Are Airlines Liable for Information Provided By AI?
There is something deceptively simple about the decision in Moffatt v. Air Canada, 2024 BCCRT 149. A passenger asked an airline chatbot whether a bereavement fare could be claimed retrospectively. The chatbot said yes. The airline subsequently said no. The passenger relied upon the answer, purchased his tickets and was later denied the refund he had been led to expect. The British Columbia Civil Resolution Tribunal found Air Canada liable for negligent misrepresentation and awarded Jake Moffatt $650.88 in damages, plus $36.14 in pre-judgment interest and $125 in tribunal fees, for a total of $812.02.
Yet beneath this modest monetary award lies a question of considerable importance to the future of aviation: when an airline delegates communication with its customers to an artificial intelligence system, where does the responsibility for what that system says reside?
The answer given by the Tribunal is both uncomplicated and profound. The machine does not become the legal actor merely because it has become the communicative intermediary. It remains a tool deployed by the airline. If the tool speaks inaccurately to a customer, the airline cannot ordinarily wash its hands of the consequences by pointing to the machine and saying, in effect, “the computer said it.”
The importance of Moffatt therefore lies less in the amount of money awarded than in the jurisprudential principle underlying the award. The decision brings the ancient law of responsibility into contact with a technological phenomenon that is still in its infancy. It asks whether automation changes responsibility. The Tribunal’s answer was that automation does not, by itself, alter the allocation of responsibility.
The Facts
The facts are almost literary in their simplicity. In November 2022, following the death of his grandmother, Moffatt was required to travel from Vancouver to Toronto. He went to Air Canada’s website and consulted its chatbot concerning bereavement fares. The chatbot informed him that if he needed to travel immediately, or had already travelled, he could submit his ticket for a reduced bereavement rate within 90 days of the date on which the ticket had been issued.
That information was wrong.
The airline’s own “Bereavement Travel” webpage contained the correct policy and stated that the policy did not apply to requests for bereavement consideration after travel had been completed. Thus, within the same digital environment, Air Canada was effectively speaking with two voices. One voice was static and formal. The other was interactive and conversational. One gave the correct answer. The other gave the wrong answer. The customer, understandably, was left to assume that the airline’s own chatbot represented the airline.
Moffatt relied on the chatbot’s representation and purchased the tickets. When he subsequently sought the bereavement fare, Air Canada rejected the application. When Moffatt produced a screenshot of the chatbot’s advice, an Air Canada representative acknowledged that the chatbot had used “misleading words” and indicated that the issue had been noted so that the chatbot could be updated.
The Relevant Issue
The legal question was consequently not whether the chatbot was intelligent. Nor was it whether the chatbot possessed artificial consciousness, independent volition or legal personality. It was whether an airline which placed an automated system before the public could be responsible for a misleading representation made by that system.
The Tribunal approached the matter through the well-established Canadian tort of negligent misrepresentation. Drawing upon Queen v. Cognos Inc., (1993), it identified five requirements: a duty of care arising from the relationship between the parties; an untrue, inaccurate or misleading representation; negligence in making the representation; reasonable reliance by the person to whom the representation was made; and resulting damage.
The first element presented little difficulty. The relationship between an airline and its customer is plainly commercial. The airline invites the consumer onto its website to obtain information and purchase transportation. When the airline provides information concerning the conditions under which its services may be purchased, it is difficult to maintain that it has no legal responsibility for the accuracy of that information.
The second element was equally straightforward. The chatbot’s representation was inaccurate. The third was more interesting because negligence requires something more than error. A person or corporation may make an incorrect statement without necessarily being negligent. The law asks whether reasonable care was exercised in the circumstances.
It was here that the Tribunal found Air Canada wanting.
The Tribunal observed that Air Canada had failed to take reasonable care to ensure that its chatbot was accurate. Its reasoning is particularly significant because Air Canada argued that the correct information existed elsewhere on its website. The Tribunal rejected that proposition, asking why the webpage containing the formal policy should be regarded as inherently more trustworthy than the chatbot. More fundamentally, it asked why a customer should be expected to cross-check one part of the airline’s website against another in order to discover which of the airline’s own statements is true.
