What does the built-in AI score mean?

It is a likelihood estimate from the classifier paired with Human Writes, meant to show whether a draft still reads statistically machine-like after an edit. It is not a guarantee of what Turnitin, GPTZero, or a client tool will say.

Why does my Human Writes score disagree with GPTZero?

Different training data, thresholds, and text length. Disagreement is normal. Treat both as review signals. Fix highlighted stiff spans and your substance gaps.

Should I only trust the built-in score?

No. If a school or client names a third-party detector, pre-check in that tool when you can. Use the built-in score to iterate faster between edits.

How many times should I humanize based on the score?

One full pass, then manual edits on remaining flagged lines. Chasing zero across five tools usually hurts voice and meaning.

Built-in Detection Score vs GPTZero and Others

Two scores show up in the same afternoon: the one inside Human Writes after you rewrite, and the one from GPTZero, Turnitin, Originality, or Copyleaks. People treat the mismatch as a bug. It is a category error.

Built-in score = iteration instrument.
Third-party detector = external judge your audience might actually open.

Accuracy limits: How Accurate Are AI Detectors in 2026?. Edit habits: Best Practices for Humanizing AI Content.

Disclosure: we build the built-in score. We still tell you to respect the detector your school or client named.

What each score is for

Score	Best use	Bad use
Built-in (Human Writes)	Did this pass still sound machine-smooth?	Proof you will pass Turnitin
GPTZero / Winston / Grammarly AI	Quick self-check; span highlighting	Sole evidence in a misconduct case
Turnitin (school)	What faculty may see	Ignoring syllabus process
Originality / Copyleaks (client)	What the SOW named	Scanning only in a different free tool

If the numbers disagree, read the spans, not the rivalry.

The iteration loop

Lock facts, citations, and claims. No score fixes invented sources.
Paste into Human Writes. Note the before score if shown.
Humanize once.
Read the after score as a hint. Fix leftover stiff lines by hand.
If a third party matters, run that tool on the same final text.
Spot-edit what they highlight. Stop.

Do not ping-pong: Human Writes → GPTZero → another humanizer → Originality → despair. You will erase the only specific detail in the draft.

Why disagreement is normal

Detectors estimate probability from patterns (often related to perplexity and burstiness). They do not share one scale. Short text is noisy. Formal ESL can false-positive. Heavy editing can false-negative. See false positives.

A lower built-in score means the rewrite moved rhythm. It does not mint a passport.

Practical pairings

Audience	Iterate with	Confirm with
Student, Turnitin course	Built-in + your judgment	School report if visible; keep drafts
Freelancer, Originality SLA	Built-in	Originality in the named mode
Agency QA	Built-in during edit	Client-named tool before delivery (agency guide)
Fiction	Built-in lightly	Human readers; detectors fit poorly

What the score never measures

Whether the argument is true
Whether you followed the syllabus
Whether the client will like the brand voice
Whether you disclosed AI when required

Substance first. Score second. Voice third.

When to stop iterating

Stop when:

You can explain every claim out loud.
The third-party tool your audience uses no longer highlights whole pages of mush (a few spans are normal).
Another pass would only synonym-swap.

If the built-in score is mid and the third-party score is high, fix substance and specifics before you buy another rewriter. If both are low and the draft still sounds like a stranger, read it aloud and cut throat-clearing by hand.

Bottom line

Use the built-in score to decide if another manual tweak is worth it. Use the third-party detector your audience trusts to decide if you are ready to submit. Neither is a verdict.

Paste the locked draft on Human Writes, humanize once, then confirm in the tool that actually matters.