Why Does AI Never Quite Get You? Because It's Missing Your Identity Core

I’ve got a pretty bad cold today, but here’s a short one for you anyway.

A lot of the time when we use AI, we come away with the same feeling: it doesn’t understand us. When it makes decisions and choices, it often doesn’t run the way we really wanted.

Our first instinct is to go tune the prompt, add memory, write more rules. But I’m increasingly convinced that a big piece of what’s missing is its grip on our identity core—it doesn’t know what our real ordering of values is.

The Stanford experiment

A while back I came across a Stanford experiment with 1052 participants. The question behind it is a fundamental one: what does it take to really define a person?

The traditional way we define someone is by labels, which is also how ad targeting works: stick a set of sociological and behavioral labels on a person—education, age, gender, occupation, even MBTI—and infer their values and lifestyle from there.

But that approach turns out to be inaccurate. It only gets you to about 74%.

Stanford’s approach: run roughly a two-hour deep interview, get the subject to narrate their own life story, keep probing details, and use one thought experiment after another to force out their real ordering of values. That gives you a far more accurate “identity core” model, one that reaches 86%—a full 12 percentage points higher.

Someone has already turned this into a business: building “digital avatars” of real consumers at scale. So when you’re planning a marketing push or a PR project, you can run these avatars to predict how a specific audience will react to a given campaign or public event.

I built a small app myself

I was curious enough to dig up the details of the experiment online and spend a few hours building a small app. It came out pretty interesting.

Once you’re in, it interviews you the way Stanford did, across 11 domains, following up until the conversation gets genuinely deep. Friends who’ve tried it usually spend anywhere from 20-odd minutes to half an hour. The deeper, the more detailed, the more honest you go, the more accurate the result.

At the end it hands you a profile of your “identity core”—honestly, even I was surprised at how cleanly it teased apart and ranked my values (that part gets into some private territory, so I won’t publish it).

Then, based on that profile, it also generates an “identity core portrait” using the avatar you upload, with your keywords and one line that sums you up printed on it. Quite a few friends of mine have made that image their social media profile picture after taking the test.

The line summing me up came back as a neat little couplet: “Strive, and prove yourself; stay true, and keep your heart.” It hit me harder than I expected.

The key step: feed the core to your agent

But the genuinely interesting thing about this app isn’t the image. It’s this:

You can copy the test result out, save it as a Markdown file, and hand it to your agent—Codex or Claude Code, either one.

Once it has that “identity core,” it knows how you rank things when a real choice comes up. So when AI is helping you plan something or pick between options, its judgment comes out much closer to what you actually want—closer, even, than you understand yourself.

Getting a result like that is a bit like looking in a mirror. It’s kind of fun.

One note

This test runs on top-tier models plus GPT Image 2 for the portrait, so it costs real money. If you want to look into this mirror yourself, come find me—cover the cost and you can try it.

All right, that’s it for today. See you tomorrow—and here’s hoping I feel better by then. Bye.

Source: EP0062_audio.mp3 · ASR model gemini-2.5-pro (chunked parallel) · full text of the original recording

[00:00] I’ve got a pretty bad cold today, so I’ll keep this one short. A while back I came across an experiment Stanford ran. They recruited over a thousand people. What were they after? Traditionally, when we define a person, we use labels. Even when we’re buying ads we slap sociological or behavioral labels on someone: education, age, gender, job, even their MBTI—and the idea is that those labels tell us what a person stands for, or how they live and how they behave.

[00:27] But that approach isn’t accurate. The traditional way only gets you to something like 70-something percent. But with Stanford’s experiment—it’s a two-hour deep interview where you keep getting the subject to tell their life story, keep pushing on the details, and through those thought experiments you get at how a person really ranks their values. That gets you a very accurate model, an identity core model that hits 80-something percent. And some people have turned this into a business, building lots of digital avatars of real

[00:54] consumers, so that when you’re running a marketing or PR project you can use the avatars to work out how a particular group will react to a campaign or to something in the news. That’s already a business. I got really curious, so I went looking online for information about the experiment, spent maybe a few hours, and built a fun little app. Once you’re in, it keeps asking you all sorts of questions, just like the Stanford interview—11 categories of different questions, with follow-ups.

[01:19] So the subject ends up going deep. Right now the friends around me typically take twenty-something minutes to half an hour to get through it. And the deeper you go, the more detail and the more of yourself you put in, the more accurate the result. At the end it gives you—and this genuinely surprised me—at the end it gives you a profile of my identity core. It touches on some private stuff so I won’t put it out here, but it really does sort out and lay out how I rank my values.

[01:44] And then, based on that identity core profile, it uses the avatar photo I uploaded to generate a portrait of my identity core. So you can see which keywords came out for me: grinding it out, proving it yourself, fairly thorough, inside matches outside. It also uses one—it also uses one line, “strive to prove yourself, stay honest to hold your center,” to sum up the whole identity core. I thought it was pretty interesting. A lot of friends around me, after taking the test, made that image

[02:10] their social media avatar. When we use AI, a lot of the time it feels like AI doesn’t get us—like when it’s deciding something or making a call, it isn’t going by what we asked for or what’s in its memory. And I think a big part of that is it has no read on my identity core. So if we copy the answers out of this test, we get a core—a Markdown file—and once we hand that to our agent, Codex or Claude Code, it can actually understand

[02:35] how we rank our values when we make choices in specific situations. So once it has that, when AI is planning something out or making a call, it understands us even better than we understand ourselves, and what it picks comes out much closer to what I actually want. It’s pretty interesting. This test runs on top-tier models plus GPT—GPT Image 2 for the picture, so it’s on the expensive side. If you want to try it, I’ll drop the link in the group; chip in for the cost and you can take it. It’s like looking in a mirror. Getting a result like this

[03:00] is genuinely interesting. That’s it for today. See you tomorrow—hoping I feel better by then. Bye.