What is Artificial Intelligence?

Can a company disown what its chatbot tells a customer? Air Canada tried. Its website chatbot had told Jake Moffatt, who was booking a flight after a death in his family in November 2022, that he could buy a ticket at the normal price and apply for the reduced bereavement fare within 90 days. The airline's own policy page said the opposite: bereavement fares could not be claimed after travel. When Moffatt took the refused refund to British Columbia's Civil Resolution Tribunal, the airline suggested, in the tribunal's words, that the chatbot was "a separate legal entity that is responsible for its own actions". The tribunal called that "a remarkable submission" and ruled on 14 February 2024: "It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot." The airline was ordered to pay C$812.02.

The decision never asks whether the chatbot was "really AI", what model it ran or how it was trained. What mattered was what the chatbot's answers were allowed to do. Air Canada had built something to inform customers, and the law treated its words as promises. That gap, between what a system was designed to be trusted with and what its outputs actually do, is where many of the costliest AI failures live.

Catching that gap in a design review takes a precise vocabulary: knowing which of three respectable definitions of "AI" you are using, describing a system on six independent axes, and pricing the decision the Air Canada case was really about, which is how much authority to give a system's outputs and when human review pays for itself.

Before you start

This is a Beginner chapter. It builds on Chapter 1, especially the idea that a behavioural test answers only the question its conditions define and the advice to restrict what a component is allowed to do. No programming is needed; the Python and Java programs are optional.

By the end you should be able to:

• apply the operational definition of an AI system used by the OECD and the EU AI Act, and explain why a minimax chess program is AI in textbooks but likely outside the Act;
• describe any system on six axes (field, method, model family, capability, application and authority) and spot a description that mixes them up;
• place a system's authority on the Sheridan–Verplank levels of automation, from informing a person to acting alone;
• decide, with numbers, when reviewing an assistant's answers costs less than the errors it would catch, and state when that answer flips.

Three definitions that disagree

Ask three groups what counts as AI and you get three answers, each correct for its purpose.

The research field is broad. The 1955 Dartmouth proposal you met in Chapter 1 set its scope with the conjecture that "every aspect of learning or any other feature of intelligence" could be simulated by a machine, and textbooks such as Russell and Norvig's Artificial Intelligence: A Modern Approach cover search, logic, planning, probabilistic reasoning, machine learning, perception and robotics. By that standard a chess program built on minimax search, which looks ahead through possible moves on the assumption that each side will pick its best reply, and scores the resulting positions with a hand-written evaluation function, is classic AI, whether or not it learned anything.

Standards bodies and regulators need a test that decides whether a law applies to a product. In November 2023 the OECD revised its definition, and the EU AI Act adopted almost the same words in Article 3(1):

"'AI system' means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and 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 environments."

Read it clause by clause. The system is machine-based and pursues objectives, which may be implicit. Autonomy may vary and adaptiveness is optional ("may exhibit"), so a model frozen at deployment still counts. The load-bearing word is infers: the system works out, from its input, how to produce its output. Recital 12, one of the numbered explanations at the start of the Act, says what that excludes: the definition "should not cover systems that are based on the rules defined solely by natural persons to automatically execute operations". A hand-coded tax calculation, however intricate, is not an AI system under the Act. The European Commission's guidelines on the definition, published on 6 February 2025, go further and name categories that may fall outside because of their "limited capacity to analyse patterns and adjust autonomously their output": basic data processing, mathematical optimisation, simple prediction and systems based on classical heuristics, with a minimax chess program as the example. The textbook's classic AI is, in the Commission's reading, likely not an AI system at all.

The distinction has consequences. The Act entered into force on 1 August 2024; its prohibitions have applied since 2 February 2025, its rules for general-purpose AI models since 2 August 2025 and its transparency duties since 2 August 2026. The Digital Omnibus on AI, Regulation (EU) 2026/1744, in force since 27 July 2026, moved the obligations for high-risk systems listed in Annex III, such as credit scoring and recruitment, to 2 December 2027 (dates as of September 2026). None of that applies to software that is not an AI system in the first place, which is why a requirement document should say which definition it uses.

