AI did the research, writing, fact-checking, and illustration. A person designed the process, tested it, made the editorial decisions, and decided what was ready to publish. Here is what that involved—and what it cannot guarantee.
The first lesson of this project was that AI can produce a convincing account of a tradition without getting the facts right.
The original starting point was a set of AI-generated surveys covering roughly thirty possible traditions. They looked useful. But when the records were tested, every record checked contained errors. The surveys were treated as leads, not evidence, and none of their claims was allowed into the finished work without a supporting source.
That experience shaped the rest of the project. Rather than ask an AI model to produce a website full of articles and trust the result, Jim Davis, the site's creator, built a process in which claims could be traced to published research, drafts could be challenged, and corrections required a human decision. The software, research agents, writing instructions, and images were themselves produced with AI. They were also tested, revised, and sometimes rejected.
This was not a one-prompt exercise. The difficult work was deciding what the system should do, finding where it failed, and changing the process when the results were not good enough.
Starting with the traditions, not with AI's answers
The site focuses on living traditions in which psychoactive plants or fungi have a role in the wellbeing of people and communities. For inclusion, a tradition must still be practiced—at least in the sense that a practitioner could be spoken to today. The creator chose the criteria and made the inclusion decisions, with additions and exclusions recorded alongside their reasons.
The selection began as a list for a map, and it remains a curated collection. It is not meant to represent every tradition, region, or substance in the world.
Once a tradition is selected, its research begins afresh. A new AI research agent receives a short orientation, but it is not handed the earlier survey's assertions as facts. Its job is to find the evidence for itself.
The research gives priority to published accounts from Indigenous peoples and the communities concerned, followed by firsthand ethnography, peer-reviewed studies, institutional publications, and reliable journalism. Reference works can help identify leads, but they are not accepted as the sole support for a claim. When the available material is mostly secondary reporting, that weakness is recorded rather than concealed.
Language matters. A search conducted only in English can miss how a community describes its own practices—or overlook research carried out nearby. Searching in relevant regional languages is therefore required, not optional. In one test, searches conducted only in Spanish uncovered community statements and recent fieldwork that the English searches had missed. This site used research articles written in English, Spanish, French and Portuguese.
The goal is not to collect as many citations as possible. It is to find sources that help explain what a practice means in its particular cultural setting, while making clear where the evidence is thin or contradictory.
From a source to a sentence
Finding a publication is one thing. Demonstrating that it supports an article's claims is another.
For the tradition articles, the research system saves the full text of the sources it relies on. Roughly 240 source texts had been collected at the time this account was written. These are working copies retained for verification, not texts republished on the website. Each source has an identity the writing system can cite, and checks look for mismatches between a saved text and the work it is said to represent.
This matters because AI can produce perfectly plausible-looking references. A real book or paper is not enough: the specific statement attributed to it must actually be there.
After research, a review of the saved material checks whether the reported uses appear in the cited sources, whether claims about a tradition's current status are dated, and whether sensitive information has been identified for protection. In an early spot check, 152 of 153 sampled statements were correctly cited. That is encouraging evidence about that sample—not proof that every claim in every article is correct.
The first method of turning the research into articles did not survive testing. It gave the writer short, extracted claims. The resulting pieces were technically organized but read like flattened academic summaries, with too little of the context that makes a tradition understandable. A comparison showed that writing from the complete saved sources worked much better, so the claim-extraction approach was abandoned.
The writer now receives the source texts, their identifying numbers, and their citation information. It is instructed to write from those materials rather than from its general knowledge, and to attach numbered references to the resulting paragraphs. The draft is meant to be readable by a curious non-specialist without blurring important differences between traditions.
Getting that balance right required repeated work on the writing instructions. Each change was made in a plain-text instruction file, tested on actual output, and read before another adjustment was made. The exact instructions and responses from each run were retained. If a change made the writing worse, it could be identified rather than quietly becoming the new standard.
A draft is not a checked article
Once a tradition article has been drafted, it goes through two different kinds of review.
First comes an editorial reading. This AI pass works on clarity, structure, and readability without having the sources in front of it. It is forbidden to invent supporting details or silently resolve uncertainty. If it sees a broad statement that may not hold up, it flags the passage for checking rather than rewriting it into something that merely sounds more certain.
Then comes the evidence check. Another AI pass examines the article and its supporting materials alongside the saved sources. It is asked to identify problems that could give a careful general reader a false impression: a statement stronger than the evidence permits, a misleading generalization, a claim attached to the wrong source, or a breach of the site's rules about privacy and neutrality.
Its output is deliberately specific. A finding must identify the article's wording, explain what a reader might wrongly understand, provide the relevant source passage, and suggest a correction. Software verifies that the quoted article text actually appears in the draft. That prevents a familiar AI failure: confidently proposing a fix for words the article never said.
The fact checker was tested, and its limits are known. In one adoption test it found all ten deliberately planted high-severity errors and roughly three-quarters of the medium-severity ones. But a confirmed serious problem found in one run disappeared on a repeat run. The findings it raised in that test held up when reviewed, yet different runs could miss different real problems. At other times it has called a claim unsupported simply because it could not access a source.
The checker was therefore adopted by human judgment, not because it passed every pre-set test. It is an error-finding tool, not a certificate of accuracy.
