Written and maintained by CASRAI Editorial Board
Last updated
Stanford STORM is a research writing system from Stanford that does the pre-writing work for a Wikipedia-style article: it researches a topic on the internet, asks questions from several perspectives, and produces an outline and a cited draft. It is a tool for gathering and structuring material, not a source you can cite in place of the sources it found. This page is for researchers, students and librarians who keep seeing “STORM AI” and need to know what it does, what its authors say it gets wrong, and how to use its output responsibly.
“Storm” is also the name of unrelated things, including STORM super-resolution microscopy. This page covers only the Stanford system, the one people search for as “stanford storm ai”.
What Stanford STORM is
STORM stands for Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking. The Stanford STORM Research Project describes it as a writing system that focuses on the pre-writing stage: the research and outlining that happen before anyone drafts prose. It was published at NAACL 2024 in the paper “Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models” (arXiv:2402.14207; first posted 22 February 2024, revised 8 April 2024).
According to the project page, STORM works in three moves:
- It discovers different perspectives on the topic.
- It simulates conversations in which writers holding those perspectives ask questions of a topic expert, with the expert’s answers grounded in internet sources.
- It curates the collected information into an outline from which the article is written.
Under the hood it is built on a large language model combined with search, an approach related to retrieval-augmented generation. The authors also released FreshWiki, a dataset they built for evaluating this kind of article generation.
Who runs it and what it covers
STORM is run by the Stanford STORM Research Project. The paper’s authors are Yijia Shao, Yucheng Jiang, Theodore A Kanell, Peter Xu, Omar Khattab and Monica Lam, all listed with Stanford. The project page links to the demo, the paper and the code repository (github.com/stanford-oval/storm), and its navigation also points to a related system called Co-STORM, which this page does not cover.
The scope is deliberately narrow. The target output is a long-form, encyclopedia-style article on a topic, written from scratch, rather than a literature review for a defined research question, a systematic review, or a grant or manuscript section.
How to reach and use it
The project page carries a “Try it now!” link that leads to the hosted demo at storm.genie.stanford.edu. The demo is presented behind a terms-of-service agreement, which you accept before entering a topic. The research code is published separately in the project’s GitHub repository for people who want to run or study the system themselves; consult the repository’s own README for installation and configuration, because those details change between releases.
The sign-in and terms screens are controlled by Stanford and can change, so follow the current instructions on the demo itself rather than a third-party walkthrough.
What the authors say it gets wrong
This is the part most explainers skip, and it is the part that matters for research integrity. The paper reports that expert editors judged STORM articles better organized than those from an outline-driven retrieval-augmented baseline, with an absolute increase of 25% in organization and 10% in breadth of coverage. The same editor feedback identified two problems: source bias transfer (bias in the retrieved web sources carries into the article) and over-association of unrelated facts (the system links facts that do not belong together).
The project page adds that the main issue is “red herrings”, meaning shaky links or irrelevant content, rather than outright factual hallucination. That distinction changes how you check the output. A reference list that exists is not evidence that each sentence is supported by the source next to it.
A verification checklist drawn from those error types
- Open each cited source and confirm it contains the specific claim attached to it, not just the general topic.
- Look for sentences that join two true facts with a causal or comparative link neither source makes.
- Check whether all the sources lean the same way; if the topic is contested, the draft may reflect the sources’ bias.
- Replace any web-page citation with the primary study, standard or policy where one exists.
If you use STORM output in anything you submit, check the target journal, funder or institution policy first. Our AI-assisted drafting best practices and AI disclosure statement examples cover how to record and declare that use.
Strengths and limits
| Strength (per the authors) | Limit (per the authors) |
|---|---|
| Targets the pre-writing stage: research and outline, not only prose | Output is a draft for human checking, not a finished or citable article |
| Multi-perspective questioning aims at broader coverage; reported +10% coverage versus the baseline | Source bias can transfer from retrieved pages into the text |
| Reported better organization than the baseline (+25% absolute) | Unrelated facts can be wrongly linked (“red herrings”) |
| Code and the FreshWiki evaluation dataset are released | The project page states no software licence; check the repository before reuse |
Alternatives
STORM generates an outline and article from open web research. Other AI research tools start from the scholarly literature instead. Each of these has its own CASRAI guide:
| Tool | CASRAI guide |
|---|---|
| OpenRead (AI paper search, Q&A and comparison) | OpenRead explained |
| Paperguide (AI research workspace) | Paperguide guide |
| AnswerThis (research gap finder) | AnswerThis guide |
For managing the references any of these tools produce, see citation management with AI writing tools.
FAQ
What is Stanford STORM AI?
A Stanford research system that performs the pre-writing stage for Wikipedia-style articles: it gathers information through simulated multi-perspective conversations grounded in internet sources, then builds an outline and draft (NAACL 2024).
Is STORM the same as Co-STORM?
No. Co-STORM is a separate system linked from the same project site; this page describes the original STORM system and the paper cited above.
Does STORM hallucinate?
The project page says its main problem is not factual hallucination but red herrings: shaky links between facts and irrelevant content. The paper also reports source bias transfer. Both mean you must check each claim against its source.
Can I cite a STORM-generated article?
Treat it as a lead generator. Cite the underlying sources you verified, and disclose AI assistance where your journal, funder or institution requires it.
Where is the code?
On GitHub at stanford-oval/storm, linked from the project page.
Sources
- Stanford STORM Research Project, “STORM” project page: storm-project.stanford.edu/research/storm/ (read 2026-11-08).
- Shao, Jiang, Kanell, Xu, Khattab, Lam, “Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models”, arXiv:2402.14207, NAACL 2024: arxiv.org/abs/2402.14207 (read 2026-11-08).
