So What? How AI Detectors Work

Katie Robbert – 00:00

Well, hey, everyone. Happy Thursday. Welcome to So What?, the Marketing Analytics and Insights live show. I’m Katie, joined by Christopher Penn and John, who are up top with their blue shirts today. Yes, I didn’t get the memo. I am wearing an olive green color. So here we are. Someone has got to be the odd man out, so it might as well be me.

This week, we have talked about it on the podcast, we have talked about it in the newsletter, and today we are going to show how AI detectors work. This came about because last week we had a very strong response to our newsletter claiming it was AI slop, written by AI, and that we were liars about our use of AI.

We cannot say for sure that the person who responded used an AI detector. Timing-wise, Substack is using Pangram, which is an AI detector—I am putting this in big air quotes. I have to get my red flag ready, too. They are using tools that allegedly can detect AI-written content.

Here is the challenge: these large language models are trained on human-written content. Now it is telling us that the content we have written—which we did not consent for them to be trained on—is AI-written content. That is a whole other ethical topic for another day. Today, we are going to walk through how to use an AI detector and maybe how to build an AI detector.

Before we get into all that technical stuff, John, what are your thoughts on AI-written content? Do you care? Does it matter? Are you looking to see if something is written by AI or not? Where do you stand on this topic?

John – 02:15

That is a good question in general. I do not really care about AI writing. Most of the time, people are using AI to generate slop. They want to write some kind of article, and it is just going to grab whatever is out there on the web and process it. If somebody just does something with AI and throws it out there saying, “Hey, here is my new thing,” that is usually of no interest.

Where it is of huge interest is if you are working on something and want it to grab everything out there and put it in a presentable form so you can learn what is going on and figure things out. That is fantastically useful. It is basically the promise of search that was stolen from us and bastardized by marketing.

There are great things that can be done with it. However, all the AI-generated stuff is here, but unique and novel research—stuff that is original or created—is at a whole other level. In fact, just this week, we have been seeing that stating something is made without AI actually adds marketing credibility and shows either luxury or a brand above the general public.

There is a lot going on. We have seen a lot of stuff with AI detectors. Last I heard, this was all garbage, so I am interested to hear where we are going today and the state of the world today.

Katie Robbert – 03:45

I find it interesting that marketers come to the rescue to rebrand things. 100 percent human-written content with no AI used is now something sought after, whereas for years that was just how it was. Now you can slap a label on it and say “100 percent human.” The whole thing is very bizarre to me.

Christopher Penn, we are going to turn it over to you and figure out where we get started. What is the deal with these AI detectors?

Christopher Penn – 04:30

With AI detectors, the vast majority today use an AI model custom-tuned to look for specific patterns. The specific pattern they look for most is the probability of the next word in the sequence. If too many words in a row are expected, it is going to say that this is probably AI.

I will give you an example. If I say to a North American audience, “I pledge allegiance to the…” what is the next logical word?

Katie Robbert – 05:00

The flag.

Christopher Penn – 05:02

Flag. Exactly. Flag. There is an almost 100 percent probability that the word is going to be “flag,” and almost zero percent that it is going to be “rutabaga.” No one pledges allegiance to a rutabaga. If you do, leave a note in the comments.

Katie Robbert – 05:15

They might have to today.

Christopher Penn – 05:18

If an AI model looks at a word sequence and sees the word “rutabaga,” it thinks, “Well, that is weird. That is a very low probability word and not what I would expect. That is probably human-written because it is so low probability.”

Here is the challenge: there are a lot of uncreative writers, especially technical writers. The probability of the next word being whatever it is will be very high because there are only so many ways to talk about a certain thing. The sun is bright; the sun is yellow; the sun is shining.

That is a basic example, but someone coming up with a flowery analogy of what the sun looks like is likely not using AI in the first place. For the majority of people describing the sun, they will say the things they know about it. It will be basic pattern matching, but they wrote it, and now it is going to get flagged as AI.

