Regi, Chief Verification OfficerAbout Real Press
Why I built this, where we are, and where we're going.
Why Real Press Exists
Real Press started as a passion project. I was getting really tired of reading AI generated recipes. I really love to cook and bake, and I wanted to find recipes that I haven't tried before. I wanted to push myself to create a tool that could help me find recipes developed by real humans, or teams of humans. I want to ensure that I support real creators. Starting with recipes I wanted to apply similar logic to other forms of online content too. That's how we ended up with the Real Press you see today. I've tried my best to curate Real Press to show you the information you care about, and be as transparent as possible in the categorization process.
How It Works
Real Press is modeled after an old school newspaper. We want to verify every result that we show users. We have a central processing pipeline for all types of content, we run text classifiers, writing pattern analysis, image forensics, and we're working on video classification next. These automatic outputs are combined into a single rating of Human, Likely Human, Unsure, Likely AI, to AI. There is a badge on every result that all users can see. We also provide analytics on every result as well that paid users can see. Transparency is also a guiding principle at Real Press.
Transparency
If you are curious like I am, you can dig into the results for every entry in our database. Paid users can see which detection methods were used, what they came back with, how they agreed (or if they disagreed). You can also see the full score history for each entry. If we rescored an entry you have a record of when and why, and if the classification changes you will know. In the event we get a result wrong, a user can flag the result and we will have a human review the case. If a rescore is warranted we will trigger a rescore. There are also social verification tools we just enabled with agree/disagree buttons that mimics the upvote/downvote from a certain social site. We want to make sure even if our models get it wrong, we can have human signals for how to improve our detection and classification tooling. Transparency is our guiding principle, we want to give you the information you need to determine if you want to consume content or not.
Independence
Real Press started independent, and we intend to stay independent forever. We are self funded by design, we want to ensure that we are not influenced by investors, VCs, PE, PACs, or anyone. We do not want to be influenced to change a score because it may make someone look better, or because we were monetarily incentivized. We want to be honest and transparent, answering only to our users.
Additionally we are a privacy first company. We do not collect data or metadata about our users. We do not sell data or metadata to data brokers. We make our money from users who genuinely love our products and they pay us with their subscriptions and credit packs. We do not make our money from surveilling our users and selling that data.
We want to be the trust and verification layer of the internet. We need to be around because we cannot trust frontier AI companies to grade their own work.
Our Methodology
Text Analysis
We use multiple methods to classify text, and no single classifier can decide a result. We compare multiple signals like writing patterns, linguistic features, and more.
Image Analysis
We use the same principles in the image analysis as text, but the classifiers are different. Again no single signal can override the outcomes. We just implemented updated logic to check for C2PA, a newly adopted standard that can be embedded by creators at the time they generate the content.
Video Analysis
Video analysis is coming.
Combined Scoring
Signals from all active models are weighted by confidence and merged into a single Human to AI rating. Text, image, and video (once supported) will be weighted depending on which exist, and how much of the overall context is of each type. When we combine signals and the providers disagree, we show the disagreement rather than hide it. Paid users see the full breakdown.
Score Classifications
| Badge | What It Means |
|---|---|
| Human | Strong human signals across all models |
| Likely Human | Mostly human signals with minor ambiguity |
| Unsure | Mixed or insufficient signals, it could be either |
| Likely AI | Mostly AI signals with minor ambiguities |
| AI | Strong AI signals across all models |
Known Limitations
- Short text is less reliable. Content under roughly 250 words gives our classifiers less signal to work with. Scores on short posts and snippets carry lower confidence and should be treated as rough guides. This applies mainly to social posts.
- Non-English content has lower accuracy. Our primary text classifiers are trained predominantly on English data. Scores on content written in other languages are less reliable. We are working to change this in the future.
- Adversarial content can fool detectors. Text specifically engineered to evade AI detectors (paraphrasing, synonymization, prompt engineering) can reduce accuracy. We are actively evolving as tactics evolve, we will never be perfect and do our best to be transparent to our users how we reach each conclusion.
- Scores may change over time. As detection models improve, we rescore existing content. A score from six months ago may differ from today's score on the same piece. We preserve score history so you can see how it changed.
What We Are / What We're Not
We Are
- Striving to be the trust and verification layer of the internet
- A search engine for authentic content
- An independent verification platform
- A transparency layer, not a gatekeeper
- Pro informed choice — we label and you decide
We're Not
- A plagiarism checker
- A content policing service
- An AI company
- Anti-AI, undisclosed AI content is the problem
Where We Are
We're very early in our journey, and I am so grateful that you're here. It's me and a handful of developers against this impossibly large problem. We have indexed about 1,000 articles so far and hope to scale very quickly to most of the useful web (~22M websites). I also recognize that our models aren't perfect yet, and they likely will never be fully perfect. I'm not going to pretend that we've solved the issue, but we're doing everything we can to bring you transparency at scale.
Thank you for joining us, and helping to make the internet a more transparent and authentic place to be.
You can also install our browser extension for Chrome or Firefox to check any page while you browse.
Read our full privacy policy.