OCC AI glossary: every lexicon term in plain English

Published Nov 21, 2026 · Last updated Nov 21, 2026

Appendix A of the OCC's AI supervisory framework (Bulletin 2026-54) defines 51 terms, from adversarial AI to version control. 47 are starred, meaning the definitions come from the AIEOG AI Lexicon with "AI" changed to "SI". This page explains each one in plain English, grouped by theme.

TRACKED DOCUMENTS: OCC Bulletin 2026-54

What is the lexicon in the OCC's AI framework?

Appendix A of the OCC's supervisory framework is a lexicon: a list of 51 defined terms that examiners and banks can use as a shared vocabulary for AI. This page explains every one of them in plain English.

The framework was issued with OCC Bulletin 2026-54 on November 4, 2026. It is a resource for examiners. The bulletin says it “does not establish supervisory expectations for banks”. The OCC calls it “Super Intelligence (SI)” where everyone else says AI, so this page writes AI in its own sentences and keeps SI only in the names of terms.

  • How many terms: 51, by our count of Appendix A.
  • What the star means: 47 terms carry an asterisk. The appendix says definitions marked with an asterisk come from the AI Lexicon workstream of the Artificial Intelligence Executive Oversight Group (AIEOG), with “artificial intelligence” and “AI” in those definitions changed to “Super Intelligence” and “SI”.
  • The four unstarred terms: Human-on-the-Loop (HOTL), Output validation, SI model/system exploitation and Super Intelligence (SI). The appendix gives no source for them.
  • Not final: the OCC says the lexicon “should be treated as a living Lexicon” that is expected to change as technology, use cases and supervisory understanding mature.

The tables below group the terms by theme. The grouping is ours. The appendix lists them alphabetically. A star (*) after a term marks the AIEOG terms. Every “For example” is our own illustration, not the OCC's. For how the terms are used across the framework, see the section-by-section reference guide.

What does the OCC mean by Super Intelligence?

In the OCC's usage, Super Intelligence means AI: the same technology banks already call AI, under a new name. It does not mean a new, more powerful kind of machine.

The name comes from footnote 2 of the framework, repeated in the bulletin. It says “Super Intelligence” refers to “the technologies and systems previously encompassed by the term ‘Artificial Intelligence’ as defined in 15 USC 9401(3) as well as the promise, potential, and rapidly advancing capabilities of those technologies”. It cites Executive Order 14434, 91 Fed. Reg. 63129 (Oct. 2, 2026). Footnote 3 adds that, consistent with that order, the OCC supports the reasonable adoption of SI. Put simply, the OCC uses a new name for the same technology.

The lexicon's own definition (unstarred) describes SI as “a machine-based system that, for a given set of human-defined objectives, can make predictions, recommendations, or decisions that influence real or virtual environments”. It goes on to say these systems take in machine and human inputs, turn what they perceive into models through automated analysis, and use those models to produce options for information or action.

In practice, a fraud-scoring model, a chatbot and a document-summary tool are all SI in the OCC's sense. Footnote 1 adds that the model risk principles in Bulletin 2026-13 apply to non-generative, non-agentic models, so the framework is more applicable to generative and agentic AI. The plain-English overview explains how the documents fit together.

What are the basic AI terms in the OCC lexicon?

These eleven terms name what AI is and how the common kinds work, from machine learning to chatbots that answer from the bank's own documents.

