A Non-AI Researcher’s Guide: Publishing Top AI Papers Using Your Domain Expertise

To be totally honest with you, you don’t need to understand the underlying architecture of large language models to get your name on a top-tier AI paper right now, you just need to be a true expert in your own field.

Author: Koutian Wu; GitHub: ktwu01

Here is the story. I have been thinking about this specific question over the last couple of days. As a Geoscience PhD student at UT Austin with a huge interest in Agentic AI, I am not traditionally trained in deep learning. So I wondered if there was a deterministic way to participate in top-tier AI paper collaborations for free, just by investing my time, without needing a hardcore ML background.

I had a chat with Gemini about this, and the answer kind of blew my mind = =

I realized that the AI research community is currently facing a massive bottleneck. They desperately lack high-quality, human-original, challenging data. Think about it. As models get stronger, the main constraint isn’t compute power anymore. It is finding questions that can actually stump the most advanced AI. And for that, they need domain experts from all walks of life, not just folks who know how to tweak a Transformer.

In the community, this is generally known as crowdsourced benchmarking or community-led research.

Following this thought process, I mapped out a few actionable paths for non-AI experts. I am still exploring this myself, and some of you might be wondering if this actually works, but after laying it all out, I am genuinely excited.

Let’s skip the fluff and get straight to the practical stuff.

The first path is joining top-tier crowdsourced benchmark projects.

There is a very popular initiative right now called Humanity’s Last Exam. The name sounds a bit dramatic, right? But it is a real joint effort by the Center for AI Safety and Scale AI. They are actively seeking graduate-level, highly difficult questions.

If you can design a question in your specific domain, like geosciences or environmental modeling for me, that only an expert would understand and that current AI will definitely fail, you are in. Once your question passes peer review and gets accepted, you are typically listed as a co-author on the final technical report or paper.

That is the deterministic part. Besides this, there is GAIA backed by Meta, which tests AI execution in real-world tasks, and GPQA, which focuses on scientific questions that are hard to answer even if you have Google.

The second path is taking the initiative to build domain-specific benchmarks.

A lot of people might not know this, but the AI community is extremely hungry for evaluation data in vertical domains. Tech giants are aggressively building large models for meteorology and geosciences, like NVIDIA’s Earth-2 or Google’s GraphCast.

My personal take is that if you find open-source, domain-specific large models in development on GitHub and volunteer to build an evaluation benchmark dataset for them, they will welcome you with open arms. Providing high-quality labeled data is incredibly valuable, and contributors usually secure a prominent spot on the author list. This is way more feasible than trying to train a model from scratch yourself.

The third path is diving into hardcore open-source research communities.

You probably know about decentralized AI research orgs like EleutherAI. If you hop into their Discord channels and demonstrate a deep understanding of a specific niche, while offering the data support they urgently need, it is very easy to get pulled into a paper working group. It is essentially trading your domain knowledge for equity in a research project.

Speaking of this, if you are not just looking to contribute for the sake of research, there are mature channels for monetizing your intellectual assets.

I looked into the outsourced research ecosystem while I was at it. Platforms like Kolabtree or ScienceExchange provide a gig economy explicitly for independent experts with PhDs. This isn’t about churning out low-quality papers, it is about solving real problems in biostatistics, clinical trials, or your specific niche. Your knowledge has a clear price tag on these digital procurement platforms.

This actually highlights a fascinating phenomenon. In this era where large models are everywhere, many people are anxious about being replaced. But think about it. When AI models converge and memorize all the public data on the internet, what becomes truly scarce?

It is the raw experience in your head that hasn’t been written into Wikipedia yet. It is your intuition and those incredibly tricky questions you’ve developed after years of wrestling with real problems in your industry.

A machine can generate ten thousand coherent literature reviews.

But it cannot fabricate a question where only a geologist knows where the hidden trap lies.

This is exactly the opportunity for ordinary people like us. Leveling out the information gap.

If you are intrigued by this, you can go check out the Center for AI Safety’s website or Twitter today, they frequently post open calls for experts. Pick the domain you are best at, and write a question that you think will completely baffle an AI.

Don’t worry about whether it gets selected or not just yet. Just give it a try.