emerging art research that treats AI as an ontological other or a virtual collaborator in the creative process [38]Choi & DiPaola (2023) - AI as other: art-as-research approach to generative AI.

Based on our methodological framework, we expect to provide a detailed account of professional screenwriters' current attitudes, adoption levels, usage patterns, perceived strengths, concerns, and expectations towards LLMs in German-speaking screenwriting ecosystems. In addition, we aim to shed light on the assumptions and attitudes that underlie LLM usage and help scholars and industry stakeholders understand the relationship between human creators and AI. Rich qualitative accounts from auto-ethnographic practices will further elucidate the experiential dynamics of engaging in AI-assisted screenwriting. These phenomenological findings will shed light on how the boundaries of authorship, authenticity, and artistic originality are negotiated, contested, and potentially redefined through everyday practice. Finally, we aim to generate actionable insights for AI developers, policymakers, professional associations, and creative practitioners by outlining emerging user needs, tool improvements, ethical considerations, and guidelines for effectively integrating AI tools into screenwriting practice while safeguarding authenticity and creative autonomy.

Results will be disseminated through peer-reviewed scholarly publications and presentations, primarily to communities active in human-computer interaction (e.g., ACM CHI), computational creativity, media studies, creative industries research, and practitioners in the creative fields.

Possible Risks

Several potential risks should be considered:

1. Recruitment and participation: Engaging professional screenwriters—especially those who are skeptical of or indifferent to AI technologies—may be challenging. If recruitment proves difficult, additional outreach methods, incentive structures, or collaborations with industry partners will be implemented.
2. Methodological resistance: Practitioners may perceive auto-ethnographic documentation as overly demanding or irrelevant to their workflow, risking low participation rates or participants dropping out during the study. To mitigate this, participants will be actively involved in designing the data collection procedures, ensuring methods align closely with their routines and professional comfort levels.
3. Technological change: Given the rapid evolution of AI technology, the landscape of available tools and perceptions of the creative role of AI may shift significantly during the study period. To address this, we will maintain flexibility in our data collection and analysis frameworks to allow for adjustments as new technological developments emerge.
4. Data overload and analysis: The mixed-methods approach generates significant amounts of data, making timely and coherent analysis challenging. To mitigate this risk, we will establish clear analysis protocols in advance and spread the data collection and analysis phases evenly over the duration of the project.
5. Bias risk: Self-selection of participants and auto-ethnographic reflection are subject to subjective bias. To address this, we will practice transparent and reflexive analytic strategies, clearly delineate findings derived from different methods and acknowledge potential limitations.
6. Ethical and privacy risks: Participants' reflections and disclosures about creative processes and AI integration may contain sensitive professional information. Strict confidentiality measures, anonymization protocols, secure data handling, and informed consent forms that clearly articulate participants' rights and data use will be strictly adhered to.

Potential Impact of the Research Project

The research promises to have impact in three main areas:

1. Scholarly Contribution

This study provides a significant scholarly contribution to ongoing discussions in human-computer interaction (HCI), computational creativity, media and cultural studies, and creative industries research. By systematically focusing on the lived experiences of professional screenwriters, this research fills an important empirical gap, shifting the focus away from theoretical speculation or laboratory-style experimentation toward real-world practice and nuanced reflection. In doing so, the project advances a deeper theoretical understanding of creativity and authorship in the age of generative AI.

2. Practical and Professional Implications

Findings from the study are expected to help inform professional screenwriting practices by providing practitioners with a clearer understanding of effective strategies for integrating AI into their creative workflows. Industry stakeholders (writers' associations, creative guilds, policymakers, and AI software developers) will benefit from clear, actionable recommendations that help shape effective and ethical AI integration strategies that protect the rights and interests of creative workers while supporting technological innovation.

3. Social and Cultural Implications

By openly addressing both the hopes and fears associated with AI integration, the research will foster a more nuanced public discourse on creativity, authenticity, and technological change. It will provide a clearer understanding of the broader cultural implications of AI, enabling audiences, artists, industry professionals, and policymakers to make more informed and reflective decisions about the directions, boundaries, and ethical frameworks of future creative technologies.