A Taxonomy of LLM-Supported Tasks for Critical AI Literacy in Journalism

Synthesis: Dierickx, Bjerknes, Opdahl, & Lindén (2026) develop a task-based taxonomy for integrating large language models (LLMs) into journalism education, organized around the four stages of the news workflow — newsgathering, sensemaking, editing, and publication/distribution. Drawing on task definitions from machine learning and natural language processing alongside journalism studies, the taxonomy translates journalistic activities into computationally-informed tasks, each linked to a baseline prompt and an explicit risk-and-mitigation strategy. By framing prompting as a situated, reflexive professional practice rather than a technical skill, the paper advances critical AI literacy — foregrounding bias, hallucination, overreliance, and the enduring need for human editorial oversight.

Dierickx, Bjerknes, Opdahl, & Lindén (2026) address a gap in journalism education: prompting is increasingly recognized as an educational tool but remains underdeveloped there, with existing work focusing on AI use in newsrooms rather than translating journalistic work into structured pedagogical tasks. They ask: How can prompting be structured as a pedagogical practice in journalism education, drawing on task definitions from natural language processing and machine learning?

The taxonomy: four stages of the news workflow

The paper organizes journalistic work into four stages that mirror the news production process, each with tasks that can be meaningfully supported by LLMs while remaining compatible with journalism's norms of verification, editorial judgment, and ethical responsibility:

Each task is defined, given a model-agnostic baseline prompt, and paired with a risk-and-mitigation strategy (e.g. linking's risk of over-interpretation, false authority, hallucination and confirmation bias is mitigated through source citation, cross-checking with primary sources, and human validation). The same computational operation (e.g. generation) recurs across stages because it serves different editorial objects with different risks.

Prompting as situated professional judgment

A central contribution is reframing prompting. Rather than treating it as a neutral technical act, the paper frames task definition as a form of journalistic judgment that must be learned, articulated, and critically assessed. Three ideas anchor this:

• Prompting is a form of writing. Because Large Language Models (LLMs) outputs result from the interplay of instructions, training data, and system processes — and because language, thinking, and knowledge are mutually constitutive in writing — a prompt shapes its output in the same way writing shapes thought. Task definition is therefore not neutral but a constitutive element of knowledge production.
• Task definition makes assumptions visible. Expressing journalistic work as explicit prompts surfaces assumptions, priorities, and ethical considerations that are more visible than in traditional newsroom practice. Defining tasks thus becomes a method for teaching editorial responsibility and awareness of the limits of automation.
• High/low-task divides are not fixed. The distinction between automatable low-level tasks and human-only high-level tasks is not a property of the tasks themselves but a relationship shaped by available technologies. As LLMs reshape what can be delegated, the question becomes how journalistic judgment, domain knowledge, and algorithmic capability work together.

Addressing the four challenges of GenAI in postdigital education

The taxonomy practically addresses Jandrić's (2026) four challenges of generative AI in postdigital education:

• Detection — exposing students to the possibilities and limitations of LLMs through repeated task formulation and output evaluation
• Sourcing — grounding outputs in evidence, highlighting that LLM output is based on probabilistic patterns rather than verified knowledge
• Writing — requiring critical evaluation of the reasoning behind generated text
• Attitude — promoting responsible and reflexive engagement, and justifying the decision to delegate tasks to LLMs

In the classroom, the approach shifts assessment from evaluating the "quality" of AI-generated text to articulating tasks, justifying prompt choices, and reflexively evaluating AI behavior. Rather than positioning human review as a final corrective step after an autonomous AI process, each task is treated as an ongoing negotiation in which journalistic judgment, professional norms, and algorithmic output are shaped together through iterative prompting and revision.

Transferability beyond journalism

While grounded in journalism, the taxonomy's underlying logic is transferable to other knowledge-intensive professions — law, medicine, public policy — where accuracy, source evaluation, and accountability are at stake, and where practitioners increasingly decide what can be delegated to algorithmic systems. It understands AI literacy as a form of professional reasoning learned through the practice of defining tasks, formulating prompts, and critically evaluating outputs in context, and it argues the ability to articulate why a task is or is not suitable for LLM support is more valuable than familiarity with any particular tool.

What this means for practice

• Instructors. Teach prompting as situated professional judgment, not a technical trick: have students define the task, justify their prompt choices, and evaluate the output against journalism's norms of verification and editorial responsibility.
• Instructors. Structure exercises around the four news-workflow stages (newsgathering, sensemaking, editing, publication/distribution) so every task is paired with an explicit risk and mitigation — for example, linking paired with source citation, cross-checking, and human validation.
• Instructors. Move assessment away from the quality of AI-generated text toward articulating tasks, justifying prompts, and reflexively evaluating model behavior.
• Instructors. Require students to argue why a task is or is not suitable for LLM support, under what conditions, and with what consequences for accuracy and accountability.
• Instructors. Address the four challenges of GenAI in postdigital education — detection, sourcing, writing, and attitude — through repeated task formulation rather than one-off tool tutorials.

Limitations

• This is a conceptual, task-based taxonomy built from literature review and theoretical analysis; the authors state that empirical work — classroom interventions, comparative studies, or think-aloud protocols — is needed to validate, refine, or challenge it.
• It was developed for journalism education and grounded in journalism's verification and editorial norms; its extension to law, medicine, and public policy is argued by analogy, not tested.
• The baseline prompts and risk-and-mitigation strategies are model-agnostic examples, not evaluated interventions: the paper reports no classroom study, student sample, or learning outcome.
• Its four-stage news-workflow frame simplifies journalistic practice, which the authors note varies across organizations, genres, and contingent circumstances, so some tasks span stages.

Citation

Dierickx, L., Bjerknes, F., Opdahl, A. L., & Lindén, C.-G. (2026). A Taxonomy of LLM-Supported Tasks for Critical AI Literacy in Journalism. Postdigital Science and Education. https://doi.org/10.1007/s42438-026-00690-0