Output, Not Answer

Of all the questions I have been asking myself about AI, this may be the most personal one.

I have a child.

So when I read discussions about AI affecting cognition, critical thinking or our ability to learn, I cannot look at them only as another technology debate. I find myself wondering what this means for her. What should she learn? What abilities will still matter ten or twenty years from now? And, perhaps more importantly, which abilities does she need to develop herself before she starts handing parts of them over to a machine?

I don't know the answer.

But I have seen something similar before.

When social media arrived, we adopted it much faster than we understood what it might do to us. There were concerns, of course, but for most people the attraction came first. Only later did we begin having much more serious conversations about attention, comparison, mental wellbeing, children, misinformation and the strange experience of looking at everyone else's carefully selected version of life.

I wonder whether we are doing something similar with AI.

Not because I think AI is another social-media problem. It is a very different technology. The similarity for me is the sequence: adoption first, understanding later.

And even that analogy needs some caution.

For years we have heard concerns that phones and screens are reducing reading and attention. There is some evidence behind parts of this argument. A 2019 meta-analysis of 33 experimental studies found a small comprehension advantage for paper over screens, particularly for informational material. But more recent research looking specifically at children is much less absolute: a 2024 systematic review of 88 studies found that the most common result was actually no difference between print and digital reading, although print did better more often for some forms of comprehension while digital formats sometimes improved engagement and vocabulary.

That complexity matters to me because it matches my own experience.

I spend a lot of time looking at screens. But I also read far more today than I did years ago. I read news, politics, business, technology and subjects I would probably never have encountered otherwise. I listen to podcasts. I follow questions from one source into another. I feel that access to digital information has expanded rather than reduced what I know.

So perhaps the useful question was never simply whether screens are good or bad for us.

It was what we do with them.

AI makes that question much harder because it does something previous information technologies did not do quite so easily: it can perform part of the intellectual work for us.

If I want to understand something, I can ask.

If I need an argument, it can construct one.

If I need a paragraph, it can write one.

If I am struggling to organize my thoughts, it can organize them.

And most of the time, I am not sitting in front of AI with unlimited time and a philosophical commitment to improving my critical-thinking skills. I am trying to get something done.

That matters.

Because the easiest way to use AI is often also the one requiring the least effort from me.

This is where I have slowly developed one small distinction that has changed the way I work with it.

I try not to think of what AI gives me as an answer.

I think of it as an output.

That may sound like semantics, but for me there is a meaningful difference.

An answer feels like the end of a question.

An output is something I still have to do something with.

Is it right? What is it assuming? What has it left out? Is there another way to see the problem? What evidence supports it? What would someone who disagrees say?

I did not develop this habit because I understood the science of large language models particularly well. Part of it probably came from years of consuming news and watching the growth of misinformation and competing narratives. I became increasingly uncomfortable trusting something simply because it was written confidently.

There can be more than one reasonable interpretation of the same event. Different stakeholders can look at exactly the same situation and see something different. Two apparently contradictory things can sometimes both contain truth.

Once you get accustomed to looking at information that way, the first plausible explanation is rarely enough.

AI can actually be very useful for this.

I can ask it to show me another side of an argument. I can ask what I have missed. I can ask it to challenge something I believe. I can ask what evidence would weaken my position or how another stakeholder might see the same issue.

In those situations, AI isn't necessarily doing my thinking for me.

Sometimes it is making me think about something I would not otherwise have considered.

That is why I struggle with arguments that AI will simply make us cognitively weaker.

There is evidence that should concern us. A 2025 study from Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real examples of using generative AI. One of its more interesting findings was that greater confidence in AI was associated with less critical-thinking effort, while greater confidence in one's own ability was associated with more. The researchers also found that AI did not simply eliminate critical thinking: some of the mental work shifted toward checking information, integrating responses and supervising the task.

That distinction feels familiar to me.

If I believe the machine has given me the answer, there isn't much reason to continue thinking.

If I believe it has given me an output, the work isn't over.

Another widely discussed MIT Media Lab study produced an even more worrying signal.

Participants writing essays were divided between using an LLM, using a search engine and using no external tool. The researchers reported lower measures of neural connectivity among the LLM group, along with poorer recall of what they had written and a weaker sense of ownership over their essays.

