Two tribunals have now said the machine is you. The third question was never put.
The short answer: an organization that invites an automated intelligence to speak or act inside its operating environment does not get to disown it afterward. Two tribunals have said so. What neither of them reached is the harder version of the question, and the case that would have reached it ended before it got there.
Air Canada: was the statement accurate?
In February 2024 the British Columbia Civil Resolution Tribunal decided Moffatt v. Air Canada. A passenger asked the airline’s website chatbot about bereavement fares and was told he could buy a regular ticket and apply for the reduced fare afterward. The airline’s actual policy did not allow that.
Air Canada’s position was, in substance, that it should not be answerable for what the chatbot said. The adjudicator described the argument as suggesting the chatbot was a separate legal entity responsible for its own actions, and rejected it. The chatbot was part of Air Canada’s website. The passenger had no way of knowing that one page was more trustworthy than another. The airline paid $812. Within a week, a dispute over a refund of less than two thousand dollars was national and international news.
That is the first rung, and it is the easy one. The system said something untrue. The company owned it.
Aesthetify: was the claim verifiable?
On 12 May 2026 the Higher Regional Court of Hamm decided a case brought by the consumer association of North Rhine-Westphalia against Aesthetify GmbH, a German aesthetic medicine business.
Its website chatbot told people that the two doctors behind the business were specialist physicians in plastic and aesthetic surgery. Some of those titles they did not hold. Some do not exist in the German medical training regulations at all. The consumer association demanded that the company sign a binding promise, backed by a financial penalty, not to repeat the claims. The company switched the chatbot off — but would not sign. So the matter went to court.
The court held that the chatbot’s answers were the company’s own commercial conduct. A chatbot is not a third party whose misbehavior relieves the business that deployed it. The court also rejected the argument that people know AI makes mistakes and check what it tells them anyway — a generic warning that the system can be wrong is not a defense.
And here is the part worth stopping on. The ruling holds the company liable even though the system had been fed exclusively correct data. Correct inputs. Wrong output. No defense.
That is the second rung. The system generated a claim about the external world that nobody had checked against the body that actually issues the credential. The data going in was clean. Nothing in the process asked whether the sentence coming out was true.
An appeal was permitted, on the ground that attributing generative-AI statements raises a novel question. The principle is not yet final.
The third case: was accurate information used legitimately?
In February 2025 the American Alliance for Equal Rights filed suit in the Northern District of Texas against American Airlines and Qurium Solutions, doing business as Supplier.io. The complaint alleged that the airline’s supplier-diversity program excluded contractors on the basis of race, in violation of §1981. Supplier.io was named as a co-defendant for its role in operating the program.
On 16 May 2025 the parties settled. Judge Mark Pittman signed a final judgment dismissing the case with prejudice. The defendants paid the plaintiff’s legal fees. Under the settlement, American agreed it would not require or consider whether a business is owned by people of any particular race or ethnicity, and would say so on its website.
I take no position here on supplier diversity as policy. What matters for this argument is structural, and it is this.
In the first two cases the machine said something false. In this one, nothing about the data needed to be wrong. Every supplier identity could have been verified. Every certification could have been current and correctly matched to the right legal entity. The dispute was never about whether the attribute was accurate. It was about whether using it that way was defensible — and if it was not, who was answerable for the use.
The complaint quotes a term in which American agreed to indemnify, defend and hold Supplier.io harmless. Responsibility had been allocated by contract before anyone asked whether it could be.
That question never got an answer. The case settled before discovery reached it. There is no ruling on how a verified classification becomes an eligibility criterion, or on where the data provider’s responsibility ends and the client’s begins. The practice changed. The question stayed open.
It is still open today, and every organization running verified attributes through an automated selection process is operating inside it.
Before, during, after
Strip the three cases down and the same thing is missing in each.
Before. Nothing established that the statement about to be made was checkable against an authority — the fare rules, the medical register, the policy under which a criterion is lawful. A claim about the external world is not something a model is entitled to settle. It has to be checked against whoever actually holds the answer.
