AI Detection, Authorship and Academic Integrity: The Strange New Science of Detecting AI (#685)

I recently did something that, a few years ago, would have made absolutely no sense.

I took one of my own blog posts and submitted it to an AI-detection program.

The result came back:

7% of this text appears to be AI-generated.

The accompanying graphic was reassuringly precise. Seven percent of the text resembles AI text. Ninety-three percent showed no AI text patterns.

My first reaction was amusement.

My second was curiosity.

What exactly does 7 percent mean?

Did artificial intelligence write 7 percent of my blog? Which 7 percent? Was it a paragraph? Several sentences? A collection of phrases? Was it my vocabulary? My sentence structure? The organization of the argument?

Perhaps I had simply written something that sounded sufficiently orderly, predictable or polished to resemble the statistical patterns associated with artificial intelligence.

And then another question occurred to me.

What if the number had been 27 percent?

Or 47 percent?

Or 87 percent?

At what point would curiosity become suspicion?

And at what point would suspicion become accusation?

That is where an apparently entertaining little experiment becomes a serious educational question.

Imagine that this was not my blog.

Imagine instead that it was an essay submitted by a university student.

The instructor runs the paper through the same detector and receives a result:

7% AI-generated.

Probably nothing to worry about.

But suppose it says:

37%.

Or:

73%.

Has the student cheated?

We do not know.

Yet something important has already happened.

A number has been attached to the student's authenticity.

And numbers carry authority.

They look objective. Scientific. Quantifiable. Impartial.

73% AI-generated somehow sounds much more compelling than:

I have a feeling this student may have used AI.

But are those percentages really measuring authorship?

That is the question at the heart of the rapidly expanding use of AI-detection programs in education.

For centuries, education has depended upon a remarkably simple assumption: when a student submits a piece of work, the work represents the intellectual effort of that student.

The essay may not be brilliant. The argument may be incomplete. The prose may be awkward. There may be spelling mistakes, clumsy transitions and conclusions that do not quite follow from the evidence.

But those imperfections have traditionally meant something.

They are evidence of a mind at work.

Generative artificial intelligence has complicated that assumption.

A student can now ask an AI system to explain a concept, construct an argument, produce an outline, rewrite a paragraph, improve grammar, identify weaknesses in reasoning - or generate an entire essay.

The educational world has responded predictably.

If machines can generate writing, perhaps machines can detect it.

And so, we have entered the age of the AI detector.

At first glance, the logic seems irresistible.

AI creates the problem.

AI detects the problem.

Academic integrity is restored.

Except there may be a more interesting possibility.

Perhaps the detector is not detecting who wrote something at all.

Perhaps it is detecting how much human writing resembles what a machine has learned human writing should look like.

And that is a very different proposition.

Commentary

I particularly like the progression from 7% → 37% → 73%.

It exposes a central problem beautifully:

Where, exactly, is the threshold at which an algorithmic probability becomes an accusation of academic misconduct?

#bunyipsbooksbeyond