Making AI Science More Accessible at NeurIPS
By Julien Besset, Science Communicator at Mila and David Rolnick, McGill University and Program Chair at NeurIPS 2026
Authors of accepted papers at NeurIPS are now asked to write a short overview of their work for a non-technical audience (also called a lay summary). This initiative aims to increase the visibility of AI research by encouraging researchers to explain their work simply but still accurately. In that spirit, NeurIPS will also offer in-person science communication training in Sydney this year.
Nuance in research can easily become lost in the hype and public debates surrounding AI. Researchers have a responsibility to help non-specialists better understand a technology that has an increasingly large impact on society. Writing in a more accessible way will also help your work reach a broader audience (media, policymakers, future employers or collaborators from other fields).
ICML already requires all authors to include a lay summary of their research. You can find their advice for authors here.
How to write a plain language summary
The goal of a lay summary is to make people want to know more about your work. Think of it as a movie trailer for your research paper: it should not be too vague nor too detailed (this is not just a shorter abstract). Find the right balance to make your work accessible and engaging without oversimplifying it. You won’t be able to say everything, so focus on the essentials.
Here are a few things to keep in mind before you start writing:
Most people don’t know as much as you do. You don’t know what your audience doesn’t know. Your expertise is the result of years of experience working on a very specific topic, and no one has the exact same knowledge as you do (even in your own lab). Always keep your audience in mind and try to put yourself in their shoes.
Jargon is your frenemy. Jargon words can be useful as they allow experts to quickly communicate ideas without having to repeat longer concepts. But by using excessive jargon and acronyms, you risk losing the attention of your audience if they don’t know what you mean. Reduce jargon as much as you can, and if you can’t, explain it first (you can use metaphors to illustrate your point).
Storytelling will make your work stand out. Humans are wired for stories, and good storytelling is what separates what people remember from what they forget. But if you don’t do the narrative work upstream, you leave it up to your audience to figure out what information matters most (they won’t). Focus on curating the facts you want people to remember the most: human attention and memory are limited, so stick to a maximum of three core ideas.
Remember that it is always easier to make a text more complicated than to make it simpler. The more technical your subject, the harder it may seem at first to simplify it without betraying its depth. The bad news is that you have to choose what to leave out. The good news is that it gets easier.
Guidelines for NeurIPS
Your summary should not exceed 150 words, should be understandable by a journalist, and should answer the following questions:
- What problem does your research aim to solve? Why is it necessary?
- How do you tackle the problem? What makes your method unique?
- Why does your research matter? What impact might it have on your field or on society?
Example for the NeurIPS 2025 paper From Noise to Narrative: Tracing the Origins of Hallucinations in Transformers
Generative AI models are increasingly being used in critical sectors like healthcare, law, and finance, but their tendency to invent answers makes them unreliable and hard to trust. Traditional AI safety methods focus on inputs and outputs to evaluate models, but we decided instead to peer directly into the “black box” of popular chatbots to better study their inner workings. We discovered that models infer meaning from random signals, much like humans “seeing” shapes in clouds. Understanding this process allows us to predict hallucinations before they can reach the end user, paving the way for safer and more reliable AI systems.