During. Nothing held a conflict open long enough for anyone to see it. In each case the output went straight to the person it affected. There was no point at which a disagreement forced escalation instead of delivery. In all three cases, the first person to inspect what the system decided was the customer — the passenger, the website visitor, the excluded contractor.
After. Nothing could be reconstructed. Who introduced the criterion. How it was weighted. What was challenged and by whom. Who approved it and accepted the consequence. Air Canada could not show it. Aesthetify could not show it. And in the third case nobody was ever required to try.
An organization that cannot answer those three questions has not deployed an AI system. It has published one and hoped.
A guardrail is a rule the system has to remember. A firm fence is a path the architecture does not let it leave.
A correction to my own record
In March 2025 I published a piece asking whether a different kind of black box would bring Supplier.io down. One sentence in it was wrong.
I wrote that intent is not required and that disparate impact might suffice. Under §1981 it does not. The claim requires intentional discrimination and but-for causation, and the complaint in this very case pleaded intentional exclusion, not impact. The reasoning built on that sentence was accordingly too confident about where liability might land.
That post stands as published. Nothing in this archive is edited after the fact. The correction travels forward, dated, which is the only way a nineteen-year record stays worth reading. An archive in which every claim survives two decades intact is not a contemporaneous record. It is polished recall.
The prediction in that post was that the black box question would be decided. It was not. It was settled around.
One question
Not for me. For you.
Your organization has an automated system speaking or acting in front of someone right now — a customer, a supplier, a candidate. If it says something today that you have to defend in eighteen months, what in your process would let you reconstruct how it got there, and who authorized it?
In our upcoming lab, we will introduce the dialogue model behind ARA™ RAM 2025™, powered by SLAP OS™, and show how dialogue-driven models governing the reasoning process can surface that question before the answer travels.
Friday, September 25 at 9:30 AM Eastern. The lab is free, and everyone who attends will receive a copy of Thinking With the Machine.
Register here: https://www.linkedin.com/events/7504567838724997120/
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Related
Air Canada, Aesthetify, and the Case That Settled Before Anyone Asked
Posted on September 21, 2026
0
Two tribunals have now said the machine is you. The third question was never put.
The short answer: an organization that invites an automated intelligence to speak or act inside its operating environment does not get to disown it afterward. Two tribunals have said so. What neither of them reached is the harder version of the question, and the case that would have reached it ended before it got there.
Air Canada: was the statement accurate?
In February 2024 the British Columbia Civil Resolution Tribunal decided Moffatt v. Air Canada. A passenger asked the airline’s website chatbot about bereavement fares and was told he could buy a regular ticket and apply for the reduced fare afterward. The airline’s actual policy did not allow that.
Air Canada’s position was, in substance, that it should not be answerable for what the chatbot said. The adjudicator described the argument as suggesting the chatbot was a separate legal entity responsible for its own actions, and rejected it. The chatbot was part of Air Canada’s website. The passenger had no way of knowing that one page was more trustworthy than another. The airline paid $812. Within a week, a dispute over a refund of less than two thousand dollars was national and international news.
That is the first rung, and it is the easy one. The system said something untrue. The company owned it.
Aesthetify: was the claim verifiable?
On 12 May 2026 the Higher Regional Court of Hamm decided a case brought by the consumer association of North Rhine-Westphalia against Aesthetify GmbH, a German aesthetic medicine business.
Its website chatbot told people that the two doctors behind the business were specialist physicians in plastic and aesthetic surgery. Some of those titles they did not hold. Some do not exist in the German medical training regulations at all. The consumer association demanded that the company sign a binding promise, backed by a financial penalty, not to repeat the claims. The company switched the chatbot off — but would not sign. So the matter went to court.
The court held that the chatbot’s answers were the company’s own commercial conduct. A chatbot is not a third party whose misbehavior relieves the business that deployed it. The court also rejected the argument that people know AI makes mistakes and check what it tells them anyway — a generic warning that the system can be wrong is not a defense.
