Let’s talk about a wild AI chatbot case with Air Canada that could’ve been straight out of a sitcom (but with legal consequences). A customer, seeking a bereavement fare, relied on their chatbot’s promise that they could apply for a refund within 90 days. Turns out, that policy doesn’t exist. Classic case of AI hallucination—except it’s not funny when a refund is on the line. Here’s where AWS’s automated reasoning tools could’ve been helpful. Automated reasoning ensures chatbots don’t just make things up by cross-checking responses with actual policies. Imagine if Air Canada’s chatbot had been equipped with these tools: it would’ve said something like, “Sorry, bereavement fares must be arranged before travel” instead of leading the customer down a path of frustration and legal drama. Instead, Air Canada argued the chatbot was its own legal entity—yes, really. Spoiler: the tribunal didn’t buy it. The company was held accountable for the AI’s output on its own site. The takeaway? If you’re building AI tools, don’t skip the safeguards. Tech like AWS’s contextual grounding and automated reasoning isn’t just about avoiding PR nightmares—it’s about creating AI that’s trustworthy, user-friendly, and, most importantly, doesn’t lie to your customers. And yes this isn’t the only way to solve for this issue. Vertex AI from Google has made progress here with toolchaik approaches. Microsoft Corrections addresses this with less emphasis on automated reasoning. I think AWS is taking a smart approach here. AI might be smart, but without the right guardrails, it can turn into a liability faster than you can say “refund denied.” #product #productmanager #productmanagement #startup #business #openai #llm #ai #microsoft #google #gemini #anthropic #claude #llama #meta #nvidia #career #careeradvice #mentor #mentorship #mentortiktok #mentortok #careertok #job #jobadvice #future #2024 #story #news #dev #coding #code #engineering #engineer #coder #sales #cs #marketing #agent #work #workflow #smart #thinking #strategy #cool #real #jobtips #hack #hacks #tip #tips #tech #techtok #techtiktok #openaidevday #aiupdates #techtrends #voiceAI #developerlife #aircanada #awsreinvent #reinvent #aws #amazon #bedrock #guardrails

Do you remember the Air Canada chatbot massive debacle and lawsuit last year? If you recall, the passenger was trying to travel on Air Canada to a funeral for a loved one and asked about the bereavement policy on Air Canada, but did so using a chatbot, and the chatbot responded with an entirely made-up bereavement policy. The passenger tried to fly on that policy, Air Canada refused, and then the passenger sued them and successfully won because, as it turns out, you can't disclaim responsibility for the chatbot on your own website, which is what Air Canada's legal team tried to do. This could all have been avoided if Air Canada had been more thoughtful about the construction of their chatbot, but until recently, it's been difficult to actually get hallucinations fully out of chatbots for enterprise use cases, and AWS made a really big step in that direction today when they announced automated reasoning in Las Vegas at the AWS reInvent conference. And what that is, is it takes a grounding document, like a bereavement policy, and it extracts from it logical statements, if this, then this, and it constructs its own set of ground truth rules based on that policy document that you can review, understand, assess, edit, etc. And then, from that grounding, it checks the responses that a large language model gives in real time to make sure that the large language model is not going to just go off the deep end and produce a policy that actually isn't there. And so, it used to be that when you wanted to do policy checks, let's say you had shipping rules, if this, then this, for a bunch of different items in a cart that have to be shipped in a package, that's a big automated reasoning problem. And you need to get that correct or else your shipping costs go through the roof and you charge customers more than you should or too little, etc. Automated reasoning is a huge step forward when it comes to LLMs, because it means that in that situation, you could talk to a large language model, get automated reasoning to assess the correct answer and check it, and that large language model would be able to come back with a checked verified response on a question like a bereavement policy or a shipment cost for a returned item. So, I think it's super applicable. It's an example of the kind of unglamorous, but really important AI development I want to call out. There are tool chain options for this. So, for developers in my audience, yes, there are definitely ways you can use tool chain approaches. I know Gemini has been working on this with their Vertex AI to get better at sort of how you handle long tool chains with responses in business critical scenarios. So, there's other options out there, but I love the elegance of automating the reasoning portion and making it sort of a callable service that you can integrate right in. So, there you go, AWS reInvent, been watching for cool stuff, and I think automated reasoning is definitely something to keep an eye on. No, this is not the same thing as the reasoning I talked about in the Apple paper yesterday, where it was like, do large language models reason? Yes or no? I said they do. I still think they do, but in this case, it's a very technical definition of reasoning. It is, what's a great example? The cat is wet. The cat lies on the couch. The couch is wet. You can then infer the cat made the couch wet. That's sort of an example of a chain of reasoning, and that's what we're looking to extract from these complex policy documents for companies and automate checks on for large language models. So, automated reasoning, I think it's a good thing. We need more of this so we have less hallucinations and less air candidate disasters.