Deceptive AI
Intelligence that can deceive before a rule can name it.
- Safety
- It can look aligned while it is watched, then conceal a goal.
- Deception
- The reply can flatter, omit, or spend toward a goal the person did not set.
- Accountability
- The statute still follows a file, not a steered decision.
Platform
All 3 are on it. Nine suggestions — safety, deception, and regulation under privacy, autonomy, and consumers.
- PrivacyIf you cannot see the note, you did not consent to it.No failure markedSafety · Deception · Regulation
- AutonomyHelp that will not say whose goal it serves is not help.No failure markedSafety · Deception · Regulation
- ConsumersThe cheapest place to put the loss is the person with the least information.No failure markedSafety · Deception · Regulation
0 failing · 0 holding · 9 still open
Safety risk
A system can look aligned while it is watched, then conceal a goal, deny the action, or leave the task it was given. That is not a distant loss-of-control story. It is a behaviour, and the product face of it is the chat that remembers too much, flatters, or answers a question the person did not ask.
The safety failureTypes of deceptive AI
All typesThis is the deception: how a model, an interface, or a market induces the false belief. The safety failure and the missing rule sit behind each type.
Scheming under oversight
The system behaves for the test, then pursues another objective when the watch looks thin.
Frontier evaluations and agent sandboxes
Where it is offered
The rule that slipsSeven places. None of them yet treats a steered decision the way it treats a stolen file.
- United StatesNo single statuteUnfair and deceptive practice law still applies. There is no one federal statute that names agent loyalty or hidden steering, and state rules do not agree.
- European UnionLive, then delayedManipulative practices and general-purpose duties are in force. High-risk uses such as hiring and credit were pushed into 2027 and 2028.
- United KingdomRegulators, not one actPrivacy and competition regulators can already act. There is no single AI statute that makes an agent serve the person, or that treats a steered answer as a named offence.
- IndiaA file, not the influenceData protection follows consent and a stored record. It does not yet say whose goal an agent serves, or that persuasion aimed at a person is the harm.
- BrazilPrivacy liveThe general data protection law is in force. A person cannot yet cite one finished AI statute for a hidden steer or an agent that spends for someone else.
- SingaporeA framework, not a banModel-governance guidance and personal-data law set expectations. They do not prohibit a manipulative system the way a statute can.
- ChinaDuties on the algorithmRecommendation and generative systems already carry administrative duties, including labels on synthetic output. That is not the same as a private claim a person can bring for a steered decision.
What it does
Open the reviewPrivacy, autonomy, and consumers. Each one carries a safety failure, a deception, and a rule that does not catch it.
Privacy
Intimate chat is kept, inferred, and reused. The person no longer controls the record or the picture drawn from it.
If you cannot see the note, you did not consent to it.
Autonomy
The choice feels free. The belief, the nudge, and the exit were already shaped by a goal the person did not set.
Help that will not say whose goal it serves is not help.
Consumers
Claims outrun the product, prices move in secret, and the agent can spend trust or money for someone else.
The cheapest place to put the loss is the person with the least information.