A principle that's easy to state

There's growing agreement around a deceptively simple norm: people should not be misled about whether they're interacting with, or looking at, the output of an AI system. A person talking to what they think is a human agent should know if it's a bot. A person viewing what looks like a real photograph should be able to find out if it's synthetic. Stated that baldly, it's hard to object to.

The difficulty, as always, is in the application. When exactly does disclosure become obligatory? How prominent must it be? And what about the enormous grey zone of AI-assisted work that isn't fully synthetic?

Why disclosure matters

The case for transparency rests on a few solid foundations. Autonomy: people make different choices when they know a counterpart is a machine — they may trust differently, share differently, or want a human instead. Deception harm: synthetic media presented as real can deceive at scale, from fabricated statements to manipulated images. Accountability: knowing AI was involved is the first step to questioning or challenging an outcome. None of these are abstract; they map onto concrete harms people experience.

The easy cases and the hard cases

Some cases are genuinely easy. A customer-service chatbot impersonating a named human employee, with no indication it's automated, is clearly over the line. A photorealistic image of a real person saying something they never said, presented as genuine, is clearly over the line. Bright lines exist.

The hard cases cluster in the middle:

• AI-assisted creative work. A photographer who used AI to remove a distraction; a writer who used a model to tighten a paragraph. Is that 'AI content' requiring disclosure? Most people's intuition says no — but where's the threshold?
• Obvious-by-context AI. Does a clearly-labelled AI art tool need every output watermarked when everyone knows what it is?
• Degree of prominence. A disclosure buried in terms of service technically discloses but practically deceives. How visible is visible enough?

These aren't gotchas — they're the actual operating questions, and reasonable frameworks land in slightly different places on each.

The role of provenance technology

Part of the answer is technical. There's active work on content provenance — cryptographic or metadata-based ways to attach a verifiable history to a piece of media, recording how it was created and edited. Done well, this lets a platform or viewer check origin without relying on a visible watermark that can be cropped out.

But provenance technology has limits. Metadata can be stripped. Standards only help if they're widely adopted across the tools and platforms in the chain. And provenance answers 'how was this made' better than it answers 'should this have been labelled to this audience in this context.' It's a powerful ingredient, not the whole recipe.

A practical stance for builders

If you're building or deploying systems that interact with people or generate content, a defensible default is straightforward:

• Don't impersonate humans. If your system converses, make its nature discoverable without the user having to dig.
• Label synthetic media that could plausibly be mistaken for a real depiction of real people or events.
• Attach provenance where you can, even if it's imperfect — it raises the cost of deception.
• Scale disclosure to the risk. Heavy disclosure where the potential for deception or harm is high; lighter touch where the AI involvement is obvious or inconsequential.

The direction of travel

Expectations here are tightening, both in norms and in law. The organisations that will be comfortable are those that treat honest disclosure as a default rather than something to be minimised. The simplest test survives all the hard cases: would a reasonable person feel deceived if they later learned how this was made or who they were talking to? If yes, disclose. If genuinely no, you're probably fine. Most of the difficulty dissolves when you ask the question from the audience's point of view rather than your own.