How can you tell if a monk in a video is actually an AI avatar? The checks we use for every profile in our database

We maintain the AI Monk Database, a public, source-linked directory of monk-like video personas. When we re-read the live directory and its public data file on 11 November 2026, it displayed 76 profiles. Only 6 of those were evidence-reviewed video personas. The other 70 were still pending research candidates, and our wider 78-record research archive kept 2 more accounts outside the directory altogether because they publish illustrated storytelling, not realistic video avatars.

That gap — 6 reviewed, 70 pending — is the honest answer to the question in the title. In most cases, you cannot tell from the video alone, and we do not pretend otherwise. A robe, a temple backdrop, an old face, and a calm voice establish nothing about production. What establishes an AI avatar is evidence outside the frame, and this article is the checklist we apply, profile by profile, before a candidate moves into the reviewed set.

What does not count as evidence of an AI avatar?

Three things we see constantly, and score at zero:

• The look. Robes and temple scenery are a format choice. A real teacher, a hired actor, and a synthetic persona can all wear the same orange robe in the same setting.
• The audience size. Follower counts tell us where to look next, never what we are looking at. In our database, the largest candidate profiles sit in the pending set precisely because reach arrived before evidence did.
• A screenshot or a name. A screenshot-visible handle seeds an unverified lead in our archive. It does not establish AI production, a video format, a creator, or a launch date, and we leave those fields empty, with a reason, until a source fills them.

If a profile has only those three things, it stays a candidate. That is not a judgment that it is real. It is a record that we do not know yet.

What is the first check? Find a source that names the production

Our strongest inclusions start with a source that speaks about how the persona is made, not how it looks.

Yang Mun is the clearest case in the directory. HeyGen's customer story identifies Yang Mun as an AI character created by Shalev Hani, and Columbia News Service independently describes the character as AI-generated and names the same creator. Provider evidence plus independent reporting, with sampled videos linked on the platform itself, is what places Yang Mun in the reviewed set.

A platform label can do part of the same work. SifuChen California is included because the first-party profile discloses AI-generated content and an account-attributed Instagram Reel displays Instagram's AI content label. Maitre Yun is included on the same pattern: the first-party profile identifies an AI-generated avatar and fictional digital character, and a Reel carries the platform's AI label. In both cases we record the limit in the same breath — this supports inclusion as a reported AI video persona. It is not an independent production audit, and we do not upgrade it into one.

What is the second check? Separate reported from unverified, and keep both labels visible

Every record in our archive carries an AI status, and on 11 November 2026 the split was 7 reported and 71 unverified across the 78 records. Reported means a creator, a platform, or a credible investigation makes the AI claim and we can link to it. Unverified means the claim has not been independently checked. Candidates remain visible with that status attached, rather than being quietly promoted or quietly deleted.

The investigation we lean on most is Maldita's July 2026 study of AI-generated health accounts, which collected 250 AI-generated health profiles. Ming San and Yan Chen enter our reviewed set through that investigation's classifications, and we keep the wording at that strength: Yan Chen's entry notes that the investigation attributes a Reel still of a realistic monk persona to the account, and that we have not independently watched the full clip. A reader should be able to see exactly how far each claim goes, because the distance between "an investigation reported this" and "we verified this" is where most directories mislead.

What is the third check? Preserve the conflict instead of voting it away

Conflicting claims are data, not noise. Yang Mun's record carries an identity conflict we have never resolved away: a first-party page linked from the Columbia report says Yang Mun is a real teacher and only the voice is AI-translated, which contradicts both HeyGen's and Columbia's descriptions of an AI-generated character. Both sides stay in the record, labelled by source.

Ming San carries a smaller version of the same pattern: Maldita classifies the profile as AI-generated, while another directory source says the public evidence does not establish whether the persona is AI. We preserve that disagreement in the profile instead of picking the answer that makes the database look tidier. If we merged or flattened conflicts like these, the directory would read more confidently and be worth less.

What is the fourth check? Confirm there is actually a video persona to review

Discovery evidence is not video evidence. Our pending reason, repeated across the 70 candidates, is deliberately plain: discovery and profile evidence does not yet establish a realistic synthetic monk video persona. A profile can be large, active, and commercially interesting — Chen Tao, a candidate, showed 976K followers in a dated Instagram snapshot on 31 October 2026 — and still fail this check, because we have not established the production or the video format.

The same check excludes a neighbouring format entirely. Kai Tsukimi and Buddha Would Say sit in the archive as excluded records because they publish illustrated and animated storytelling. They are monk-themed, they are not realistic talking video avatars, and counting them as AI monks would inflate the field we are trying to measure.

How do we record numbers so they do not harden into claims?

Every number in the database is stored with its platform, its source, and its capture date, and different kinds of numbers are never merged. Instagram followers, YouTube subscribers, video views, seller-displayed counters, and creator revenue statements are different measures. When the live Yang Mun profile shows a creator-reported figure of more than $213K gross over three months, we display it as creator-reported and not audited, with its source and date, alongside a note that the creator's own later claims do not reconcile to a single Yang Mun-only total. A displayed counter is a lead for the next investigation. It is never presented as orders, profit, or proof.

The same discipline is why our follower figures are dated snapshots — the counts behind the current directory were refreshed on 31 October 2026, and accounts without a newer measurement keep their 23 October 2026 snapshots, labelled as such. An undated follower count looks current forever, which is exactly the failure we are trying to avoid.

Where can you apply these checks yourself?

Start from the searchable AI Monk Index: filter by language, platform, or region, open a profile, and follow its sources before you repeat any claim about it. The full methodology and the profiles behind this article are written up in The Art of AI Monks, our sourced English book — 71 pages, 10 chapters and 4 worksheets, $9.99 in PDF and EPUB as listed on 11 November 2026, with the full manuscript free to read on GitHub. If, after checking, you want to build an original character of your own rather than copy anyone above, AI Monk Factory is our production route — $199 per month as listed on the same date, with the first 3 completed videos free and no card required. More case studies in this series are collected on the AI Monk Avatars hub, and our earlier field guide, The Art of AI Monks — AI僧侶列伝, works through the first reviewed profiles in depth.

Is a large follower count proof that a monk is an AI avatar?

No. Follower counts decide where we investigate next, not what a persona is. Several of the largest profiles in our archive are still pending candidates because no source has yet established their production.

Does an Instagram AI content label settle the question?

It helps, and it does not settle it. In our rubric, a platform AI label together with a first-party disclosure supports inclusion as a reported AI video persona. It is not an independent audit of how the videos are made, and we record it at that strength.

Can a real monk use an AI avatar?

Yes, and that is why appearance alone fails. Our classification separates three cases that look alike on a phone: a synthetic persona, a real religious teacher speaking through a digital character, and an account whose origin has not been established. The production evidence — not the robe — decides which case a profile belongs to.

What should you do when the evidence conflicts?

Keep both claims, labelled by source, and say that they conflict. That is what we do for Yang Mun and Ming San. Resolving a conflict by deleting the inconvenient source makes a directory look certain at the exact moment it stops being trustworthy.