/Voice check

See what ChatGPT writes for you, and how far it sits from your voice

You type your website. Fractl Agents reads your blog posts, keeps three of them aside, and has ChatGPT write those three again from a brief of what each one says. Then a judge that has read your other posts is shown your post beside ChatGPT's and asked which one you wrote. You see the three posts next to both ChatGPT versions, and how often the judge told them apart.

The address of the site with your blog, such as 'example.com'. A full link works too; Fractl Agents keeps only the site's name. It needs at least eight posts of 300 words or more to measure your voice.

How long should the run take?
  • A run takes two, three or four minutes, your choice. Most finish sooner. You can start two runs a day, and a site is checked at most three times a day.
  • The run is free. Only you see the results. The three posts are shown as a short opening and a link; the ChatGPT versions in full.
  • Keep this tab open. Your results live in this tab until you send them to your email.
/Voice check Ready

SiteYour website appears here

Fractl Agents has not started. Type your website, then press Check my voice.

What Fractl Agents does during the run

  1. Fractl Agents reads your blog posts

    It lists the pages on your site from its sitemap and its home page, reads the ones that look like posts, and keeps up to 25 that run to 300 words or more, in English, and are not podcast episodes. Bylines, dates and share lines are trimmed away. This takes about half a minute.

  2. Three posts are set aside and briefed

    Three posts are held out, chosen by a fixed rule so the choice is not ours. For each, a small model writes a brief of what the post says: its points, in order, with the numbers and names it uses, and nothing about how it is written. From the other posts, Fractl Agents also builds a description of your writing style.

  3. ChatGPT writes each post twice

    From the brief alone, the way most people prompt it. Then again with the style description added, the way a careful marketer would. Company names are masked on every side, so the judge cannot cheat by spotting a name.

  4. A judge that has read your other posts picks the one you wrote

    The judge is a model shown your other posts, nothing else. For each held-out post it sees your post and a ChatGPT version side by side, twice with the sides swapped, and says which one you wrote and how sure it is, out of 100. Twelve judgements in all.

  5. You read the verdict, then the evidence

    First the verdict in plain words: whether ChatGPT could pass as you, what the judge kept noticing about your writing, and what that means. Then, if you want it, the three posts beside both ChatGPT versions with the judge's reason under each.