Statements we make about ourselves, and what AI changes

I’ve spent a career in banking and one of the most consistent things I’ve heard from the people around me has nothing to do with money. It’s a statement, in different forms:

“I’m not technical.”

“I’m not a numbers person.”

“I never got my head around X.”

They are made matter-of-factly, as if reporting the weather. They sound like statements of fact. They aren’t, or at least not in any straightforward way. Statements like these are self-beliefs, and the way they form is more complex than the words themselves suggest.

What these statements actually are

Decades of research on self-efficacy beliefs¹ have established that people’s judgements of their own capability are not simply based on their experiences. They consolidate over time, shaped by our social and economic circumstances, by repeated outcomes, by what we observe in others, and by what we are told to expect of ourselves. Eventually those beliefs settle into stable accounts of what we are and aren’t.

Those accounts then become part of a life story we tell ourselves. Research on narrative identity² describes how people construct internalised stories that integrate their successes and failures into a coherent sense of self. Statements like “I’m not a numbers person” are compact summaries of much longer narratives. They organise the past; they also protect against future disappointment.

Sometimes the narrative is shaped by social context in ways that have little to do with the individual. Research on stereotype threat³ has documented how people’s accounts of their own capability track the social environments they have been operating in. In those cases the statement is doing identity work as much as descriptive work, and the patterns underneath it are social rather than individual.

And underneath all of this, the conditions under which someone tries are of course unevenly distributed. Class shapes who has the time to study, the cultural confidence to begin and the relationships to sustain it.⁴ Most failures that become self-narratives happen inside this structure, not separately from it.

So most adult self-narratives, including the everyday statements I started with, sit at the intersection of several factors. But they are not facts about people. They are accounts that have settled, drawn from the experiences a person has had, the contexts they had them in, and the resources they had to draw on at the time.

The impact of access to knowledge

Within that broader picture there is one contributor that has been particularly under-recognised relative to the others: immediate access to knowledge. Not the only contributor, and not always the main one, but one that is often obscured because the failure is experienced as a personal limitation rather than a missing support at the moment it was needed.

Not formal teachers, books or courses. The ‘someone-to-ask-right-now’ type of access. The person who would walk you through the bit you didn’t follow, at the moment you didn’t follow it, in the form that fits how you happen to think.

Moments of asking (the small windows when a real question has formed and the asker is open to following the answer) are rarely free to coincide with someone able to help. They are often blocked by other people’s calendars, by social cost, by geography, by the working conditions of the asker, by the limits of available books and courses, by the cost in time of finding someone qualified.

The issue is access: access to knowledge.

Provide that access and you can open a door to competence. Block access and the opposite happens; we walk away from things we could have done.

And where access was the main blockage, the statement that follows misnames its cause. It reads as aptitude. It says “I’m not the kind of person who can do this.” What it is actually a record of is a sequence of moments when the asking failed at the front door.

I’ve walked away from things like that. Many times. I suspect most people have.

What AI changes about access, and what it doesn’t

Immediate access to knowledge has changed in the last eighteen months or so.

Whatever you make of AI in general, the practical fact is this: there is now a kind of always-on interlocutor available to anyone with the means to reach it. Of course, that access is uneven, shaped by devices, connectivity, cost, literacy, language and confidence in using the tool at all. But for those who can reach it: it is there at any hour. About almost anything. With seemingly infinite ‘patience’ for the ignorant question, the half-formed question, the question that exposes a misunderstanding the asker did not know they had. The moment of asking is no longer the constraint it was. It happens at 6am on a Sunday, in fifteen minutes on the bus, at midnight when a thought hits us.

But the change comes with two complications.

The first is that AI itself is not always reliable. Far from it. Large language models confabulate, agree with users for reasons that have nothing to do with what is true, and are most confidently wrong in the very places a beginner is least equipped to notice. Using them well requires bringing scepticism, expecting to be misled in unfamiliar territory, holding the answer at a slight distance until it has been triangulated. The asking has been unblocked. The judging of the answer is still on the asker. So what AI has opened is not access to truth but access to provisional explanation: useful, often powerful, but still requiring judgement.

