trust & safety

Resolve builds tools that make probabilistic estimates — about whether media was manipulated, about how a game might go, about what's in the news. This page explains what we do to keep the service safe, what we hold, what we don't, and how to reach us when something is wrong.

What we don't keep

Media you upload for analysis is deleted after it's scored. We keep the result — a score, a label, a version number, a timestamp — and not your file. If you ask us for a copy of your uploads, we can't produce one, because we don't have them.

There are two exceptions, both narrow, and both described below.

What we don't build

We don't identify people. Our video analysis locates and crops faces to examine image properties. It does not compute or store facial templates, faceprints, or voiceprints, and it cannot recognize or match a person. We are not building a face recognition system and we don't intend to.

We don't train on your content. Uploads and prompts are not used to train models: detector uploads are deleted after scoring, DeepDive's analysis runs on Anthropic's commercial API (whose terms do not permit training on our inputs), and its optional Gemini grounded-search cross-check runs on Google's paid tier, which does not train on prompts.

We don't sell your data, and we don't share it for advertising.

Prohibited uploads

Some content may never be submitted, with no exception for research, journalism, curiosity, or testing:

• Child sexual abuse material, in any form, including synthetic depictions
• Non-consensual intimate imagery, including material altered to depict someone sexually without their consent
• Content depicting or facilitating serious violence, terrorism, or trafficking
• Malware or content designed to compromise systems

The full list is in our Acceptable Use Policy.

Content screening

Uploads are screened against hash sets of known child sexual abuse material, using Microsoft PhotoDNA. When a submission matches:

• We do not score it, and you receive an error rather than a result
• We preserve the file and associated records securely, for at least the period federal law requires
• We report the matter to the National Center for Missing and Exploited Children (NCMEC) CyberTipline, as 18 U.S.C. §2258A requires
• We restrict the account

Two honest limits, because a screening claim that overstates itself is worse than no claim:

This screening detects known material only. Hash matching compares against catalogs of previously identified content. It does not detect material that has never been catalogued.

It applies to images and to frames extracted from video. It does not cover audio. There is no equivalent hash service for audio, so our audio pipeline has no comparable control. We rely on the prohibition above and on reports from users.

If our screening service is unavailable, we refuse to score rather than proceeding unscreened.

Non-consensual intimate imagery

We prohibit it, and we act on reports. We don't currently have an automated control for it — no hash service covers this category for a service like ours — so enforcement depends on reports. If you find your own intimate imagery has been submitted, contact us and we will act.

Who we serve — and who we don't

Resolve is a private service, and we choose who uses it. We will never turn you away or remove you because of your sex, gender, sexual orientation, race, ethnicity, religion, disability, age, or any other protected characteristic. That's a line we don't cross.

What we do act on is behavior. If you harass or abuse other users or our team, act in bad faith, or set out to make this a worse place for everyone else, we may remove you — and we won't owe the internet a debate about it. We judge what you do, never who you are.

How enforcement works

Some enforcement is automated. Content screening and abuse detection run without human review. Automated systems make mistakes, and ours will too.

If we restrict your account, we'll tell you it was for a violation of the Acceptable Use Policy. We won't speculate about you or describe what was detected — partly because the law limits what we may say, and partly because we would rather say nothing than say something wrong about someone.

Restricting an account is not deleting records. Anything we're legally required to keep, we keep.

Appeals. Contact us and we'll review it. For automated content screening, our review covers technical and account records only — we do not review the content itself, and in some cases the law prohibits us from doing so. That constraint cuts both ways and we can't waive it.

What our tools can and can't tell you

Our detector estimates a probability. It produces both false positives and false negatives — it will sometimes flag authentic media as manipulated, and sometimes miss real manipulation. Audio performance is materially weaker than video, and both degrade on sources unlike those used in evaluation. In evaluation the audio model achieves roughly 0.76–0.81 AUC and the video model roughly 0.77 AUC. Full methodology is available under NDA.

We do not certify, authenticate, or verify anything. If someone tells you Resolve confirmed a file is fake, they are misdescribing what we do. The correct description is that our detector returned a probability.

Our analysis tools generate text using large language models. That text is unverified, sometimes wrong, and never a factual record — particularly where it concerns real people.

Corrections

When we learn that something we produced contained a material factual error, we record the correction and fix the underlying behavior where the error is systemic. Report an error and we'll tell you what we did about it.

Copyright

We respond to notices under the Digital Millennium Copyright Act and terminate repeat infringers in appropriate circumstances.

Designated Agent Resolve Research LLC Copyright Office registration: DMCA-1076294 (Active; registered 2026-07-25; renew by July 2029) Nathan Humphrey Annapolis, MD 21403 nathan@resolveresearcher.com (443) 510-9101

A notice needs the elements listed in 17 U.S.C. §512(c)(3). If your content was removed and you believe that was wrong, you may send a counter-notice to the same address.

Security

Passwords are stored using bcrypt with a random per-password salt, over a SHA-256 pre-hash so that long passphrases contribute in full. Traffic is encrypted in transit. Production access is limited to the owner.

Found a vulnerability? Report it to nathan@resolveresearcher.com and we'll respond. We won't pursue anyone acting in good faith who reports a genuine issue and gives us a reasonable chance to fix it before disclosing.

Reporting something

We read everything sent to these. You'll get a human, because there's only one of us.

Related documents

Terms of Service · Privacy Policy · Acceptable Use Policy · Subscription & Billing Terms · Contributor Terms · Cookie Notice