The Reasoning
The reasoning goes to the heart of digital commerce. The website is not a library in which the customer is required to conduct an archaeological excavation to discover the truth. It is the airline’s principal interface with the consumer. If an airline elects to place an artificial conversational agent at the front door of that interface, it cannot easily disclaim responsibility for the agent’s words while retaining the commercial benefit of the interaction.
The principle is older than artificial intelligence.
Long before generative AI, courts were confronted with computers acting on behalf of businesses. In Century 21 Canada Limited Partnership v. Rogers Communications Inc., (2011) the British Columbia Supreme Court considered the legal consequences of automated computer activity. The court accepted the proposition that an electronic agent or computer program employed by a person is ordinarily a tool of that person, and that liability is not avoided simply because the relevant action has been automated.
The reasoning has an attractive simplicity. A computer does not possess independent legal volition merely because its output appears autonomous. Someone designed it, someone programmed it, someone deployed it, someone supplied it with information, someone determined its function and someone benefited from its operation.
The same proposition appeared decades earlier in the United States in State Farm Mutual Automobile Insurance Co. v. Bockhorst, (1972), where an insurer was held bound by an insurance renewal generated through its pre-programmed computer. The court observed that holding a company responsible for the actions of its computer was not an expression of hostility towards modern business practice. The computer operated according to the information and instructions supplied to it.
This is an important distinction. The law does not have to decide whether a computer “thinks.” It must determine who introduced the computer into the legal and commercial relationship.
This becomes increasingly important as artificial intelligence evolves from deterministic software into systems capable of producing outputs that were not expressly written by a human programmer. The traditional computer followed a rule. The modern AI system may infer, predict, generate, rank and formulate. The output may be unforeseen even by the person who deployed the system.
But unpredictability does not necessarily mean legal independence.
The fact that a system generates an unexpected answer does not mean that the legal consequences of the answer automatically migrate from the corporation to the machine. A corporation cannot ordinarily create an artificial intermediary, give it access to customers, allow it to speak in the corporation’s name and then invoke the intermediary’s autonomy when the conversation becomes inconvenient.
That is the significance of Moffatt.
Aviation Law
The Moffatt reasoning also accords with the broader trajectory of contemporary aviation law. Aviation has never been technologically static. From the radio navigation and computerized reservation systems to automated flight controls, satellite navigation, drones, biometric identification and advanced air mobility, the industry has continuously incorporated technologies capable of performing functions once undertaken by humans. The law has consequently evolved not by creating a separate legal universe for each technological innovation, but by asking who controls the technology, who benefits from it, who is expected to supervise it and who should bear the consequences when it fails.
This is a recurring theme in the legal analysis of artificial intelligence in aviation. Algorithms increasingly constitute the infrastructure through which aviation decisions are made. They influence air traffic management, passenger processing, cybersecurity, predictive maintenance, airline revenue management and customer communications. The technological architecture may be invisible to the passenger, but invisibility does not diminish its legal significance. As I have argued elsewhere in considering the relationship between algorithms and air law, the increasing penetration of algorithmic systems requires aviation law to reconsider how traditional concepts of liability, responsibility and governance apply to systems that can make decisions without immediate human intervention.
Moffatt represents the same movement at the consumer level.
There is, however, a caution which must be entered into the jurisprudential ledger.
The Tribunal’s conclusion that Air Canada was liable is easier to understand than its explanation of precisely why Air Canada was negligent. The record contains remarkably little technical information about the chatbot. The Tribunal noted that Air Canada did not provide information about the nature of the system. We do not know from the decision whether the chatbot was a conventional scripted system, a retrieval system, a machine-learning model or something approaching what would now be called generative AI. Indeed, calling it an “AI chatbot” risks reading today’s technological vocabulary backwards into a system whose architecture was never established in the proceeding.
This matters because negligence is contextual.
Suppose an airline had deployed a highly controlled chatbot whose answers were generated exclusively from an authoritative database, had tested the system extensively, had undertaken periodic audits, had established escalation procedures and had maintained extensive records. Suppose further that an unexpected answer emerged because of a previously unknown interaction between components. Would the airline necessarily have been negligent?
