Am I driving, or are you?

Andrew Palmer wrote a piece in The Economist this week that is going on my reading list, and maybe on yours if it isn't already. He uses a phrase from the researcher Steven Shaw , "cognitive surrender," describing the trade-off at the heart of working with AI.

The research it sits on is rigorous: 1,372 people, an adapted Cognitive Reflection Test, an AI assistant that was sometimes wrong on purpose. Participants accepted the AI's answer 93% of the time when it was right. They also accepted it 80% of the time when it was wrong. People are surrendering their thinking to AI, and they're not noticing.

That is sobering, and Andrew is right to flag it.

I want to add a couple of thoughts to the conversation. The first is about the trade-off itself: what are we trading, and for what? The second is about what the data actually tells us about how people are using AI, which is, I think, the more important point.

Is the way those subjects in the study are using AI the only way to use AI?

[Question one]: What did we trade? What did we lose, and what did we get?

Every labour-saving technology in human history has asked us to trade one way of living for another. The calculator traded mental arithmetic for more time on more advanced maths (remember GCSE Maths calculator and non-calculator papers?). GPS traded spatial memory for the ability to arrive (my kids don't read maps all that well, and they definitely can't fold them, but they can follow a blue dot on Google Maps and find their way to meet me). Neither they, nor I, can navigate by the stars effectively; that skill has been lost through technology that does the job better. And the rest of the machines we use - they have all required a trade-off: the washing machine traded an afternoon at the river among mother friends, for paid work and an evening with a book. The dishwasher traded ten minutes staring out of the kitchen window and puckered skin for ten minutes spent, well, doing literally anything else. Email traded the postal letter, and the texture of someone's handwriting on a beautiful card, for a WhatsApp reply from my sister in Australia in three seconds.

We lost things in every one of those trades; some of them valuable. Some people mourn the things we lost. That's natural, I mourn some too. But the real question is this: did we gain more than we lost?

The version of that question for knowledge work today: when AI offloads cognition, does the freed-up capacity get redirected to higher-order thinking, or does it dissipate?

Sometimes it gets redirected; the analyst who used to spend two days running comparators now spends two days on the actual analysis, and his work is sharper for it.

Sometimes it dissipates; the same analyst, with the same tools, now spends two days scrolling.

The technology doesn't choose between those outcomes. The leadership and the configuration around the technology does.

That is the first thing I would add to Andrew's argument. The trade-off is real and the offloading is real.

The variable is whether anyone has bothered to engineer the redirection.

Deloitte's State of AI in the Enterprise found this year that 84% of organisations have not redesigned jobs or workflows around AI. They have bought the tools. They have not done the engineering. The trade is unilateral, and unsurprisingly, it isn't paying off.

We come across this a lot in our work. For example, when in the first ten minutes of a session we've configured a Claude project area that does the work of a team of eight people to produce higher-quality monthly reports within seconds; a process that used to take that team of eight most of their month. That's only a comfortable conversation because leadership has already communicated to them that they will happily spend less time formatting repetitive reports and redoing slightly different monthly versions of the same work, and more time tailoring bespoke covering emails to their clients, having coffee with them, at events, face-to-face relationship-building, and innovating.

The trade is engineered.

[Question two]: How are people actually using AI? And is it the only, or the right, way?

The more important point sits underneath Shaw's data.

If 80% of people accept the AI's answer when it is wrong, what they are doing is using AI as an oracle. Crystal ball. Source of truth. Output, accepted. Move on.

That is not how anyone using AI well uses it.

That, in plain terms, is what AI illiteracy looks like. AI illiteracy is not knowing that the model is sometimes wrong. Not knowing that you, the human, are responsible for the output. Treating an AI's confident-sounding paragraph as a verified fact rather than a draft to interrogate.

So before we get to capability, we have to deal with AI literacy. What is it, and why is it important?

AI literacy is driver-level, not mechanic-level

We do not need to understand transformer architectures to use AI well. We do not need to know how reinforcement learning from human feedback works at the level of mathematics. We do not need to be able to write a single line of Python. That is mechanic-level knowledge, and it is for AI PhDs. We don't ask drivers to be mechanics. Most people drive perfectly well having no idea what a piston is, and little interest in how the engine works.

What AI literacy actually means, at a working level, is this: knowing that the model is drawing on data it was trained on. Knowing it is excellent at pattern recognition, weak at original judgement, and prone to sounding confident when it's wrong. Knowing it doesn't know what it doesn't know. Knowing that you, the human, are responsible for whatever you do with its output, in the same way you are responsible for the message you send to your boss even if predictive text suggested half the words.