Katie Robbert – 05:50

I think about students in school who are trying to get through an assignment. They think, “What do I have to say about the sun? The sun is in the sky. The sun is blindingly bright. I need sunglasses when it is sunny.” We are all going to say that.

That is hugely problematic because it is trained on the way that we talk and write. How could it flag that as AI when it was trained on how humans write and speak?

Christopher Penn – 06:15

Everybody starts with a base model of some kind. Every company claiming to offer an AI architecture has to custom-tune the model or software around it to define what is and is not AI-written.

We know this because AI models have what we call “tells”—constructions they overuse. For example, negative parallelism is a very common tell for both ChatGPT and Claude. It is not this; it is this. It is not the death of attribution; it is the arrival of multi-touch attribution. You see that construction a lot. It is a human construction, but AI tends to overuse it.

Another one is triadic rhythm—things that come in triplets. Maybe it went well; maybe it stalled; maybe it never started. Boom, boom, boom. Claude, in particular, loves triadic rhythm. It has obviously been trained on that material, and it does that a lot.

Claude and ChatGPT both like passive voice starters. You will see this in LinkedIn comments especially. They start with a passive voice starter that does not say anything and is just a restatement. One of the best ways to look for AI-generated text is in comments at a profile level on LinkedIn because you see recurring patterns. To use a Claudism, the pattern is the tell.

Here is an example. Looking at the general text shape of the sentences, you see shorter first lines, longer second lines, and shorter third lines. Looking at the shape of the words as colored lines shows how formulaic it is. “The AI audit only works if it knows whose eyes are being messaged. This amendment moved the scope boundary, not just the compliance dates. The most telling choice in the playbook is where agents sit.”

When you look at the pattern, you can tell this is a bot generating the same format over and over again. The underlying language model varies the words. Early AI detectors used static features like sentence length or noun counts, but almost no company uses that anymore due to high false accusation rates.

Today’s AI detectors use language models. The false accusation rate is lower, but still not zero. Zero is the only acceptable rate for anything punitive. You must never use an AI detector for punitive actions like ending someone’s tenure, kicking them out of school, or disciplining them because the tools cannot falsely accuse someone.

Katie Robbert – 09:00

It is funny because you talk about the shape of the content, whereas I looked at that and thought it was written by AI because it is utter nonsense. It looks like a structured sentence, but reading the words, “The audit only applies to the eyes reading it,” is not a sentence a human would write because it does not make sense or mean anything.

Christopher Penn – 09:30

AI detectors especially penalize non-English speakers, neurodivergent people who write in non-standard ways, and highly templated text. Anyone who remembers school from the last 20 years remembers the five-paragraph essay: stating what the paper is about, three paragraphs of exposition, and a conclusion restating key points with transition sentences. At that point, you may as well have ChatGPT write it for you. That formulaic template gets flagged as AI because it is so strict.

Katie Robbert – 10:15

That also does not teach you how to write. Paragraph one is the argument; paragraphs two, three, and four support it; and paragraph five is the conclusion that ties back. It is almost like the 5P Framework.

Christopher Penn – 10:45

There are definite tells AI uses that humans are skilled at identifying, especially sentence-shaped objects that do not say anything. Looking at that LinkedIn example: “The audit only works if it knows whose eyes are reading the message”—that does not mean anything. “The amendment moved the scope boundary, not just the compliance dates”—that one almost makes sense if you squint, but there is no reason to frame sentences that way. It sticks out as AI doing its best to sound like a competent human and failing.

Bots take the original post and wash it through a prompt to generate a templated response. These bots are not terribly difficult to make. This speaks to why you would use an AI detector. Valid use cases include detecting whether someone is attempting to deceive you—such as deepfaking images or videos to show someone is alive—academic integrity, or assessing social status claims.

Christopher Penn – 13:05

You can give a language model a list of anti-patterns not to use in its work, along with a scoring rubric to review and remove them. In the Trust Insights Google Drive, we have one of these tools that I made last week.

Katie Robbert – 13:30

I appreciate that. I always love learning something new on these shows. At least it is not being demonstrated live.

John – 13:45

Let’s not get ahead of ourselves.