Not necessarily.
An inaccurate answer and a negligent answer are not synonymous.
The law of negligence ordinarily asks whether the defendant exercised reasonable care in the circumstances. Technology cannot be treated as a strict-liability machine in which every error automatically creates liability. The standard must take account of the nature of the system, the foreseeable risks, the seriousness of the consequences, the cost and feasibility of safeguards, the degree of human supervision and the extent to which the enterprise invited reliance.
This is why the technological silence in Moffatt is simultaneously its strength and its weakness.
It is a strength because the decision establishes a technology-neutral principle: an airline cannot evade responsibility merely by inserting a chatbot between itself and its customer.
It is a weakness because future litigation involving genuinely generative systems will require much more sophisticated analysis.
The Challenge
A generative AI system may produce an answer that no programmer specifically intended. The question will then become whether the corporation exercised reasonable care in selecting, configuring, testing, monitoring and supervising that system. Liability may potentially implicate several actors: the airline as deployer, the AI developer, the provider of underlying models, a systems integrator, a vendor responsible for the knowledge base, or even an employee who configured the system improperly.
The customer, however, should not ordinarily be required to untangle that technological supply chain before obtaining a remedy.
This is where aviation law’s traditional concern with allocation of risk becomes relevant. The passenger interacts with the airline. The passenger does not negotiate with the AI developer. The passenger does not know which large language model is being used, which database feeds the chatbot or which contractor wrote the integration code. The airline, by contrast, chooses the technology and decides whether it is sufficiently reliable for deployment.
The law has long recognised that the person who creates or controls a risk may be required to bear its consequences. Artificial intelligence does not abolish that principle; it makes its application more complicated.
The Paradox
There is another important dimension to Moffatt: reasonable reliance.
Air Canada argued, in substance, that Moffatt should have discovered the correct information elsewhere on the website. But the very purpose of a conversational chatbot is to make information retrieval easier. A consumer who asks a direct question of an airline’s customer-service system is not behaving unreasonably by accepting the answer.
Indeed, there is a paradox here. The more sophisticated the chatbot becomes, the more likely consumers may be to rely upon it. A system that speaks fluently, responds instantaneously, understands context and presents itself within an airline’s branded environment creates an impression of institutional authority. Artificial intelligence therefore has a peculiar legal characteristic: its very convenience may increase the reasonableness of reliance upon it.
The more human the machine sounds, the greater may be the expectation that the corporation stands behind what it says.
This creates a difficult question concerning disclaimers. Many AI systems now contain warnings that information may be inaccurate. Such disclaimers may have legal relevance. But their effectiveness cannot be assumed merely because the words exist.
A warning at the bottom of a webpage saying that an AI system “may make mistakes” cannot necessarily neutralise a precise affirmative representation that tells a customer how to exercise a contractual right. There is a substantial difference between a general warning concerning technological fallibility and a specific contractual or commercial representation.
An airline saying “our chatbot may occasionally make errors” is not necessarily equivalent to an airline saying “the answer you have just received is not information on which you may rely.”
The distinction becomes especially important where the airline itself encourages the passenger to use the system as a customer-service channel.
The contractual defence in Moffatt is equally instructive. Air Canada referred to provisions in its Domestic Tariff but did not provide the relevant provisions themselves. The Tribunal consequently found that the airline had not established contractual defense. This is a procedural lesson with technological significance: corporations using AI must preserve not merely the algorithm but the legal environment in which the algorithm operates. Terms of carriage, tariffs, fare rules, customer-service scripts, model instructions and knowledge bases may all become evidence.
In future litigation, the question will not merely be “What did the AI say?” It will be: hat model was being used? what version was deployed? What data informed the answer? what policy was in force? what instructions had been supplied to the system? what safeguards existed? was the system tested? were known error rates documented? were high-risk subjects excluded from automated responses? who had responsibility for monitoring the system? when was the airline first informed that an answer was wrong? How many customers received the same answer?
And, perhaps most importantly, what did the airline do once it knew?
These questions transform AI liability from a philosophical debate into an evidentiary one.
