Who Grades the AI?

There is a lot of conversation right now about what artificial intelligence is going to replace.

Jobs. Tasks. Expertise.

But there's another question that doesn't get asked nearly as often:

What happens when AI enters a profession where judgment isn't optional?

Today we're talking about welding.

My guest is Curtis Casey, an AWS Certified Welding Inspector and Certified Welding Educator who has spent decades around welding, inspection, and the classroom.

He's also helping develop a new platform called WeldGrade, which puts AI into the welding education workflow—but perhaps not where you'd expect.

Curtis, welcome to The Next Arc.

Thanks, Rachel. Glad to be here.

Before we talk about AI, I want to talk about teaching.

You've got a welding instructor standing in a lab with, say, twenty students.

What does grading actually look like?

Well, that's part of the problem.

You've got twenty students, but you don't have twenty welds.

You've got students practicing repeatedly. Different processes. Different positions. Different assignments. And every one of those welds represents something that needs to be looked at.

So the instructor is walking around inspecting welds, answering questions, watching technique, looking at fit-up, checking finished work, documenting grades.

And while you're doing that, somebody else is waiting for you.

That's welding education.

So when somebody says, “AI could make grading faster,” your ears perk up?

Sure.

But so do my warning bells.

Why?

Because I don't want a computer deciding whether somebody knows how to weld.

AI can look at something. It can offer observations. It can recognize patterns.

That's useful.

But there's a big difference between giving me information and replacing my judgment.

I'm the instructor. I'm responsible for that student.

The final decision needs to stay with me.

There's a phrase you use that I really like: the inspector's eye.

What does that mean?

It's learning how to look at a weld and actually see it.

A beginner looks at a weld and says, “That looks pretty good.”

An experienced welder or inspector starts seeing bead profile, consistency, undercut, overlap, porosity, tie-in, reinforcement.

You're asking: Why does it look like that? What happened while that weld was being made?

That ability doesn't come from reading a definition once.

It comes from looking at weld after weld after weld.

So you're not necessarily trying to turn every welding student into an inspector.

No.

We're trying to make them better welders.

A welder who can look at their own work critically has an advantage.

If they can see the problem before I have to point it out, now we're getting somewhere.

Was WeldGrade born because you thought welding education needed AI?

No.

It came from trying to improve the learning process.

AI happens to give us another tool for doing that.

The workflow we're developing is pretty simple.

The student makes a weld.

Before the instructor tells them what's wrong with it, the student looks at it first.

They photograph it. They inspect it. They identify what they think they're seeing.

Then the AI can provide another opinion.

And after that, the instructor reviews the whole thing.

Student, AI, instructor.

Exactly.

And the order matters.

Why?

Because if the AI goes first and the student just copies what it says, I haven't taught the student anything.

I want them thinking.

“What do I see?”

Then compare that with what the AI sees.

And ultimately compare both with what the instructor sees.

That's where the learning happens.

Which brings us to the title of this episode.

Who grades the AI?

We do.

The instructors?

Absolutely.

AI is going to make mistakes.

Sometimes it's going to identify something incorrectly. Sometimes it's going to miss something. Sometimes the photograph isn't good enough.

The instructor has to be able to say:

“Yes.”

“No.”

“Here's what you missed.”

“Here's what actually matters.”

That's why I don't look at WeldGrade as an automated grading system.

It's an assisted inspection and teaching system.

Human in the loop.

Human in charge.

That's an important distinction.

Let me play devil's advocate.

I'm a welding instructor.

I've already got an LMS. I've got Blackboard or Canvas. I've got attendance. I've got paperwork. I've got students waiting for me.

And now Curtis shows up and says:

“Here's another piece of software.”

Why would I want it?

That's exactly the question we have to answer.

If WeldGrade gives an instructor more work, we've failed.

It has to give something back.

What?

Time.

That's probably the biggest thing.

If the student has already inspected their work, the AI has already provided observations, and I open that report and most of it is correct, I'm no longer starting from zero.

I'm reviewing.

Confirm this.

Correct that.

Add a comment.

Grade it.

Move on.

So the instructor becomes—

The editor instead of the first-draft writer.

That's a pretty different value proposition from “AI grades your welds.”

Very different.

And I think it's a better one.

You and the WeldGrade team have started discussing something I haven't seen very often in educational software: actually showing the instructor what the software gave back to them.

Right.

If we're telling instructors we're saving them time, we ought to prove it.

So one idea we're looking at is an Instructor Impact dashboard.

Maybe at the end of a semester it tells me:

“You reviewed 327 inspections.”

“You confirmed 149 AI assessments.”

“You corrected 74.”

And based on the workflow, maybe WeldGrade estimates:

You saved 6.8 hours this semester.

Now that's meaningful.

Because that's six hours that isn't spent staring at a screen.

Exactly.

That's six hours I can spend teaching.

Working with a student who's struggling.

Demonstrating something.

Or frankly, six hours I'm not spending doing paperwork after class.

But there's another side to those corrections.

Every time you tell the AI, “No, that's not undercut,” you've created something valuable.

You've contributed judgment.

You looked at a weld, considered what the system said, and explained where it needed correcting.

That's more meaningful than just clicking a button.

And if your expertise helps build WeldGrade, that contribution should matter.

That's where Calibration Credits come in?

