The AI Enforcement Glossary

The compliance enforcement model family.

Live oversight of every verdict.

Send any AI output, get a verdict.

Enforcement where AI communicates.

Where your company's rules live.

The runtime's decision on a message: pass, fix, block, or escalate.

A pre-built, maintained set of enforceable rules for one regulation area.

FINRA's rule on communications with the public. Every retail communication from a broker-dealer must be fair, balanced, and not misleading: no promissory language, no exaggerated claims, no performance projections.

Requires broker-dealers to maintain a supervisory system, including written procedures and review of correspondence and internal communications. Supervision covers an AI-drafted message the same as a human one.

Business communications sent over channels a company does not capture or supervise, such as personal text, WhatsApp, or an unapproved AI tool. U.S. regulators have levied more than $2.8 billion in penalties for it since 2021.

The SEC's Regulation Best Interest. When a broker-dealer recommends a security or strategy to a retail customer, the recommendation must be in the customer's best interest, with disclosure, care, and conflict-of-interest obligations attached.

The SEC's recordkeeping rule for broker-dealers. Business communications must be retained in a compliant, tamper-evident format and produced on request. Most off-channel communications fines cite it.

Advisers Act Rule 206(4)-1. Governs how registered investment advisers advertise: no untrue or unsubstantiated statements, strict rules on presenting performance, and conditions on testimonials and endorsements.

Unfair, Deceptive, or Abusive Acts or Practices, enforced by the CFPB under Dodd-Frank. Applies to any consumer-facing financial communication, including an AI chatbot's answer about a fee or a rate.

The HIPAA Privacy Rule limits how protected health information is used for marketing. Most marketing communications need written patient authorization, and paid promotion has to disclose the payment.

Filters and classifiers wrapped around a model's inputs and outputs to catch unsafe or off-policy content. Guardrails detect and flag. They do not read a regulation, and they do not fix the message.

A record of an enforcement decision a regulator can review as-is: the message, the rule cited, the verdict, any fix, and the timestamp. Produced at the moment of enforcement, not reconstructed later.

A proxy between an application and its model providers. It handles routing, keys, logging, rate limits, and pattern-based redaction. A gateway can mask a phone number. It cannot tell whether a sentence violates the Marketing Rule.

Material nonpublic information: facts about a company or security that a reasonable investor would act on and that are not yet public. Sending it outside an information barrier can be insider trading, whether a person or a model wrote the message.

An attack that hides instructions inside a model's input, such as a document or web page, to make the model ignore its rules. An enforcement runtime enforces the output, so a hijacked model's message still has to pass before it ships.

Applying written rules to AI behavior while it happens, in the request path, instead of in a review queue or an archive afterward. This is the market category ZeroDrift operates in.

A language model with a fraction of a frontier model's size, post-trained for one job. It is faster and cheaper per call, and more accurate inside its domain. Anchor is an SLM.

The HIPAA Privacy Rule limits how protected health information is used for marketing. Most marketing communications need written patient authorization, and paid promotion has to disclose the payment.