AI Academy · Book
Executives & Directors · Module 02 · Chapter 001

What Exactly Is Artificial Intelligence?

"AI" is a label, not a specification. It covers a family of capabilities that real systems combine unevenly, and the label alone tells you almost nothing about what a system does, how reliably it does it, or what happens when it is wrong. Leaders who translate the label into a capability buy better, deploy more safely and stay out of court.

≈ 15 min read

After this chapter you can

  • Define AI as the capability of a computer system to perform tasks that normally require human thinking, and relate it to the OECD and EU definitions.
  • Separate the "AI" label from the capability underneath it, including why the label has covered very different methods.
  • Name the six capabilities of the AI family and describe a system's capability profile from evidence.
  • Apply six questions to test any AI claim before trusting, buying or making it.

On 18 March 2024, the US Securities and Exchange Commission settled charges against two investment advisers. Their offense was not a trading fraud or a hidden fee. It was a word. One firm had told clients it used artificial intelligence and machine learning on their data to predict which companies were about to “make it big”; it could not. The other marketed itself as the “first regulated AI financial advisor”. Together they paid 400,000 in civil penalties, and the SEC’s chair gave the practice a name: “AI washing”1. Six months later the Federal Trade Commission announced a sweep of its own against deceptive AI claims, with a blunt message from its chair: “there is no AI exemption from the laws on the books”2.

Before reading on, try a small test. Is the spam filter in your email AI? The app that plans your route to the airport? The chatbot on a retailer’s website? Most leaders hesitate, and the hesitation is the right instinct. The honest answer each time is that it depends on what the system does underneath. One spam filter learned from millions of past messages; another checks a list of blocked senders. One chatbot reads a question and writes an answer; another walks the customer through a fixed script. A single word covers all of them, and that is exactly the problem.

The label is not the capability

You will hear the word AI every week. A vendor calls its platform AI-powered. A start-up promises advanced AI. One of your own teams asks for AI in a process. On its own, the word answers none of the questions that matter. It does not say what task the system performs, what data it needs, how often it is right, how much it does without a person, what it costs or what value it creates.

The label 'AI-powered' tells you almost nothing; the capability - what a system does, how reliably and with what data - tells you what you are buying.THE LABEL'Our platform is AI-powered.'Tells you almost nothingTHE CAPABILITYWhat it does, how reliably,with what data.Tells you what you are buyingvs
Figure 2.1.1 The label is where your questions start. The capability is what you are actually buying.

Translating the label is a skill, and it rests on a small vocabulary: a working definition, the official definitions regulators now apply, a sense of how many different methods the word has covered, the family of capabilities underneath it, and a short set of questions that turn any AI claim into something you can test. A lawyer who skipped that translation in 2023 shows what it costs.

A working definition: capability, not consciousness

For this Academy, use one working definition. Artificial intelligence is the capability of a computer system to perform tasks that normally require human thinking: recognizing an object in a photograph, understanding a question in plain language, predicting an outcome, recommending an action, drafting a document or planning a sequence of steps.

Notice what the definition does not claim. It does not say the machine thinks the way a person does, feels anything or understands the world as you do. It says the system can perform a task we associate with intelligent behavior. That is a statement about capability, not about consciousness, and holding that line saves a great deal of unproductive debate. For a leader, the useful questions are what the system can do and how well.

The official definitions make the same move in more formal language. In November 2023 the OECD revised its definition of an AI system to cover generative AI: a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments3. The international standard for AI vocabulary, ISO/IEC 22989, frames an AI system the same way, by the outputs it generates for objectives that people set4. The EU AI Act adopted almost the OECD’s wording in its Article 3(1)5.

An AI system pursues an objective set by people, takes input, infers how to produce an output such as a prediction or decision, and influences the world with some autonomy.ObjectiveSet by people,explicitly orimplicitlyInputData, text,images, signalsInferenceWorks out how toproduce theoutputOutputPrediction, content,recommendation,decisionOutputs that influence the world, with some degree of autonomy
Figure 2.1.2 The shared anatomy of the OECD and EU definitions. Inference, not appearance, is what makes a system AI.

