In 2022, Jake Moffat sadly lost his grandmother as she died on Remembrance Day. He went to do what any grieving grandson would, he went online the same day to book travel arrangements. But, after interacting with a support chatbot; Moffat found himself in a court case with Air Canada.
Jake provided the court screenshots of the conversation where the bot said, “If you need to travel immediately or have already travelled and would like to submit your ticket for a reduced bereavement rate, kindly do so within 90 days of the date your ticket was issued by completing our Ticket Refund Application form.” And Air Canada’s defence? They can’t be liable for information provided by their “agents, servants or representatives – including a chatbot” and even claimed it was “a separate legal entity that is responsible for its own actions.” Yeah… Air Canada paid Jake $812 to cover the difference between the full-price and bereavement rates (Proctor, 2024).
The Reality of AI Hallucinations
How many times have you heard that you needn’t be confident, just act confidently? Turns out AI is very human in that regard and spouts falsehoods more than 1 in 3 times according to a report done by US news rating company Newsguard (Desmarais, 2025). There’re 3 sobering takeaways from that report:
- Almost all AI models tested lied more from 2024 to 2025.
- Even with insufficient information, AI will confidently fabricate a response like an under pressure Jay Cartwright at a party.
- This report included ChatGPT, Copilot and Claude; the ones likely to be underlying your AI solution.
Isabel Allende must have been delighted when she featured in The Chicago Sun-Times and Philadelphia Inquirer’s reading list this summer. Well, it would have been until she noticed it was for the title ‘Tidewater Dreams’, a non-existent AI hallucination (Mihalopoulos, 2025).
Mitigating Algorithmic Bias
The lying isn’t even the worst of it. A lack of information is a seemingly easy fix, but what is providing the information becomes the problem? This is a common headache in the world of hiring.
In the US, AI hiring decisions have been the most prominently litigated. The University of Washington provided 3 AI models with identical CVs in all respects other than the name. The AI treated them all fairly with no preference… I lied, the models preferred white-associated names 85% of the time and Black-associated names only 9%, plus male names were selected more than female ones. These kinds of biases have meant Workday, iTutorGroup and HireVue AI hiring tools have all been taken to court; HireVue even fed back to a deaf job applicant that she needed to ‘practice active listening’ post AI conducted video interview (emanuel, 2025).
The law can get off its high horse though, historical biases meant that training data was unrepresentative when creating facial recognition software. A study on the US’ law enforcement image database showed that people of African descent were more likely to be wrongly singled out.
Historic systemic racism meant that people of African descent were overrepresented in police photograph databases. The self-reinforcing nature of machine learning is making predictions about the future on the basis of data from the past (K.P., 2024).
Understanding Compliance and Fines With AI Regulations
This may seem like an issue for the big fishes but you’re not too small to get caught. The EU have implemented the ‘AI Act’, which prohibits certain AI systems. That kicked off back in August 2024, and the initial grace period ended in February 2025 that banned certain practices and required staff AI literacy. The Full Monty comes in August 2026 where ‘high-risk’ AI systems come under scrutiny (High-level summary of the AI Act, 2024).
There is no distinction between a one-man band and an enterprise with a £1 billion a year revenue. All must meet the same standards. Your AI solution that is driving data decisions to streamline departments (aka redundancies)? That’s in the high-risk category exposing you to fines up to €35 million or 7% of your worldwide annual turnover. Cartify, that AI chatbot citing company policies on your e-commerce page? That’s limited-risk but you need to make it clear it’s a chatbot not a human. Get that wrong and a bill of €15 million or 3% of turnover will be in the post.
Even if you weren’t that one who built it you’re the one liable if you’ve deployed it. Not the vendor, not the “separate legal entity” AI, you are.
Not based in the EU? If your company trades at all in the EU borders you must comply. And whilst the UK are yet to have a formal act, it is a widely debated subject in Parliament.
A Framework for Ethical AI
Implementing ISO Principles
Using the principles outlined by ISO (the International Organisation for Standardisation) the foundations of building ethical AI can be laid:
- Fairness – Datasets used in training your AI need to be checked to avoid any future bias.
- Transparency – The derivation of AI decisions needs to be understood by the users.
- Non-maleficence – AI systems should not harm individuals, groups or the environment.
- Accountability – Developers, organisations and policymakers are responsible for developing principled AI.
- Privacy – Personal data must be protected; individuals should be able to control how their data is collected and used.
