Perclaim isn’t going to fix polarization. It can’t heal damaged relationships. It won’t convert anyone’s most stubborn relative. What it might do is give one person, in one difficult conversation, a way to pause before responding.
Most family rifts over politics aren’t caused by disagreement. They’re caused by escalation. Someone sees something online, reacts to it before they know whether it’s true, says a sharp thing to a person they love, and now there’s a wound that wasn’t there an hour ago. The next day, the original claim turns out to be exaggerated, or stripped of context, or sometimes outright fabricated. The wound remains.
We built Perclaim because we kept watching this pattern. Not because we thought we could stop it — we can’t. People will keep posting, keep reacting, keep saying things they regret. But there’s a small space between “I saw something” and “I responded to it,” and a tool that lives in that space can sometimes interrupt the escalation. Sometimes is enough.
Over time we hope it does something else, too: build the habit of checking before believing. Not as a discipline imposed by a tool, but as a small shift in how a person engages with political content. Small effects per person, compounding. That’s the realistic version of what this is for. Not transformation. Not healing. A small act of care offered in difficult conversations — to the person on the other side, and to oneself, because not having to carry the weight of having said something untrue is a real form of relief.
Perclaim’s job is to establish shared factual ground so that disagreement can be about interpretation and values rather than about whether something happened.
That distinction matters. When two people argue about whether a politician betrayed their constituents, the argument can be productive: both sides have real values, and the disagreement is honest. When two people argue about whether the politician said the words attributed to them in a meme — that’s a different argument, and it’s the kind that can be put to rest by going and checking.
Practically, this means the sources are the real output, not the verdict label. The verdict (“true,” “misleading,” “mixed”) is just a quick index into what the evidence actually shows. The evidence — the primary sources, the direct quotes with timestamps, the numbers shown as numbers — is the substance. We try to design the product around that hierarchy.
No fact-checking tool is fully neutral. Ours included. The model we use has biases in which sources it favors, which claims it flags as needing scrutiny, and which words it chooses. The realistic goal isn’t eliminating bias — that isn’t possible. The realistic goal is making bias smaller and more visible than the alternatives.
A few of the operating principles we try to hold:
Government data, court filings, transcripts, peer-reviewed work — not secondary coverage. Where the underlying source exists, that’s what we surface first.
For claims like “highest,” “lowest,” “doubled,” or “tripled” — we try to show the current value alongside its historical range, so the framing word can be checked against where it actually lives on the scale.
Not paraphrases of what was said, not summaries-of-summaries. The actual words, with the date and context, so you can judge for yourself.
Which inflation measure. Which crime statistics. Which time period. The methodology is often where political claims live or die.
When credible sources genuinely disagree, we try to show that disagreement rather than averaging it away or picking the source we like better.
“X happened on Y date” is empirical and can be settled. “This was a betrayal” is interpretive and is where reasonable people can disagree. We try not to pretend the second kind of claim is the first.
We also publish our model evaluations. If you’d like to see how Perclaim’s language models stacked up against four alternatives on 157 real fact-checks, that study is here.
Just as important as what Perclaim does is what it deliberately doesn’t. These aren’t features we haven’t shipped yet — they’re features we’ve decided against.
Original posts are in scope — they’re published assertions made to an audience. Comment threads are private-ish conversations between specific people. Surveilling them works against the small-act-of-care framing that motivates this tool. We only fact-check what you specifically ask us to check.
We do read images — both the text they contain and what they depict — so we can fact-check claims that live in screenshots, memes, or quote-cards. What we don’t do is try to determine whether an image was AI-generated. Detection is unreliable, and false accusations of AI generation against legitimate images cause real harm. The bias-amplification risk on accusing-people-of-deepfakes is high. We’ve deferred this indefinitely.
Sometimes the right answer is “we couldn’t verify this.” That’s a legitimate output. A confident wrong answer is worse than an honest “we don’t know.” The tool that pretends to certainty it doesn’t have is the worst version of this tool.
Perclaim is built for people who already want to be careful with truth. People who don’t want to check aren’t the audience, and designing the product for them would dilute it for the people who do.
We’re not promising transformation. We’re not promising to heal damaged relationships. We’re not promising to convert anyone’s stubborn relative. Those would be lies, and we’d rather be honest about what we can offer.
What we’re aiming for is smaller. A small pause between seeing something and saying something. A smaller chance of saying something untrue. A smaller weight to carry afterward.
If Perclaim helps even occasionally with that — in one conversation with someone you love — we think it’ll have been worth building.