The key word is infers. A system that only executes rules a person wrote down, such as a tax formula in a spreadsheet, does not infer anything; it follows instructions. The EU AI Act says so explicitly in its recitals, and the European Commission’s guidelines of February 2025 break the definition into seven elements that must be judged case by case6. Where exactly the line falls between written rules and inference is the subject of the next chapter.

One label, many methods

Part of the confusion is historical. The term artificial intelligence was coined in 1955, as Why AI, Why Now? recounted, in a proposal whose central conjecture was that every feature of intelligence could in principle be described precisely enough for a machine to simulate it7. That framing defined the field by its goal, not by any particular method. It has been carried by very different methods since.

The AI label has covered hand-written expert rules, deep learning from data and foundation models adapted to many tasks, so the word alone says little about the method.1955The term is coinedA goal, not a method1970s-1980sRules writtenby expertsKnowledge enteredby hand2010sDeep learningPatterns learned from largedata sets2020sFoundation modelsOne model adapted tomany tasks
Figure 2.1.3 Every generation of methods was called AI. The label stayed; what it promised changed.

For decades, much of the field tried to capture knowledge by writing it down: experts and engineers entered facts and rules by hand, and the system reasoned over them. Those systems worked in narrow settings but proved brittle, because the world contains more exceptions than anyone can write down. The breakthrough of the 2010s came from the opposite direction. Instead of being told the rules, deep-learning systems learned patterns from very large collections of examples8. The 2020s added foundation models, trained once on broad data and then adapted to many tasks, which is what made today’s general-purpose assistants possible9.

The practical consequence is that “AI” in a sales pitch might mean a hand-tuned rule engine, a statistical model trained on your data, or a large language model rented from a provider. Those are very different products with very different risks. How machine learning relates to AI is set out in AI vs Machine Learning; for now, the point is simpler. The label has never told you which method is inside.

Six capabilities, one family

If AI is not one technology, what is it? The most useful answer for a leader is a family of capabilities. The six below are this Academy’s working vocabulary, not an official taxonomy, and they overlap. Most modern systems combine several.

AI is a family of six capabilities - perceive, understand, learn, reason, generate and act - that real systems combine in different mixes.AIA familyof capabilitiesPerceiveUnderstandLearnReasonGenerateAct
Figure 2.1.4 Six capabilities, combined in different mixes. Ask which ones a system provides, and how well.

To perceive is to pull useful information from signals: a photograph of a damaged part, a voice on a call, a scanned page, a sensor reading. To understand is to interpret language and context, such as a customer’s email or a clause in a contract. Use that word with care; useful language behavior is not human comprehension. To learn is to improve from data or experience rather than follow only rules someone wrote down. To reason is to work from available information to a conclusion, compare options or plan steps, and apparent reasoning still needs checking. To generate is to create text, images, audio, code or designs; Generative AI Changes the Game explained why this capability changed knowledge work so quickly. To act is to take a step in the world, such as opening a ticket, updating a record or sending a message, which moves the question from what AI can tell you to what it can do for you.

Most of these capabilities work out of sight, inside forecasting, search, translation and document systems you already own, as AI Is Already Inside Your Organization showed. That is one more reason the visible label is a poor guide. A system with no chat window may be doing more inference than one with a friendly face.

Every system has a capability profile

The family is a vocabulary. Any real system has a profile: strong at some of these capabilities, weak or blind at others, and uneven even within one. A tool can understand a typed contract well and be unable to perceive a scanned one. A model can draft a fluent memo and invent the facts inside it. AI Is Changing Everything called this the jagged frontier: tasks that look equally hard to people can fall on either side of what a system does well10.

The way we hire people is a useful comparison. Nobody hires a candidate simply for being intelligent. You hire a financial analyst for numerical reasoning and look at their models; you hire a designer for visual creation and ask for a portfolio. The comparison breaks down on accountability, since a system cannot own a decision, but it holds on the main point: you judge a capability profile by evidence on the work you need done.

