Two Communities Spent the Week Arguing About an AI Label Neither of Them Read

On Monday night a developer posted a Safari extension to Hacker News that hides YouTube videos carrying the platform's Made with AI label. It took 158 points and 71 comments. In the same week, a Lobsters thread was picking apart zuzai, a proposed word meaning free of modern AI, and had gone 31 points and 34 comments deep into an argument about Norwegian adjectives.

Two communities, two artifacts, one question: how do you mark the absence of machine generated content so people can filter on it. Both threads are worth reading. Neither of them, with one exception we will get to, checked what the label they were arguing about actually certifies. So we read YouTube's policy and the extension's 377 lines of source, and the answer changes the argument.

What the extension actually does

Weedout is a Safari extension, $1.99 on the Mac App Store, with the source published at masteranza/weedout-for-youtube. We cloned it. The whole filter is one content script of 377 lines, plus a 59 line popup, running as a Manifest V3 extension with exactly two permissions: storage, and host access to youtube.com. The privacy claim on the landing page checks out at the manifest level, which is more than most extensions can say.

The detection is a single regular expression tested against YouTube's own badge data:

const AI_LABEL_RE = /(^|\s)(AI|IA|KI|\u0418\u0418)(:|\s|$)|made with AI|synthetic|altered|generat|.../i;
// tested against metadataBadgeRenderer.label and accessibilityData.label

Fourteen language forms, matched against the label YouTube itself attaches. There is no classifier, no heuristic, no blocklist. The author is upfront about this in the README, under a heading reading "Honest limit": Weedout can only filter videos YouTube labels, and unlabeled AI content still gets through.

So the extension is exactly as good as YouTube's label. Which sends the question one level up.

What the label certifies

YouTube's disclosure rules require a creator to declare altered or synthetic content in three cases: it makes a real person appear to say or do something they did not, it alters footage of a real event or place, or it generates a realistic scene that did not occur.

The exemption list is the interesting half. No disclosure is needed for AI assisted scripts, outlines or thumbnails, for captions, for beauty and colour filters, for blur and vintage effects, for audio repair and upscaling, for cloning your own voice for your own voiceover, or for anything clearly unreal such as animation and fantasy scenes.

The badge is a deepfake disclosure. It was never a quality signal, and filtering on it removes the videos that told you the truth.

Read those two lists together and the label describes realism, not effort. A fully machine written, machine narrated, machine thumbnailed twelve minute video about a topic the uploader has never thought about carries no obligation to be labeled, because nothing in it depicts a real person or place deceptively. A careful documentary that used AI to reconstruct one historical street scene has to be labeled and gets filtered.

The takes, and where they land

The Hacker News thread found the edge of this without quite going over it. Schnitz put it plainly: "I've found that the AI label doesn't necessarily mean AI slop, it just means there's a non zero amount of AI in the video. I want to avoid AI slop but not all videos that have an AI title screen or whatever."

tadfisher replied with the incentive problem: "Now that the AI label exists, the creators using AI for an intro graphic or background music are probably going to stop using it." Which, per the policy, would be a rational response to a rule they were never subject to.

singpolyma3 aimed at the gap the label leaves: "I'm more concerned about the stuff that is fraudulent then possibly legit stuff explicitly labelled as AI." The extension's author, masteranza, answered honestly rather than defensively: "True, but that's a way harder problem to solve. For me what was important was reducing the noise as much as possible as fast as possible consistently."

thatmf wanted the tool everywhere: "I wish I had this but for every website and media surface." embedding-shape supplied the joke that answers it: "Ironically, if you had this for every website, including GitHub and HN, you wouldn't have seen this HN submission / repository."

Over on Lobsters, the same instinct was aimed at vocabulary instead of software. nemin, at 18 points, made the marketing case for a single word: "Would the market for organic food be the behemoth that it is if the choice of label had been something like 'synthetic pesticide-free'?" tinsmith, at 3, made the case against this particular one: "Neologisms only work when they are comprised of familiar word parts... Zuzai gives us an unfamiliar prefix portmanteau'd with an acronym." travisgriggs proposed stealing Norwegian "ekte", meaning real, and half the thread discovered that German, Hungarian and English already have it. The two highest scoring comments in the whole thread, tied at 35, were scraps posting the single word "clankless" and an_origamian posting "No thank you."

What the two platforms disagreed about

Same week, same underlying question, and the communities barely overlapped in what they found worth arguing about.

Hacker News argued economics and distribution. Why Safari and not Chromium. Whether people really pay for browser extensions, with 0x1d7 pointing out the $100 a year Apple signing fee that explains the $1.99. Whether the fix is a tool at all, with mikeweiss arguing that "the EU, or California or some other large jurisdiction" should require labels by law.

Lobsters argued morphology. Whether a recursive acronym can catch on, whether "zuzai" reads as a prescription drug, whether it collides with the Polish name Zuzia or a Czech vulgarity. kevinc summarised the structural objection: "it's not built from common parts like 'non-', so you must expand the acronym to understand it."

The gap is the story. One community treats this as a filtering problem to be solved in code, the other as a naming problem to be solved in language, and both are downstream of a definition nobody has agreed on. zem, three points and buried, put a finger on the part both threads skipped: the distinction between AI in the production pipeline and AI in the output. "I really don't care if someone's video game engine or blogging software uses AI code, but I do care if the game assets or blog posts are AI generated."

The comment nobody built on

Two levels down a subthread, Muskwalker did the thing the rest of both threads did not, and quoted the policy directly, listing the three disclosure triggers verbatim and noting the oddity that follows: "So oddly an intro graphic only has to be tagged as AI if it shows a real/realistic person, place, or event."

The author's reply was three words plus an ellipsis: "If so, that's indeed a real concern..." And that was the end of it. The one comment that checked the primary source got a nod and no thread.

Our read

Weedout is a good piece of work aimed at the only machine readable signal that exists, and its author says so in the README more clearly than the threads discussing it did. We would install it. We would also not expect our feed to get much better, because the videos it removes are disproportionately the ones made by people who followed a disclosure rule.

The uncomfortable part is that the labeled set and the slop set are close to disjoint. Anyone who wanted to hide low effort machine content this week and reached for the AI badge to do it built a filter for honest deepfakes. That is a real result, it is just not the one anybody wanted.

What would change our mind is YouTube splitting the disclosure into two flags, one for realism and one for generation, and letting viewers filter on the second. Until then the argument stays where both threads left it: people want a badge for effort, and effort is the one thing a platform cannot detect.