Misleading AI Claims: The $100 Million Reason to Check What Your AI Says About You
Updated: Aug 30
Your AI is making claims on your behalf every day. Can you prove all of them?
In my consulting work, I see this regularly. A business uses AI to draft website copy, a proposal or a client email. The writing is clean and professional. But buried somewhere in the output is a claim the AI generated entirely on its own.
It wasn't in the brief. It isn't accurate and, in some cases, it's been sitting in front of clients for weeks before anyone notices.
And nobody checked it.
What used to be an embarrassing quality problem can now be a $100 million one.
From 28 March 2026, maximum penalties increased for many contraventions (breaches) of the Australian Consumer Law (ACL), including false or misleading representations. For corporations, the maximum is now the greater of $100 million, three times the benefit obtained, or 30 per cent of adjusted turnover during the relevant period when that benefit cannot be determined.
That doesn't mean every inaccurate sentence attracts a $100 million penalty. Courts consider the conduct, its duration, the harm caused and the size and financial position of the business. But the direction is unmistakable: unsupported claims can carry very serious consequences.
AI-washing is only part of the problem.
One emerging risk is AI-washing: overstating or misrepresenting the role or capabilities of artificial intelligence in a product or service. That might mean claiming a product is AI-powered when AI plays only a minor role, promising capabilities that do not exist, or using the AI label to make an ordinary service appear more innovative than it is.
The ACCC’s 2026–27 priorities include manipulative and false practices in digital markets. But the risk for businesses is broader than AI-washing. AI-generated content can introduce unsupported claims about products and services that have nothing to do with AI at all.
The ACL is technology-neutral, which means it does not excuse a misleading claim because a machine drafted it. If your business publishes the content, your business remains responsible for ensuring its representations are accurate.
The rule is the same. What changed is everything else.
Marketing compliance training has been a constant in my career. From the early days at Pacific Brands where I managed cheeky advertising and PR campaigns through to promoting flash memory products when digital cameras were taking off, and more recently at Fujifilm DMS, where managing communications for large government and banking clients meant compliance was non-negotiable. The industries and products may have changed, but the ACCC's position never did: a business must be able to prove any claim it advertises.
Ultimately, you are responsible for what you say, and you had better be able to back it up.
The process wasn't perfect, even back then. Reebok found that out in 2014 when the Federal Court ordered Reebok Australia to pay $350,000 after finding that claims about the muscle-toning benefits of its EasyTone shoes were false and misleading and were not supported by reasonable grounds. The claims appeared on boxes, swing tags and in-store material. They were written by humans, reviewed by humans and approved by humans, and the process still failed.
AI changes the shape of that failure. It can infer, embellish and fill gaps with plausible-sounding content that nobody briefed and nobody consciously approved. Because the result is polished, it doesn't trigger the alarm bells that a rough draft might. It looks like finished work, it reads like something a professional wrote, and so it goes out the door.
The risk lives in the plausible ones.
AI’s obvious errors are relatively easy to catch. The harder problem is the content that is almost right, or the claim that sounds like something you might have said, or something a busy reviewer would nod at and move past.
A flash-memory class action filed in California in 2004 comes to mind. I had just started working in the industry at the time, and the case was a regular topic of conversation. The plaintiffs argued that, although manufacturers used the decimal definition of a gigabyte, a widely used industry convention, customers could interpret the advertised capacity differently and receive less usable storage than they expected.
The case settled without a trial ruling on the underlying allegation. The settlement required clearer disclosures explaining how capacity was calculated. That detail matters: the industry convention could be technically defensible while still leaving room for a different consumer interpretation.
That is also what makes AI-generated claims dangerous. The output may be fluent, confident and nearly correct, while concealing a gap between what your source material supports and what the system produced independently. You cannot see that gap unless you are looking for it.
Confident output is not a green light. It is a prompt to ask better questions.
So what does responsible practice actually look like?
The good news is that it does not require a new compliance team or a technology audit. It requires a habit.
In my work I use a verification step called the A/B/C method. Before AI-generated content goes to a client, customer or public audience, each material claim in the output gets a label:
A | DIRECTLY SUPPORTED Present in the source material you provided. Verified. It can stay. |
B | IMPLIED BUT NOT STATED The AI inferred this. Reasonable, but needs a human judgement call. |
C | NOT SUPPORTED AI generated this independently. Verify it, or cut it. |
That invented capability claim buried in a client's website copy? It was a C that looked like an A. That is the whole problem, and it is exactly what the A/B/C step is designed to catch.