- Robustness – They must be resilient both to output errors and unexpected inputs.
- Inclusiveness – Everyone should have a chance to voice concerns and efforts made to address them.
Which all sounds like common sense, but fairness means you are proving that training data is scrutinised not just at conception but continuously. Transparency means you can explain why your AI rejected an applicant for a loan and not shrug it off saying it was the AI’s decision. Accountability means that if your AI says your returns policy is 90 days and not the true 30 days, then guess who has to honour that return on day 80.
Driving Cultural Adoption and Employee Trust
The building of the framework above is only part of the story. To actually make it meaningful it must become an organisation-wide initiative not just something imposed upon employees from the ivory tower. Collaboration and conversation between departments to get the full spectrum of technological and business knowledge is key to ensure nuanced perspectives. An education plan needs to be created and delivered to ensure best practices are maintained whilst using the systems. And there needs to be full transparency in the full AI process, stakeholders need to be able to trust the AI solution and understand how and why it is being used. Employees are scared that their livelihoods are at risk from AI, so for the best chance of adoption they need to be part of it not just a consideration.
The “Human-in-the-Loop” Necessity
Once you have your AI pillars in place and have the troops on board, you need to build sustainably. The data it’s trained on needs scrutinising well before it can even smell a model. Collection methods, data cleansing and sampling are not your IT department’s concern, they’re a critical foundation. Once in training use multiple different metrics to assess overall performance: false positives, false negatives and feedback from a diverse group of users. Understanding the limitations of what the model should and shouldn’t be used for, as outlined above, AI is confident in the face of uncertainty. Knowing where the line is where AI starts guessing is the difference between useful tool and a liability.
But it’s never actually over. Even when built, tested and deployed the job continues. Regular feedback using ‘human in the loop’ mechanisms continually assess the AI outcomes. This ensures that as the world changes around it, the AI’s system performance degradation doesn’t go unnoticed. The training data that once was representative of the real-world scenarios the AI would face may become outdated. Unless you’re monitoring and realigning you become the next Workday.
This constant need for oversight is exactly why deploying AI is a full-on risk management strategy, far beyond the scope of a standard IT project. Navigating that risk without a clear roadmap is a fast track way to regulatory fines.
How Triangle Can Help You Build a Safe AI Strategy
AI isn’t going away. The once rose-tinted view has faded. It was once Harry Potter level sorcery, but with rigorous use it’s less Potter and more Longbottom. But that doesn’t mean it has become less useful, it means that we were too quick to let it into the wild.
The companies getting it right and reaping the benefits are taking aim before firing. They carefully considered the use cases, the potential risk that the AI could pose and taking full responsibility for its actions.
If you need help implementing your AI strategy to avoid stepping on something you shouldn’t, that’s what Triangle is here for. With years of data experience and deep knowledge of the business challenge in adopting technology we can help you toward a successful future in the AI space.
Bibliography
Building a responsible AI: How to manage the AI ethics debate. (n.d.). Retrieved from ISO: https://www.iso.org/artificial-intelligence/responsible-ai-ethics#:~:text=Responsible%20AI%20is%20an%20approach,issues%20such%20as%20AI%20bias.
Desmarais, A. (2025, September 5). Which AI chatbot spews the most false information? 1 in 3 AI answers are false, study says. Retrieved from Euronews: https://www.euronews.com/next/2025/09/05/which-ai-chatbot-spews-the-most-false-information-1-in-3-ai-answers-are-false-study-says
emanuel, q. (2025, August 18). Lead Article: When Machines Discriminate: The Rise of AI Bias Lawsuits. Retrieved from quinn emanuel: https://www.quinnemanuel.com/the-firm/publications/when-machines-discriminate-the-rise-of-ai-bias-lawsuits/
High-level summary of the AI Act. (2024, May 30). Retrieved from EU Artifical Intelligence Act: https://artificialintelligenceact.eu/high-level-summary/
K.P., A. (2024, June 3). Contemporary forms of racism, racial discrimination,. Retrieved from https://docs.un.org/en/A/HRC/56/68
Mihalopoulos, D. (2025, May 5). Syndicated content in Sun-Times special section included AI-generated misinformation. Retrieved from Chicago Sun Times: https://chicago.suntimes.com/news/2025/05/20/syndicated-content-sunday-print-sun-times-ai-misinformation Proctor, J. (2024, February 15). Air Canada fines
