They are not trusted, but they are used. They are not entirely reliable and accurate, but journalists are experimenting daily with large language models (LLMs) that are likely to be used at every stage of the news process. After a series of blog posts aimed at identifying where AI is being used in the production and distribution workflows, and defining a set of twenty tasks that can be used with LLMs – based on task definitions in machine learning and natural language processing – this blog post aims to practically outline how LLMs can support different stages in journalism process, assess the risks at each step, and suggest strategies for safe and meaningful use.
Supporting story discovery, document exploration, and early research
Risk Level: Moderate
LLMs help journalists scan news sources, forums, databases, and social media for emerging topics. They assist in exploring large sets of documents by extracting entities, detecting patterns, and summarising content, which makes mining and early analysis more accessible. However, the risks here are substantial. LLMs do not « understand » material; instead, they identify patterns statistically, which means they might miss key insights, overlook contradictions, or highlight superficial trends. Complex investigative leads could be ignored. Over-reliance on AI may flatten narratives, introduce bias, or lead to shallow interpretations if journalists fail to re-engage critically with the sources.
Research and interview preparation
Risk Level: Low to moderate
LLMs help brainstorm broader perspectives, including compiling background briefings, suggesting expert contacts, anticipating interview questions, and simulating potential dialogues. However, the main risks involve outdated, incomplete, or biased suggestions. Therefore, there is a need for independently verifying sources or background information, as erroneous or skewed insights may be embedded early in the process. Moreover, since LLMs merely reproduce patterns rather than generate original ideas, they should be seen as an instrument to stimulate creativity rather than rely on them for novel insights. On a more practical level, using LLMs with sensitive or confidential information is not recommended due to security concerns.
Drafting and refinement
Risk Level: Moderate
LLMs assist in structuring drafts, suggesting headlines, and exploring stylistic variations. They can propose multiple narrative approaches quickly, helping journalists transition from idea to written content faster. However, the risk of « hallucination » — the confident fabrication of false information — is a persistent problem. AI-generated outputs can also reflect hidden biases, prioritise specific framings over others, or fail to deliver a balanced narrative. Thus, human journalists rework any text generated by an LLM to ensure factual accuracy, consistency, and stylistic integrity.
Editing, fact-checking, and verification
Risk Level: High
While LLMs can suggest clarifications, detect inconsistencies, or propose potential fact-checking targets, they should never be relied upon as verifiers. LLMs are prone to inventing quotes, misrepresenting facts, or subtly distorting the context of information. Every factual claim flagged by AI must be independently verified through primary sources, trusted references, or direct reporting. Editorial oversight at this stage is non-negotiable. Furthermore, LLMs will likely create additional work, as fact-checking generated content often requires more scrutiny and correction. While they can assist in assessing collected evidence, they cannot replace thorough, human-driven verification.
Final editing, language checking, and risk of misinterpretation
Risk Level: High
LLMs can support final editing by checking grammar, punctuation, and style, but they struggle with more nuanced aspects of language, particularly when context is essential. They often fail to grasp polysemy (words with multiple meanings) fully or homonyms (exact words but with different meanings), leading to potential misinterpretation or ambiguity. Their lack of contextual understanding means they might miss how specific phrases, idioms, or terminology should be used in a particular journalistic context. Human review is therefore essential to ensure that meanings are conveyed as intended, avoiding errors that could lead to confusion or misrepresentation of the story.
Publishing and distribution
Risk Level: Low to Moderate
LLMs can help optimise content for SEO, generate catchy headlines, create social media posts, and draft meta descriptions, thereby maximising the reach and visibility of stories. Here, the risks are lower, but they still exist as LLMs can unintentionally foster sensationalism, clickbait tactics, or misframe the ethical context of a story. Therefore, human editorial oversight is needed to ensure that promotional materials stay true to the story’s integrity, avoiding ethical pitfalls or misleading impressions.
Audience engagement and feedback analysis
Risk Level: Low
After publication, LLMs can summarise reader comments, detect emerging reactions, and suggest follow-up topics. This allows journalists to monitor audience sentiment without manually sifting through large volumes of feedback. However, AI misinterpreting sarcasm, irony, or polarised language can lead to incorrect assessment of audience sentiment. Thus, while AI-generated insights can be helpful, they should be taken as guidance rather than definitive answers, with human journalists interpreting and acting on the data cautiously.
Risk Matrix: Stage-based risk assessment
| Stage | Risk level | Key risks | Human oversight needed |
|---|---|---|---|
| Story discovery, document mining | Moderate | Missing insights, shallow interpretation | Full re-analysis and questioning |
| Research and interview preparation | Low to moderate | Outdated or biased suggestions | Independent source verification |
| Drafting and refinement | Moderate | Hallucinations, bias in framing | Full rewriting and fact-checking |
| Editing and fact-checking | High | Invented facts, miscontextualised information | Manual fact-checking essential |
| Final editing, language checking | High | Misinterpretation of polysemy, homonyms, contextual errors | Thorough human review to ensure clarity |
| Publishing and distribution | Low to moderate | Sensationalism, misframing | Editorial control over output |
| Audience feedback analysis | Low | Misinterpretation of tone (e.g., sarcasm, irony) | Interpretation with caution |
LLMs remain prone to errors, biases, and hallucinations. They lack understanding, ethical judgment, and critical thinking. Therefore, strategies to mitigate risks not only involve building solid AI literacy to understand the possibilities and limits of the systems but also optimising prompting strategies to reduce risks related to accuracy and reliability, as well as introducing validation steps that include human-in-the-loop verification, protocols defining AI use, and when it relates to content quality, transparency toward audients.