Marketing uses the word most loosely of all. On 18 March 2024 the US Securities and Exchange Commission settled its first "AI washing" cases: two investment advisers, Delphia and Global Predictions, paid civil penalties of $225,000 and $175,000 for claiming to use AI and machine learning in ways they could not substantiate. "AI-powered" on a product page is a claim, and like any claim it needs evidence.

One more word needs pinning down, because it has two meanings in this handbook. In the definition above, and in logic, inference means deriving an output from inputs. In machine-learning engineering it also means running a trained model to produce an answer, as opposed to training it; "inference cost" and "inference latency" use this sense. Context usually makes it clear, and this handbook says which sense it means when it could be either.

Interactive

Is it an AI system under the Act?

Read the EU AI Act's definition the way the European Commission's February 2025 guidelines read it. Which of these systems is most likely outside the definition of an AI system?

• A chess program that chooses moves by minimax search with a hand-written evaluation function.
• Gmail's spam classifier.
• An assistant built on a large language model that answers customer questions.
• A credit-scoring model trained on repayment histories.

Commit to an answer first; the explanation appears afterwards.

Six questions, six axes

The most common confusion about AI comes from arranging every term in one ladder: AI contains machine learning, which contains deep learning, which contains foundation models, which contain language models, which power generative AI. The terms do not sit on one ladder, because they answer different questions. Describe a system with six.

A few terms are worth defining precisely, because they are routinely used as synonyms.

• Machine learning is a method: producing behaviour by fitting a model's parameters to data instead of writing the rules by hand. Deep learning is machine learning with multi-layer neural networks, the networks Chapter 1's threshold units grew into.
• A foundation model, in the definition of the Stanford report that coined the term in 2021, is "any model that is trained on broad data (generally using self-supervision at scale) that can be adapted (e.g., fine-tuned) to a wide range of downstream tasks". Self-supervision means the training signal comes from the data itself, for example by hiding the next word of a sentence and asking the model to predict it; fine-tuning means training an existing model a little further on a narrower task. A large language model (LLM) is a large neural network trained to predict text; the broadly trained, widely reused ones are foundation models. Chapters 11 and 12 open both up.
• Generative AI is a capability, producing new text, code, images, audio or structured data, not a model family. Language models, image models and others all provide it.

A system has a value on every axis, and the values vary independently. The four real systems below show it.

Figure 2.1 · Interactive

Four real systems on six axes

Step through Gmail's spam filter, GitHub Copilot's inline suggestions, the Waymo Driver and Air Canada's 2022 chatbot. Compare the authority rows: similar technology can carry very different authority, and authority can be judged even when the technology is secret.

1. 1
2. 2
3. 3
4. 4
5. 5

Step 1 of 5

Gmail spam filtering

Google reported in November 2023 that RETVec, a component that turns the text of a message into numbers its neural network can read, raised Gmail's spam detection rate by 38% over the previous baseline and cut false positives, genuine mail wrongly marked as spam, by 19.4%. The filter acts on every message without asking anyone, which is tolerable because each decision is visible and easy to undo.

Field

AI: natural language processing

Method

machine learning

Model family

neural network text classifier

Capability

classify

Application

email client

Authority

execute, reversibly: moves mail to Spam; the user can move it back

← PreviousNext →

Two habits follow. First, fill in every axis, and write "not disclosed" rather than guessing: an honest blank is more useful in a review than a confident wrong label. Second, never let one axis stand in for another. "It uses an LLM" answers the model-family question and says nothing about what the system may do; "it's generative AI" names a capability and says nothing about how the behaviour was produced.

Authority: what an output is allowed to do

Authority is the axis most descriptions leave out, and the one that decides what a mistake costs. It says what happens to the world when the system produces an output, and it is set by the application around the model, not by the model. Human-factors engineers mapped this ground long before language models. Thomas Sheridan and William Verplank's ten levels of automation (1978), in the form Parasuraman, Sheridan and Wickens published in 2000, give each step a precise meaning:

Vehicles have their own scale. SAE J3016, a standard from the engineering body SAE International, defines six levels of driving automation, from Level 0 (warnings only) to Level 5 (drives anywhere a person could). The line that matters most is between Level 3, where the system drives but a person must take over when asked, and Level 4, where no human fallback is needed inside the operating domain. Waymo's claim of Level 4 is a claim about authority, and its limits (the operating domain, remote advice but no remote driving) are how that authority is bounded.