It is tempting to turn that into the headline that AI is weakening our brains.

I don't think the evidence allows us to go that far.

The main experiment involved 54 people, with only 18 continuing into a later crossover session. It examined a particular task, essay writing, under particular conditions. It is an important warning signal, not proof of long-term cognitive decline.

And there is evidence pointing in the other direction.

In a randomized experiment with university physics students, a carefully designed AI tutor produced greater learning gains than an active-learning classroom lesson, in less time, while students also reported greater engagement and motivation.

Another randomized trial found the opposite kind of effect when students used ChatGPT more freely as a study aid: 45 days later, the AI-assisted group retained less of what they had learned than students who used traditional study methods.

At first these findings might look contradictory.

Perhaps they aren't.

One AI system was deliberately designed to make students learn.

The other made it easier for students to offload some of the learning.

That difference may be more important than whether AI was present at all.

And this is where I come back to my daughter.

For an adult who already knows how to write an argument, asking AI to critique a draft may be useful. The struggle of learning how to form sentences, organize thoughts, defend a position and revise bad writing has already happened.

For a child who is still developing those abilities, asking AI to write the essay is something very different.

It may remove the very struggle through which the skill is formed.

I have written elsewhere about the importance of boredom for children. The more I think about AI, the more I suspect there is a related idea here: not all friction is bad.

Sometimes struggling with a paragraph is the learning.

Sometimes sitting with a question without immediately finding the answer is the learning.

Sometimes trying, getting it wrong and restructuring the thought is the learning.

If AI removes every one of those moments, the result may be more efficient while the person becomes less capable of producing the result independently.

That does not mean keeping children away from AI forever. UNESCO's guidance on generative AI in education takes a similarly cautious rather than prohibitive approach: it calls for age-appropriate use, human oversight and the deliberate development of AI competencies rather than simply placing the technology into classrooms and assuming learning will follow.

For me, the question becomes less about whether my daughter should eventually use AI and more about when, why and for what.

I would much rather have her write something herself and then ask AI, What am I missing?

Or ask it to challenge her argument.

Or explain why a particular sentence is unclear.

Or show her an opposing perspective that she then has to evaluate.

That feels very different from saying, Write this for me.

But I also have to be careful here, because it is easy to turn this into another neat prescription: use AI properly and everything will be fine.

I don't know that.

Time changes our behaviour.

Convenience changes our behaviour.

Even when we know that thinking first would be better, we will sometimes choose the shortcut because we have a deadline, because we are tired, or simply because the machine can do it faster.

I do it too.

That is where a little self-introspection becomes important.

Before asking AI to do something, perhaps the useful question is not always Can AI do this?

It almost certainly can.

The question may be:

What part of this do I still want to do myself?

This is partly where my own PAUSE for AI idea came from. Not from believing that people should spend twenty minutes carefully engineering every interaction, but from wondering whether we sometimes need a moment before automatically handing over the task.

What am I trying to accomplish?

What thinking matters here?

What can I safely offload?

And what would I lose by doing so?

The answer will not always be the same.

If I need a routine email summarized, I probably don't need the intellectual struggle.

If I am trying to understand an issue on which I have never formed an opinion, perhaps I do.

This may become increasingly difficult as AI gets better.

Poor AI forces us to question it.

Excellent AI may be more cognitively dangerous precisely because questioning it feels unnecessary.

When we know a subject well, we can usually recognize a weak answer. When we know almost nothing about it, an articulate response can be impressive simply because we have no reference point against which to judge it.

That may be one of the risks I worry about most.

Not that AI will make us stupid.

That is too simple.

It is that AI can make us feel that we understand before we have done the work required to understand.

And yet the same technology gives us extraordinary access to knowledge, explanation, criticism and perspectives that were previously difficult to obtain.

That is why I remain optimistic about it.

But not unconditionally.

Perhaps one of the important skills of the AI era will not be knowing how to produce better prompts or how to get better answers from a machine.

It will be knowing when an answer should not be enough.

For myself, I have arrived at a simple habit:

Output, not answer.

For my daughter, I am still figuring out what the equivalent habit should be.

And maybe that uncertainty is exactly where we should be right now.