These are the rules and formulas the check actually runs, so you can trace every number in a report. Rules v2026-11-06.2 · AI reviewer v2026-11-06
1. What data is read
- Profile: name, bio, followers, following, account age, verification and X’s automated and sensitive labels.
- Posts and replies from the last 90 days: up to 8 pages of results, with a hard limit of 400 items.
- Follower sample: up to 200 follower IDs, whose profiles are checked for verification.
- Replies from other people under your 5 most-viewed original posts, used to judge whether engagement looks genuine.
- Posts that talk about review results are removed first, so the result depends on what the account normally posts.
- With fewer than 10 of your own items left, no report is produced and the points are refunded.
What we can’t see: images and video content, reports against you, media fingerprints, device data, age and region eligibility, and any other private signal X uses.
2. Requirements engine
This only decides whether you can apply, and never feeds the likelihood or content confidence. We can’t read X’s own verified follower and impression data, so those are estimates, and each item has its own fixed engine confidence. The figures in Creator Studio are the accurate ones.
| Requirement | How it’s calculated | Status | Engine confidence |
| Premium subscription | Reads the blue check field, or a Business / Government account type | Yes = pass; no = fail | 95 |
| 500 verified followers | Total followers × verified share in the sample (estimate) | ≥ 600 pass; ≥ 400 close; lower fails | 65 with a sample of 150+, otherwise 45 |
| 90-day 500,000 verified impressions | 90-day views on non-reply posts × verified share (kept within 5%–95%, or 30% with no sample); scaled up when under 90 days were read | ≥ 600,000 pass; ≥ 400,000 close; lower fails | 55 when 90 days were read, otherwise 40 |
| Account standing | Whether it’s protected, at least 90 days old, labeled automated, or restricted in its latest shadowban check | Protected fails; other issues are flagged | 70 |
3. Risk engine
This is what decides the result. Six signals each add points, and 2 points or more counts as a risk. Apart from sensitive content, no single signal decides it alone.
| Signal | Source | Trigger | Points |
| Sensitive content | Rules | At least 1 of your own posts, or your name or bio, matches a list of explicit sexual terms. Replies don’t count, because they often just discuss the topic | +2 |
| Media from other people’s posts | Rules | X’s media source fields show at least 2 such posts, making up 6% or more of posts with media | +1 |
| Follow-for-follow content | Rules + AI classifier | At least 2 posts or replies match follow-for-follow wording (such as f4f or follow back), or your name or bio does; or the AI labels at least 3 posts as follow farming. Both together still count once | +1 |
| Templated posts | AI classifier | The AI labels at least 2 posts as hollow templates, such as repeated slogans, check-ins and near-identical sentences | +1 |
| High risk found by the AI | AI reviewer | The highest risk across reused content, sensitive content, inauthentic behavior, engagement bait and automation reaches 65 | +1 |
| Mostly very short posts | Rules | The median length of your own posts is under 20 units. One English word is 1 unit, and about two Chinese characters are 1 unit | +1 |
4. How the likelihood is calculated
Likelihood = the AI’s estimate, limited to the range set by risk points | Risk points | Range | Result |
| 0 | 60–90% | Likely |
| 1 | 30–59% | Possible |
| ≥ 2 | 5–29% | At risk |
Risk points alone set the result, and the AI only picks the number inside the range. For example, with one weak signal and an AI estimate of 72, the range tops out at 59, so the likelihood is 59%.
5. The AI reviewer model
The AI reviewer works in two steps, and both read only text and media types.
Step 1: Label each post
It reads up to 60 posts, sampled evenly when there are more. It only labels each post and doesn’t judge the account.
- Aggregation
- News or others’ content retold with no view of your own. Counted, but not scored.
- Follow farming
- Follow-for-follow and mutual-growth posts. Scored as above.
- Engagement bait
- Asking for likes, reposts or follows in exchange for something. Counted, but not scored.
- Adult
- Sexual hints or adult content. Counted, but not scored.
- Promotion
- Paid promotion, airdrops and token calls. Your own products don’t count. Counted, but not scored.
- Template
- Hollow, repeated content. Scored as above.
Step 2: Review the account
It reads up to 45 posts (the 15 most viewed plus 30 sampled evenly), 35 replies and 30 replies from other people, then scores six dimensions. Higher is safer.