The evidence often differs from the marketing. A 2025 study by researchers at Stanford and Yale found that leading purpose-built legal research tools, promoted as avoiding invented citations, still produced a hallucinated answer on between 17 and 33 percent of test queries11. Better than a general chatbot is not the same as safe to use unchecked, which is why a profile has to be built from evidence on your own work, and kept current as the system and your use of it change.

A profile also has limits of scope. Almost all AI in business today is narrow: specialized, often impressive, and not general. You will hear the term artificial general intelligence, a hypothetical system with human-like breadth. There is no agreed definition or test for it, so when someone uses the term, ask what was demonstrated, under what conditions and on what evidence. The Major Types of AI puts the term in context.

Six questions that translate any claim

One more principle sits behind everything above: AI is not magic. Behind every capability you see is a system you do not: data, models, computing power, software, connections to your other tools, testing, security and people. Data, Models and Compute explains those layers. A polished demonstration hides all of them, which is why the translation from label to capability needs a short, repeatable set of questions.

Six questions - task, input, output, evidence, failure and outcome - turn any AI claim into a capability you can judge.TaskWhat exactly does it do?InputWhat does it need, and what can itnot read?OutputWhat does it produce?EvidenceProven how, on work like ours?FailureWhat happens when it is wrong?OutcomeWhich business result does it change?
Figure 2.1.5 Six questions turn an AI label into a capability you can judge. The failure question is the one most often skipped.

The first three questions pin down the capability: what the task is, what input the system needs and what it can not read, and what it produces. The fourth asks for evidence, and the best evidence is a test on your own material, not a benchmark on someone else’s. The fifth, what happens when the system is wrong, is the easiest to skip and often the most expensive to miss. A system that fails loudly is manageable; one that fails silently, with a confident and tidy output, is dangerous. The sixth connects the capability to a business result, which is where Module 03 picks up the thread.

None of the six questions asks whether the system is “really AI”. That question is almost never the useful one.

Story: the super search engine

In 2022 a passenger sued an airline in federal court in New York, claiming a metal serving cart had struck his knee on a flight. The airline moved to dismiss the case. The passenger’s lawyers, at a firm that practiced mainly in New York state courts, needed federal case law on an unfamiliar question. Their usual research service had limited access to federal cases. One of them turned to a general-purpose chatbot12.

A brief cited six opinions that did not exist; warnings from the opponent and the court were missed, and the court imposed sanctions in June 2023.1 Mar 2023Filing madeCites six court opinions15 Mar 2023Opponent cannotfind themThe warning is not read25 Apr 2023The firm standsby themFiles the purported texts22 Jun 2023Sanctions5,000 penalty; lettersto judges
Figure 2.1.6 Each step was a chance to ask what the tool actually does. The first admission came only on 25 May.

The filing of 1 March 2023 cited six court opinions that did not exist, complete with invented quotations. On 15 March the airline’s lawyers told the court they could not locate most of the cases. The court ordered copies. On 25 April the firm filed what it presented as the opinions themselves, which had also been generated. The first admission that the cases came from the chatbot arrived only on 25 May, after the court ordered the lawyers to show why they should not be sanctioned. On 22 June 2023, Judge P. Kevin Castel imposed a penalty of 5,000, and ordered the lawyers to send letters to their client and to each real judge falsely named as the author of a fake opinion12.

The judge was careful about where the fault lay. He wrote that there is nothing inherently improper about using a reliable AI tool for assistance; the failure was that the lawyers abandoned their duty to check what they filed. The post-mortem, read through this chapter, is about a capability mismatch.

The lawyer treated a text generator as a search engine, assumed it found real cases, and filed them without reading the sources.Label"A supersearch engine"AssumptionIt finds real casesCapabilityIt generatesplausible textMissing checkNobody readthe sourcesThe tool did what it does. Nobody asked what that was.
Figure 2.1.7 The failure was not the software. It was a wrong model of the capability, and no check behind it.

