Our AI Policy

as of October 1, 2026

Large Language Models (LLMs) are tools, but they are so powerful and pervasive that they are also changing the wider context of our reading, writing and research. In some usages, say for search and discovery of scholarship, and processing of archival materials, they offer powerful new ways of working, and we are exploring these actively. But we urge caution in the use of AI for writing and development of arguments, at individual and collective levels.

We see risks at three levels. The use of these technologies at scale poses epistemic risks as AI-generated texts, hallucinated citations, chatbot sycophancy and unattributable claims swamp our publishing systems. At the NUS Press level we are concerned about the risks to the bond of trust we build with our community if books show signs of being heavily AI drafted. And finally we are also alive to the risk of cognitive offloading in ourselves and our authors, which the heaviest users of AI tools report.  

It is a complex picture and we look forward to working with our stakeholders to find the right path forward. Norms may change, AI tools are evolving rapidly. Please check this policy when you reach a new milestone in your project!

Policies

  1. Please do not submit texts that include wording from AI models, even if you are prompting models to simply "polish", "streamline" or "correct" your draft. Subject to the important caveat below, we support your use of AI tools to proofread and point out issues or problems in how you write or develop an argument. We accept that AI tools might be useful to help you brainstorm and trial potential arguments, but the first draft, and subsequent revisions, should be written by you.
  2. Search and discovery with AI tools is very powerful and may be very useful, particularly in literature review. Rather than relying on chatbots, we encourage you to use dedicated tools for search and discovery, including services like Elicit, Consensus or Scite.ai. These tools are grounded in citation databases, and are able to read abstracts and sometimes full text under license. The field is evolving quickly and we encourage you to seek help from your university librarian to identify the tools of most use for you.
  3. When you submit, we will ask for a detailed declaration of your use of AI tools, including which model version and number was used (ie "OpenAI GPT-5.6 Sol"), how and why. If you use models in any way on the writing, we ask that you keep previous versions of texts, and a record of any prompts you used on text. We may need these for review.
  4. At this moment, we are not accepting AI-generated images or diagrams for publication. The field is evolving quickly, and if you have a case for your usage, we are happy to hear it.

Why do we hold this view on AI as a tool for writing?

First, we do not accept AI as an author, co-author or credited collaborator. Individual scholars must be fully responsible for any work they submit. But it is no longer sufficient just to declare that if your text shows signs of AI generation.

Readers are increasingly sensitive to and suspicious of AI writing. While models are changing and improving, the text they generate displays different traits and tendencies, which sensitive readers (like peer reviewers) are starting to recognise and reject. We've never used AI-detection software in the past, believing it wholly inadequate to the task, but a tool called Pangram is proving highly effective in independent tests. More importantly, under EU law, models will be required to use "watermarking" techniques to identify AI-written text. Anthropic has recently introduced a system to do so.


As a humanities and social sciences publisher, we believe that "the writing is the research", at least in these fields. How you form and develop ideas as you write is part of the process of creating a strong book.


On a more positive note, AI tools can be highly effective at identifying problems in your writing, catching grammatical errors, tenses used improperly, or unclear modifiers and pronouns. You may find it useful to use the tools to check your writing, but again, we ask that you do not ask the models to propose alternate phrasing. This path takes a bit more time, but in our experience will not only preserve your own unique voice, it will help you improve your writing, which is after all a life-long journey. Here are two examples to illustrate the distinction:

Against Policy

We've all done this... asked the model to redraft a passage or a paragraph. 

Within Policy

 

Important caveat

While we allow (or even encourage) you to use AI for certain tasks along your journey, we ask that you not feed your work to free versions of the AI tools before submitting to us. The same goes for articles and files from your colleagues that you may be consulting in your research. Almost all of the free tools now take material you upload and reuse it to train new AI models, at least in their default settings. Terms and conditions here are complex and hard to find; settings are not easy to change.

Paid versions also require careful (and constant) checking! Terms of use can be changed very quickly. Ideally your university has offered you a version of the AI tools housed on-site, with stronger protection of material you share. You can also use an app like MSTY, which runs locally on your computer and uses APIs to access LLMs. LLM use under API generally has much stronger protection for intellectual property.

NUS Press editors may use AI to highlight issues in your text, but we will only do so on protected versions of LLM software provided by NUS.

Why do we feel strongly about this?

  • Firstly, copyright reasons. We tend to agree with the Microsoft executive who wrote in an internal memo that LLM training on published material without credit, consent or compensation for authors was "an astonishing theft of unprecedented proportions". That process continues, though it is increasingly challenged. For one perspective on this issue, see the Director’s substack.

  • Secondly, AI pre-training does not preserve attribution of ideas in any systematic way.

  • And thirdly, we don't feel it in our authors' best interest to expose an early version of their work to AI systems, before peer-review and editing. Let us work on your material together first, and get it into its best possible state. Then, with your consent, we can license your material to systems that will allow AI-aided discovery of published research material. We believe that such systems are essential to ground AI results, avoid hallucinations and to attribute specific claims and findings to specific human authors.

We welcome comment or feedback on our policy, by email to nusbooks@nus.edu.sg.