In our recent Worlds Apart paper, we made an argument that runs against the grain of the last decade of AI: the road to trustworthy AI isn’t necessarily paved with more data or bigger models, but with better context – models that are aligned and grounded in the same world we live in. Alignment and grounding are engineering terms, but there’s a more human word for what all that grounding actually buys us, and it’s worth drilling into: relatability.

We trust things we can relate to. It’s true of people, gadgets and organizations, and it turns out to be just as true of AI. So, before we talk about neural networks, let’s talk about house plants, sat-navs and the corner shop.

The friend who just gets you

Think about the friend you’d happily hand your keys to while you’re away. You don’t choose them because they hold a certificate in plant-watering. You choose them because they know you: they know the fern by the window has been half-dead since 2019 and isn’t worth worrying about, that the spare key is at the neighbor on the left because you don’t get along with the neighbor on the right, and that if the boiler makes that noise again you just thump it. You don’t leave them a two-page manual, because you share so much context that most of it goes without saying.

Now imagine handing the same keys to a stranger from an agency. Perfectly competent, well reviewed, and yet you’d write the manual, and still end up fretting from the beach. The gap between the two isn’t skill. It’s shared context.

“Trust scales with how much of your world the other party already understands without being told.”

It’s also worth noting that relatability works both ways – a lack of relatability is not always about making a leap of faith to trust someone less capable or experienced than yourself. It could be the other way around. What if your neighbor was the cleverest person on the planet, an unparalleled genius whose every thought seemed to be on a higher plane of understanding than the rest of us? The further they are away from our own world view, the less we trust them, even if we know they can out-think us on every topic.

The tech that reads the room

The same feelings subconsciously governs our relationships with our gadgets. Early sat-navs would route you, with total confidence, straight into a river. At first we laughed. But then we stopped trusting them, second-guessing every instruction, which rather defeated the point of having one. A good navigation app today behaves differently: it seems to know you avoid motorways, that it’s school-run o’clock, or that there’s roadworks on your usual route. It feels like it gets you, so you relax and free up mental energy that would normally be wasted.

A thermostat that learns you like the bedroom cool at night earns the right to make the call itself. When a tool appears to understand your intent rather than just executing a literal command, you feel comfortable delegating to it. When it feels alien and unrelatable, you supervise, and supervision spends the very time the tool was meant to give back.

Why we trust the corner shop

The same pattern holds for people and organizations. Most of us trust the local shopkeeper, or the GP who has known the family for years, more readily than a faceless call center reading from a script, and not necessarily because they’re more skilled. It’s because they hold our context: our history, our quirks, the unwritten stuff nobody wrote down.

It’s also why a large organization wins back our trust the moment it acts as though it knows us,  for example remembering our last complaint, or understanding why this particular thing matters to this particular customer. Strip all of that away and even a technically flawless service feels cold and untrustworthy. Relatability, in the end, is just this question: does this system hold enough of my world to act on my behalf without me having to spell out every last thing?

From relatable to trustworthy, explainable and governable

Grounding of that kind has consequences well beyond a warm feeling. Three things we badly need from AI are direct consequences of greater relatability.

  • Trust. We grant real authority to those who share our context. An AI grounded in our world can be delegated to, not merely prompted and double-checked.
  • Explainability. A relatable system explains itself in terms we already recognize. “I kept my distance because that was clearly a police scene” is an explanation any of us can weigh. “Token 4,417 exceeded a threshold” is not. Shared concepts are what make shared explanations possible.
  • Governance. You can only govern a system using concepts it actually represents. You can’t govern a pure pattern-matcher with human ideas like fairness, safety or legality if those ideas simply don’t exist inside it. AI we can relate to carries the very concepts we want to hold it to.

The uncanny valley of competence

The unsettling thing about today’s most impressive AI isn’t that it’s stupid. It’s that it’s brilliant and alien at the same time – an eloquent stranger who sometimes says all the right words while missing the point entirely. The fix isn’t to make AI more human on the surface. It’s to make it more human underneath: grounded in the same world, carrying the same context, reasoning over the things we actually care about.

Relatability isn’t a nice-to-have bolted on top of trustworthy AI. It is trustworthy AI, viewed from the human side. Build AI we can relate to, and trust, explanation and governance arise naturally, rather than being a complex retrospective paperwork exercise.

We don’t need AI that is one of us. We just need AI that’s on the same wavelength.

Read our POV on How world models and context will enable the next generation of trustworthy AI