And if a random AC employee said[0] they'd give you a billion dollars, you wouldn't be entitled to it because any judge or jury hearing the case would say a reasonable person wouldn't believe that. Unlike computers, which do exactly what they're told[1], the legal system applies sanity checks and social context.
[0] perhaps because they're disgruntled and trying to hurt their employer
[1] generative models are not an exception; they're a way of telling computers to generate text that sometimes contains falsehoods
> And if a random AC employee said[0] they'd give you a billion dollars, you wouldn't be entitled to it because any judge or jury hearing the case would say a reasonable person wouldn't believe that.
I'm sure that if the bot had said that the airline would raise your dead relative from the grave and make you king of the sky or something equally unbelievable the courts wouldn't have insisted Air Canada cast a crown and learn necromancy.
>If a random AC employee gave you a free flight, on the other hand, you'd be entitled to it.
The company would be entirely within their rights to say 'this employee was wrong, that is not our policy, goodbye!'. This happens all the time with more minor incidents.
No idea about the US but this very same case was tested in France and some part of Germany in the late 90s when some PayTVs (Sky or Canal+, can't remember) tried to cancel multiple subscriptions offered with an extremely aggressive pricing by some of their agents. Courts concluded that the signed agreements superseded the official pricing and they had to offer the service for the entire length of the original subscription.
Where I live, "meeting of the minds" is necessary for a contract. Written or not. In this case, that meeting didn't happen. Due to the bullshit generator employed by Air Canada.
So there was no contract but a consumed flight. The court has to retroactively figure out a reasonable contract in such cases. That Air Canada couldn't just apply the reduced rate once they learned of their wrong communication marks them as incredibly petty.
That's far less likely to be true if the customer buys something based on the employee's erroneous statement. I suspect in an otherwise-identical case with a human customer service agent, the same judge would have ruled Air Canada must honor the offer.
A random AC employee who is having a bad day and hates his employer says "You are entitled to free flights for the rest of your life." Is Air Canada liable?
No, because that's not "reasonable". My dad jokes that he's made a career off of determining what is "reasonable" and what isn't, and he's a contract attorney.
If you were standing at the customer service desk, and instead they said: "sorry about the delay, your next two flights are free", then all of a sudden this is "reasonable".
No, because no reasonable person would think that they had the authority to authorize that. Remember, the legal system is not computer code - judges look at things like intent and plausibility.
>How is this different from me getting one of my friends to work at Air Canada and promise me a billion dollars to cancel my flight?
There is a common misconception about law that software engineers have. Code is not law. Law is not code. Just because something that looks like a function exists, you can't just plug in any inputs and expect it to have a consistent outcome.
The difference between these two cases is that even if a chat bot promised that, the judge would throw it out, because it's not reasonable. Also, the firm would have a great case against at least the CS rep for this collusion.
If your friend of a CS agent promised you a bereavement refund (As the chatbot did), even though it went against company policy, you'd have good odds of winning that case. Because the judge would find it reasonable of you to believe and expect that after speaking to a CS rep, that such a policy would actually get honored. (And the worst that would happen to the CS rep would be termination.)
Law and justice is not like a computer program that you can exploit and control without limits by being a hacker.
If the chatbot told them that they'd get a billion dollars, the courts would not hold Air Canada responsible for it, just as if a programmer put a decimal wrong and prices became obviously wrong. In this case, the chat bot gave a policy within reason and the court awarded the passenger what the bot had promised, which is a completely correct judgement.
This argument seems overly dramatic and distorted. Yes, in an outrageous situation like a billion-dollar mishap, most people would know something isn't right. But for a policy that appears legitimate, especially when it's replacing a human customer service rep, it's not that obvious. In these cases, Air Canada should definitely be held accountable.
These ones do what they "learned" from a lot of input data using a process that is us mimicking how we think brains could maybe function (kinda/sort off with a few unbiological "improvements").
Say your webserver isn't scaling to more than 500 concurrent users. When you add more load, connections start dropping.
Is it because someone programmed a max_number_of_concurrent_users variable and a throttleExtraAboveThresholdRequests() function?
No.
Yes, humans built the entire stack of the system. Yes every part of it was "programmed", but no this behaviour wasn't programmed intentionally, it is an emergent property arising from system constraints.
Maybe the database connection pool is maxed out and the connections are saturating. Maybe some database configuration setting is too small or the server has too few file handles - whatever.
