Skip to content
Dustin's AI Lab
Go back

AI Will Never Go to Jail for You — I Figured Out What I Want to Teach, Then Got Told I Had Half of It Wrong

Since generative AI arrived I have been asking what AI can never do on a human's behalf. My answer is accountability, because AI does not get sentenced. Then I posted the argument and found out what I had missed about the risk carried by people at the bottom.

On this page
  1. Two years is optimistic
  2. Taste might not be the point either
  3. I finally worked out what I want to do
  4. Why I posted it
  5. What I mean is smaller than that

Two years is optimistic

Someone asked me how long before AI takes over the whole thing. From what I’m seeing right now, two years is optimistic. Five years isn’t even a sure thing.

But the timeline isn’t the real point. Everybody keeps circling back to data and judgment, and I don’t think that’s the point. Accountability is the most important human ability. However powerful AI gets, it will never go to jail for you. So the abilities tied to accountability are the ones that last.

Taste might not be the point either

The other answer people keep offering is taste. I think taste might not be the point in the end either.

Look at those AI short dramas. They reek of AI, and it doesn’t stop them from blowing up. You know it, I know it, everyone knows those are AI, and most of society doesn’t care. They still sell. Never mind that the next generation grows up alongside AI, so their taste getting assimilated is easy to imagine.

I’ve been thinking about this for a while now: what it is about a person that AI won’t replace. So far I have one answer. Responsibility. Because AI will never go to jail for a person. So the abilities tied to responsibility are the ones that last.

I finally worked out what I want to do

Since generative AI showed up we’ve been watching another industrial revolution. And I’ve been sitting with one question the whole time: will there be a day when AI replaces humans across the board? The things I used to think were safe don’t feel so safe anymore, not with how agents and robotics have moved these past few years. So I pushed the question down to something more fundamental. Is there anything AI can never do on a human’s behalf?

In the end I found a very clear answer. Being accountable. One sentence explains it: AI doesn’t get sentenced, and AI won’t go to jail for you. Responsibility always lands on a person. That simple.

So everything connected to being accountable is what I want to do my best to build in the next generation (for what those skills actually look like, I covered the verification side in Once the Alignment Gap Closes, What’s Left Is Taste and Verification). We’ve all seen too much of it on Threads already: people screenshotting AI answers and arguing with them instead of checking anything themselves. People at school turning in AI homework unchecked. People at work using AI with no idea how to sign off on the output. People vibe coding and never once thinking about security or privacy.

I came up through teaching GMAT. GMAT gave me a lot of training in logical thinking, and that’s what lets me check an AI’s summary chain and tool chain when I work with it, and find the exact place where it went wrong.

But I also know that teaching how to catch errors and how to be accountable isn’t going to be popular. It’s not attractive the way vibe coding is. So what I want to ask is this: if I taught this, would anyone actually be interested?

Why I posted it

After it went up, someone came back at me from the angle of governance and risk premium. Here’s what I said to them.

One of the reasons I put this post out is that I felt like I’d walked into deep water in my own thinking. I thought my logic had closed into a loop, but I was starting to get a faint sense that I was drifting into overconfidence, and talking to AI about it any longer would only feed the positive signal back to me. So I posted it for adversarial review (?). And now I do know there’s a lot left to think through and fill in.

On governance, I wasn’t originally trying to push it that far. My intent isn’t to train people into dedicated AI governance roles. What I want to build is the literacy and ability of ordinary people at the ground level to use AI responsibly. But the uneven risk premium that lands on those people is genuinely something I hadn’t thought about. Thank you for the reminder.

I’m going to rethink where the whole argument sits. I might also start by sharing hands-on practice instead, pushing it out at a small, careful, responsible pace while I watch what comes back, and correct direction from the actual response.

What I mean is smaller than that

Agreed, this isn’t AI governance. What I mean is smaller. It’s that everybody who ships AI output, top or bottom of the org, can check and correct their own work before letting it through.

Maybe I wrote the post too abstractly. This was a reminder for me.


15-Minute AI Adoption Diagnosis

Fill in a short form to book your free 1-on-1 diagnosis, and find out how you can work with AI without the pain and get a real productivity lift.

Book my free diagnosis
Share this post on:
Previous Post
Information Diet: I Pulled My Own Chrome History and Audited Where My Attention Goes
Next Post
NT$18,000 a Month for an AI Course: Price and Value Came Apart a While Ago