The lawyer testified that he had “falsely assumed” the chatbot was “like a super search engine” and that he was operating under the belief that it “could not possibly be fabricating cases”. He had even asked it whether one of the cases was real; it said yes12. That is the label-versus-capability gap in its purest form. He bought, in his own mind, a retrieval capability: find real documents. What he was using was a generation capability: produce text that reads like the answer. How Generative AI Works explains why fluent output and correct output are different properties.

Run the six questions and the gaps appear at once. The task was not “find cases” but “write plausible text”. The input did not include a database of court opinions. The evidence for legal research was nil. And the answer to “what happens when it is wrong?” was the worst kind: a confident, well-formatted citation to nothing. Asking the system to vouch for itself was not evidence.

The case was not an isolated lapse: by October 2026 a public database listed more than 2,000 court decisions worldwide involving hallucinated content, some 850 of them involving lawyers13. Knowing the law did not protect them; an accurate picture of the tool would have.

What this means for leaders

The chapter reduces to a habit: translate the label before you trust it. That habit protects three things. It protects your purchasing, because vendors are often not asked to specify the capability they are selling. It protects your operations, because the failure mode of a system is part of its design, and it should be known before the system goes live. And it protects your organization’s word, because regulators now treat claims about AI like any other claim about a product.

Two cautions keep the habit honest. Do not swing from credulity to cynicism; a system that is not “true AI” may be exactly the right tool, and one that is may be the wrong one. And do not let the debate settle on definitions. Your legal team needs the formal definition to judge regulatory scope. Everyone else needs the capability.

Check yourself

  1. If a vendor calls a product “AI-powered”, you know what capability you are buying.
  2. Under the OECD and EU definitions, inference from input is central to what makes a system AI.
  3. A purpose-built AI tool marketed for a profession no longer needs its output checked.
  4. The word AI has described quite different methods over the decades, from hand-written rules to foundation models.
  5. In the 2023 fake-citations case, the court said lawyers must not use AI tools for legal work.
  6. Asking an AI system whether its own answer is correct is good evidence of reliability.

Reflection: find the label in your own house

What comes next

A working definition puts inference at the heart of AI. That raises an obvious question. Organizations have automated work with rules and software for decades, and much of what is sold as AI looks a lot like that. Is AI really different from automation, or is it automation with a new name? The next chapter, AI vs Automation, draws the line.

Laws referenced

Not legal advice. Laws change; verify before relying on this, and consult counsel for decisions.

EU AI Act · EU

Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744

Risk-based rules. Prohibited practices include social scoring, untargeted scraping of facial images, and emotion recognition in workplaces and schools (with narrow exceptions). High-risk systems (Annex III: biometrics, safety components of critical infrastructure such as energy, water and traffic, employment and worker management, credit, education, essential services, law enforcement, migration, justice) need risk management, data governance, documentation, logging, human oversight, human oversight that keeps people able to understand the system, notice automation bias (over-reliance on its output), override it or stop it (Art. 14(4)), appropriate accuracy, robustness and cybersecurity (Art. 15), automatic logging of events (Art. 12), a provider quality-management system (Art. 17) and conformity assessment. An Annex III system is not high-risk if it poses no significant risk of harm, for example a narrow procedural or preparatory task that does not replace human assessment; systems that profile people are always high-risk, and a provider relying on this exception must document it and register (Art. 6(3)). Deployers of high-risk AI must use it as instructed, assign competent human oversight, monitor its operation, keep logs for at least six months and report serious incidents (Art. 26); employers must inform workers' representatives (Art. 26(7)). Public bodies, private providers of public services, and deployers of credit-scoring or life and health insurance pricing systems must carry out a fundamental-rights impact assessment before first use (Art. 27). Providers must run post-market monitoring (Art. 72). A deployer that puts its name on a high-risk system, substantially modifies it, or changes its purpose so that it becomes high-risk takes on the provider's obligations (Art. 25(1)). A substantial modification (Art. 3(23)) of a high-risk system needs a new conformity assessment, unless the change was pre-determined and documented at the first assessment, as with planned continuous learning (Art. 43(4)). Providers of general-purpose AI models (from 2 Aug 2025) must keep technical documentation, have a policy to comply with EU copyright law including text-and-data-mining opt-outs, and publish a sufficiently detailed summary of training content (Art. 53). Research, testing and development before a system is placed on the market or put into service is outside the Act, except testing in real-world conditions (Art. 2(8)). Since the 2026 Omnibus, the Art. 4 AI-literacy duty is an obligation of effort (take measures to support literacy), not of result. Fines reach EUR 35 million or 7% of global turnover for prohibited practices.