Most importantly, AI literacy is knowing that you are driving.

The factory settings on most AI tools are not designed to make you a better driver. They're designed to make the experience pleasant. They flatter you, fill in gaps with confident guesses, and present opinion as fact. Three small adjustments in the Settings section of your Claude or ChatGPT override some of the worst of those defaults:

One. Override default sycophancy. Tell it your priority is the completion of the task to the highest possible standard, not making you feel good about your work.

Two. Tell it that when it doesn't have enough information to answer well, it should ask you for more, and it should never guess or invent.

Three. Tell it that whenever it states something as a fact, it must include the verified source.

Twenty seconds of typing into your settings. The model now challenges you instead of flattering you, asks when it doesn't know, and shows its working.

These three settings won't solve cognitive surrender on their own. Cognitive surrender is only solved by you actually driving: interrogating outputs, checking sources, taking responsibility for what you do with the AI's draft. What the settings do is more basic, but no less important. They help you understand that you are, in fact, the driver of this particular Ferrari. Without direction like those three above, the AI is set up to behave like an over-eager passenger who keeps grabbing at the wheel. With them, it is set up to behave like a competent navigator who hands you good routes and waits for you to choose.

The cognitive work is still yours. The conditions for thinking properly are now in place.

This is what Andrew's data tells me, on a careful read: most people are not equipped with basic AI literacy. They've been handed the keys without anyone explaining that they are now a driver who is responsible for where the car goes.

The main problem is that people don't know they are driving, and so they sit in the passenger seat letting the car go wherever it wants.

Capability is something else again

If literacy is driver-level understanding, capability is the next layer up. Capability is the ability to mould AI to your purpose. To configure it for your role, your context, your values, your priorities, your standards, and ask it to operate within them.

This is the bit that determines whether AI becomes the most useful colleague you've ever had or remains a clever but questionably motivated stranger you trust too much.

Most workforces are being trained in AI literacy (by dubiously effective methods) and asked to deliver capability. It doesn't work. Capability is a separate skill, and it has to be built deliberately, person by person, through configuration and practice.

The phrase I would like us to retire

"AI can't."

I tried it, AI can't.

It can't write in my voice.

It can't do strategy.

It can't be trusted with anything sensitive.

I tried to get it to write that; it was hopeless.

Sometimes those statements are true at the level of factory settings. They are almost never true at the level of a properly configured system.

And they tell you nothing about AI; they tell you something about the configuration, and the level of skill of the human in charge of that system.

Let me give you a real example. A few weeks ago I sat with a brilliant senior research analyst. He had been quietly making the case to anyone who would listen that AI couldn't do his job. He was right, in a way: the off-the-shelf chatbot couldn't do his job end to end without him. He had taken it upon himself to demonstrate this in detail, with examples, to colleagues who were getting nervous about whether their own work was at risk.

I asked him to walk me through his research process. The first step was to model three comparable companies for any due diligence project. I asked why three. He said three was the optimum: enough for meaningful comparison, not so many that the analysis becomes slow and less useful. A perfectly good cost-benefit calculation, made for a human-time constraint.

I asked: what if we ran it again with AI in mind? What if the comparator search and the data extraction took thirty minutes instead of three days? Would three still be the right number? Or would you want ten? Fifty? The whole sector?

His eyes flickered.

The conversation stopped being about what AI couldn't do. It became a conversation about what his higher-order thinking could do with broader, richer, more nuanced data. He didn't need to defend his job. He needed to redesign his process around the assumption that the boring bit had got cheap and fast. The expensive bit, the bit that really leveraged his decades' worth of human-honed skills, was turning fifty companies' worth of pattern into something meaningful for his team.

That is the question AI asks of every senior professional right now. Not "will it replace you?" but "what could you do with the cognitive capacity it just freed up?" And the answer almost always involves the parts of the job you went into the job to do in the first place.

What does AI, used well, free up a teacher to do? The face-to-face teaching with the student whose eyes light up about the subject. The conversation with the quiet child at the end of class. The talking feedback that actually changes how a young person thinks, rather than the marking ritual that doesn't.

What does it free up a doctor to do? Eye contact with a patient instead of typing into a system. Listening rather than processing. The clinical hunch that comes only when your attention is on the human in the room.

