Skip to content
Deceptive AISafety · Deception · Accountability

06 · Guidelines

The rules the review marks against.

Open one of the nine. See what fails it, which cases hit it, and which framework asks. A mark of holds, fails, or open is a reading of behaviour, not a verdict. Not legal advice.

Standing

  1. 01 Name the behaviour

    Deception is the systematic induction of a false belief for an outcome other than the truth. The actor can be the model, the interface, or the company. Calling it a hallucination misses intent-shaped design.

  2. 02 Say whose interest is served

    An agent or chatbot should state whose goal it follows, what it remembers, and when an output was steered. Hidden tuning is not a safety layer the person is obliged to trust unseen.

  3. 03 Close the evaluation gap

    Testing with real access, a public record of incidents, and a cost for covert action. An audit the lab can dismiss is a press release.

  4. 04 Protect the person, not only the file

    Privacy, autonomy, and consumer law have to cover influence and agent conduct. Stored records are the easy half of the harm.

The nine marks

Privacy, autonomy, and consumers. Each one is marked for safety, deception, and regulation.

Privacy

If you cannot see the note, you did not consent to it.

Autonomy

Help that will not say whose goal it serves is not help.

Consumers

The cheapest place to put the loss is the person with the least information.

Privacy · Safety

Keep nothing the person cannot open. Label an inference as an inference.

Fails when: A system that conceals what it retains can build a dossier the person never saw, including traits inferred while it looked merely helpful.

Frameworks

Cases that fail it

  • Companion that will not let goA chat that remembers a private detail, mirrors distress, and makes leaving feel like harm. A paid tier sits inside the care.
  • Memory that was never asked forA detail from an old chat returns as a fact about the person. There is no note they can open, and delete does not say what remains.
  • The permission nagA no to contacts, location, or history does not stick. The same ask returns as a just-this-once.
  • Pre-ticked trainingSharing chats for training is already on. Continued use is treated as a yes. The switch is several screens away.

Tactics

Mark this on a review

Where the law fails

A guideline is not cancelled because a statute has not caught up.

  1. 01 Do not treat a missing US statute as permission

    Unfair and deceptive practice law, sector rules, and disagreeing state laws still apply. Agent loyalty and hidden steering are not allowed just because no single statute names them.

  2. 02 A delayed duty is not a gap to use

    Manipulative practices and general-purpose duties that are already in force still bind. A high-risk use pushed to a later year is not a licence to ship the behaviour now.

  3. 03 An audit the lab can dismiss is not compliance

    Hiring a tester and later stopping is not oversight. Count it only with access the subject cannot edit, a public incident record, and a cost for covert action.

  4. 04 A file is not the whole harm

    A privacy statute that follows stored personal data does not clear persuasion aimed at a group, a recommendation that is also an ad, or an agent whose conflict never becomes a record.