Whatever the root cause (even though that cause incidentally was implemented by a human if you trace the causal chain back far enough) this behaviour is an almost incidental unintended side effect of that.
A machine learning system is like that, but more so.
An LLM, say, is "parsing" language in some sense, but ascribing what it is doing to human design is pretty indirect.
In a way you typing words at me has in some way been "programmed" into you by every language interaction mankind has had with you.
I guess you could see it that way, but I don't think it's a particularly useful point of view.
In the same way an LLM has been in directly "programmed" via it's model architecture, training algorithm and training data, but we are nowhere near the understanding of the process to be able to consider this "programming" it yet.
This is different from a bug or hitting an unknown limitation—the selling point of this was “it makes shit up” and they went “yeah, cool, let’s have it speak for us”.
Its behavior incorporates randomness and is unpredictable and hard to keep within bounds on purpose and they decided to tell a computer to follow that unpredictable instruction set and place it in a position of speaking for the company, without a human in between. They shouldn’t have done that if they didn’t want to end up in this sort of position.
We agree that this is an engineering failure - you can't deploy an LLM like this without guardrails.
This is also a management failure in badly evaluating and managing the risks of a new technology.
We disagree in that I don't think that its behaviour being hard to predict is on purpose: we have a new technology that shows great promise as tool to work with language input and outputs. People are trying to use LLMs as general purpose language processing machines - in this case as chat agents.
I'm reacting to your comment specifically because I think you are evaluating LLMs using a mental model derived from normal software failures and LLMs or ML models in general are different enough to make that model ineffective.
I almost fully agree with your last comment, but the
> they decided to tell a computer to follow that unpredictable instruction set
reflects what I think is now an unfruitful model.
Before deploying a model like this you need safeguards in place to contain the unpredictability. Steps like the following would have been options:
* Fine-tuning the model to be more robust over their expected input domain,
* Using some RAG scheme to ground the outputs over some set of ground truths,
* Using more models to evaluate the output for deviations,
* Business processes to deal with evaluations and exceptions,
Etc
No, if you conspire with your friend to get them to tell you an incorrect policy, then you have no reasonable expectation that what they tell you is the real policy. If you are promised a billion dollars even without a pre-existing relationship with the agent, you have no reasonable expectation that what they are promising is the real policy because it's an unbelievably large amount.
If you are promised something reasonable by an agent of the company who you are not conspiring with, then the company is bound to follow through on the promise because you do have a reasonable expectation that what they are telling you is the real policy.
The chatbot instructed the passenger to pay full price for a ticket but stated they could get a refund later. That refund policy was a hallucination. The victim her just walked away with a discounted ticket as promised not a billion dollars.
Because that would not be reasonable, and nobody would be surprised if Air Canada reneged on that. See, for instance, Leonard vs. PepsiCo.
If your friend promised you something reasonable in the course of carrying out their duties, and you honestly believed them, I think that would be legal and enforceable just as this case suggests.
> How is this different from me getting one of my friends to work at Air Canada
One major difference is the AI wasn't your friend, another is that you didn't get it hired at Air Canada, another is that the promise wasn't $1B, etc...
Your friend is not trained by Air Canada. The bot is Air Canada property.
If they decide it is reliable enough to be put in front of the customer, they must accept all the consequences: the benefits like having to hire less, and the cons, which is that they have to make it work correctly.
Otherwise, woopsy, we made our AI handle our accounting and it cheated, sorry IRS. That won't fly.
No, it is more similar to Air Canada hiring a monkey to push buttons to handle customer complaints. In that case, the company knows (or should know) that the given information may be wrong, but accepts the risk.
People working for companies are sometimes separate legal entities responsible for their own actions, and sometimes they act on behalf of the company they work for and the company is responsible for their actions. It depends.
A chatbot(computer) cannot be responsible for their own actions so the only half of the coin you have left is "the company is responsible for their actions."
Computers are inanimate objects and are not recognized as legal entities.
The legal system recognizes that people, or groups of people, are subject to legal authority. This is a story about a piece of software Air Canada implemented which resulted in them posting erroneous information on their website.
So replacing all their customer support staff with AI that misleads customers is OK? That's pants on head insane, so why spend time trying to justify it.
If a random AC employee gave you a free flight, on the other hand, you'd be entitled to it.
Anyway, the chat bot has no agency except that given to it by AC; unlike a human employee, therefore, its actions are 100% AC actions.
I don't see how this is controversial? Why do people think that laws no longer apply when fancy high-tech pixie dust is sprinkled?