# Can TikTok Comments Prove That a Video Was Made With AI?

No. TikTok comments alone cannot prove that a video was made with artificial intelligence, or AI. Comments record what people say. To check the claim, find the publisher’s source version, inspect any direct disclosure about how it was made, and check available records of its origin and edits. Save each precise claim with its evidence and limits. Leave the origin unresolved when the evidence does not settle it. Missing AI labels or Content Credentials do not prove human origin.

Start with the video and its publisher’s available records. You can collect comments later if they help you list the claims. A comment collector can save public statements. It does not decide whether those statements are true.

## 1. State Exactly What You Need to Check

Two public questions show the problem. One reader began to doubt a snow video after comments called its appearance unnatural. Another could not find an original version of a boat video. These questions show a need to check evidence. They do not establish that either video used AI. See the [snow-video question, retrieved August 11, 2026; reader concern only](https://www.reddit.com/r/isthisAI/comments/1q69a8c/it_this_video_ai_it_came_up_to_me_on_tiktok_and/) and the [boat-video question, retrieved August 11, 2026; original source not found by the reader](https://www.reddit.com/r/isthisAI/comments/1t517k2/a_boat_passes_through_haulover_inlet_in_this/).

The phrase “made with AI” can refer to different claims. Generative AI produces content, such as images or video. An AI editing tool can change existing footage. A claim about the depicted event asks a separate question.

Write each claim as a statement that evidence could support or contradict:

1. The whole scene was generated with AI.
2. An AI tool changed part of recorded footage.
3. No AI tool was used.
4. The caption gives a false account of the scene.
5. The depicted event happened as shown.

Do not combine these statements into one yes-or-no question. Evidence of one AI edit does not establish full-scene generation. Evidence about the production process does not, by itself, establish that the event happened.

Save the known video link, post identifier, review date, and version under review. An identifier is a value used to distinguish one record from another. Describe the version clearly enough to compare it with another copy. For example, note the visible scene, caption, and relevant point in the video.

## 2. Inspect the Video and Available Publisher Records

Watch the actual video before you judge an allegation. Find the publisher’s available original upload, description, labels, and linked source material. Record what each source says and which version it concerns.

Use this order:

1. Open the known public post.
2. Read its description and available production disclosures.
3. Follow relevant source links supplied by the publisher.
4. Compare any source version with the reviewed video.
5. Record the match and any remaining difference.

A version match matters. A disclosure for another edit or another upload may not apply to the video you are checking. Compare the scene and the relevant edit. If the connection remains uncertain, record that limit before drawing a conclusion.

An earlier-looking upload is a possible source, not an established original. If you cannot find the original, write “original source not found in this review.” Do not replace that gap with a claim that the video must be generated.

Public posts, descriptions, and linked files can support this review. Owner-only project files, private account records, and unpublished original assets are separate evidence. Do not imply that you inspected them unless they were supplied and checked. A publisher’s statement about a private file remains a statement until the relevant file can be checked.

Keep production research separate from this claim review. A [TikTok creative example library](/en/articles/tiktok-ugc-creative-library/) serves a different purpose. A useful scene reference does not establish how the scene was made.

## 3. Collect Comments Only When They Help Define the Claims

For a few visible allegations, manual notes may be enough. Save the exact relevant wording, its public source, and any reply context needed to understand it. A small collection can help when you need a structured record of several statements.

An Actor is a hosted data collection program. The TikTok Comments Scraper documents collection from known public TikTok posts. Its documented role fits the comment-recording step. The reader must perform the evidence review.

The following is the supplied small input. It matches the documented input structure. No run was made for this guide. The sample address is a format example, not evidence of a media allegation.

```json
{"includeReplies":true,"maxCommentsPerVideo":10,"maxRepliesPerComment":2,"startUrls":["https://www.tiktok.com/@zachking/video/7631200623693729038"]}
```

3. Collect Comments Only When They Help Define the Claims

Before a proposed run, replace the whole `startUrls` list with the full public post address or numeric post identifier you intend to check. The documented input accepts 1–100 addresses or identifiers. It supports video and photo/carousel posts. Equivalent inputs are collected once. Short links are not supported in the checked contract.

The two comment limits are separate:

- `maxCommentsPerVideo` limits saved top-level comments for each post. A top-level comment is a comment posted directly under the post.
- `includeReplies` controls whether replies are also saved.
- `maxRepliesPerComment` limits saved replies for each top-level comment when replies are enabled.

For one post, the supplied limits give this possible maximum:

`10 top-level comment rows + (10 top-level comments × 2 reply rows per comment) = 30 comment rows`.

