"Ownership, Not Just Happy Talk": Co-Designing a Participatory Large Language Model for Journalism

This work addresses the complex intersection of Large Language Models (LLMs) and the journalism industry, a domain grappling with significant financial and ethical challenges in the era of AI. It articulates a critical perspective on the limitations of commercial, “one-size-fits-all” foundation models for sector-specific applications, particularly in fields requiring high levels of trust, ethical discernment, and intellectual property control. The central contribution is the co-design and articulation of a “Newsroom Tooling Alliance” (NTA)—a proposed cooperative, journalist-owned and -governed LLM ecosystem.

The significance of this research lies in its paradigm shift from simply applying AI to journalism to fostering AI by and with journalism. It moves beyond efficiency gains to confront fundamental questions of ownership, control, and societal impact. By grounding the LLM’s design in the socio-technical realities and “constitutive tensions” of the news industry, the paper demonstrates how participatory methods can yield more holistic, equitable, and sustainable AI solutions than commercially driven approaches. It highlights that the true value proposition for AI in journalism extends beyond task performance to encompass trust, predictability, fair compensation for data, and the preservation of the human element in news creation.

Key Innovations:

• Cooperative LLM Model (Newsroom Tooling Alliance - NTA): The primary innovation is the conceptualization and co-design of the NTA. This is not merely a technical LLM, but a complete organizational and governance structure. It proposes a collective where news organizations contribute their copyrighted data under shared terms, with a journalist-led steering committee overseeing the fine-tuning of transparent, open-source LLMs specifically for journalistic tasks. This model aims to protect intellectual property, ensure fair revenue sharing, enable collective decision-making on AI use cases (including auditing for bias and accuracy), and provide an alternative to reliance on large tech companies.
• Methodological Application of Participatory Design Fiction: The paper innovatively employs participatory design fiction as a research probe. By presenting a hypothetical “Newsroom Tooling Alliance” proposal to diverse stakeholders, the researchers elicited rich, nuanced feedback on complex macro, meso, and micro-level tensions before any technical implementation. This approach allowed for a deeper exploration of organizational, ethical, and power dynamics, moving beyond simple feature requests to co-design the underlying social and governance structures critical for responsible AI.
• Identification of “Constitutive Tensions”: The research systematically identifies and categorizes the inherent conflicts and dilemmas facing journalism’s adoption of AI—spanning macro (market dynamics, financial pressures, audience shifts), meso (inter-organizational competition vs. cooperation, data sharing vs. protection), and micro (individual journalist concerns about efficiency vs. human element, skill expansion vs. job displacement). The NTA design explicitly responds to these tensions, making the solution robust and contextually relevant.
• Prioritization of Ownership and Control: The paper explicitly argues for and designs mechanisms that prioritize journalists’ ownership over their data and control over the AI tools built from it. This includes provisions for data provenance, opt-out clauses, and journalist-led governance, differentiating it from existing commercial models that often operate opaquely with significant power imbalances.

Main Prior Ingredients Needed:

• Large Language Models (LLMs) and Generative AI Fundamentals: A foundational understanding of LLM architecture, training methodologies (especially pre-training and fine-tuning), capabilities (e.g., text generation, summarization), and inherent limitations (e.g., hallucination, bias, data dependency). This technological context sets the stage for both the opportunities and challenges explored.
• Participatory AI (PAI) and Participatory Design Theories: Deep engagement with the principles and methodologies of participatory design, particularly as applied to AI systems. This includes an understanding of power dynamics in technology development, the importance of stakeholder involvement (especially those impacted by AI), and approaches to sharing decision-making power. Concepts from Participatory Action Research (PAR) and design fiction are directly employed.
• Journalism Studies and Media Economics: A comprehensive knowledge of the current landscape of the news industry, including its economic precarity, the impact of digitization and platformization, intellectual property rights, labor dynamics within newsrooms, and the historical relationship between journalism and technology. This domain-specific expertise informs the identification of key challenges and the design of contextually appropriate solutions.
• Qualitative Research Methodologies: Proficiency in qualitative data collection and analysis, specifically semi-structured interviewing and grounded theory. These methods were crucial for eliciting rich insights from diverse journalists and systematically synthesizing their perspectives into the identified “constitutive tensions” and desiderata for the NTA.
• Sociotechnical Systems Thinking: An understanding that AI systems are not purely technical artifacts but are deeply intertwined with social structures, organizations, and human practices. This perspective is vital for designing solutions that address not just technical performance but also organizational arrangements, ethical implications, and human values within a specific work context.