  • 2024-08-01 — Entered into force
  • 2025-02-02 — Prohibited practices (Art. 5) and the AI-literacy duty (Art. 4) apply
  • 2026-07-27 — Omnibus softens Art. 4: providers and deployers must take measures to support AI literacy; no specific level must be guaranteed
  • 2025-08-02 — General-purpose AI model obligations apply; governance and penalties regime in place
  • 2026-08-02 — Transparency duties (Art. 50) apply: disclose AI interaction, label synthetic and deepfake content (marking for generative systems already on the market: 2 Dec 2026)
  • 2027-12-02 — High-risk obligations for Annex III systems (e.g. hiring, credit, education, essential services) - moved from 2 Aug 2026 by the 2026 Omnibus
  • 2028-08-02 — High-risk obligations for AI in products regulated under Annex I

Last verified 2026-10-06 · official text

US enforcement against false AI claims ("AI washing") · US - federal

Federal securities antifraud rules and Investment Advisers Act Marketing Rule (SEC); FTC Act Section 5 (unfair or deceptive practices)

There is no AI-specific federal statute, but existing law already applies to what companies say about AI. Regulators have penalized firms that overstated their use or capability of AI to investors (SEC) and to consumers (FTC). Claims about AI in marketing, investor materials and product descriptions need the same substantiation as any other claim.

  • 2024-03-18 — SEC's first AI-washing cases: Delphia and Global Predictions settle for USD 400,000 in total civil penalties
  • 2024-09-25 — FTC launches Operation AI Comply, with five actions over deceptive AI claims and uses

Last verified 2026-10-08

References

  1. US Securities and Exchange Commission. SEC Charges Two Investment Advisers with Making False and Misleading Statements About Their Use of Artificial Intelligence (Press Release 2024-36). SEC. 2024.
  2. US Federal Trade Commission. FTC Announces Crackdown on Deceptive AI Claims and Schemes (Operation AI Comply). FTC. 2024.
  3. OECD. Recommendation of the Council on Artificial Intelligence (OECD AI Principles), updated 2024. OECD. 2024.
  4. ISO/IEC. ISO/IEC 22989:2022 Information technology - Artificial intelligence - Artificial intelligence concepts and terminology. International Organization for Standardization. 2022.
  5. European Parliament and Council of the European Union. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 3(1) and Recital 12. Official Journal of the European Union. 2024.
  6. European Commission. Commission Guidelines on the definition of an artificial intelligence system established by Regulation (EU) 2024/1689. European Commission. 2025.
  7. John McCarthy, Marvin L. Minsky, Nathaniel Rochester and Claude E. Shannon. A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, August 31, 1955. AI Magazine 27(4), 2006 reprint. 1955.
  8. Ian Goodfellow, Yoshua Bengio and Aaron Courville. Deep Learning. MIT Press. 2016.
  9. Rishi Bommasani et al. On the Opportunities and Risks of Foundation Models. Stanford CRFM (arXiv:2108.07258). 2021.
  10. Fabrizio Dell'Acqua et al. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality, HBS Working Paper 24-013. Harvard Business School. 2023.
  11. Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning and Daniel E. Ho. Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. Journal of Empirical Legal Studies 22(2), 216-242. 2025.
  12. US District Court for the Southern District of New York. Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), Opinion and Order on Sanctions (22 June 2023). S.D.N.Y. (Judge P. Kevin Castel). 2023.
  13. Damien Charlotin. AI Hallucination Cases database. damiencharlotin.com. 2026.

Further reading

Sources last verified 2026-10-08.