Term (as the OCC writes it)Plain meaning
Super Intelligence (SI)The OCC's name for AI. A machine-based system that works toward goals people set and produces predictions, recommendations or decisions that affect the real or digital world. See the next section.
SI system*Any software, application or tool that runs wholly or partly on machine learning or other AI. It excludes everyday commercial products with AI built in, and the OCC names a word processor and map navigation as examples.
SI model*The part of a system that does the AI work: it takes inputs and uses computational, statistical or machine-learning methods to produce outputs.
Machine learning*A way for computers to learn patterns from large amounts of data instead of being programmed step by step for each task. It includes supervised, unsupervised and reinforcement learning. For example, a fraud model that learns from past fraud cases.
Deterministic (algorithm/model)*A model that always gives the same output for the same input. For example, a fixed rule that flags every wire over a set amount. In practice, generative AI is often not like this.
Generative SI*AI that creates new content (text, images, audio, video) modeled on the data it learned from. For example, a tool that drafts a reply to a customer complaint.
Foundation models*Very large machine-learning models trained on huge amounts of unlabeled data, which can then be adapted to many different tasks. Large language models are a common kind.
Large language model*A kind of machine learning trained on large amounts of data, through self-supervised learning, to recognize patterns and answer requests in ordinary language. The lexicon says these models underpin many generative AI applications in finance.
Prompt*The plain-language text that tells an AI system what task to do. For example, “Summarize this loan file in five bullet points.”
Retrieval augmented generation (RAG)*A setup where a generative model is paired with a separate library of documents. The system finds the relevant passages for each question and hands them to the model, so the model's knowledge can be updated without retraining it. For example, a staff assistant that answers from the bank's current policy manual.
Text/word embedding*A list of numbers that represents a piece of text so a computer can work with its meaning, for example to judge how similar two passages are. In practice, RAG systems commonly use them to find relevant documents.

What do agent, human-in-the-loop and human-on-the-loop mean?

An agent is AI that takes actions, not just gives answers. The oversight terms describe how close a person stays to those actions.

Term (as the OCC writes it)Plain meaning
Agentic SI*A category of AI that can make decisions, interact with its surroundings and optimize processes without a person stepping in directly.
SI Agent*A system that, in the lexicon's words, “autonomously perceives its environment, decides what to do, and takes actions to achieve its goals”. For example, a tool that reads an incoming payment request, checks it and sends the payment itself.
Human-in-the-Loop (HITL)*A control where a person sits inside the decision. The lexicon calls it “a risk-control approach for SI where a human is integrated within the SI's decision-making process”. Put simply, a person takes part in the decision itself.
Human-on-the-Loop (HOTL)The system acts on its own while a person watches. The lexicon describes an agent or system that “is operating independently and a qualified human is overseeing, overriding and/or halting actions based on predefined parameters”. For example, an analyst who monitors an agent's actions on a dashboard and can stop it.
Override*An output or input that is ignored, changed, rejected or reversed. For example, a loan officer who sets aside a model's recommendation.
Guardrails*Layers of safeguards that keep an AI system away from bad information and bad behavior. They can be policies, technical controls or monitoring, at the data, model, application or infrastructure level. For example, a filter that stops a chatbot from giving investment advice.

Autonomy itself is not a lexicon entry. Section 3.1, the illustrative risk-tiering method, describes the autonomy level as the “Degree of human involvement in the loop”, running from human-in-the-loop (approves each action) through human-on-the-loop (monitors and can intervene) to fully autonomous (“acts without real-time human checkpoint”). Section 2.2 also points to how long a system may run without “a meaningful human checkpoint” and how broad its data, tool and account access is. See human oversight of AI and the four risk tiers for the next step.

What does the OCC lexicon call the ways AI can go wrong?

Five terms describe how an AI system fails on its own, without an attacker: it makes things up, it is skewed, it gets worse over time, it cannot be seen into, or its errors lead to bad decisions.

Term (as the OCC writes it)Plain meaning
Hallucination*When AI produces something wrong or flawed that still reads as convincing. The lexicon notes this can happen even when the underlying data has no defects. For example, a chatbot that confidently quotes a fee the bank does not charge.
Bias*A systematic distortion, not a random error, that makes an AI system's outputs less accurate or less representative for their purpose. It can be accidental or deliberate and can also emerge once the system is in use, for example when the data does not reflect the people or conditions the system serves. The lexicon names statistical/computational, systemic and human bias as common sources, and says the list is not exhaustive.
SI drift/decay*A model's tendency to get worse over time once real-world conditions differ from those in its training and testing. For example, a fraud model built before a new scam pattern appeared.
Black box*A system whose inner workings are complex or hidden, so even its builders and users cannot easily see how it reaches a result or how well it will perform overall.
Model risk*The chance of harm from decisions based on model outputs that are wrong or misused. The lexicon adds that it exists for each model and in aggregate, and that aggregate risk is affected by how models interact and depend on each other and by shared assumptions, data or methods.

