I Was 98% Sure a Writer Used AI – What Jon Gillham Taught Me

I once hired somebody to create content, and I was 98% sure it came out of a machine. The giveaway, I thought, was the long, blustering introduction: I’m about to tell you how to do this thing, and then at the end, this article was about this thing. So I put the question to Jon Gillham, founder of Originality.ai, who builds AI detection for a living. Are windy intros the tell? His answer left me a lot less sure of my 98%, and if you’ve ever been sure about a writer, it should do the same for you.

My gut was the least reliable detector in the conversation

Humans are poor judges of whether a person or a model wrote something, and Jon gave me two reasons. To be clear, he never ruled on my writer, and I didn’t ask him to. He was talking about people in general, which includes you and me.

“Humans have some inabilities, some flaws and biases that make us really poor judges of whether or not we can tell if content is human generated or AI generated, and we’re massively overconfident.”

The first bias is overconfidence. Ask a room who’s an above average driver, he said, and 80% of the room puts up their hand. The second is pattern recognition: “we think we can recognize patterns when no patterns exist.” He pointed to blind studies aimed at professors, with educators “at like 60% probability” of correctly identifying AI content. Originality’s summary of those human detection studies today lists one study with college students at 60% and instructors at 70%, so hold his figure loosely. Either way, your gut misses a lot.

And before you swap your gut for a tool: Jon’s answer to why his own detector flags a piece is “we don’t know.” That’s from the man who sells it. You still have to make the call about the writer.

Staring at the abyss for AI will show you AI

I reached for a board game when Jon said this. You’ve seen somebody draw five of the same card in a row and demand a reshuffle. I’m the guy saying that’s a sign it was shuffled, because streaks are what randomness looks like.

Our pattern brains do earn their keep sometimes. You can walk along a creek, spot an odd piece of rock, and know it’s an arrowhead. Assign yourself the job of finding AI, though, and you’ll start seeing it whether it’s there or not. Jon’s addition was the better one: the gambling industry “is heavily built off of the back of people believing they have a system.”

If you review freelancers, that’s the trap. You went looking, so you found it.

I watched a two million dollar content budget disappear

When Jon said about 13% of the top 20 Google results were AI, I told him that was “a lot bigger than I thought it would be.” That figure is his from September 2024, from Originality’s own study of AI in Google results, whose written timeline now lists 17.31% for September 17, 2025 (I checked on November 6, 2026).

What I could add was one company. It had been spending at least two million dollars with a firm I worked with. All of that budget went away, and then it produced four times as much content from somewhere. I can’t prove what that somewhere was. I can tell you that much content on that budget wasn’t humanly possible.

Jon asked what you’d ask: “And how did it work out for them?”

Up, up, down. When I’d peeked in on them about four months before we recorded, they weren’t riding as high as they had been. It’s one company, so treat it as a story. It isn’t a study. Jon’s advice for you lines up with it, though: “if the thing that you’re trying to push out into the world is 750 words or 1,500 words about a specific topic” and it’s pasted out of an AI, he sees “a really, really low chance” of it lasting. Data plus words with a human in the loop gets a yes from him.

I brought the ouroboros, and Jon brought three “I don’t knows”

The snake eating its tail is how I framed it: more of the web gets written by models, then gets scraped back into data sets as if humans wrote it. What’s your read on that, I asked. This is the line to sit with.

“For the rest of humanity, all known large human data sets have been created.”

Then I noticed what he did next. He could speak to what’s on the web, because he measures it. On what that does to the models, he said “I don’t know” three separate times in one answer, while talking about synthetic data, o1 and what looked to him in 2024 like a plateau in language capability. Originality also tracks which sites block AI bots, a link he sent me afterward that we didn’t discuss on tape.

I trust people more when they do that. It’s the same reason I wrote that LLM optimization expertise is ankle deep. If somebody hands you certainty on this subject, ask what they measured.

I’ve seen Google use fear before, and Jon sees it in March 2024

Jon’s take is that the March 5, 2024 manual actions were “more of a psyops than an update,” with the manual actions timed to make the update look bigger. He’s offering opinion there, and he has a business interest in the AI question. His company did publish a study of those manual actions you can read for yourself.

It rang true to me anyway. I’ve been around long enough to know Google uses fear, “sometimes wisely, sometimes over-aggressively,” under different spam team leaders. I told him John Mueller has been fairly muted, with nothing as extreme as Matt Cutts’ days. You get slapped around to make people afraid of doing X, Y or Z.

So when a client asks me whether Google will punish them for AI, I’d start by separating the announcement from what actually got hit. That’s my advice and not Jon’s.

One study I can’t square with the tape

You should see this mismatch, since I’m the one who said “dumb it down” on the recording. Jon told me the grade eight advice for readability was “wrong high,” and that it depends on the industry. Originality’s readability study, as it reads now, advises a grade 12 reading level and reports no significant correlation with rank. I can’t reconcile those from here, so I’m not giving you a grade target. Jon also sent a how-to for checking readability in Google Docs, which tells you the score and nothing about what it should be.

Spotting AI writing: what I’d do instead of trusting my gut

If you hire writers, here’s where this conversation left me. These are my conclusions from what Jon said, so blame me for them.

1. Decide the policy before you suspect anybody. Jon’s old agency had policies and “no certainty.” Settle what’s allowed, and who gets the efficiency lift.
2. Treat a detector score as a reason to look. It can’t give you its reasons, and Jon calls false positives “painful in the world of writing.”
3. Review the page for what only a person could add. Ask what’s in it that couldn’t have come from averaging the top 10. Jon’s line: “Google doesn’t want to rank average.”
4. Keep your 98% to yourself until you’ve got more than a windy intro. That’s a note to me as much as to you.

The bigger answer is to stop publishing word-only pages, and that’s why I do these interviews. I use AI tools for ideation. Dr. Christiane Schroeter’s warning about the yes man is a good next read on how to keep them in that lane. But I had no idea, in detail, what Jon did before this call, and those 38 minutes gave me more than asking a model about fact checking would. As I told him at the end: “These LLM tools will never experience a sunny day.”

For the full episode recap, read Jon Gillham on AI content detection on Unscripted SEO, and for the agency version, Jon’s words vs data test on SEO Arcade. Jon’s at originality.ai, LinkedIn and X.

Sponsor disclosure: uSERP, Oleno, ClearLead Digital and OppAlerts are paid sponsors of the Unscripted podcast network. Links to them are marked as sponsored. Jon Gillham has no connection to them.