This is where aviation should learn from its safety culture. Aviation safety is not built on the proposition that machines never fail. It is built on the proposition that systems should be designed on the assumption that failures will occur. Redundancy, monitoring, reporting, investigation, corrective action and accountability are embedded in the architecture of aviation safety.
The same philosophy should increasingly govern AI used in aviation.
An airline should not ask whether its chatbot is intelligent enough. It should ask whether the chatbot is safe enough for the function assigned to it.
A chatbot answering “What time does check-in close?” is not in the same risk category as a chatbot answering whether a passenger is entitled to a refund, whether a visa is required, whether dangerous goods may be carried, whether a disability accommodation is available or whether a contractual limitation applies.
Risk-sensitive AI governance
The future of airline liability may involve what could be described as an AI hierarchy of reliance. The greater the economic, legal or human consequence of an answer, the greater the obligation upon the airline to ensure that the answer is grounded in an authoritative source or escalated to a human decision-maker.
There is also a profound conceptual issue concerning the word “agency.” We commonly describe AI as an “agent.” Yet legal agency traditionally presupposes a relationship between a principal and an agent capable of acting within an authority structure. Artificial intelligence does not necessarily possess legal personality merely because it performs an agent-like function.
It may be an artificial agent in a technological sense without being an agent in the jurisprudential sense.
This distinction is critical. Calling a chatbot an “agent” does not transform it into a defendant. Nor does the use of the word “autonomous” mean that the system has become legally independent. Autonomy in computer science is not necessarily autonomy in law.
The law remains concerned with the human and corporate actors standing behind the technology.
That is why the Tribunal’s rejection of the separate-entity argument is perhaps the most enduring feature of Moffatt. The chatbot was not an independent corporation. It had no assets, no contractual capacity, no employees, no shareholders and no legal identity separate from Air Canada. It was an interface through which Air Canada communicated with its customers.
The airline could not simultaneously claim the commercial advantages of the chatbot and disown the chatbot when it became inconvenient.
This principle is unlikely to remain confined to airline customer service. It may eventually apply to automated systems that sell tickets, determine eligibility for refunds, allocate seats, communicate disruption information, process claims and interact with passengers during irregular operations.
Imagine, for example, an AI system informing a passenger that a cancelled flight is eligible for a particular compensation payment when it is not. Or imagine an automated system incorrectly informing a passenger that a connecting flight will be held. Or an AI system providing incorrect information about dangerous goods, medical assistance or documentation requirements.
The financial consequences may be much larger than those in Moffatt. The human consequences may be larger still.
In these circumstances, the law will have to distinguish between errors that are unavoidable and errors that are foreseeable and preventable.
That is the real frontier.
The future case will not ask merely whether AI made a mistake. It will ask whether the airline knew, or ought reasonably to have known, that the system could make that mistake, whether reasonable safeguards existed, and whether the airline nevertheless placed the system in a position where customers could reasonably rely upon it.
This is a familiar negligence inquiry dressed in unfamiliar technological clothing.
There is also an important distinction between the Montreal Convention and the kind of liability encountered in Moffatt. The Montreal Convention of 1999 establishes a specialised liability regime concerning, principally, passenger injury or death, baggage and cargo, and delay. The Moffatt dispute concerned a representation about a fare policy rather than one of the Convention’s traditional heads of carrier liability. It therefore does not mean that every AI error made by an airline automatically becomes a Montreal Convention claim.
That distinction should not, however, obscure the larger point. The Convention operates within an aviation ecosystem increasingly mediated by algorithms. National tort law, contract law, consumer protection law, cybersecurity law and aviation treaty law will inevitably intersect with AI governance.
The future legal architecture of aviation will therefore have to accommodate two propositions at once: technological innovation should not be inhibited merely because it creates new forms of risk, but technological innovation should not become a mechanism for transferring those risks from sophisticated enterprises to unsuspecting passengers.
This is consistent with the broader evolution of aviation law. The Chicago Convention was drafted in an age when aviation was understood principally through the physical aircraft, the pilot, the State and the territorial airspace through which the aircraft moved. The contemporary aviation system is much more complex. Data, algorithms, digital platforms and artificial intelligence now constitute an invisible layer of aviation infrastructure.