Three things raise the authority a system can safely hold: how reversible its actions are, how quickly a person would notice a mistake, and how much a mistake costs. Gmail can execute because a misfiled email is cheap, visible and one click from undone. A refund, a payment or a statement a customer relies on is none of those, which is why capability alone, however well measured, never grants authority.

Interactive

Place the authority

GitHub Copilot shows a suggested line as dimmed text, and nothing changes in the file until the developer presses Tab. Which Sheridan–Verplank level describes that authority?

• Level 4: the computer suggests one alternative.
• Level 5: the computer executes that suggestion if the human approves.
• Level 7: the computer executes automatically, then informs the human.
• Level 10: the computer decides everything and ignores the human.

Commit to an answer first; the explanation appears afterwards.

Air Canada's chatbot: an authority failure

Tell the Air Canada story the way an incident review would.

Symptom. A customer relied on the chatbot's statement that he could claim a bereavement fare within 90 days of buying his ticket. The airline's policy, on the bereavement page that the chatbot's own answer linked to, said such claims could not be made after travel. The refund was refused and the customer sued.

Diagnosis. The chatbot was designed to inform, but its outputs had the authority of the company: in law, what the chatbot said was what Air Canada said. The tribunal rejected the argument that the chatbot was responsible for its own actions, and it would not require a customer to work out which part of the airline's own website was right. Nothing in the system kept the chatbot's answers consistent with the policy page, and nothing routed a question with money attached to a person. The authority axis had been set by accident.

Fix. Treat every customer-facing statement about policy as a commitment the company must honour. Generate those answers from the canonical policy text and link to it, so the chatbot cannot drift from the page; send anything that promises money or an exception to a person or to a rules engine that implements the policy exactly; and log each answer so a dispute can be checked against what was actually said. The consistency check in the chapter's program flags the Air Canada profile for exactly this reason.

The award was C$812.02: C$650.88 in damages, C$36.14 in interest and C$125 in tribunal fees. The precedent cost more. A company is answerable for what its automated systems tell customers, whatever the technology.

Pricing human review for a production assistant

Air Canada shows the cost of too much authority granted by accident. Klarna shows the same decision made on purpose, and then revisited. On 27 February 2024 the payments company reported that its OpenAI-powered assistant had handled 2.3 million conversations in its first month, two-thirds of Klarna's customer-service chats and the work of about 700 full-time agents. Customers resolved their issues in under 2 minutes instead of 11, and Klarna estimated a US$40 million profit improvement for 2024. Fifteen months later its chief executive told Bloomberg, on 8 May 2025, that "cost unfortunately seems to have been a too predominant evaluation factor" and that the result was "lower quality"; Klarna began recruiting people so that customers could always reach a human.

Put numbers on that decision for a customer-service assistant that answers policy questions. Three policies are on the table: let it answer without review, have a person review every conversation, or route only the risky conversations (refunds and policy exceptions) to review. The expected monthly cost of each is

$$C_{\text{none}} = N\,e\,a\,L, \qquad C_{\text{full}} = N\,r_c + C_{\text{none}}\,(1 - k), \qquad C_{\text{routed}} = N\,s\,r_c + C_{\text{none}}\,(1 - k\,q)$$

where N is conversations per month, e the share of answers that are wrong on a policy point, a the share of wrong answers a customer acts on, L the cost of one relied-upon wrong answer, $r_c$ the cost of reviewing one conversation, k the share of wrong answers a reviewer catches, s the share of conversations routed to review and q the share of costly errors that fall inside those routed conversations.

Take L = C$812.02, the Moffatt award, and illustrative values for the rest: N = 100,000, a = 10%, review at 1.5 minutes and C$40 an hour (C$1.00 a conversation), k = 90%, s = 10% and q = 80%. At an error rate of 1%, no review costs C$81,202 a month in expected relied-upon errors, full review C$108,120 and routed review C$32,737. Routed review wins by a wide margin because it spends reviewer time only where the costly mistakes are.

Figure 2.2 · Interactive

When does reviewing an assistant's answers pay for itself?

