If AI has the answers, why bother with training?

It’s worth asking the question directly, because dodging it does no favours to the case for learning.

If the answer to almost any technical question is a few keystrokes away, and increasingly arrives pre-explained and ready-applied through AI, what exactly is training still for?

This is not a new challenge. Search engines raised something similar years ago. What makes the AI version sharper is that it does not simply point you to an answer: it delivers one processed and in context, phrased with total confidence.

The honest answer is not that AI is wrong or unreliable, though it sometimes is – at least for the time being. It is that having an answer and knowing what to do with it are not the same thing, and the gap between them is where professional judgement lives. Training plays a vital role in developing this judgement.

Psychological perspectives

Cognitive psychology offers some important insight here. I am indebted to Paul Eccleson who has helped me to understand and apply this. Working memory, the mental space we use to process anything new, is limited. Cognitive load theory, developed by the educational psychologist John Sweller, describes how experience builds schemas: mental structures that let familiar material be handled with little conscious effort, leaving capacity free for what is genuinely new. A novice and an expert looking at the same problem are not spending the same mental effort to reach a view, because the expert has already automated most of what the problem requires and can give full attention to the part that actually needs thought.[1]

This is the real cost of learning without foundations. Consider an analyst reviewing a transaction alert that may be suspicious. AI can produce a fluent explanation of the underlying typology in moments. What it cannot do is tell that analyst whether this pattern, in this client relationship, against this regulatory backdrop, meets the threshold for escalation, or whether the write-up will hold up under scrutiny months later.

An analyst who understands the typology and has practised applying it against a live regulatory backdrop spends their attention on that judgement call. One who does not is still working out what the typology means once the decision needs to be made.

Understanding, not just knowing

Before moving on, there’s a deeper, more pervasive element to this. It is reasonable to expect AI, over time, to develop the ability to comprehensively process, evaluate and apply information about possible money laundering in context and without apparent error. At this point the question changes from: how do we ensure that humans are able to handle AI-generated insight, to why do we need humans handling this information? This is a larger and I believe more urgent issue, but not the subject of this short article.

None of this makes AI irrelevant to learning, and this is where our thinking in the ICA is so strong. Effective training needs to build on foundations of understanding underpinned by relevant knowledge. This understanding in turn is challenged and refined through practical, scenario-based exercises, debate with colleagues, long form assessment and Socratic dialogue with experienced trainers.

That is a different design brief to the one most training is built around. Training should spend less time transmitting facts that can be looked up in seconds, and more time building the judgement to use them well and the scepticism to test them properly.

So the more useful question is not whether learning still matters. It is what learning needs to spend its time on now that recall is cheap and judgement is not. Training built around the old assumption – that the scarce resource was information – will keep losing ground to AI because it is fighting on the wrong terrain. Training built around the new one – that the scarce resource is now judgement – has a case that AI cannot make for itself.

[1] Daniel Kahneman's distinction between fast, intuitive thinking and slow, deliberate thinking points to something similar: with enough grounding, the familiar parts of a problem can be handled quickly, so the harder, deliberate thinking is reserved for the part that genuinely warrants it.