Wikipedia:Responsibly using large language models

Further information: Wikipedia:Large language models

For the responsible use of LLM-assisted machine translation, see Wikipedia:LLM-assisted translation.

Per Wikipedia:Writing articles with large language models (LLMs), the use of LLMs to generate or rewrite article content is prohibited. However, if used responsibly, LLMs can still serve purposes outside of content generation and this essay seeks to walk editors through their responsibilities relating to this content guideline and the usual core content policies.

Editors should only employ LLM assistance if they can do so competently and this essay also intends to set out some of the relevant pitfalls to be aware of, rooted in established policies and guidelines that have consensus within the Wikipedia community. It is not intended to prescribe a course of action beyond adhering to Wikipedia's policies and guidelines.

Research

Finding sources

AI tools such as ChatGPT have seen increasing use as search engines and are able to rewrite a conversational prompt into a keyword-based web query, returning links cited to their answers. However, empirical research has suggested that at least half of LLMs' responses are not fully supported by the online sources they cite.

Therefore, to ensure that you are not adding content that contains original research, you should manually read these sources and write the content yourself such that it reflects them as accurately as possible. Per the guideline above, you should not use the LLM-written summary cited to the source and you may choose to ignore it completely per the Direction and self-education section below. Given that Wikipedia policy requires encyclopedic content to represent all majority and significant minority views, editors may still be required to read widely around a subject to ensure that what they end up writing is neutral. An LLM may not recommend the best sources available, which can potentially introduce bias into an article. LLMs can also be used to find sources across languages, though editors should at least be "reasonably certain" about the accuracy of any translated material. Editors should ensure that they read and verify these sources themselves, following the guideline on LLM-assisted translation if used.

Direction and self-education

A Wikipedia editor of 26 (right) who has adopted ChatGPT (left) as their tutor

Building an encyclopedia can entail having to write authoritatively about vast and complex topics. Given that Wikipedia is a project driven by volunteers who aren't necessarily experts in the subjects they write about, it may sometimes be alluring to turn to an LLM to clarify a key concept or a piece of peripheral context to help you understand and accurately paraphrase a source.

Empirical research on the use of LLMs in education has been mixed. Some studies have found LLMs to improve students' engagement while others highlight their potential to provide inaccurate or biased information that can mislead students or instill cognitive biases. Other research has found that, even when aided by real-time links, learning through LLMs can lead an individual to develop a shallower understanding of a topic compared to traditional web search. The Wikipedia community considers content generated by LLMs to be "generally unreliable".

In the course of their research, editors may turn to an LLM to help orientate them around the topic they're intending to write about, or to analyse an article for improvements and opportunities for expansion. Even if you consult the best sources available, having only used an LLM as an initial guide, they will still sometimes commit original syntheses or hallucinations. If these mistakes are not caught, this has the potential to significantly skew the direction of your research and undermine your understanding of a topic. Inaccurately paraphrasing a source such that it no longer reflects the meaning of the original text contravenes Wikipedia's core policy prohibiting original research. If you do not have the time to do the appropriate due dilligence around checking and correcting for the mistakes LLMs often make (or you don't think white hair would suit you), then it may be best to stick to human-driven methods of research that are more grounded in reliable sources.

Source-based LLMs

Source-based LLMs such as NotebookLM allow users to upload PDFs or webpage links, with the LLM then being able to use and cite the source(s) to respond to conversational prompts. Users can copy-paste the cited text and search for it in the uploaded source to find, or re-find, material relevant to their query. Editors intending to use these LLMs to orientate them around a source or to find relevant information should be aware that they will still sometimes synthesise or hallucinate information, meaning that they possess many of the same pitfalls as LLM-assisted web research. Editors should not use the LLM-written response and should ensure that they read the source(s) themselves and write any content such that it reflects them as accurately as possible.

Copyediting

Editors are permitted to use LLMs to suggest corrections to their own writing, and to incorporate them after human review. This is limited to spelling, punctuation, capitalisation, grammar, and other simple mistakes. Editors must ensure that any text copyedited in this manner is still supported by the sources cited.

— Wikipedia:Writing articles with large language models

Prompt: (draft text author added in section). Review above text. List any errors but do not make suggestions.

---

ChatGPT: The passage is generally clear, but these are the main issues I notice:

• Incorrect apostrophe: "LLM's" uses an apostrophe for a plural. Standard English uses "LLMs".
• Comparative construction: "more readable to readers less familiar with the work that you are as the author" contains an awkward comparison ("that you are") that is grammatically strained.
• Possessive form: "LLM's are prone" again uses an apostrophe where a plural is intended.
• Word repetition: "readable to readers" repeats the same root word in close proximity, making the sentence less smooth.

— on 22 July 2026

LLMs can be very helpful in reviewing draft text. Sometimes they suggest rewording that is clearly more readable to those unfamiliar with the topic. At the same time LLMs are prone to include promotional language and other slop. For this reason the editor community currently favors imperfect prose written by humans to smooth-sounding but questionably accurate text. Generally, you'll do well to avoid copy-pasting text.

Editors should show their work when using an LLM for fixing mistakes in their writing. First make your edit without LLM assistance. Then make another with the corrections applied by your LLM. This will help bring transparency that the machine generated text is not being added to Wikipedia.

Wikipedia is not a technical manual or a how-to guide (WP:NOTGUIDE). For this reason this essay will not offer guidance on which LLM to use. If you are uncertain, it is far better you try copy editing using your own human abilities only.

See also

• Wikipedia:Large language models and copyright
• Wikipedia:WikiProject AI Tools § Existing projects
• Wikipedia:Artificial intelligence resources

Bibliography

• Melumad, Shiri; Yun, Jin Ho (2025). "Experimental evidence of the effects of large language models versus web search on depth of learning". PNAS Nexus. 4 (10) pgaf316.
• Reyna, Jorge (2025). The Potential of Google NotebookLM for Teaching and Learning. eLearn 2025 Conference. Bangkok.
• Wang, Jin; Fan, Wenxiang (2025). "The effect of ChatGPT on students' learning performance, learning perception, and higher-order thinking: insights from a meta-analysis". Humanities and Social Sciences Communications. 12 621.
• Wu, Kevin; Wu, Eric; Wei, Kevin; Zhang, Angela; Casasola, Allison; Nguyen, Teresa; Riantawan, Sith; Shi, Patricia; Ho, Daniel; Zhou, James (2025). "An automated framework for assessing how well LLMs cite relevant medical references". Nature Communications. 16 3615.