The second is that AI changes the ‘asking’ but not the ‘becoming’. Expert competence is built through years of effortful practice with feedback,⁵ not through information transfer. The asking has been unblocked. The work of becoming someone who can do the thing is often still the work it was: months of practice, the slow building of pattern recognition, the patient absorption of rules that feel arbitrary at first.

In the end, what AI has changed is one specific thing: the asking. And to the extent that this has been one important obstacle behind many statements about ourselves then it is no longer the obstacle it was, although the effect is likely strongest in domains where progress is often blocked by unresolved confusion rather than by physical practice or long formal apprenticeship.

What that means for how we see ourselves

What does that imply for the “I’m not the kind of person who…” statements? It depends on what each particular statement was actually about.

For some of these statements, AI does not help. The forces that produced them, things like aptitude, identity or the broader inequalities AI does not reach, are still in play. The statement remains about what it was always about.

Other statements formed because someone tried, hit a wall when the asking failed, and walked away. For those statements, one of the conditions that helped produce them may no longer hold. The statement was accurate when it was made. It may not be now.

What has changed, then, is the practical basis for revisiting these conclusions, if we want to.

So if we have a statement about ourselves that starts with “I’m not the kind of person who…” it may be worth asking whether the conditions that produced that statement still hold. If they do, the statement stays true. If they don’t, in particular if the asking part of those conditions was blocked, then the conditions in front of us have changed, and the narrative is worth re-examining.

What this looks like in practice

So for anyone who wants to try this, what does that kind of ‘asking’ actually look like?

For many people, their experience with AI has been mostly transactional: a recipe, a weather query, a definition. The agent answers, you take the answer. That’s not the kind of use being described here. This is something different: closer to ‘working through something with a colleague who happens to know a lot’, where the value comes from the back-and-forth, rather than from any single answer. Of course it should be acknowledged that this is not specific to AI; these are the moves of any working conversation with someone who knows more than you about something. AI is just newly available as that ‘someone’.

It looks something like this:

• Start with a half-formed topic, not a polished question. Transactional use rewards a clean factual query. In contrast, this kind of use rewards bringing the topic you are actually trying to understand, even if you can’t yet phrase it well. “I’ve been trying to make sense of X and the part I can’t get past is…” An LLM can work with that.

• Don’t settle for the first answer. When the answer doesn’t quite work, say so rather than accept it and move on. “I don’t quite follow that, can you put it differently?” “I’m not sure where Y fits in.” And don’t take answers as wholes: find the specific phrase or step you don’t understand and ask about that directly. “When you said X, what do you actually mean?” “How does that follow from Y?” Questions like that can take you all sorts of places. Most of the genuine learning seems to happen in those follow-up exchanges.

• Insist on the evidence. Models can produce confident factual claims without much underneath them, and even worse they can take a small or contested concept and build an entire edifice of reasoning on top of it. Both are correctable if you ask: “Where does that come from?” “Is that actually established or are you just making that up?” Refusing a confident answer until you can see what it is built on tends to change the answer materially.

• Push back on the reasoning. And even when the evidence is sound, the logic can be dodgy. Plausible-sounding reasoning is not necessarily good reasoning. LLMs are great at producing confident chains of argument when actually the logic is shaky. Challenge it: “I don’t think that’s right, here’s why.” See whether it defends well or backs off. This is not the same as catching factual errors; this is about the integrity of the reasoning itself.

What you are doing across all of these is iterating. Going through the questioning multiple times. Coming back to it again with further questions. Challenge - question - challenge - learn. The aim is not to come away with a clean definition you can repeat. It is to get to the point where the next question becomes obvious because you have actually understood the previous one.

But note: this kind of engagement unblocks the front door, but the ‘becoming’ is still what it was. It often needs a slow building of pattern recognition, the practice that has to happen for any real competence, the discomfort of being a beginner again: none of that is shortcut. But the front door used to be blocked for lack of someone to ask. That is no longer the case.

A lot of what gets said about AI is about its broader impacts on the world. But there is also this, available to all of us individually: a particular and narrow change in one specific component of access to knowledge. Worth knowing about, particularly if we have statements about ourselves we have stopped questioning.