Trust, limits and history
This page says what exists at Real Press today, where a record stops short, what happened to AI detection, and who to write to about your data or a security problem.
What exists today
As of . The date moves when this list changes.
Live
- The interactive demo, including a fingerprint check that runs in your browser. It compares content only: it does not check the signature, the log proofs, identity or time.
- A synthetic sample you can check offline with the open-source verifier. The sample states its own limits: a test key held by Real Press signed it, no identity is asserted, and time is not established.
Not live yet
- Identity checks. Nothing checks who holds a signing key.
- Public receipts. There is no page where an editor can look up a real receipt.
- Time evidence. No receipt carries an external timestamp.
- The pilot. It is planned, by invitation, and has not started.
- The private prototype. It exists and is not running.
AI detection is not part of this list. It is retired. See what happened to AI detection.
Limits
These limits apply to every Real Press record, including the synthetic sample.
- Authorship
- A Real Press record does not prove human authorship, and we will never claim that it does.
- Attestation
- A record links a signing key to one exact version of a text. It says nothing about who composed the text or how.
- Witnesses
- Version 1 has no independent witnesses. Real Press runs the log, so readers rely on Real Press to show everyone the same log.
- Custody
- In the prototype, Real Press holds signing keys on its own servers. A record says which custody applied. A server-held key shows that Real Press signed for an account; it does not show that a writer personally held the key.
- Identity
- Account or domain control is not legal identity.
- Absence
- When no record exists for a text, Real Press says no attestation found. That is not evidence that anything is wrong with the text.
What happened to AI detection
Real Press scored web pages for AI likelihood. It ran text classifiers and writing-pattern checks on each page, and image checks where a page had images, and combined their outputs into one label on a five-step scale, shown as a badge beside every result. The labels were estimates. They did not establish who wrote a text or how it was produced, and they were less reliable on short text, on writing in languages other than English, and on text reworked to avoid detection.
| Event | Date |
|---|---|
| Decision to stop AI detection | |
| Crawler stopped and scheduled scoring jobs removed | |
| Public checks closed | |
| Developer API closes | About 30 days after account holders are notified |
Badges on other sites
Badge snippets that other sites embedded may still carry the alt text “Real Press Verified Human”. We cannot edit copies on other websites. That wording was wrong, and Real Press no longer uses it.
Transparency reports
We said we would publish quarterly transparency reports. We have not published one. The current Terms of Service still mention those reports; revised Terms will remove the promise after the notice period.
Scoring defects and corrections
These are the defects we can date from our own records. It is not a complete list.
Very short pages were given a placeholder result. Pages of 200 to 249 characters were below the length every detector needed, so no detector could score them. The system stored a neutral placeholder score instead, and that score fell in the range for an AI label. One further path stored a result as human with a score of zero when no detector reported any confidence. From this date, pages too short to score get the status “insufficient signal” and no score, and they no longer appear in search. Results of this kind already stored for 6,794 articles were not removed.
Image scores carried no information. The image detector returned almost the same value for every image, for about five months, and nothing flagged it. That value counted as a vote for AI, so pages with images were labelled human about a thirtieth as often as pages without images. On this date the image output was set aside and 1,091,119 pages that contained images were rescored. The label changed on 149,093 of them. The gap between pages with and without images narrowed but did not close.
A placeholder score became an AI label for shop products. Product descriptions shorter than the detectors’ minimum were given a neutral placeholder score, and the threshold for the AI label treated that placeholder as AI. 4,579 of 12,713 products, more than a third, carried a label that came from a placeholder, not from an analysis of the text. On this date the code was fixed and the placeholder scores were removed, so those products carry no score.
Percentages looked like probabilities. Results showed a percent-human and a percent-AI figure. They were the composite score converted to a percentage, which made an estimate look like a probability. On this date they were replaced with the five-step label and a three-level indicator of how well the detectors agreed.
The classifiers are kept privately. They are not part of the product.
Data and removal requests
To ask us to remove a scan of your page, or to ask about data we hold about you, email privacy@real.press.
What the AI-detection archive stored
Its records included:
- For each page it scored: the web address, the page text, the title, the author name and publication date when the page gave them, and the scores with their history.
- The images it found on those pages, with a score for each.
- An average score for each author name taken from page bylines.
- Scores for online shops and their products.
- The web address or text that a signed-in person submitted for a check, linked to their account.
- Screenshots that people uploaded for a social-media check, stored as images.
- A record of each submission attempt: the visitor’s IP address and the time, kept for rate limiting.
- For people who signed in with Google or GitHub: the name, email address and profile image those accounts provided.
Closing the archive pages did not delete the records behind them. The privacy policy describes how deletion requests are handled.
Corrections
Corrections to this site and to what Real Press has said, newest first.
AI-detection wording retired from this site. Pages, badges and messages that described Real Press as an AI-detection service were retired or rewritten. The About page and blog posts had also carried a line that was wrong: “Every word on this page was written by a real, verified human.” It has been retired. Real Press cannot show how a text was composed, including its own pages.
Security
To report a security problem, email hello@real.press with “Security” in the subject line. Say what you found and how to reproduce it.