Which lexicon terms cover testing and checking AI?

These seven terms describe how a bank can understand an AI system and confirm it works, before launch and afterward. Validation means confirming with evidence that a system does its intended job.

Term (as the OCC writes it)Plain meaning
Explainability*Whether a given audience can understand why the system behaved as it did, and why a given input led to a given output. For example, being able to tell a declined applicant the main reasons.
Interpretability*Visibility into how the system works inside, relative to what it was designed to do. It helps users judge how the system functions and how far to trust it.
Validation*Confirmation, backed by objective evidence, that an AI system or model meets the requirements for its intended use and works in the environment where it will actually run.
Output validationA systematic check that an AI system's outputs meet set requirements for accuracy and quality before anything downstream uses them. For example, checking an AI-drafted customer letter against a rule list before it is sent.
Benchmarking*Comparing a model's inputs and outputs with estimates from other data or other models, inside or outside the bank. For example, running a second, simpler model on the same loans to see if results line up.
Performance monitoring*Ongoing checks that a system was put in place correctly, is used as intended and keeps working as intended over time.
Performance threshold*A value or range of a performance measure that decides whether a model's performance is accepted or rejected. For example, a minimum accuracy rate set before launch.

Note the two kinds of check. Validation looks at the whole system for its intended use. Output validation checks individual outputs before they are used. More on both in model risk management for generative AI.

What data terms does the OCC lexicon define?

Four terms cover the data an AI system learns from: what it is, where it came from, whether it is fit for use, and whether it is real.

Term (as the OCC writes it)Plain meaning
Training data*The sample of data used to teach a machine-learning model.
Data lineage*The history of a piece of data: where it flowed from point to point and what was done to it along the way.
Data quality/validity*How useful, accurate and correct data is for the job it is being used for.
Synthetic data*Artificial data made by a purpose-built model or algorithm so that it is statistically realistic. It can be used to build and train models. For example, made-up transaction records that behave like real ones but belong to no customer.

For how examiners may look at data, privacy and fair lending, see data, privacy and consumer compliance.

What security attacks on AI does the lexicon name?

Five terms describe deliberate attacks on AI and the protection of a model against tampering. Prompt injection is the one most specific to chatbots and agents.

Term (as the OCC writes it)Plain meaning
Adversarial SI*Techniques and attacks that trick an AI system into wrong or unintended predictions or decisions, often by subtly changing its inputs, training data or interactions.
Data poisoning*An attack that corrupts training data so the system performs badly.
Prompt injection*An attack that exploits the way an application mixes untrusted input with instructions written by a more trusted party, such as the designer, so that the system follows the untrusted instructions. For example, a customer email containing hidden text that tells an email-sorting assistant to forward the inbox.
SI model/system exploitationAttacks that use weaknesses in an AI model or system to make it perform badly, cut off access to it, or reach restricted or proprietary information.
Model integrity*Protecting a model from improper changes or destruction, and making sure its information is authentic and cannot later be denied (non-repudiation).

For containment, kill switches and the security risks specific to agents, see resilience, kill switches and cybersecurity.

What governance and record-keeping terms are in the lexicon?

Seven terms describe how a bank organizes, lists and records its AI: who decides, what is in use, and what changed when.

Term (as the OCC writes it)Plain meaning
SI governance*The policies, rules, roles and oversight that direct how a bank adopts, builds, deploys and monitors AI, so AI risks are found and managed throughout the system's life.
SI lifecycle*The phases an AI system goes through: plan and design, collect and process data, build and use model, verify and validate, deploy and use, operate and monitor. The lexicon notes these often repeat and need not run in order.
SI use case*One specific way AI is designed, built, bought or used to reach a goal. For example, “AI drafts first responses to card-dispute emails” is one use case.
SI use case inventory*A maintained list of the bank's AI use cases, recording where and how AI is used, for what purpose and with what outputs. It supports governance, transparency and risk management.
SI risk assessment*A process to identify, estimate and rank the risks of operating and using an AI system, including threats and weaknesses, and taking account of the controls in place or planned.
Documentation*The records describing an AI system: its purpose and intended use, key design choices, the data it uses and where that came from, test results, limitations and version history.
Version control*Tracking, managing and documenting changes to AI assets as they are built and deployed. For example, knowing exactly which model version gave a customer an answer last March.