Jim Davis reviews each finding with the cited passage beside it. He accepts the proposed change, rejects it, or supplies his own wording. Only after those decisions is the article marked as checked. This is human oversight of the flagged issues, not a claim that a specialist has independently verified every sentence.
A different approach for the substance pages
The shorter pages about individual plants, fungi, and other substances use a different process, and that distinction is important.
Unlike the tradition articles, these pages begin with AI drafting from general knowledge and the site's checked tradition summaries. A separate AI with web search then checks factual claims and wording against sources. The pages cover such topics as names, identity, psychoactive effects, use within traditions, and present-day context. They deliberately exclude doses, preparation instructions, legal status, safety advice, and health claims.
On the substance pages reviewed so far, the checker reported between five and eleven issues per page. Many were small but consequential: a date stated too confidently, a botanical name applied to the wrong species, or a health claim that had slipped into the text.
This process also produced examples of why the human review remains necessary. The checker missed an error about which part of the ayahuasca vine is used; Davis noticed and corrected it from his own experience. In another case, the checker labeled a claim unsupported even though the saved source did support it. That error was caught by a person, too.
Those episodes point in opposite directions, but teach the same lesson: an AI review can both miss a mistake and invent one. A proposed correction is evidence to examine, not an instruction to obey.
Even the illustrations were researched
Every image on this site is AI-generated. They are illustrations, not documentary photographs or botanical identification plates, and they have not been reviewed by a professional botanist or mycologist.
For the plant and fungus images, the project did not simply ask an image model to draw a named species. Such models can make a less-familiar plant look plausible while getting the leaves, flowers, fruit, or growth habit wrong. A picture can be visually convincing and botanically misleading at the same time.
The process therefore begins with botanical research. One AI agent collects descriptions from published sources and preserves the passages supporting each detail. Software checks that the quoted passages exist in the saved material. Disagreements between sources are recorded instead of being resolved by guessing. A second AI then builds the image description from those checked facts, and the description is checked again before generation. Several image models were evaluated before one was selected.
The source research has sometimes been more instructive than the picture-making. Two reference works disagreed about the number of ribs on a San Pedro cactus and the elevation at which it grows. Rather than insist on a single exact figure, the description used an approximate range. An early description of Salvia divinorum, based on dried specimens, reported blue flowers; a later paper clarified that the flower is white within a violet calyx. The illustration followed the correction. During work on ololiuqui, the botanical research also identified a species-name problem in the site's own text and sent it back for correction.
Once an image exists, a different AI system with web access examines it against the real species. That extra distance is intentional: the system that described the picture should not be its only judge. One review caught seed pods drawn far too large, prompting a change in how the instructions specify the size of small plant parts.
These checks improve the odds of a faithful illustration. They do not make the images reliable for identification, and they cannot replace an expert's eye.
Rules that do not depend on an AI remembering them
Some failures are better handled by software than by another request to an AI model to be careful.
Each tradition has information marked as unsuitable for publication, including private individuals, protected sacred locations, ritual texts, and other sensitive details. Automated name checks can stop a piece containing a protected name; one caught a name embedded in a quotation. The system also checks that citations refer to sources actually supplied to the writer.
The substance pages have explicit subject restrictions. Software flags terms associated with excluded topics for review. Those flags are not treated as automatic proof of a violation: ordinary, harmless phrasing can trigger them too.
The technical process is designed to preserve its own history. AI calls pass through a logged interface that keeps the requests and replies. Runs are numbered rather than overwritten, and later automation is not allowed to erase the creator's hand edits. The software has more than 500 automated tests. Other AI models also examined the process design itself; their criticisms were considered individually, with accepted and rejected suggestions recorded along with reasons.
That is a substantial amount of software and quality-control work for a content website. It does not show that every article is right. It shows that the result did not come from accepting the first plausible output and moving on.
Who made the decisions—and why the connection matters
Jim Davis created the project. He owns Hummingbird Healing Center, an ayahuasca retreat center near Iquitos, Peru. Jim selected the traditions and inclusion rules, set the editorial direction, revised the prompts, read trial articles, and decided which corrections to make. AI produced the software, research, drafts, fact-checks, documentation, and images. The roles are different: AI generated the material; the human creator designed and governed the process.
Because of the connection to a retreat center—and because ayahuasca appears in several articles—the writing rules require a neutral tone. The articles mention ayahuasca tourism only where the research describes its effects on a tradition; they do not name, describe or recommend retreat centers. Claims about ayahuasca are meant to face the same scrutiny as claims about any other subject on the site.
Getting the facts right
Considerable effort went into researching, verifying, and reviewing these articles to make them as accurate as possible. Nevertheless, AI can make mistakes, and some errors may remain. We welcome corrections, particularly from researchers and members of the communities described here.
If you find an error, please contact us.
A starting point for discovery
Despite these limitations, we believe these articles offer real value by bringing together research that is often scattered across academic publications and languages.
They are not intended as authoritative references for academic research. Instead, they offer an accessible introduction to these traditions and a starting point for further exploration. The references accompanying each article provide a path to the original sources for anyone wishing to investigate more deeply.
Above all, we hope this site encourages curiosity and a greater appreciation of the remarkable diversity of these living traditions.