That is a limit calculation, not a result or a promise of 30 rows. Replies are separate billable rows. The documentation says replies increase runtime and cost. No price or measured runtime is established here. See the [Actor build 0.1.17 contract and README, checked October 25, 2026; documentation only, no run](https://api.apify.com/v2/acts/chronometrica~tiktok-comments-scraper/builds/default).

The [Export TikTok Comments to CSV Task, checked October 25, 2026; public page presence only](https://apify.com/chronometrica/tiktok-comments-scraper/examples/export-tiktok-comments-to-csv) is a published Task page. A Task is a saved input preset. Its saved settings were not inspected through an authenticated account, and it was not run. Check the inputs yourself before using it.

## 4. Save Comment Context and Build a Claim Ledger

A claim ledger is a table that keeps each claim beside its evidence and conclusion. Build it yourself. It is not an analysis output promised by the collector.

First, preserve the documented comment fields needed to trace each statement:

| Field | What to preserve |
| --- | --- |
| `inputUrl`, `postUrl`, `postId` | Supplied input and the post identity |
| `commentId`, `text` | Comment identity and exact available text |
| `parentCommentId`, `replyToCommentId` | Available parent and direct reply context |
| `isReply`, `replyDepth` | Whether the row is a reply and its documented depth |
| `createdAt`, `scrapedAt` | Comment creation time when available and collection time |
| `authorId` | Public author identifier when available |
| `replyCount`, `status` | Available public reply total and saved row outcome |

The output fields are documented fields, not results from a current run. A reply row is another statement in a conversation. It is not automatically an independent source of evidence.

Keep missing values distinct from zero. A missing `likesCount` does not mean zero likes. A missing `authorId` does not identify a new person. A missing `text` value leaves the statement unavailable for text review. Do not replace it with an empty statement and then classify it.

The documented `replyCount` is a public reply total for a top-level comment. It is null for a reply row. It is not the number of reply rows you saved. Likewise, comment rows are not a count of distinct people.

Then create these reader-owned fields:

| Ledger field | What to save |
| --- | --- |
| `claim_id` | A stable identifier for this claim |
| `exact_allegation` | One precise statement to check |
| `source_comment` | Comment reference, wording, and required context |
| `evidence_needed` | The evidence needed to decide this claim |
| `checked_source` | Source link, checked version, and review date |
| `what_it_establishes` | The narrow finding supported by that source |
| `limit` | Missing evidence, uncertainty, or scope restriction |
| `conclusion` | Supported, contradicted, or unresolved |

Group repeated versions of the same allegation. Preserve their comment references, but do not turn repetition into separate supporting evidence. A highly liked comment can help identify a question to check. Its likes do not verify the answer.

## 5. Check Available Content Credentials

Provenance means information about a file’s source and history. Content Credentials can carry signed provenance records. C2PA is the standards group behind the referenced Content Credentials specification.

When a publisher offers an original asset with Content Credentials, inspect that asset with a tool that can read and validate them. An asset is the specific media file being checked. Validation checks the signed record and its connection to that file.

Save the following details:

1. The asset link and version.
2. Why it matches the reviewed video.
3. The inspection tool and its version.
4. The validation result as reported.
5. The named signer or disclosed action.
6. The finding that applies to the precise claim.
7. Anything the record leaves unresolved.

Do not invent a validation result when you have only seen a label. If the asset is unavailable, record “asset unavailable.” If you cannot inspect its credentials, record “credentials not checked.” If a tool reports a validation problem, preserve that result without changing it into an AI-origin verdict.

A disclosed editing action can support that stated part of the production history. It does not establish that every frame was generated. A signed provenance record also does not prove that every depicted event is true.

C2PA describes provenance as optional and potentially incomplete. Its presence or absence alone does not establish media truth. See the [C2PA Content Credentials Explainer 2.2, retrieved October 25, 2026; provenance and trust limits](https://c2pa.org/specifications/specifications/2.2/explainer/Explainer.html).

## 6. Compare Evidence With Each Exact Claim

Use three conclusion states consistently:

- Supported: The checked evidence supports the precise claim within its stated scope.
- Contradicted: The checked evidence conflicts with the precise claim.
- Unresolved: The available evidence cannot settle the claim.

Keep the source check separate from the conclusion. “Source unavailable” describes an evidence gap. It does not mean the claim is false. “No disclosure found” describes the scope of your search. It does not establish that no AI tool was used.

Compare possible explanations before you decide. An unusual-looking scene might lead you to examine full generation, an AI edit, or a real scene with an unusual camera effect. Record the evidence needed to distinguish those possibilities. Do not choose one because it attracts the most comments.

For each source, ask whether it actually addresses the claim. A publisher may disclose one removed object. That finding addresses an editing action. It does not settle whether the whole scene was generated or whether the event occurred.

