Why style-spotting fails now
People look for tells: overly even sentence length, hedging, generic structure. Those heuristics worked on raw model output, but a light human edit erases most of them. They produce false alarms too. Plenty of careful human writing is even and well-structured.
A detector looks at statistical properties of the text rather than vibes, which is more consistent than style-spotting. It still isn't certain, but it doesn't get fooled by a confident tone or thrown off by clean prose.
Reading verif·ai's text result
verif·ai puts a soft tint behind passages it judges likely AI-written and scores each one out of 100, with a high / medium / low level alongside. Because the tint is on the text itself, you can see which parts of an article are flagged and which read as human, rather than getting a single verdict for the whole page.
Treat a high score as a strong reason to dig deeper and a low or medium one as inconclusive. The point is a calibrated read on what to trust, not a stamp of real or fake.
how it works
- 01
get verif·ai in your browser
With verif·ai in Chrome it's available on any page.
- 02
open the page
Navigate to the article or page whose text you want to check.
- 03
hit scan
Open verif·ai and scan. It evaluates the page's text on-device.
- 04
read the tint
Likely AI passages get a soft tint and a confidence score out of 100.
frequently asked
Does verif·ai check the whole page or one passage?
It scans the text on the page and tints the passages it judges likely AI-written, so you see which parts are flagged rather than a single overall verdict.
Can it be fooled by edited AI text?
Editing makes any detector less certain, which is why verif·ai reports a confidence score and level instead of a yes/no. Use it as a signal, not proof.
Is my reading sent to a server?
No. verif·ai runs on-device in your browser, with no sign-in and no server-side history of what you scan.
What's a high/medium/low level?
It's a plain-language band for the numeric score, so you can act on 'high confidence' without parsing the exact number.
Last updated June 7, 2026