I think I know central London unusually well for someone who doesn’t drive a cab – especially the West End – and it’s mostly because I used to freelance across different offices with no smartphone, just a pocket-sized A-Z to find my way.
I’d arrive at a station, stand by the exit with the right page open, trace a route with my finger to the office, and navigate by checking road names. Over time I learned the feel of the city as much as the layout: how quickly you can cross town and improve your mood if you cut through the parks, when it’s worth crossing the Thames twice to go in a straight line, and how you can slip from Charing Cross Road to Covent Garden through Goodwin’s Court and that weird passageway by the Lamb and Flag instead of dodging tourists on the bigger roads.
I know I’ve lost some of you with the niche London content, but imagine I’m talking about where you lived as a teenager – when you first started getting yourself to school, your friends’ houses or the newsagent. You can picture the view from the bus even now.
I still rely on that learned map a lot, but I also get in my car and confidently drive somewhere I’ve never been, trusting my phone to pull an optimal route out of thin air. I’m perfectly happy to turn left when the arrow tells me to, whatever the road might be called. Do streets even have names anymore? Maybe not. If I ever wrote an address on an envelope I’d know, but… what is “post”?
In other words, I’ve lived on both sides of the line, like many of you have: building my own mental map and happily outsourcing the job. Not because I’ve become lazier (although…) but because that’s what humans do. Most of us no longer remember phone numbers. We don’t carry atlases. We don’t memorise facts we can look up in seconds any more than we know how to navigate by the stars.
Outsourcing thinking to the machines
Psychologists call this “cognitive offloading” and I’ve written about it before: the entirely ordinary habit of handing parts of our memory and thinking to tools that can take the load, and trusting the output (because “automation bias” is also a thing). Experiments suggest that when people expect information to be stored online, they remember less of the information itself and more about where to find it again – the so-called “Google effect”. Writing and notebooks are just older versions of the same trick.
What is changing at pace is the type of thinking we’re offloading. With the old sat navs, we mostly outsourced memory, time and attention. With generative AI, we’re increasingly outsourcing first drafts, summaries, risk identification, recommendations and probably a few “options for ministers” if you work in government. That nudges us from offloading effort – make this quicker, neater, shorter – towards offloading judgement: tell me what matters here, what the options are, what’s safe to ignore.
The line between the two is fuzzy, and we don’t yet have great language for it, which is awkward when the tools sound so confident. Studies of trust in AI chatbots suggest that fluency, speed and a human-like tone all increase perceived expertise, even when accuracy is mixed. Which brings us to another very human habit.
The voice of authority
Humans already have a bad habit of mistaking fluency for competence. A smooth, confident voice feels trustworthy. Unconsciously, it’s something like the voice of a newsreader from your childhood, a teacher, or a senior colleague. A well-structured email feels authoritative even if the reasoning underneath is a bit wobbly.*
I know this all too well because I wrote quite a few essays at university where the effort I put into the writing was inversely proportionate to the effort I put into the research, and it worked all too well. Generative AI leans heavily into this tendency: it writes fluently, answers immediately and rarely says, “No idea, mate – you should talk to a real expert”. It doesn’t pause, mumble, hedge or contradict itself in quite the same way that humans do. That can make the output deeply attractive. It sounds like the person who has read all the papers, attended all the meetings and reached an informed view.
Unsurprisingly, research suggests that greater confidence in generative AI is associated with lower levels of critical engagement with its output, particularly when people are busy or under pressure. It’s the technology-enabled version of believing something because The Man said it.
When we cross the line
That, for me, might be the biggest risk with generative AI; not that it will suddenly become sentient and launch the missiles, but that it will become so embedded in our habits that we stop noticing where the handovers happen.
First we ask it to tidy our prose, then we ask it to draft the first version, then to suggest arguments, frame options and identify the most important points. At some point we’re still editing and approving in some way, but the shape of the thinking – what is foregrounded, what is backgrounded and what never appears at all – has been influenced by a system we didn’t consciously interrogate.
There are already examples in journalism and publishing where AI-generated copy, references or quotes have slipped into print and later had to be withdrawn because the sources didn’t exist (awkward…). Usually, nobody set out to deceive anyone. People simply accepted confident-sounding output and assumed somebody else had already done the thinking.
Tools that keep you awake
The good news is that not all AI tools encourage the same behaviour.
One reason I’ll tell anyone who listens to do searching via Perplexity is that it shows its workings really well. You can see the searches it ran, the sources it considered and, crucially, the citations behind the answer. Even when I agree with the conclusion, I’m encouraged to glance at how it got there. And I’m much quicker to notice when it’s not got it right.
The same principle sits behind the AI tools we’ve built at Leading AI. When our tools answer a question, they don’t simply produce a confident paragraph and ask users to trust it: they provide citations back to the underlying reference materials. People who never looked at company policies when they were buried in folders are suddenly reading them. Not because they’ve become policy enthusiasts overnight, but because the tool and their sense of responsibility have given them a reason to click through and check.
That may sound like a subtle distinction (or a sales pitch) but I think it’s an important one. One of the reasons generative AI can be so persuasive is that it hides an enormous amount of work behind a single polished response. Showing sources, exposing reasoning steps and encouraging users to explore alternatives (by telling the tool its job is to help in a particular way, not just to keep you talking) creates a useful amount of friction. Not enough to make the tool annoying, but just enough to keep your brain switched on.
Perhaps the future isn’t simply about building smarter AI. Perhaps it’s also about building AI that keeps humans usefully engaged in the process.
The real shortcut for better outcomes: conscious thought
Some AI uses are clearly about effort: “Tidy this paragraph.” “Turn these bullets into a table.” “Summarise this annex.” Others edge much closer to judgement. “What’s the best option?” “Which risks matter most?” The first category can save time. The second category still needs you at the wheel.
My own rule of thumb is simple: if the task is about expression, AI gets the first go. If it’s about selecting evidence, options or priorities, I treat it like a provocative colleague and sounding board rather than a decision-maker. And if the consequences affect real people, particularly people with less power than me, the judgement stays firmly human.
Which brings me back to the maps. Using AI well is less like putting the A-Z away forever and more like learning when to glance at it and when to look up at the street. Offload the grunt work, certainly. But keep hold of the bit that notices where you are, who is affected and what kind of shortcut you’re taking. The real shift we need is not from distrust to trust, but from unconscious to conscious offloading: knowing when we’re outsourcing effort and when, almost without realising it, we’ve started outsourcing judgement and lose the hard-won thinking skills that make us… us.
* Of course, the opposite mistake exists too. Humans sometimes stop trusting expertise altogether and start favouring confidence, authenticity or outrage instead. That’s not really the opposite of authority bias. It’s just authority bias wearing a different suit.
