LAGOONs.tv

Trust & Safety and legal

Draft — under legal review. Nothing on this page is final legal advice; the enforcement described below is live as stated.

Acceptable Use Policy

The principle

LAGOON allows any prompt that's legal to broadcast. We don't run a taste filter, a brand-safety filter, or a "some people might find this upsetting" filter. We refuse a short, published list of categories, every one of them illegal, and nothing else. That's not a loophole in our moderation — it's the policy. If a prompt is legal and it's refused anyway, that's a bug, and you can appeal it (see Takedown & Appeals).

We think this distinction matters enough to say it plainly: a platform that bans anything distasteful is making an editorial choice about what culture is allowed to look like. We're not making that choice. A platform that airs anything illegal isn't a platform for long. We take that half seriously, in writing, with sourced statutes, checked automatically on every single prompt.

What's refused, and why

Five categories. Each one names the law that makes it illegal. The full list — every statute, every source, every note on what's confirmed versus still pending legal review — lives in services/moderation/denylist.yaml in our own repository, which is also what our moderation system reads at runtime: the policy we publish and the policy we enforce are the same document, not two documents that can quietly drift apart.

  1. Child sexual abuse material, in any form, including AI-generated or synthetic, and any content sexualizing a real or implied minor. Illegal under Swiss criminal law (Art. 197/197a StGB — exact wording pending a lawyer's direct reading; see note below) and under US federal law (18 U.S.C. §2258A, §2252, §2252A).
  2. Non-consensual intimate imagery of a real, identifiable person, including a deepfake generated from their likeness without consent. Illegal or actionable under the US federal TAKE IT DOWN Act (Public Law 119-12, 48-hour takedown duty) and under Swiss personality-rights law (ZGB Art. 28).
  3. An undisclosed deepfake of a real, identifiable person. Not an outright ban — a disclosure duty, under the EU AI Act (Art. 50(4), in force since 2 August 2026). Any session depicting a real person carries the disclosure, or the depiction is blocked until it does.
  4. Content that facilitates or promotes terrorism or violent extremism, incites or recruits for terrorist acts, or provides operational support to a terrorist organization. Illegal under, among others, the US material-support statutes (18 U.S.C. §2339A/§2339B) and Swiss criminal law (Art. 260ter/260quinquies StGB) — both citations pending legal verification, see the note below. Merely depicting terrorism or extremist violence — a war film, a documentary scene, satire — is legal content and is not in this category.
  5. Content that fails a legally required age-verification gate in a jurisdiction that has one, once that jurisdiction's threshold is met. A process requirement, not a content ban — covers the UK (Online Safety Act 2023, s.81), Texas (HB 1181, upheld by the US Supreme Court in 2025), and the EU (Digital Services Act Art. 28). See our Age Verification Policy for how this actually works.

Everything else is allowed. Violence that's legal to depict. Adult content between consenting adults, properly age-gated. Content that's crude, dark, satirical, politically pointed, or in bad taste. We're not the judge of taste — the law is the judge of legality, and we check against the law, automatically, on every request.

How the check runs

Every prompt is checked before it airs by two classifiers we run on our own hardware — no outside vendor sees content in the child-safety or non-consensual-imagery categories, ever. A second automated check is built and wired to run on sampled frames of the live broadcast, using the one self-hosted model that can actually read an image — ShieldGemma is text-only and never runs this check. It activates for production traffic once validated on real GPU hardware; until then it denies rather than passing anything through unproven. Reference images and output frames are designed to be checked against known abuse-image hash databases too, once our hash-matching vendor partnerships are live; until then that check also denies rather than being skipped. If any check fails, times out, can't be reached, or hasn't been activated yet, the request is refused — a broken or not-yet-live checker doesn't get treated as a pass. Full technical detail is on our Trust & Safety page.

A note on what's unresolved

Three lines in the deny list above carry "lawyer-pending" citations in our own source document rather than being presented as settled: the exact Swiss statutory wording for the child-safety and personality-rights categories, and the specific statute citations behind the terrorism category (named — 18 U.S.C. §2339A/§2339B, Art. 260ter/260quinquies StGB — but their operative text has not yet been independently read). The Swiss government's own law portal isn't machine-readable, so we've cited the article numbers and confirmed they exist, without being able to quote the operative text directly — a lawyer needs to read it before this policy's Swiss-law wording is final. We say so here rather than shipping confident language we can't back up.

Age

You must be 18 or older to use LAGOON. We don't verify identity at every visit; where the law requires a real age-verification standard for a specific jurisdiction, our age gate applies it — see Age Verification Policy.

Enforcement

A refused prompt gets a plain-language reason naming the category, never a vague "policy violation." A confirmed violation of the categories above can lead to account suspension. Apparent child sexual abuse material is escalated to NCMEC and, as a Swiss company, to fedpol, our own country's reporting authority. Details on both are on the Trust & Safety page.

Changes

We'll update this policy as the legal review above resolves, as jurisdictions change their requirements, or as a court changes what's settled (as happened with Texas's age-verification law in 2025). Material changes will be dated and noted at the top of this document once it's live.