Three minutes, applied consistently before anything goes out with your name on it. It is the same rigour we once applied to a product-performance claim on a swing tag, adapted to content an AI system can produce in thirty seconds.
Algorithms do not sit outside consumer law.
In 2022, the Federal Court ordered Trivago to pay $44.7 million in penalties for misleading representations about hotel-room rates. Trivago’s algorithm gave significant weight to the fees paid by booking sites and, as a result, often did not highlight the cheapest available rate, even though consumers were led to believe it would help them find the best deal.
That case demonstrated something every business using AI should understand: algorithmic outputs are subject to the same consumer law scrutiny as human-generated claims. The ACL does not turn on whether a human or an algorithm produced the representation. It asks whether the overall impression conveyed to consumers was false or misleading. The business making the representation remains responsible for it.
Qantas provides another measure of the stakes. The case did not involve AI, but that is precisely the point: the law is concerned with the misleading representation, not the technology or process that produced it. In 2024, the Federal Court imposed aggregate penalties of $100 million after Qantas admitted continuing to sell tickets for thousands of flights it had already decided to cancel and continuing to display some cancelled flights to existing customers. The penalty was imposed under the previous framework, demonstrating how serious the consequences could already be before the maximum penalties increased in 2026.
Reebok paid $350,000 in 2014. Trivago paid $44.7 million in 2022. Qantas received aggregate penalties of $100 million in 2024. Since March 2026, the statutory maximum for many relevant contraventions has increased again.
The regulatory reach is extending further.
From 1 July 2027, the Competition and Consumer Amendment (Unfair Trading Practices) Act 2026 will introduce a general prohibition on conduct that manipulates consumers or unreasonably distorts their decision-making environment and causes, or is likely to cause, consumer detriment. The law is technology-neutral, so it can apply to AI-enabled conduct when the full statutory test is met.
In practice, this could include AI-generated urgency messages telling a customer there are only two seats left when there are twenty, or personalised promotional copy that uses someone’s browsing history to pressure or exploit rather than inform them. A false availability claim may also breach existing prohibitions against misleading representations.
The prohibition could also apply to AI-optimised website design that obstructs consumers from making or carrying out a decision, such as making the cancellation process significantly harder to find or complete than the sign-up process. The conduct does not have to involve AI-generated words. If AI is used to engineer a customer experience that manipulates consumers or unreasonably distorts their decision-making environment, and actual or likely detriment results, the prohibition may apply.
The discipline has not changed. The context has.
What I took from those years in marketing was a habit of mind rather than a checklist. You treat every claim as something you might have to defend, you ask whether you can prove it before it goes anywhere, and you build that check into how you work. The discipline came from experience and from seeing what happened when the process failed.
That habit is what many businesses using AI in external communication are missing. AI-generated content does not look like a draft that needs scrutiny. It looks like finished work. And that is precisely where the risk lives.
The A/B/C method brings the discipline back. Label what the AI knew because you supplied it. Label what it inferred. Treat anything it generated independently as something to verify before it stays.
One more thing is worth acknowledging. This article went through the A/B/C verification process before it was first published, and it still contained an inaccurate definition of AI-washing. The check was run against research notes rather than the ACCC’s own published material, so a secondary source produced a false A label. That is not a flaw in the framework. It is a reminder of what the framework requires: the source you check against has to be authoritative. Checking AI output against an AI-generated summary of a regulator’s position is not the same as checking what the regulator actually published. That is precisely the gap the A/B/C method is designed to close.
The marketing discipline hasn't changed. The context has, and so have the consequences of getting it wrong.
A note on how this article was produced.
The ideas, observations and case research in this article are mine. Claude helped me organise and articulate them, and I ran every claim through the A/B/C method you just read about. That felt like the right thing to do, given what this article is about.
For what it's worth, the Australian Government's own AI transparency guidance openly states it was produced with AI assistance, with full review and editorial control retained by its team. Helping people get clear on their thinking and making it land for an audience is what I do. As it turns out, AI can help with that too.
If this applies to your team
The AI in High-Stakes Communication workshop I run covers responsible AI use in practice, including the A/B/C verification method. It is built around real scenarios from your own environment and has been delivered to parliamentary offices and professional services teams.
Every session produces the same moment: the room becomes quiet when people first see the gap between what they gave the AI and what the AI added entirely on its own.
If you want to explore what responsible, practical AI use could look like for your organisation, I would love to have that conversation.
Sources
This article provides general information and is not legal advice.
Comments