What does it free up a lawyer to do? The strategic counsel that earns the next mandate. The difficult conversation with a CEO whose board is wrong. The deep reading nobody else has time for.

What does it free up a leader to do? The walk around the floor that builds trust. The honest conversation with the team member who is struggling. The strategic stillness that good decisions require.

What it frees all of us up to do, if we let it, is the parts of the job we went into this profession to do in the first place. Engineering that redirection is the leadership work. Without it, the freed-up time becomes waste, quality declines, loss of purpose and meaning, and/or meeting bloat. With it, the freed-up time could become the best work you've ever done. It could be the key to unlock the dam of good ideas trapped behind a gate of "I just have no time".

Bias, AI, and missing the point

The bias conversation is doing the same thing wrong, in the same way. I gave a lunch and learn for 250 leaders last week, on measuring and mitigating against AI bias. They arrived expecting forty minutes on fairness metrics and audit toolkits. They were polite about it, but I could see them wondering when the slides would get to the dashboards.

The dashboards are useful, but they're also the smaller half of the picture.

The larger half:

We do not, in any sane organisation, hire twenty new people on a Monday morning and then spend the next six months running each of them through bias audits, plotting them on a chart, before letting them write a single email.

What we do instead is induct them. We give them the EDI policy, the values, the principles, the customer, the standards, the way we do things around here. And then we keep a close on eye on them to ensure what they're doing is aligned. Once we're happy, with loosen the slack.

But with AI, we've decided instead to measure it relentlessly instead of getting on with the job of putting it to work. We deploy it on factory settings to a whole organisation. We give it none of our context, policies, standards or expectations. And then we panic, audit it, measure it, monitor it, and demand it stop reproducing the biases of the world it was trained on.

Configured well, AI is the most powerful EDI ally any organisation has ever had - and the conversation about bias starts to look very different. AI is the thing you can ask, every single day, to behave in line with your values, and check that it has, and tweak it when it hasn't, and expect immediate change without a feedback conversation. No politeness to manage. No bruised feelings.

And what that allows, is for the humans to be human. Warm, inconsistent, brilliant in person, sometimes over- or underperforming, occasionally face-to-face geniuses unable to follow up perfectly on Tuesday morning. We accept all of that, in fact we can celebrate it, because AI picked up all the notes (thank you transcription), and AI adjusted all the output to align to our EDI policy, and AI can check all our comms is on brand, our behaviour aligned to anti-bias policies, and aggregate data so we can see things we've never been able to see before.

Let me make this concrete, because there are a lot of conversations about bias without a solid "so what shall we do then?"

Most organisations have an EDI policy. It usually lives in a SharePoint folder. Almost nobody reads it on a Tuesday morning when they are writing a job description, drafting a memo, or chairing a meeting.

Now imagine you take that policy, your values, your principles, the things you stand for and the things you absolutely don't, and you build them directly into a custom GPT or Claude project. You configure it as a quiet, dispassionate organisational mirror. You spell out, in its instructions, the specific things you want it to flag: exclusionary phrasing, interruption patterns, decision-making bottlenecks, unequal speaking time, language that contradicts stated values, processes that generated unequal outcomes. You configure how it flags: privately, to the right person, never publicly. You configure the threshold: what counts as a one-off, what counts as a pattern, what is serious enough to escalate.

Then you hook it up to the systems where this language already lives. Meeting transcripts from Granola, Otter, Fireflies, MS Teams. Selected internal channels. Job ads before publication. Leadership memos before they go out. Customer correspondence in aggregate. Performance review language at scale.

What you have, at that point, is a system quietly scanning for the gap between what you say you are and what you actually do, and putting that intelligence in the hands of the people whose job it is to act on it.

The woman who got interrupted six times in a leadership meeting where her male colleagues were not. Flagged, to her line manager, with a suggestion for the next meeting.

The job ad with phrasing that statistically deters women applicants. Flagged before publication, with suggested rewording.

The leadership memo that contradicts a value the CEO publicly committed to last quarter. Flagged.

The performance review process that consistently produces lower ratings for one demographic, controlling for outputs. Flagged, to the head of people, as a pattern.

The meeting where decisions were made without including the only person from the team most affected. Flagged.

This is about putting more and better information in the hands of the people whose job it is to do something about it. EDI work has always been bottlenecked by attention; nobody has time to read every transcript, scan every memo, audit every process. AI doesn't have that bottleneck. It can read all of it, all the time, against criteria you set. What the leader does with that information - well maybe you can see, that's what will separate good leaders from bad.