And here is the part worth stopping on. The ruling holds the company liable even though the system had been fed exclusively correct data. Correct inputs. Wrong output. No defense.
That is the second rung. The system generated a claim about the external world that nobody had checked against the body that actually issues the credential. The data going in was clean. Nothing in the process asked whether the sentence coming out was true.
An appeal was permitted, on the ground that attributing generative-AI statements raises a novel question. The principle is not yet final.
The third case: was accurate information used legitimately?
In February 2025 the American Alliance for Equal Rights filed suit in the Northern District of Texas against American Airlines and Qurium Solutions, doing business as Supplier.io. The complaint alleged that the airline’s supplier-diversity program excluded contractors on the basis of race, in violation of §1981. Supplier.io was named as a co-defendant for its role in operating the program.
On 16 May 2025 the parties settled. Judge Mark Pittman signed a final judgment dismissing the case with prejudice. The defendants paid the plaintiff’s legal fees. Under the settlement, American agreed it would not require or consider whether a business is owned by people of any particular race or ethnicity, and would say so on its website.
I take no position here on supplier diversity as policy. What matters for this argument is structural, and it is this.
In the first two cases the machine said something false. In this one, nothing about the data needed to be wrong. Every supplier identity could have been verified. Every certification could have been current and correctly matched to the right legal entity. The dispute was never about whether the attribute was accurate. It was about whether using it that way was defensible — and if it was not, who was answerable for the use.
The complaint quotes a term in which American agreed to indemnify, defend and hold Supplier.io harmless. Responsibility had been allocated by contract before anyone asked whether it could be.
That question never got an answer. The case settled before discovery reached it. There is no ruling on how a verified classification becomes an eligibility criterion, or on where the data provider’s responsibility ends and the client’s begins. The practice changed. The question stayed open.
It is still open today, and every organization running verified attributes through an automated selection process is operating inside it.
Before, during, after
Strip the three cases down and the same thing is missing in each.
Before. Nothing established that the statement about to be made was checkable against an authority — the fare rules, the medical register, the policy under which a criterion is lawful. A claim about the external world is not something a model is entitled to settle. It has to be checked against whoever actually holds the answer.
During. Nothing held a conflict open long enough for anyone to see it. In each case the output went straight to the person it affected. There was no point at which a disagreement forced escalation instead of delivery. In all three cases, the first person to inspect what the system decided was the customer — the passenger, the website visitor, the excluded contractor.
After. Nothing could be reconstructed. Who introduced the criterion. How it was weighted. What was challenged and by whom. Who approved it and accepted the consequence. Air Canada could not show it. Aesthetify could not show it. And in the third case nobody was ever required to try.
An organization that cannot answer those three questions has not deployed an AI system. It has published one and hoped.
A correction to my own record
In March 2025 I published a piece asking whether a different kind of black box would bring Supplier.io down. One sentence in it was wrong.
I wrote that intent is not required and that disparate impact might suffice. Under §1981 it does not. The claim requires intentional discrimination and but-for causation, and the complaint in this very case pleaded intentional exclusion, not impact. The reasoning built on that sentence was accordingly too confident about where liability might land.
That post stands as published. Nothing in this archive is edited after the fact. The correction travels forward, dated, which is the only way a nineteen-year record stays worth reading. An archive in which every claim survives two decades intact is not a contemporaneous record. It is polished recall.
The prediction in that post was that the black box question would be decided. It was not. It was settled around.
One question
Not for me. For you.
Your organization has an automated system speaking or acting in front of someone right now — a customer, a supplier, a candidate. If it says something today that you have to defend in eighteen months, what in your process would let you reconstruct how it got there, and who authorized it?
In our upcoming lab, we will introduce the dialogue model behind ARA™ RAM 2025™, powered by SLAP OS™, and show how dialogue-driven models governing the reasoning process can surface that question before the answer travels.
Friday, September 25 at 9:30 AM Eastern. The lab is free, and everyone who attends will receive a copy of Thinking With the Machine.
Register here: https://www.linkedin.com/events/7504567838724997120/
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Share this:
Like this:
Related