AI Labels Need Controls, Not Just Disclosures
2026-06-06 · 4 min read · Janaina Maia
A label tells me what something is. A control lets me decide what to do with it. AI products are still confusing those two things.
The Verge argued this week that platforms should let people filter AI-generated content, not simply label it. YouTube, TikTok, Instagram, Spotify, Pinterest, and others have all moved toward disclosure systems that identify some AI-generated images, videos, music, or posts. But disclosure does not give users much agency if the feed still decides what they see.
I think this is a much bigger product design lesson than social media moderation. It is a warning for every AI product team building labels, warnings, badges, confidence scores, provenance notes, or policy banners and assuming that counts as trust.
Labels are information. Controls are agency.
Labels can be helpful. If a video was generated, edited, or heavily altered by AI, people should know. If an answer came from an internal document, a model guess, a third-party source, or a synthetic dataset, the product should make that visible.
But a label is still passive. It puts the burden on the user to notice, interpret, and mentally compensate. That is weak design when the system is already making the stronger decision: what appears, what is ranked, what is recommended, what is hidden, and what gets repeated until it feels normal.
Control is different. A control lets a person say, “show me less of this,” “exclude this category,” “use only verified sources,” “do not use synthetic examples,” or “ask before bringing AI-generated material into this workflow.” That turns transparency into actual choice.
The enterprise version is even more important.
It is easy to treat this as a consumer-feed problem, but enterprise AI will face the same pattern. A product may label an output as AI-generated, show a confidence score, cite a source, or add a warning that the answer should be reviewed. Useful, yes. Sufficient, no.
If the user is making a decision in engineering, finance, healthcare, HR, legal, operations, or geoscience, the design question is not only “did we disclose the AI?” The better question is “can the user shape how much AI enters this work at all?”
That might mean letting teams restrict answers to approved knowledge bases, exclude low-quality sources, separate human-authored evidence from generated synthesis, require human review before AI-drafted content becomes part of the record, or turn off generative suggestions in high-risk moments.
Trust is not a sticker.
What worries me is the rise of decorative trust: a badge, a disclosure, a friendly explanation, a confidence meter, and then the same opaque experience underneath. That is not governance. It is labelling the box while keeping the steering wheel hidden.
Good AI design should make three things visible: what the system knows, what it is doing with that knowledge, and what the human can change. Most products are slowly improving the first part. The second and third are where the real work is.
Design implications.
- Pair every label with an action: if the product says something is AI-generated, let users suppress, filter, inspect, or route it differently.
- Make preferences durable: people should not have to repeat the same instruction every session, feed, project, or workflow.
- Separate provenance from permission: knowing where something came from is not the same as agreeing it should be used.
- Design for context: playful AI content in a feed is not the same as AI-generated evidence in a work decision.
- Give admins policy controls: enterprise teams need organisation-level settings, not only individual toggles.
My take.
The next maturity step for AI products is not more labels. It is better user control over when AI appears, what it can influence, and how much authority it has inside the workflow.
Transparency without agency becomes homework for the user. It asks people to read the warning, understand the risk, and adapt around a system they cannot steer.
If a product can identify AI-generated content well enough to label it, it should be honest enough to let people control it.