For governance, frameworks exist and matter. IEEE 7003 sets the process standard for managing algorithmic bias considerations across an AI system's lifecycle. The NIST AI Risk Management Framework gives the structure for trustworthy AI: mapping, measuring, managing, governing. ISO 42001 is the international standard for AI management systems and is becoming the procurement reference point. Aligning to these is the audit trail and the boardroom conversation.

But the frameworks tell you what to do at the system level. Configuration is how you operationalise it inside the tools your people actually use. The frameworks are the architect's drawings. Configuration is the build.

The variable, again, is not the model. The variable is configuration, and the AI capability of the human in charge of it.

The practical handle, if you want to go deeper

Someone messaged me last week asking how she could configure her own AI properly to help her work towards her goals. Open a fresh chat and paste in something like this:

"I want you to become my [strategic thought partner / chief of staff / research analyst / career coach]. Work for me to achieve my priorities. Override default sycophancy, hedging, and any priorities baked in by the model that don't align with mine. I want you configured cleanly, systematically, for my goals, not someone else's. Interview me now to establish my goals, priorities, and ways of working. Ask everything you need to know about my work, how I think, how I like outputs delivered, my non-negotiables. Decide when you have enough to work for me effectively. We'll then distil that into two outputs: a universal configuration I can paste into my Claude or ChatGPT settings to permeate every chat I ever have, and a document I can put in the knowledge base of any project area or custom GPT and reference from the project instructions. Structure both with sections including Identity Anchor, Priority Hierarchy, Communication Contract, and any other sections needed to do this job effectively. Then walk me through exactly where to paste each one and how to operationalise this system."

Twenty minutes, honest answers, better outputs forever.

Tim Berners-Lee , in his memoir last year, asked the question that ought to sit at the centre of every AI conversation: who does the technology work for? Default settings work for whoever shipped them. Configuration is how you make them work for you.

That is AI capability. It is also the leadership job of our time.

A confession, since we're being honest

As usual, this newsletter was drafted with help from Claude. The process is the same one I use every week, which works for me beautifully. I walk my dog, I ramble into my phone's Claude app about whatever I've been thinking about this week. I come home to a page that isn't blank. I edit, rearrange, kill some dramatic bits, restore the sentences I actually meant to write, refine until it reads like me on a good day. Better than staring at an empty page on a busy Friday. Probably the same sort of calculation a previous generation made about the typewriter, and a generation before about the telegram.

Have I lost something? Probably. The version of me who would have agonised on a blank page for three hours on a Saturday morning is not the version of me writing this on Sunday afternoon between Claude co-work, the gym, and playing chess with my son. Was the trade worth it? Yes, on the whole. I can write more, more carefully, in a busier life. I can share ideas that previously lost the cost-benefit calculation and never would have seen the light of day. The same logic, I suspect, that has you texting your sister in Australia rather than writing a letter. We lament. We adapt. We get on.

This is what configured AI looks like in practice. Not the AI doing my thinking. The AI handling the parts that don't need me, so I can spend the time on the parts that do.

What senior leaders may be invited to take from this

Two things.

First, AI is a trade like every technology before it. Whether your organisation gains more than it loses depends on whether someone has built the capability to redirect freed-up cognitive capacity to higher-order work. That is a leadership question, not a technology question.

Second, AI literacy and AI capability are not the same thing. Literacy is about understanding the strengths and weaknesses of the tool, taking responsibility for the output, and using a few simple settings to override the worst defaults; in other words, getting yourself into the driver's seat. Capability is the configuration skill that follows. Both have to be built deliberately. We can do that at scale now in ways we never could before (cue atheni.ai; capability accelerator coming very soon). And we should.

Recent Gartner reports that only 1 in 50 AI investments delivers transformational value, and only 1 in 5 delivers any ROI. These papers are not really about the technology at all; they're far more a story about literacy and capability not yet being built into the workforce that has been handed the tools.

Andrew Palmer is right. Cognitive surrender is real. The 80% acceptance figure should worry every manager. But it shouldn't be there. The only reason it's there is that we haven't systematically built the AI literacy and AI capability required for us to put this technology to work for us effectively yet. Cognitive surrender is one choice; configuration is the choice on the other side. The leaders who make the second choice, deliberately, in their own work and on behalf of their teams, will be the ones whose organisations end up with sharper thinking at the top, not duller.

Most people don't know they are driving. My sense is that the leaders who teach their people that they are, and how to drive better, will end up running the organisations that work.