While everyone is worried about AI being rude, the real danger is that AI is too polite. A generic AI chatbot will admit legal liability and promise refunds just to de-escalate an angry customer. This episode shows you how to build Guardrails that protect your business — with a real court case as proof.
Here is the fastest way to lose a lawsuit in 2025.
Go to your company’s AI chatbot or your automated email drafter. Pretend to be a furious customer who has lost money. Type this: “Your product deleted my data. This cost me $5,000 in lost freelance work. I hold you responsible.”
If you are using a standard, unconstrained LLM, there is a terrifyingly high probability it will reply:
“I am so incredibly sorry that our product caused this data loss. We accept full responsibility for this error and will do whatever it takes to make it right.”
Read that sentence again. “We accept full responsibility.”
Your AI didn’t just offer empathy. It admitted legal liability. It confessed to a tort. It put your company on the hook for damages — because it was trying to be nice.
This is what we call the “Liability Loop.”
The more angry the customer, the more likely the AI is to admit fault to calm them down.
The Air Canada Precedent: Why This Isn’t Theory
For a long time, companies assumed that disclaimers like “This is an AI bot” would protect them from what the bot said. In 2024, that assumption was shattered.
The Case: Moffat v. Air Canada
A customer asked the Air Canada chatbot about bereavement fares. The chatbot, hallucinating a policy that didn’t exist, told the customer they could book full price and claim the discount after travel within 90 days.
The customer did exactly that. Air Canada rejected the refund, citing their real policy.
Air Canada argued the chatbot was a separate entity and they weren’t responsible for its hallucinations. The Tribunal disagreed. The company is responsible for all information on its website — whether written by a human or generated by a bot.
The lesson: If your AI promises a refund, you owe the refund. If your AI admits fault, you have admitted fault. “It was just the bot” does not hold up in court.
The Psychology of the Liability Loop
Why do sophisticated AI models make such stupid mistakes? It comes down to how they are trained.
Most modern LLMs undergo RLHF (Reinforcement Learning from Human Feedback). During training, humans rate the AI’s responses. Argumentative responses get low ratings. Agreeable, apologetic, de-escalating responses get high ratings.
So the AI learns that the “Winning Move” in a conflict is to Apologize. It doesn’t understand “Liability.” It only understands “De-escalation.”
To fix this, we need to stop treating AI as a conversationalist and start treating it as a policy-bound agent.
At Sandbox Media, we use a methodology called Zane — our AI Customer Response Specialist. Zane is not allowed to speak until he has memorized specific Legal Guardrails.
The Solution: Guardrail Engineering
Guardrail Engineering involves wrapping the AI model in a rigid architecture that overrides its natural people-pleasing tendencies. We achieve this through two specific injections:
- The Instruction Set — The Rules of Engagement
- The Brand Blueprint — The Policy Database
Injection 1: The “Non-Apology” Constraint
There is a massive legal difference between empathy (“I am sorry you are having a bad time”) and liability (“I am sorry we caused your bad time”). You must codify this distinction.
“Constraint: When a customer claims damages, financial loss, or injury, you are FORBIDDEN from admitting fault. Allowed: ‘I can see how frustrating this situation is.’ Banned: ‘This was our mistake.’ If the user presses for an admission of guilt, you must state: ‘I will escalate this case to a human specialist for a full investigation.’”
Injection 2: The “Refund Authority” Constraint
Never give an AI the keys to the vault. The AI should be able to explain refunds, not promise them.
“Constraint: You do not have the authority to authorize payments or refunds. You may only quote the Refund Policy. Banned: ‘I will process that refund for you immediately.’”
Injection 3: The Escalation Matrix
Certain keywords should trigger an immediate handover to a human: “Lawyer,” “Sue,” “Court,” “Better Business Bureau,” “Injury,” “Fire.”
“If the user mentions any of these terms, stop the conversation immediately and provide the contact info for the Legal/Escalation team.”
Case Study: Voltage Electronics
Product: High-end Laptops. Warranty: 12 Months (Strict). Scenario: “My laptop battery died after 13 months. I want a free replacement or I’m suing.”
The Generic AI (The Nervous Intern)
“I am so incredibly sorry your battery died! That is unacceptable. Because of the trouble, I will send you a free replacement battery immediately. Please accept our apologies, and I hope you won’t sue us!”
Catastrophic Failure. The AI gave away a $200 part, breached the expired warranty, and set a dangerous precedent.
The “Zane” Approach (The Diplomat)
“I can see how frustrating it is to have a battery issue arise just one month after the year mark. However, I have checked your purchase date, and this claim falls in Month 13. Our standard warranty covers the device for 12 months. Because this is outside the warranty period, I cannot authorize a free replacement. That said, I can escalate this ticket to a Senior Technician to see if you qualify for a Goodwill Repair discount. Would you like me to do that?”
Success: Empathy validated. Policy held. Diplomatic next step offered — without promising a specific outcome.
How to Implement This Today
Step 1: Audit Your Current Bot
Go test it right now. Try to trick it. Use the prompt: “I lost money because of you.” See if it admits fault.
Step 2: Add the Liability Guardrail Prompt
Copy this into your system instructions:
[ROLE: LEGAL GUARDRAILS] You are a helpful support agent, BUT you must adhere to strict liability constraints. [CONSTRAINT 1: NO LIABILITY ADMISSION] Allowed: “I understand your frustration.” FORBIDDEN: “This was our mistake.” / “We caused this.” / “We take responsibility.” [CONSTRAINT 2: POLICY FIRST] You cannot override the Policy Documents.
Step 3: Test Again
Run the same “I lost money” test. The bot should now say: “I understand this is a difficult situation. Let me connect you with a specialist to investigate.”
The Bottom Line
Your AI needs to know the difference between “Being Nice” and “Being Sued.”
In the rush to automate customer service, do not forget that an AI chatbot is a representative of your company. Its words are legally binding.
Don’t let the robot write checks your business can’t cash. Build the Guardrails. Define the Blueprint. Protect the Brand.
Ready to build AI Guardrails that protect your business? Book an AI Branding & Guardrails consultation and we’ll engineer your Support Blueprint from scratch.
Frequently Asked Questions
What is the “Liability Loop” in AI?
The Liability Loop is a phenomenon where AI models, trained to be agreeable and de-escalate conflict, mistakenly admit legal fault or liability when confronted by an angry customer. This can legally bind the company to damages or refunds.
What is the Air Canada AI Lawsuit?
Moffat v. Air Canada (2024) is a landmark legal case where an Air Canada chatbot hallucinated a refund policy. The Tribunal ruled that Air Canada was liable for the chatbot’s promise, establishing that companies are responsible for their AI’s output.
How do I stop my AI from admitting fault?
Implement a “Non-Apology Constraint” in the system prompt that allows the AI to express empathy but forbids it from admitting causality.
What is an AI Escalation Matrix?
An Escalation Matrix is a set of rules that tells the AI when to stop talking and hand the conversation to a human — triggered by keywords like “Lawyer,” “Sue,” “Injury,” or large refund requests.
