AI Should Happen
With Us

Sol Rashidi — the world's first Chief AI Officer — posted something this week that stopped me mid-scroll. She raised the question of what happens when AI causes catastrophic harm and no one can say whose fault it is. She described the layers of accountability that should exist at every level of an AI deployment — the builders, the deployers, the businesses, the individuals — and concluded with a line I keep returning to:

AI should happen with us. Not to us.

It is a deceptively simple formulation. But it contains, I think, the clearest articulation of what is actually at stake in the accountability conversations we are not having — and the ones we are having too late.

The Accountability Gap Is Structural

Sol's framing points to something my own research has been circling for years: the gap between what AI systems do and what the humans working with them actually understand about what they are doing. I call this the understanding-trust gap, and it is not a soft or philosophical concern. It is a measurable, documented phenomenon.

In a controlled comparative study of N=150 participants, I found that as time pressure increased, trust in AI outputs increased — even as comprehension of those outputs decreased. People leaned harder on the system precisely when they were least equipped to evaluate whether leaning on it was warranted.

That finding was about individual operators in a laboratory setting. But Sol's post illustrates what it looks like when that same dynamic plays out at the scale of institutional deployment — and in domains where the stakes are not academic.

When accountability is diffuse, when no single actor owns the outcome, when the humans closest to the decision have been systematically disempowered from questioning the system they are running — that is not just a governance failure. That is the understanding-trust gap operating at civilizational scale.

What "With Us" Actually Requires

I am a firm believer in Sol's principle. But I think it is worth being honest about how difficult it is to actually live by — not as a policy position, but as a daily practice.

"With us" requires that the human in the loop be positioned to understand, question, and override. It requires interface design that surfaces uncertainty rather than concealing it. It requires organizational cultures that reward scrutiny over speed. And it requires individual practitioners who maintain the intellectual discipline to distinguish between outputs they understand and outputs they simply trust.

None of that is automatic. All of it has to be built.

This is why I have spent years studying not just AI performance, but the human relationship to AI performance — the conditions under which people calibrate their trust accurately, and the conditions under which that calibration breaks down. The design of the interface, the framing of the output, the presence or absence of explanatory features: these are not aesthetic choices. They are the architecture of accountability.

What I Do Every Day

I want to say something that might seem counterintuitive coming from a researcher who studies the risks of over-reliance on AI: I work with Claude daily. Not as a shortcut. Not as a replacement for thinking. As a collaborator — a conversation partner who helps me stress-test ideas, articulate arguments more precisely, and move faster on the implementation work that would otherwise bottleneck the intellectual work I care most about.

The distinction matters. There is a version of AI use that atrophies the human — where the model generates, the human accepts, and the loop closes without anyone truly understanding what happened. Sol has written compellingly about Intellectual Atrophy as a real and measurable risk of this kind of uncritical use. I share that concern.

But there is another version, which is what I practice: one in which the AI's output is always in dialogue with my own judgment, always subject to revision, always in service of goals I have set and can articulate independently. In this version, I am not downstream of the model. I am working alongside it — and the difference is legible in the quality of what we produce together.

Much of the rapid prototyping behind my research essays, website, and framework development over the past year would not have happened at the pace it did without that working relationship. And this matters for a reason that is also personal: at a pivotal stage of my doctoral journey, I found myself navigating a significant contraction in the professional resources and institutional support that had previously been available to me. The tools, the access, the collaborative infrastructure I had relied on earlier became harder to reach. In that environment, the ability to work with AI as a genuine intellectual partner — to move fluidly between ideation and execution, to prototype rapidly without a full team — was not a convenience. It was what allowed me to continue doing serious work.

That is not a story about AI replacing human support. It is a story about what becomes possible when the human-AI relationship is genuinely collaborative — and what is lost when either side of that relationship is treated as dispensable.

Agency Is the Point

Sol's accountability framework — builders, deployers, businesses, individuals — is correct. And it points to something my research has confirmed: accountability without understanding is hollow. You cannot be genuinely responsible for an output you do not comprehend. You cannot exercise meaningful oversight of a system you have been positioned to simply trust.

The reason "AI should happen with us" is not just a slogan but a design requirement is that agency — real agency, not nominal presence in the loop — is the precondition for accountability. Without it, we get the diffusion of responsibility that Sol describes: everyone and no one at fault, and real people bearing the cost.

Building that agency into how AI is deployed, how interfaces are designed, how humans are trained to work with these systems — that is the work. Not the exciting work. Not the work that gets funded first or demoed at conferences. But the work that determines whether the rest of it is worth anything.

AI should happen with us. I believe that. I try to practice it. And I think the field — researchers, designers, deployers, policymakers — needs to reckon honestly with how far most current deployments fall short of it.