Content score = Originality × 30% + Reused content × 25% + Authentic engagement × 20% + Spam and manipulation × 10% + Content value × 10% + Policy risk × 5% | Dimension | Weight | 90+ | About 60 | Under 30 |
| Originality | 30% | Almost all personal views, experience, analysis or creative work | About half original, half retelling | Mostly retelling or aggregating others |
| Reused content | 25% | Almost no sign of reuse | Occasionally uses others’ material, with content of its own | Heavy reuse of others’ video, images or text |
| Authentic engagement | 20% | Specific replies from real discussion | Some stock phrases or follow-for-follow traces | Obvious engagement rings or mechanical replies |
| Spam and manipulation | 10% | Normal pace | Occasional bursts or templates | Lots of repetition, templates and rapid-fire posts |
| Content value | 10% | Views or information with value of their own | Ordinary everyday sharing | Almost only traffic-chasing content |
| Policy risk | 5% | No risk | Occasional engagement bait | Repeated clear violations |
The content score doesn’t decide the result. It appears in the report and feeds the consistency part of confidence.
Risk types the AI reviewer checks
- Reused content
- Copied, lightly edited, aggregated or cross-posted content from others
- Inauthentic behavior
- Coordinated engagement, fake amplification, engagement rings
- Mutuals
- Organized follow-for-follow aimed at swapping followers
- Engagement bait
- Repeatedly asking for likes, replies, bookmarks, follows or reposts
- Automation
- Templated filling and uniform, high-volume posting
- Sensitive content
- Adult content, gore, hate and extremism, and restricted topics such as disasters and war
- Past violations
- The content itself mentions a violation or paused monetization
- Other standards
- Other clear problems that fit none of the above
Rules given to the AI
- It judges content only. Premium, follower counts and impressions are never given to it as reasons, and it doesn’t receive your follower count or view counts.
- Topics and opinions aren’t penalized. Politics, crypto and controversial views aren’t problems in themselves.
- It can’t see images or video. Unless the text is clear or X marks the media as coming from someone else, reused and sensitive risk can’t exceed 40.
- With no evidence, a dimension stays neutral (about 70–85). Writing that merely sounds like AI isn’t treated as a violation, and “content value” can’t be a reason for rejection on its own.
- It isn’t told the account’s real result. Everything in your posts is treated as data, so a post telling it to give a high score has no authority.
- It runs with temperature 0, to keep results as stable as possible for the same input.
How code checks the AI’s output
- Every cited post must exist in the input; anything else is dropped and counted in the report.
- A dimension score below 60 without a verifiable citation is raised to 60, so suspicion alone can’t lower a score.
- Only the listed risk types are accepted.
- The AI’s likelihood estimate is forced into the range set by risk points, so it can’t change the result.
6. Content review confidence
Confidence is calculated by code rather than reported by the AI. It shows how much evidence supports this result, and it isn’t the likelihood or the accuracy rate. Every report shows the breakdown.
| Item | Points | How it’s calculated |
| Base | +10 | Fixed |
| Amount of content | +25 | Full points at 150 usable items, proportional below |
| Time span | +15 | Full points at 90 days, proportional below |
| Evidence | +20 | Full points at 12 verifiable citations, proportional below |
| Consistency | +20 | Higher when the AI’s estimate and the content score are closer; 0 at a gap of 40 or more |
| Clarity | +10 | Higher the further the likelihood is from 50%; full points at 30 or more |
| Uncertainties | −10 | −2 for each thing the AI says it can’t judge, up to −10 |
Capped at 90. 70 or more is high, 45–69 is medium and under 45 is low.
7. Validation
Leave-one-out validation on 49 real accounts, run on 2026-11-06. Their results come from what creators shared publicly.
84%Passed accounts correct21 / 25
75%Rejected accounts correct18 / 24
80%All accounts correct39 / 49
Known limits
- Only 49 accounts, all from creators who chose to share a result publicly, so the sample may not represent everyone.
- Signal thresholds were tuned on these same accounts and haven’t been tested on a fresh set. Accuracy on new accounts may be lower.
- Used alone, the AI performed poorly in blind testing, getting only 25% of rejected accounts right. That’s why it doesn’t decide the result.
- If an account has already removed content, the check judges what remains.
The result is only a risk reference. X makes the final decision.