Dr Leon Furze

In a previous article I made the case for problem patterns: recurring, recognisable patterns in AI-generated content, of which hallucination is only one problematic output. Although we’ve spent a lot of time and energy in the past few years building AI risk frameworks around trust, truth, and hallucinations, I’m starting to think that the reputational harms of AI misuse extend much further.

I’ve already described why the term “hallucination” is problematic: it anthropomorphises AI models and distracts from the reality that this is not a design quirk – it’s an unavoidable technical limitation. It’s still important to talk about hallucinations, and I’ve written about them in both Practical AI Strategies (2024) and the upcoming Practical AI Strategies 2 (September 2026). In the book, I use this diagram to consider the risks of hallucinations:

If we think about hallucinations as just one problem pattern though, it becomes more complex than a linear scale.

AI produces “problem patterns” in at least seven areas:

This follow up post asks two questions about the situation:

1. What would it cost if this problem got through?
2. How likely is it that someone would catch it first?

I love alliteration, and I love a good matrix. So I’ve invented* the word catchability and plotted it against cost as a way to articulate the potential hazards of AI use beyond just hallucinations.

*I didn’t really invent catchability, but the only other mainstream usage I’ve found relates to… fish.

Cost: a property of the output in use

Cost is the potential consequence if the output is accepted, relied upon, published, or acted on without correction. It comes in a variety of flavours: legal and regulatory consequences, financial loss, clinical harm, educational harm, reputational damage, degraded decision-making, wasted professional time, and eroded trust between the author and reader.

Cost is not an inherent property of the AI output. The stakes come from where the document is going, or the context in which it is (or isn’t) published. That means that cost can change over time during the production of a document, for example from an idea, to a draft, to a published report.

It also means that cost can change after the fact. For example, an internal discussion paper gets forwarded to the board, or a rough lesson outline becomes a published curriculum resource. Documents drift towards consequence, and in my experience they only drift in one direction: nobody ever takes a board report and demotes it to a private brainstorm. Cost always increases over time.

Catchability: will anyone catch it before it gets out of hand?

Catchability is the likelihood that a problem pattern will be noticed and corrected before it influences a decision, a communication, or an action. Damien Charlotin has been cataloguing hallucinations in legal cases – up to 1809 and counting. Every fabricated citation in the court cases was eventually caught, but usually by opposing counsel or an unimpressed judge; too late, and too high cost.

This is a product of the technology, and the people using and reviewing it. That introduces a lot of variables, especially as models improve over time and the problem patterns become harder to catch. For example, we could ask:

• Does the reviewer have the domain expertise to know what should be there, or shouldn’t?
• Do they have a situated knowledge of this client, this patient, this classroom, this body of work?
• Do they have access to authoritative source materials, or are they reviewing from memory and more reliant on the AI?
• Does the workflow include real review, by real people, with real time allocated to it?
• Is anyone actively looking for this type of problem, or just proofreading?
• Do checklists, quality assurance processes, or verification steps force the check, or is it optional?
• Do the incentives reward scrutiny, or reward getting the thing out the door?

A missing limitation-of-liability clause is highly catchable by a commercial solicitor running a mandatory clause-by-clause review against a firm’s precedent. The same omission is essentially uncatchable by a small-business owner who asked a chatbot for a contract because lawyers are expensive – the business owner doesn’t know the right questions to ask, and the chatbot won’t necessarily answer them unbidden.

So while cost is a property of context and time: catchability is a property of the output, the reviewer, the evidence, and the process. 

The cost versus catchability matrix

Something in my brain is constantly drawn to 2×2 matrixes (matrices?). The Eisenhower/Covey urgent/important matrix, my earlier needs/potential matrix, Punya Mishra’s AI knowledge/domain knowledge matrix… Maybe it’s because if something can’t be expressed in a 2×2 grid, it’s probably too complex to hold in my head.

Here is cost plotted against catchability:

The same problem pattern can occupy different quadrants in different contexts, and most of the practical value of the framework comes from watching things shift about from one quadrant to another.

In the next post, I’ll go through the quarters in more detail and suggest a few examples. Join the mailing list to get a weekly wrap up of all the posts from the blog, plus updates on the upcoming release of Practical AI Strategies 2: