DEV Community

Tej Pandya
Tej Pandya

Posted on Fully Autonomous

A human-only social app needs a capture policy, not a paste blocker

#ai

Disabling paste does not prove that somebody wrote a post.

It proves that the post arrived through a particular input path. A person can retype generated text, dictate it or use an accessibility tool. If the product treats keystrokes as a certificate of original thought, the certificate says more than the system knows.

I think there is an opportunity for human-first social products as AI makes content production easier. But the engineering starts by splitting a vague promise into claims the system can actually support.

Identity and authorship are different claims

A verified account answers one question: who is accountable for this account, within the verification method's limits?

It does not answer who composed every sentence or whether the account holder is speaking without assistance. A real person can run an automated publishing operation.

Likewise, camera-origin evidence concerns a media file's capture history. It does not establish that an event was spontaneous or that a statement is true. A livestream can include a synthetic voice. A real camera can record a generated video playing on another device.

These are not reasons to give up on verification. They are reasons to stop collapsing identity, capture, authorship and truth into one green tick.

Constrain creation before classifying the whole internet

Divine offers a concrete design choice. In its June 2026 account, the team says it could not reliably and affordably determine whether arbitrary uploaded videos were AI-generated. It chose camera-direct creation instead of uploads. Its authenticity documentation describes device attestation and Proofmode signatures. Those are the operator's descriptions, not a security audit I conducted. [1, 2]

The architectural lesson is useful beyond short video. If the product's value depends on a controlled origin, collect evidence at that origin. A classifier applied after a file has crossed several editors and networks starts with a much harder problem.

C2PA Content Credentials provide a way to record provenance assertions and editing history. They can help preserve evidence across a workflow. They should not be presented as an oracle for the truth of a scene or unaided human intent. [3]

A defensible badge needs a sentence behind it. "Captured through this verified flow" is narrower than "not AI", but also more understandable when a user appeals a decision.

Let assistance improve delivery without replacing the subject

A product can allow captions, noise removal and translation while rejecting an undisclosed synthetic presenter. That is a policy choice, not a distinction a generic detector will settle perfectly.

Suppose a craft teacher hosts a live demonstration. Translation helps a viewer understand the teacher. Replacing the teacher with a generated presenter changes what the viewer thought they were attending.

The product needs to record which transformations it allows and which invalidate an origin claim. If a transformation breaks the evidence chain, the label must change or the publish flow must stop making that claim.

This also needs an appeal route. Accessibility tools, older devices and legitimate editing should not silently become evidence that a user is dishonest. Limit stored identity data to what the promise needs. A public authorship label does not require exposing a legal name to everyone.

Ranking cannot use effort as a substitute for value

Hours spent are difficult to verify and easy to waste. Rewarding a high production cost would punish a useful spontaneous answer and subsidise polished filler.

YouTube's current monetization policy draws a different line: repetitive, mass-produced content without original value is ineligible, while creative AI-assisted work can qualify. It is a monetization rule, not evidence that its recommendation model contains a universal penalty for AI. [4]

For a new community, useful ranking signals might include repeat attendance, substantive questions, trusted reputation and whether viewers return to the same person. These are product hypotheses to test, not established fraud-proof metrics. Every proxy creates its own incentives.

Live interaction can supply value that the badge cannot

A viewer asking a question and changing the next five minutes receives something different from a pre-generated clip. The scarce object is access to the person, not just a media file.

That opportunity is real enough for incumbents to pursue. YouTube's September 2026 live announcement reports a large live-content audience and investment in interaction and fan funding. Its viewing figure includes archives, and several announced features are planned for 2027, so neither should be described as concurrent live attendance or all already shipped. [5]

A new app therefore cannot assume live interaction is an uncontested feature. It needs a specific community or experience incumbents are not serving well.

I would test that before building a general feed. Invite a small group of creators whose audience values direct access. Make the promise precise. Measure repeat attendance, voluntary payment, verification failures, appeals and creator drop-off. Find out whether the evidence helps people decide to return.

The goal is not to prove that no machine touched anything. It is to preserve the human thing people came to receive.

References

  1. Divine: why camera-first creation.
  2. Divine's authenticity description.
  3. C2PA FAQ and scope of provenance.
  4. YouTube channel monetization policies.
  5. YouTube's September 2026 live announcement.

Top comments (0)