Expected monthly cost of three policies for a customer-service assistant: no review, review of every conversation, and review of only the conversations routed as risky. Start at the defaults, then slide the error rate down to 0.1% and up to 8% and watch which policy is cheapest.

Reset

TryDefaults: Moffatt award as the cost per errorKlarna scale: 2.3 million chats a monthRouting that cannot tell risky from routine

Conversations per month

100,000

Share of answers wrong on a policy point

1%

Share of wrong answers a customer acts on

10%

Cost of one relied-upon wrong answer (C$)

812.02

Reviewer minutes per conversation

1.5

Reviewer cost per hour (C$)

40

Share of wrong answers a reviewer catches

90%

Share of conversations routed to review

10%

Share of costly errors inside routed conversations

80%

Cost of reviewing one conversation (C$)

1

No review: expected relied-upon errors (C$ a month)

81,202

Review everything (C$ a month)

108,120

Review routed conversations only (C$ a month)

32,737

Error rate above which full review beats no review

1.37%

Error rate above which routed review beats no review

0.17%

Slide the error rate and the cheapest policy flips twice: below about 0.17%, no review is cheapest; between 0.17% and 6.2%, routed review wins; above 6.2%, reviewing everything is worth its cost. Volume does not move the flip points, because every term grows with N; Klarna's 2.3 million conversations multiply every cost by 23 and leave the ranking unchanged. And routing is only as good as the router. When the routed tenth of conversations holds only a tenth of the costly errors, the third preset, routed review costs C$83,894 at a 1% error rate and loses to no review at all.

The model also shows what you must measure before deciding. The flip points depend on the error rate, and Chapter 1's arithmetic says how expensive that measurement is: to show that a rate is below 0.17% at 95% confidence takes about 1,750 consecutive clean answers. A team that has not run that evaluation is not choosing between these policies; it is guessing. When the costs a model cannot price, such as a regulator's attention or a lost customer, are large, the honest answer is to start at draft authority, measure, and raise authority only as the measured error rate earns it. Klarna's reversal, from replacing agents to promising that a customer can always reach one, is the same lesson learned in public.

Interactive

Make the call

With the calculator's defaults, suppose the assistant's policy answers are wrong 0.5% of the time. Which policy is cheapest, and what would change the answer?

• No review, because 0.5% is small.
• Routed review; it stays cheapest until the error rate passes about 6.2%, when full review wins, and it loses to no review below about 0.17%.
• Full review, because every error should be caught.
• It depends mainly on how many conversations there are.

Commit to an answer first; the explanation appears afterwards.

When not to use AI at all

The six axes also make the opposite decision easier. Prefer ordinary software when a formula, a ledger or a database query produces the authoritative answer: an account balance is computed from the ledger, and a language model may explain it but must never be its source. Prefer it when the rules are complete and change only through a controlled policy process, when no one can define how the system's quality would be measured, or when nobody owns the outcome if it is wrong. In each case the "infers" clause is the tell: if nothing needs to be inferred, an AI system adds cost and uncertainty and removes nothing.

Summary

"AI" means different things to researchers, regulators and marketers. The research field includes search, logic and planning, so a minimax chess program is classic AI. The OECD and EU AI Act definition turns on whether a system infers how to produce its outputs, excludes rules written entirely by people, and, in the Commission's February 2025 reading, likely excludes classical heuristics such as minimax chess. Which definition applies decides whether a law applies, so requirements should name it.

Describe any system on six independent axes: field, method, model family, capability, application and authority. Machine learning is a method, a large language model is a model family, and generative AI is a capability. Four real systems show the axes varying separately, and the Air Canada chatbot shows that authority can be judged, and can fail, even when nothing about the technology is known.

Authority runs from informing a person to acting alone, and the Sheridan–Verplank levels give each step a meaning. It is set by the application, and it should follow reversibility, detectability and cost, never capability alone. For a customer-service assistant, the cheapest review policy flips from none to risk-routed at about a 0.17% error rate and from routed to full review at about 6.2%, under the chapter's illustrative costs, and none of it can be chosen without measuring the error rate first.

Chapter 3 takes the terms people use to sort AI systems into types (narrow and general, agentic and autonomous) and asks what evidence each type claim needs. It separates capability from autonomy, the dial this chapter called authority, and tests a famous benchmark claim against its fine print.