State of Psychedex (September 2026)

After starting development in March 2026, Psychedex was judged ready for beta launch on September 2026. Using a LLM pipeline that researches, generates, fixes and checks citations, articles for around 300 substances, 80 practices, 250 effects, and 500 concept/glossary pages were generated, with more planned.

Rather than just letting the LLM generate prose from its training data—which is generally accurate but also not reliable in itself—we designed a multi-step LLM pipeline that has agents conduct literature research, write prose, match citations, and review its own output. The final product is then presented for human review and approval before publication.

The pipeline is quite complex and compute-heavy due to the research and fact-checking steps involved. Each substance takes about an hour to generate using frontier LLM models. Reviewing the pipeline’s output during RnD, we found that the general accuracy is good—on-par with human editors—and the LLMs lean towards caution with regards to safety info, which is vital for a harm reduction website.

However, there is one major barrier to achieving exceptional accuracy and that is the fact that the research agents cannot bypass research article paywalls. Instead, they have to rely on public paper abstracts. As a result, some claims are not fully backed by the citation attached to them. I have done my best to be fully transparent about where they fail by leaving an audit log.

We also found that while the LLM can be slightly off about specific claims such as exact data values, they are capable of understanding the major effects and risks of each substance, including life-threatening ones. As a result, we judged that we could ship the content as is and rely on contributors to backfill citations with full paper links, or if not, suggest the claim be rewritten. We have included multiple surfaces to provide feedback on incorrect article information.

As the landscape evolves, we plan to continue fine-tuning our article generation and review processes.

If you are interested in contributing to Psychedex articles, the best thing you can do right now is to find and fix citations. This is easier than in the past because the LLM doesn’t usually say something totally wrong, so it’s more about confirming it by reading the quoted abstract’s paper. Safety specific data get their own citation needed flags and are the highest priority.

We will continue to keep you updated over time. In the meanwhile, please enjoy the site. As a former operator of PsychonautWiki, I am quite pleased at how useful and accurate the articles are overall. If you are interested in the origins of Psychedex, please read this blog post.

Thank you for participating in the beta and please send any questions, concerns, or comments to contact@psychedex.org.