The aircraft remains visible.
The algorithm does not.
Yet the invisible algorithm may determine what the passenger pays, what information the passenger receives, how the passenger is processed and, increasingly, what operational decisions are made.
Law must therefore learn to see the invisible.
Moffatt provides one small but important lens through which to do so. Its central lesson is not that airlines are strictly liable for every utterance of every machine. Its lesson is that a corporation cannot escape ordinary principles of responsibility merely because the medium through which it speaks is artificial.
The technology may be new.
The legal principle is not.
And perhaps that is why the decision is more important than its $812.02 award suggests. It tells the aviation industry that digital transformation is not a transformation of accountability. The adoption of artificial intelligence may transform the means by which an airline communicates, but it does not necessarily transform who is responsible for the communication.
My Take
Was the Tribunal’s decision legally and logically acceptable?
On the principal issue, yes.
The conclusion that Air Canada could be held responsible for a misleading statement made through its own chatbot is legally coherent and logically persuasive. The Tribunal did not invent a new species of AI liability. It applied the established Canadian doctrine of negligent misrepresentation to a new factual circumstance. The five-part framework identified in Queen v. Cognos remained intact: duty, misrepresentation, negligence, reasonable reliance and damage.
The Tribunal was also correct in rejecting the proposition that the chatbot somehow constituted a separate legal entity. The chatbot was part of Air Canada’s website and was placed there by Air Canada for the purpose of communicating with its customers. The fact that the communication was automated did not sever the legal relationship between the airline and its customer.
The more difficult question is whether the Tribunal sufficiently demonstrated negligence.
Here I would enter a qualification.
The decision establishes the result more convincingly than it establishes the technological pathway to the result. We are told that Air Canada did not provide information about the nature of the chatbot. We are therefore not given sufficient technical evidence concerning its architecture, training, testing, safeguards, monitoring or failure mechanism.
That omission matters because negligence should not become strict liability merely because the word “AI” has entered the facts.
An airline that deploys an AI system which unexpectedly produces a false answer after having exercised extensive and demonstrable reasonable care should not necessarily be treated in the same manner as an airline that deploys an inadequately tested system, knows that it is producing inaccurate answers and nevertheless leaves it operational.
In Moffatt, however, the circumstances went beyond an isolated inexplicable technological failure. The chatbot contradicted the airline’s own policy, the contradiction appeared within the airline’s own website, the passenger reasonably relied upon the chatbot, and Air Canada subsequently acknowledged that the chatbot had used misleading words. The airline was therefore not dealing merely with an abstract possibility that AI could hallucinate. It was dealing with a concrete inconsistency in information supplied to its customers.
The Tribunal’s conclusion on reasonable reliance is particularly sound. A passenger should not ordinarily be required to act as quality-control officer for an airline’s website. If an airline provides two contradictory answers to the same question, it is difficult to say that the customer acted unreasonably by relying upon the answer delivered through the very channel the airline had provided for interactive assistance.
The case nevertheless should not be read as establishing that every AI-generated statement is automatically binding upon an airline. That would go further than the decision itself. Liability must continue to depend upon the ordinary principles of duty, accuracy, reasonable care, reliance and damage.
The real significance of Moffatt therefore lies in a more modest proposition: automation is not an escape from responsibility.
An airline cannot delegate communication to an artificial system and simultaneously delegate accountability to that system.
That is both legally defensible and logically necessary.
If the opposite rule prevailed, the more technologically sophisticated the corporation became, the less responsibility it might bear for what it tells its customers. Such a rule would produce the perverse result that a company could increase its reliance upon automated systems precisely because automation created an opportunity to disclaim responsibility.
The law should not reward that paradox.
The better principle is that an airline which places an AI system between itself and its customer assumes responsibility for deploying that system with reasonable care, subject always to the particular facts, the nature of the technology, the foreseeability of the error, the reasonableness of reliance and the actual loss suffered.
Moffatt does not make the chatbot a person.
It does something more important.
It keeps the person—and the corporation—behind the chatbot legally visible.
