The End of Academic Writer's Block? USC Scientists Launch GRAIL to Draft Papers in Under an Hour
Across labs and departments, finished experiments stall at the writing stage. Not because the ideas aren't there, but because formatting, citations, and ping-pong revisions eat months. Two USC researchers think that cycle can end.
They built GRAIL, an AI platform that turns rough notes, spreadsheets, and voice memos into submission-ready papers in under an hour. It integrates with LaTeX and handles the grunt work that usually slows a paper to a crawl.
The Mudpile Problem
Mayank Kejriwal keeps a folder called "obsolete-mudpile." It holds more than 50 half-written papers from past projects-solid results that never became manuscripts. Experiments done. Insights ready. Time ran out on writing.
That story is common across academia. The bottleneck isn't the science-it's formatting, citation wrangling, and rewriting for shifting reviewer and journal requirements.
David vs. Goliath Timing
On January 27, Kejriwal and his co-founder, PhD student Zhisheng Tang, released GRAIL. OpenAI announced Prism, a competing scientific writing tool, just one day earlier. Unexpected timing put a two-person team up against the biggest name in AI.
Early traction suggested real demand. In 48 hours, GRAIL processed over 300 million tokens. One PhD student wrote through the night for 12 hours. Another uploaded an old Excel full of literature notes and produced a formatted survey paper in under an hour.
Built From Nights and Weekends
Tang and Kejriwal met in a text analytics class back in 2021. After their first journal paper, they kept asking the same question: why does writing cost so much focus that should go to experiments?
In early 2024, Tang prototyped a chat-style science writer over a free weekend using Cursor. That quick build became the backbone of GRAIL. He then spent two months of nights and weekends polishing it-sometimes five hours a night.
How GRAIL Works (and Why It Feels Different)
GRAIL brings AI straight into LaTeX instead of forcing copy/paste loops with generic chatbots. You work in the same typesetting system journals expect, with automation layered in for structure, edits, and formatting. For context, LaTeX remains the standard in many technical fields for professional typesetting and math-heavy documents.
The Citation Fix Most Tools Miss
Citations are where AI writing often collapses. Models invent references. Reviewers catch them. Trust evaporates.
GRAIL takes a different approach. Kejriwal says they never ask the model to generate references directly. If a citation can't be verified, it's left out. He claims "zero percent hallucination" in citations with this method and argues this is a key difference from Prism's approach.
From Mudpile to Published
Tang uses GRAIL on his own research, and the results have surprised the team. One of Kejriwal's master's students, previously blocked by writing, used GRAIL and produced a draft strong enough for journal submission. The work didn't change; the writing barrier did.
This could surface more negative results and incremental findings that usually get shelved. It may also help underrepresented researchers and faculty with heavy teaching loads get their work out. The content is there-access to clean, consistent writing has been the gate.
Funding, Access, and What's Next
The founders initially funded development with $15,000-$20,000 of personal money. GRAIL is free for now and backed by Gutter Capital, a selective accelerator. The focus today is computer science and mathematics-areas where most work is computational.
The plan for this summer is bolder: an "autonomous scientist" that can suggest improvements to a paper, run computational experiments, return results, and draft the next version. If it ships, the writing step won't be the only thing on rails.
Authorship, Disclosure, and Responsibility
Who gets credit, and who is accountable? Kejriwal's stance is simple: the person using the tool is the author and responsible for verifying results. Tools can help, but they don't carry accountability.
AI is already showing up in accepted work. One analysis found hallucinated references in over a hundred NeurIPS papers. Conferences are adjusting policies to allow AI with author responsibility. The question isn't if AI is in the loop-it's how to use it without compromising integrity.
How to Put This to Work (Today)
- Collect assets: notes, figures, tables, code outputs, and prior drafts. Messy is fine-volume helps.
- Outline first: section headers, key claims, and the narrative arc. Clear structure makes better drafts.
- Feed sources: point to specific papers, datasets, and results you want cited. Verify every citation.
- Iterate in LaTeX: keep edits, comments, and formatting in one place to reduce context loss.
- Disclose AI use per venue policy. Final responsibility sits with you and your co-authors.
Why This Matters for Labs and PIs
Backlogs waste grant cycles and stall careers. If drafting time drops from months to hours, you can ship more work, publish negative results, and get students credit faster. That's practical leverage.
If you're building a lab-wide workflow around AI tools and training, a structured curriculum helps. See our AI courses by job for targeted upskilling paths.
Bottom line: GRAIL won't think for you, but it will remove the friction between finished experiments and a clean manuscript. Less time formatting. More time doing science.
Published on February 5th, 2026
Last updated on February 5th, 2026
This article may feature some AI-assisted content for clarity, consistency, and to help explore complex scientific concepts with greater depth and creative range.
Your membership also unlocks:
