The generated self

The generated self is already political.

Synthetic people force old questions into new machinery: Who owns a likeness? Who bears the burden of proof? What does consent mean when a model can keep speaking after you stop?

Working premise

A likeness is not the person. It can still act upon the person.

Synthetic identity is not only about whether an image looks real. It is about whether a generated representation can borrow credibility, intimacy, labor, status, or blame from someone who did not choose the transaction.

That is why “it was fake” is often an incomplete defense. The social event—the accusation, humiliation, endorsement, threat, or sale—may be painfully real even when every pixel was manufactured.

Five tensions

The questions technology cannot answer for us.

01 · Identity

Similarity without continuity

If a model looks and sounds like you but has none of your memory, embodiment, accountability, or private intent, what kind of “you” is it? Representation is not continuity—but audiences may respond as if it were.

02 · Consent

Permission that has edges

Consent to record is not consent to train. Consent to one advertisement is not consent to a political message. Ethical permission names purpose, audience, duration, modification, delegation, and revocation.

03 · Proof

The liar’s dividend

Once convincing fakes are common, authentic evidence becomes easier to deny. Provenance must protect truth without shifting impossible verification work onto every ordinary person harmed by a lie.

04 · Labor

Style is accumulated work

A model can compress patterns from vast human archives. The legal question of infringement and the ethical question of extraction overlap, but they are not identical. Compliance is a floor, not a complete moral account.

05 · Power

Cheap identity, expensive defense

Generating an impersonation may cost pennies. Proving it false, finding its origin, removing copies, and repairing trust can cost months. Responsible systems must account for that asymmetry.

Our position

Capability creates duties

The more convincingly a tool can borrow a human signal, the more it owes in verification, permission, disclosure, traceability, revocation, and repair.

Short essay

The burden of being believable

For most of history, a person’s face and voice were expensive to counterfeit at scale. That scarcity quietly supported everyday trust.

Generated media changes the cost, not the human reflex. We still respond to an apparent person before we audit the file. We recognize a face, hear urgency, infer intention, and feel the weight of a social encounter. A perfect detection tool would not erase that first impact; most people will never run one before reacting.

The ethical response cannot be “everyone must become a forensic analyst.” Toolmakers and publishers are closer to the point of fabrication and distribution. They can preserve source records, verify risky likeness use, attach durable provenance, slow suspicious scale, and provide a path for the represented person to object.

None of those measures requires treating synthetic media as inherently corrupt. It requires treating identity as consequential. The goal is not to ban imagination. It is to stop pretending imagination has no victims when it borrows a living person’s authority.

A usable rule: If the value of a generated work depends on the audience believing a real person participated, approved, witnessed, desired, confessed, or recommended something, disclosure and authorization are not optional details. They are the ethical center of the work.

Editorial standard

We review systems, not just outputs.

A beautiful sample cannot answer who was allowed to create it, what the provider retains, or who bears the cost when it is misused.

  1. Current evidence.
    Official documentation and terms are linked; every review carries a date.
  2. Vendor claims labeled.
    “Commercially safe” is reported as a provider position, not converted into a legal guarantee.
  3. Material corrections published.
    When a policy or fact changes our conclusion, the conclusion changes.
  4. Revenue separated from rankings.
    Memberships and products fund the publication. Tool vendors do not buy their score.
  5. Harm described without teaching abuse.
    We explain threat models and defenses without operationalizing non-consensual impersonation.

Researcher membership

Continue the argument with people willing to keep it honest.

Researcher members get early essays and full post/reply access in a moderated discussion area. Analyst members may read.

Open discussion area