AI Sounds Confident About Herbs. It Can't Cite Sources.

AI Sounds Confident About Chinese Herbs. It Usually Can't Tell You Where It Got That.

The problem I kept running into

I started building MateriaTrace for a simple reason: when I tried to actually check what a Chinese herb was recorded to do, I could almost never find anything I trusted.

The information online is a mess. A lot of it is people saying whatever sounds right, with nothing to back it up. The same herb gets described three different ways on three sites, and none of them cite a source. A block of text about a herb's "uses" gets copied around until nobody knows where it came from.

Okay, so go to the actual classical text. But searching a thousand-page book by hand is brutal. You know the answer is in there somewhere. Finding the exact line means flipping pages until you give up.

So you ask an AI, or you search. Same problem, different shape. You get back a wall of confident, professional-sounding text that doesn't match any real source — and to confirm which book and which line it's from, you're back to digging through the original yourself.

A good book is too slow to search. AI can't find the exact line. That gap is the whole reason MateriaTrace exists. I didn't set out to invent new knowledge. I just wanted finding one trustworthy line of original text to be fast, and checkable.

"Fluent" and "trustworthy" are not the same thing

The AI era makes this worse, not better.

Ask any AI assistant what astragalus does. It'll rattle off a clean list — tonifies qi, secures the exterior, promotes urination. Sounds like an expert wrote it. Now ask the follow-up: which book, which line? Most of the time you get a vague "this is the general understanding in traditional Chinese medicine." Sometimes you get a citation that's completely made up.

That's the whole problem in one exchange. Fluent isn't the same as trustworthy. A passage can read perfectly and still be worthless if you can't trace it to a real, checkable source.

For trivia, who cares. But this is health information — what search engines call YMYL, "your money or your life." Here, a sentence with no source and a sentence backed word-for-word by the Bencao Gangmu (the 16th-century Compendium of Materia Medica) are not the same thing at all. And AI trains on the sourceless version, then hands it back to you in fluent prose. The errors don't go away. They just get harder to spot.

What MateriaTrace does the slow way

My fix is simple to describe and tedious to build: every field in the database points back to an exact spot in a classical text, with the original quote attached so you can check it yourself.

Here's what that feels like. When MateriaTrace says astragalus (huang qi) "tonifies qi and secures the exterior," you can click down and see exactly which scroll and passage of the Bencao Gangmu those words come from — and whether it's attributed to the Shennong Bencao Jing, the Mingyi Bielu, or a specific physician. It's not my summary. It's the original text. Every record carries a content fingerprint, so anyone can open the actual book and confirm it's there, word for word.

I've done this across all sixteen sections of the Bencao Gangmu — every herb in it, traced line by line. Then I added two classical formula texts, the Shanghan Lun and the Jingui Yaolue, and traced hundreds of formulas — their ingredients, doses, and preparation — straight back to the Song-dynasty originals, cross-checked against a separate public-domain edition. I'm still adding more. The recipe for Gui Zhi Tang — cinnamon twig, peony, licorice, ginger, jujube — is the exact wording of clause 12 of the Shanghan Lun.

This is the thing I wanted and couldn't find: not a tome to flip through, not an AI guessing, just a click straight to the original line.

One rule I don't break: I report what the texts say, not what they cure

This is the most important thing about MateriaTrace, and the easiest to misread.

I record a historical fact — "the Bencao Gangmu records this herb was used for that condition." I do not tell you the herb treats anything.

Those sound similar. They're not. The first is a sourced statement about a book. The second is a medical claim. And the modern science here is genuinely unsettled — the NIH's own center for this, the NCCIH, is blunt about it: the studies are mixed, nothing is conclusive, and you shouldn't use TCM to replace or delay real medical care. I take that seriously. So MateriaTrace stops at "what the text records" and never crosses into "what you should take."

Where the classics note toxicity, the traditional incompatibility rules, or cautions in pregnancy — I record those too. Where there's a plain modern fact about a substance (cinnabar is mercuric sulfide, for example), I note it. But all of it is "here's what's recorded, here's what's known," never "here's what to take."

Why so careful? Because traceability only works if you're honest about where every sentence comes from — what an ancient author wrote, what's modern fact, and what nobody actually knows. A database that blurs those lines and starts handing out advice is exactly the dangerous kind in the AI era.

The cheaper content gets, the more a checkable source is worth

Now that AI can spit out fluent text about Chinese herbs on demand, the text itself is nearly worthless. Anyone can mass-produce professional-sounding explainers for free.

So the scarce thing has changed. It's not the claim anymore. It's the source you can actually check. A reliable answer — whether a person finds it or an AI gives it — doesn't need one more fluent retelling. It needs facts that trace back to the primary text. That's what MateriaTrace is trying to be: every field nailed to an original you can verify line by line.

When anyone can generate content, that's one of the few ways left to earn trust.

This is just the start

The Bencao Gangmu is MateriaTrace's first source, not its limit. I'm steadily adding more public-domain classics — the theory in the Huangdi Neijing, more formula texts — held to the same standard: traced word for word, cross-checked, honest about the line between record and recommendation, and never made up.

Don't take my word for any of this. Try it: open astragalus and click down to its classical sources. See what it's like when a single statement traces all the way back to the original.

It's the thing I wanted and searched the whole web for and couldn't find. If it's useful to you too, see what else is in there.

MateriaTrace is a source-traceable, bilingual knowledge graph of Chinese medicine. It records how the classical texts describe each entry, with every field traceable to the primary source, and makes no treatment recommendations. Every classical record in this article can be checked word for word on its entry page.