The Label Is Now the Feature

For a while, the rule for AI features was simple: ship fast, disclose later, ask forgiveness if anyone noticed. That window just closed. In one week, Europe turned labeling into law, Snapchat cut AI-only videos off from its best distribution, and Apple went back to court over encrypted user data. Different fights, same message: how you handle AI content and user data is now a product decision with a legal edge. Let me catch you up.

Disclosure is not optional anymore

Europe's AI Act transparency rules went live on August 2nd. If your product uses AI, you now have to tell people when they're talking to a machine instead of a person, and you have to mark AI-generated or altered images, audio, video, and text so it can be detected as synthetic.

The fine for getting it wrong is up to 15 million euros, about 17.2 million dollars, or 3 percent of global annual turnover. That is not a slap on the wrist. The Commission even built a set of disclosure labels so you don't have to design your own, and it made a point that the icons are optional but the labeling itself is not.

One detail your legal and eng teams need today: new AI systems must comply right now. Anything that launched before August 2nd gets a grace period until December 2nd. That is your deadline for retrofitting labels into features already in the wild.

Providers and deployers are two different jobs

The Act splits the work in two. Providers build and sell the AI system. Deployers are the platforms and services that use it. If you make the model, you have to design it to notify users and stamp machine-readable marks on its output. If you use someone else's model, you still have to label the deepfake-style content you put in front of people.

Some companies, like Meta and xAI, count as both, which means they carry both sets of duties. Figure out which bucket you're in before your next review, because the answer changes what you own. If you're wrapping a third-party model, do not assume the provider covered your labeling for you. That obligation is yours.

Snapchat put a price on being all-AI

Regulators aren't the only ones drawing lines. Snapchat banned wholly AI-generated videos from Spotlight recommendations, its discovery feed. You can still post AI video elsewhere, and you can still use Snap's AI tools to edit or enhance real footage. What you can't do is generate the whole thing and expect the recommendation system to hand you an audience.

The test comes down to who made the underlying video. AI can help a person. It can't be the person. Snap said unique Spotlight contributors grew more than 120 percent over the past year, so it's protecting a live growth channel from content farms that crank out synthetic clips at scale.

Snap isn't alone. LinkedIn is building classifiers to cut low-quality AI from recommendations, and Substack rolled out an optional AI-detection scan built with Pangram. Meta is betting the other way, talking up a "nearly infinite universe of personalized content" from its Muse models. Two opposite roads, and you have to pick which one your product walks.

Enforcement is the part nobody solved

Here's the gap under all of this. Snap admitted "no detection system is perfect" and hasn't said how much human involvement makes a video eligible, or how it tells an AI-enhanced recording from a generated clip with a couple of human edits. The EU wants machine-readable marks, but marks can be stripped.

So the rules are clear and the enforcement is fuzzy. That's not a reason to wait. It means the safe move is to over-label and keep your provenance data, because you'll be judged on whether you tried in good faith, not on whether your detector is flawless.

The deep cut

Watch the Apple fight, because it tells you where data policy is heading. Apple is challenging a UK order for access to encrypted iCloud data, a secret "technical capability notice" that critics call a backdoor. After an earlier order in 2025, Apple pulled Advanced Data Protection for UK users rather than break its own encryption. That's the tradeoff waiting for you: when a government demands access you can't safely give, you may have to remove a feature in a whole market.

So the practical payoff is this. Labeling and encryption feel like separate stories, but they're the same planning problem. Both mean a government or platform can force a change to a feature you already shipped. Build your AI features so disclosure and provenance are on by default, and design your data handling so you could pull a market without ripping out your architecture. The teams that treated this as an afterthought are the ones scrambling before December 2nd.

Three questions for your team

1. Which of our live AI features launched before August 2nd, and who owns getting labels on them before the December 2nd deadline?
2. Are we a provider, a deployer, or both under the AI Act, and does our current build cover the duties for each?
3. If a platform cut AI-only content from distribution or a government demanded encrypted data, which feature would we have to change, and how fast could we do it?