For who is accountable and what the board sees, see board, management, audit and reporting.

What do the vendor and concentration terms mean?

Six terms cover AI that the bank gets from someone else, and the danger of leaning too heavily on one provider. A third party here means any outside company the bank relies on.

Term (as the OCC writes it)Plain meaning
Third-party SI risk*Risk that arises when the bank relies on another company to build, provide, host, run or support AI systems or key parts of them, such as models, data or infrastructure.
Service level agreement (SLA)*Binding contract terms between a provider and a customer that set the services to be delivered and measurable commitments such as uptime and response times. SLAs usually also cover each side's duties, reporting, how problems are fixed and remedies when targets are missed.
Service Provider Concentration (Financial Institution)*How much one bank depends, directly or indirectly, on a single provider, especially for critical activities.
Service Provider Concentration (Financial Sector)*The same idea across the industry: how much financial institutions as a group depend on one provider.
Service Provider Concentration Risk (Financial Institution)*The chance that trouble at a provider stops one bank from carrying on its activities, especially critical ones, or causes it significant harm. For example, a bank whose chatbot, fraud screening and document search all run on one AI vendor that goes down.
Service Provider Concentration Risk (Financial Sector)*The same chance across many institutions at once, with possible system-wide impact on the financial sector.

The four concentration terms come in pairs: one for a single bank, one for the whole financial sector. For what examiners may look at when the AI belongs to a provider, see vendor and third-party AI. The tracked Bulletin 2026-54 page and the OCC hub list the official links.

Is Super Intelligence different from AI?

No. Footnote 2 of the framework says Super Intelligence refers to the technologies previously called Artificial Intelligence under 15 USC 9401(3), citing Executive Order 14434. It is a new name for the same thing. A bank's existing chatbots and models are SI in the OCC's sense.

What does the asterisk next to a term mean?

It means the definition comes from the AI Lexicon workstream of the Artificial Intelligence Executive Oversight Group (AIEOG). The OCC changed "AI" to "SI" in those definitions to match the framework. 47 of the 51 terms are starred.

Which terms did the OCC not take from the AIEOG lexicon?

Four terms carry no asterisk: Human-on-the-Loop (HOTL), Output validation, SI model/system exploitation and Super Intelligence (SI). The appendix does not say where those definitions come from.

What is the difference between human-in-the-loop and human-on-the-loop?

Human-in-the-loop puts a person inside the decision, so in practice a person approves before anything happens. Human-on-the-loop lets the system act on its own while a qualified person oversees it and can override or halt it. Section 3.1 adds a third level, fully autonomous, with no real-time human checkpoint.

Does a bank have to use these definitions?

The framework is a resource for examiners, and the bulletin says it does not establish supervisory expectations for banks. The lexicon gives examiners and banks a common vocabulary. Using the same words can make an exam conversation easier, but the text sets no requirement to adopt them.

Is a word processor with AI features an SI system?

No, by the lexicon's definition. An SI system excludes any common commercial product with AI embedded in it, and the OCC names a word processor and a map navigation system as examples.

Will the lexicon change?

The OCC says it should be treated as a living lexicon, expected to change as technology, use cases and supervisory understanding mature. It gives no date for an update.

  1. Super Intelligence: Supervisory Framework (full text, PDF), Appendix A: Lexicon — Office of the Comptroller of the Currency, Nov 4, 2026 (occ.gov) · our page on OCC Bulletin 2026-54
  2. OCC Bulletin 2026-54: Super Intelligence: Supervisory Framework — Office of the Comptroller of the Currency, Nov 4, 2026 (occ.gov) · our page on OCC Bulletin 2026-54
  3. OCC Bulletin 2026-13: Model Risk Management: Revised Guidance — Office of the Comptroller of the Currency, Apr 17, 2026 (occ.gov) · our page on OCC Bulletin 2026-13

When this rule moves, the next morning's brief says so

when one of these regulators moves, the next morning's brief says so · six sourced stories · 7 am PT · free

plus every regulation, bank and agent page update, the morning after · leave any morning