Blogs

Your people are signing off on AI errors they can’t see.

Shane Cooper, Founder - AI Accuracy Lab ·

You have the tool, the policy and the risk register. The step closest to your clients/customers is the riskiest, and one you’ve probably overlooked.

A typical AI rollout follows a familiar sequence. Choose a tool. Run the security and privacy reviews. Approve an acceptable-use policy. Hold a session on prompting. Add AI to the risk register. Switch it on.

Every one of those steps matters. But look at where AI risk reaches those outside of your organisation. An email, report, judgement, or advice drafted with the help of AI, sent to a client or the regulator, or put in front of the board.

That moment is the last step in every AI workflow. It’s also the one most organisations overlook.

Many AI policies include some version of “a human must review the output”. It’s the right instinct. But it quietly assumes the human knows what to look for.

The evidence suggests many don’t. In the University of Melbourne and KPMG global study of trust in AI (2025), 66% of employees who use AI at work said they had relied on its output without evaluating its accuracy. More than half (56%) said they had made mistakes in their work because of AI.[1]

That isn’t carelessness. AI errors are hard to see. The writing is fluent, the tone is confident and the formatting is clean. A fabricated reference looks exactly like a real one. A rule that changed last year reads as if it’s current. The signals we normally use to judge quality, such as polish, confidence and structure, are the things AI produces best.

The Australian Government’s Guidance for AI Adoption, published by the National AI Centre in October 2025, sets out essential practices for using AI responsibly. One such practice is “Maintain human control”. Under it, organisations are asked to make sure the people overseeing AI are equipped to do it.[2]

The cost of not adequately training employees to spot and correct AI errors is already on the public record. In 2025, an A$440,000 report prepared for a federal government department was found to contain references that didn’t exist and a made-up quote attributed to a Federal Court judgment. The consultancy agreed to repay the final instalment of its fee, and the story ran in national and international media.[3]

The lesson isn’t that AI was used. It’s that nobody between the draft and the client caught what it got wrong. Closing the gap doesn’t need another platform or a day out of the business. It needs three things.

  1. Shared understanding of how AI gets things wrong. AI errors fall into recognisable types. Once people can name these, they start to see them, everywhere.
  2. Short, focused self-paced online lessons that fit around the working day. Realistic examples of AI errors, drawn from the kind of material people actually handle, build the instinct to catch errors, without slowing down productivity.
  3. Evidence that the company takes responsible use of AI and governance seriously. If your board, a client or a regulator asks how you manage AI risk, “we told people to check” isn’t an answer. Organisation-wide reporting on lesson completion is. It shows who has been trained and where the gaps are, and it demonstrates a visible commitment to the responsible use of AI and to good governance.

The natural home for this is onboarding, alongside cyber security, privacy and conduct, with short refreshers to keep the habit sharp.

So here’s one question to put to your AI lead, your head of learning and your risk team this week: “Who has shown our people what an AI error looks like?” And, can your people spot them instinctively?

If the answer is a pause, you’ve found the missing step in your AI governance and risk management frameworks.

Shane Cooper is the founder of AI Accuracy Lab, Technology Attorney, and Law Lecturer at Bond University. AI Accuracy Lab provides focused online learning that teaches your employees to instinctively identify and correct AI errors before they matter. Learn more at aiaccuracylab.com.

Sources

  1. Gillespie, N., Lockey, S., Ward, T., Macdade, A. and Hassed, G., University of Melbourne and KPMG, Trust, attitudes and use of artificial intelligence: A global study 2025 (over 48,000 people in 47 countries). https://doi.org/10.26188/28822919
  2. National AI Centre, Department of Industry, Science and Resources, Guidance for AI Adoption, 21 October 2025. https://www.industry.gov.au/publications/guidance-for-ai-adoption
  3. Carly Page, “Deloitte refunds Aussie gov after AI fabrications slip into $440K welfare report”, The Register, 6 October 2025. https://www.theregister.com/software/2025/10/06/deloitte-refunds-australian-government-over-ai-in-report/722000

Originally published on LinkedIn

← All blogs