AI Tell Scan

Find the UI defaults costing you trust.

Read-only source evidence, context review, and no authorship guess. You get only the three findings worth fixing.

Public GitHub repository Read-only source scan React / Next.js

Not an AI detector. Not an authorship verdict. Not a shame score.
Controlled evidence

Precision comes from refusing the easy answer.

In a controlled 30-repository gold corpus, the deterministic pass found 23 correct candidates with zero false positives and zero false negatives. Context review confirmed 20 and correctly rejected 3.

23correct candidates
20confirmed findings
3context rejections
0FP / FN

If every candidate were auto-confirmed, precision would be 86.96%. The review pass is the product, not cleanup. This is a regression corpus result, not a population-wide claim.

It does not detect AI. It detects defaults.

A gradient headline is not a finding. A pill is not a finding. Even a glass card is not a finding on its own. AI Tell Scan looks for composite signals: repeated, co-located choices that flatten product meaning and make unrelated interfaces converge on the same visual language.

Every candidate is checked against nearby content, components, and framework structure before it can appear in the final report. Context can confirm a tell, or reject it.

Ten families. No single-property gotchas.

Each family requires a compound pattern before it becomes a candidate.

  1. 01

    Gradient-clipped display headings

    A decorative treatment repeated where plain hierarchy would carry more meaning.

  2. 02

    Aurora-centered hero composition

    Soft atmospheric fields, centered copy, and no product-specific visual anchor.

  3. 03

    Floating glass navigation

    Detached translucent navigation used as a default shell rather than a product choice.

  4. 04

    Icon-card triptychs

    Three equal feature blocks with interchangeable icons and identical information weight.

  5. 05

    Repeated micro-kickers

    Uppercase labels above every heading until every section announces itself the same way.

  6. 06

    Round metric ornaments

    Numbers presented as decorative proof without enough source, unit, or decision context.

  7. 07

    Glass-card field systems

    Four or more translucent framed panels competing with the marketing task they surround.

  8. 08

    Pill role overload

    Tags, controls, statuses, and navigation all reduced to the same rounded capsule.

  9. 09

    Spring hover everywhere

    The same lift-and-scale motion attached to unrelated controls, content, and decoration.

  10. 10

    Multicolor card wash

    A rainbow of low-contrast panels substituting palette variety for information structure.

A finding has to survive three gates.

  1. 1

    Candidate scan

    Deterministic rules inspect source structure and produce file-and-line evidence. No browser theater, no vibes-only classifier.

    status: candidate
  2. 2

    Context review

    The agent checks the surrounding interface and product meaning. Weak or intentional matches are rejected with a reason.

    confirmed | rejected
  3. 3

    Digest-bound finalize

    The report is finalized against the exact scan digest. Changed evidence requires a rescan, and only the Top 3 ship.

    ats-1 / final

Enough evidence to act. Not enough noise to hide in.

The final report stops at three findings. Each one names the composite tell, points to the source, explains the credibility cost, and suggests the smallest product-specific correction.

Evidence
file:line, signal family, observed components
Decision
confirmed or rejected with context rationale
Priority
Top 3 visible credibility costs only
Safety
read-only; no source modification

Find the three defaults your launch can do without.

Read-only. At most three context-confirmed findings.

View the independent scanner source