Use the same discipline as a [checked page evidence record](/en/articles/web-agent-evidence/): preserve the source, version, relevant statement, and support limit. This review adds the need to match the media version and separate production history from event truth.

## 7. Worked Example: Six Comments and One Disclosed Edit

Original fiction. No real video, creator, inspection result, or detector result is represented below.

Suppose a manual sample contains six comment rows. Four claim that the whole scene was generated with AI. One calls it ordinary footage. One is unclear.

The claim share is:

`4 matching comment rows ÷ 6 sampled comment rows × 100 = 66.7% of sampled comment rows`, rounded to one decimal place.

The unit is sampled comment rows. This does not measure the share of distinct people, all comments, or independent evidence sources. It is not a 66.7% probability that the video was generated.

In this fictional review, you find a publisher disclosure stating that an AI tool removed a background object. You check that the disclosure concerns the reviewed version. No independent event source is available, and Content Credentials are unavailable for inspection.

| Claim | Checked evidence | Limit | Conclusion |
| --- | --- | --- | --- |
| The whole scene was generated | Publisher discloses removal of one background object with an AI tool | One editing action does not establish full-scene generation | Unresolved |
| An AI tool removed a background object | Matching publisher disclosure states that action | Supports the stated editing disclosure only | Supported within the disclosure’s scope |
| No AI tool was used | The same matching disclosure states that an AI tool was used | Depends on the checked source and version match | Contradicted by the disclosure |
| The depicted event really happened | No independent event source is available | Production disclosure does not verify the event | Unresolved |

The four repeated comments do not establish full-scene generation. The “ordinary footage” comment also does not establish that no AI tool was used. Its wording needs a precise claim before review.

The next evidence depends on the remaining question. A full-scene generation claim needs evidence that addresses the whole production process. The event claim needs independent evidence about the event. Unavailable Content Credentials answer neither question.

A careful report would say: “The matching publisher disclosure states that an AI tool removed one object. This review does not establish full-scene generation or verify the depicted event.”

## 8. Check the Review Before You Share It

Use these checks on the complete ledger:

1. Confirm that every conclusion refers to one exact claim.
2. Confirm that each source applies to the reviewed version.
3. Separate comment statements from production evidence.
4. Keep repeated statements grouped without calling them independent support.
5. Preserve missing values instead of changing them to zero or false.
6. Keep unavailable sources and unchecked credentials visible.
7. Remove conclusions based only on appearance, confidence, or absent labels.
8. State the next evidence needed for each unresolved claim.

If the original remains missing, retain that gap. If a disclosure applies to another version, do not use it to settle this version. If reply context is unavailable, leave any context-dependent interpretation unresolved. If a proposed collection has not run, report the proposed input rather than invented row counts.

Share redacted records or clearly marked fictional examples when practical. Do not accuse a named person of deception from visual oddities. Report what the checked sources establish and where the review stops.

## Take the Next Evidence Step

Write down one precise claim and check the publisher’s available source version first. Complete a ledger row with the evidence, limit, and conclusion. An unresolved result is useful when it states exactly what remains unknown and what evidence could resolve it.

If a small public comment sample will help list the allegations, review the [TikTok Comments Scraper Actor, contract checked October 25, 2026; documented collection fit only, no run verified](https://apify.com/chronometrica/tiktok-comments-scraper). Replace the sample address and inspect the settings before a proposed run. Use the collected statements as inputs to your review. Keep responsibility for the evidence decision with the reviewer.

## Sources

- [Public snow-video question, retrieved August 11, 2026; evidence of reader uncertainty, not AI origin](https://www.reddit.com/r/isthisAI/comments/1q69a8c/it_this_video_ai_it_came_up_to_me_on_tiktok_and/)
- [Public boat-video question, retrieved August 11, 2026; reader reports a missing original, not a verified origin finding](https://www.reddit.com/r/isthisAI/comments/1t517k2/a_boat_passes_through_haulover_inlet_in_this/)
- [TikTok Comments Scraper build 0.1.17 contract and README, checked October 25, 2026; documented inputs and outputs, no runtime result](https://api.apify.com/v2/acts/chronometrica~tiktok-comments-scraper/builds/default)
- [Export TikTok Comments to CSV Task, checked October 25, 2026; public page presence only, saved settings and execution unverified](https://apify.com/chronometrica/tiktok-comments-scraper/examples/export-tiktok-comments-to-csv)
- [C2PA Content Credentials Explainer 2.2, retrieved October 25, 2026; optional and potentially incomplete provenance does not establish media truth](https://c2pa.org/specifications/specifications/2.2/explainer/Explainer.html)