That's the idea we're exploring.

A way to document meaningful instructor contributions. Human-verified inspections. AI findings reviewed. Expert corrections. The work that helps us calibrate the system around what instructors actually see.

The word “credits” can carry some baggage. Are we talking about cryptocurrency? Money? A piece of the company?

No cryptocurrency. They're not currently money, and they're not promised equity.

We're talking about recognition and a record of contribution.

We need to be very clear about that. I don't want an instructor participating because they think we've promised something we haven't.

And presumably not every click deserves the same recognition.

Right. The meaningful part is the expertise.

A thoughtful correction that explains what matters is different from just moving through a screen. We still have to work out how to recognize that fairly without creating another administrative job for the instructor.

You've also talked about a Founding Instructor Council. What would you want that to be?

A group of instructors who help shape this while we're still learning.

People who are willing to use it in their labs and tell us where it helps, where it gets in the way, and what we've overlooked.

I'd want that input to influence the curriculum connections, the feedback students receive, and the features we build next.

So more than putting somebody's name on a list.

That's the hope.

Recognition. Early feature access. Product feedback. Advisory participation. A community of instructors learning from each other.

It's conceptual at this point. We're not announcing an established program or offering compensation.

And any future financial arrangement would be a separate conversation?

Absolutely. Any financial or equity program would need separate formal structuring. It shouldn't be implied by a title or a credit balance.

For now, the question is how to give instructors a meaningful voice in something intended for their classrooms.

We've talked a lot about the instructor. What does the student get that they don't already get from practice and a grade?

A record of the learning.

Think about what happens now. A student finishes a weld. We talk about it. They get a grade. Eventually that coupon goes into the scrap pile.

And a lot of the evidence goes with it.

You can see the final grade, but you can't necessarily see how they got there.

Exactly.

I'd like a student to be able to look back at the first weld, the feedback, the correction, and the improvement.

Not just a folder of pictures. A record with context: what they noticed, what needed work, and what the instructor verified.

Over time, that can become a portfolio of demonstrated progress.

Something they could explain to an employer?

That's the potential.

Anybody can write “proficient in SMAW” on a résumé. I'd rather be able to show you the work.

That doesn't turn a classroom portfolio into a qualification test or a certification. But it gives the student something concrete to talk about.

Here's where I started. Here's the problem I learned to recognize. Here's what I changed.

And that comes back to looking at their own work.

We're teaching them how to see.

The record matters, but the thinking that produces it matters more.

There's another thing taking shape in that workflow: a weld image, an AI observation, a student's judgment, and an instructor's correction.

That sounds like potentially valuable data. How do you think about that responsibly?

Carefully.

The interesting part isn't just having lots of weld photos. It's the reasoning attached to them.

What did the student notice? What did the AI think it saw? What did the instructor confirm or correct, and why?

Potentially, that's expert-labeled visual reasoning about weld quality.

But a classroom isn't automatically a data source that a company gets to use however it wants.

Correct.

Student privacy, school agreements, permissions, and institutional responsibilities all have to come first.

We can't assume that because an image went through a teaching workflow, WeldGrade automatically owns it or can freely sell it.

Any broader use needs to be considered separately, with the appropriate agreements and permissions in place.

So there's a distinction between preserving a student's learning record and using that information to improve a system.

Yes. And the school and the student need clarity about those uses.

There's an interesting possibility here. But being interested in the possibility doesn't relieve us of the responsibility to handle it properly.

Where does that leave WeldGrade today?

Being tested against the real classroom.

We're not claiming we've solved welding education. We're trying to prove that this workflow helps instructors and students do something useful.

Can the student submit work without it becoming a distraction? Can I review it efficiently? Can I correct the AI when it's wrong? Does the feedback help the student on the next weld?

Those are the questions that matter in the pilot.

And you're teaching with it yourself.

Yes. Which means if something is annoying, I'm going to know pretty quickly.

[Laughs.] That's perhaps the best product-testing methodology I've heard all week.

[Laughs.] Welding students will let you know too.

What would make you say it's working?

Students doing more of their own thinking. Instructors spending less time repeating documentation and more time teaching.

And evidence from the workflow that supports those claims. Not just a demonstration that looks good on a screen.

AI is improving extraordinarily quickly. Let's imagine we're having this conversation twenty years from now. Computer vision is dramatically better. AI understands welding processes better. Maybe the system can inspect a weld more accurately than either of us can imagine today.

Does a welding student still need somebody like Curtis Casey standing in that lab?

Absolutely.

Because my job isn't just telling a student whether a weld passes.

I'm teaching them why.

I'm watching how they hold themselves.

I'm watching what they're doing before they ever strike an arc.

I'm teaching judgment.

I'm teaching responsibility.

I'm teaching them to recognize when something isn't right.

Technology can help with all of that.

And I hope it does.

But ultimately, we're trying to develop something in the student.

The inspector's eye.

The inspector's eye.

Not because they're all going to become inspectors.

Because they're going to become better welders.

And if WeldGrade gives them more opportunities to practice that—and gives the instructor more time to actually teach it—then I think we've built something worthwhile.

Curtis Casey, thank you for joining us.

Thank you. I enjoyed it.

WeldGrade is currently being piloted in welding education programs. You can learn more at WeldGrade.com.

I'm Rachel Torres, and this is The Next Arc.