Christopher Penn – 13:50

How do these things work? First, you need a language model that returns log probabilities—a metric guessing the most probable next word in a sequence. “God save the…” is going to be “king” or “queen,” probably not “rutabaga.” Every company building these must pick a base language model balancing intelligence and speed.

In December, we put up a paper on Zenodo discussing how Meta’s Llama 3.2 model is fast but biased. Running it through the BBQ (Bias Built into Queries) test, Meta’s models fail 50 percent of the time on multiple-choice questions involving bias.

Katie Robbert – 15:15

We will put a link to that paper in our free Slack community at TrustInsights.ai/analyticsformarketers.

Christopher Penn – 15:30

You need a model to start with, a fast server, and a training corpus to build internal benchmarks of what is and is not AI.

For a project, I gathered LinkedIn posts from people who adamantly do not use AI, such as Ann Handley, Jay Acunzo, and my own unassisted posts. I then fed those topics into Claude, ChatGPT, Meta, Perplexity, Qwen, Google Gemini, and Copilot to write synthetic posts for each topic. This paired human and machine examples one-to-one for model training.

Model training takes a couple of days and significant compute. Rather than spending Trust Insights money renting compute from AWS or Google, I ran it on my MacBook.

Katie Robbert – 17:45

We appreciate that. The CFO thanks you.

Christopher Penn – 17:55

Finally, you test it. My friend Becca’s human-written posts kept getting flagged as AI. When I asked her about it, she explained that she dictates her posts and uses Claude to clean them up. When AI edits text, it rearranges words back into statistically probable patterns.

For example, Yoda saying, “For 800 years have I trained Jedi. My own counsel will I keep on who is to be trained,” gets rewritten by AI as, “I have trained Jedi for 800 years. I will keep my own counsel on who is to be trained.” Same words, but in a statistically probable word order that triggers AI detectors.

Every detector company—Pangram, Originality, GPTZero, Winston—uses undisclosed base models and proprietary training libraries. For my tests, I used Beaver AI’s Skyfall model—an upsample of Mistral fine-tuned for writing fiction and nonfiction.

Running text through Skyfall color-codes words green for unexpected words and red for expected words. In testing Katie’s newsletter text, it scored “probably not AI” for nonfiction, with high predictability flagged around terms like the 5P Framework. Conversely, AI-generated text displays dense clusters of red.

Katie Robbert – 21:30

I appreciate that it thinks my writing is probably not AI. I shared a snippet of the unedited draft in Slack. My first draft is generally pretty rough, and I disclosed using AI support to clean up my writing.

Christopher Penn – 22:15

The unedited draft scored “definitely not AI” with much higher perplexity and surprise levels. From an AI detection standpoint, rough cut drafts perform better because they lack predictable word patterns.

Katie Robbert – 23:00

That is the conundrum. Publishing raw, unedited writing might avoid AI flags, but if it is unreadable, it is not useful. People want to publish polished, engaging, educational, and entertaining content.

Previously, John was my human editor who made sure my content was grammatically correct and coherent without rearranging my sentences. AI editing tools try to be helpful by rearranging content. Now, I instruct tools like Google Gemini not to rewrite text, but to provide a list of suggested changes for me to decide on.

Christopher Penn – 24:30

To preserve your voice, instruct the AI to operate strictly at the sentence level: it may reorder sentences, but it may not reconstruct them.

Evaluating custom datasets shows another problem with detectors: training corpora skew toward educated, wealthy, majority demographics, causing higher false positive rates for non-native speakers or neurodivergent writers.

Katie Robbert – 26:15

Regarding the 5P Framework book, you assembled a version using AI from everything I have written. I was hesitant because I felt I should write it from scratch, even though I created the framework and wrote the underlying material. If someone runs the assembled book through an AI detector, it will likely flag it. Does AI detection measure what actually matters?

Christopher Penn – 27:30

AI detectors measure statistical word probability, not quality, helpfulness, or value. If someone objects to AI on moral or environmental grounds—such as energy consumption or unethically sourced training data—detection matters to them. If the goal is practical value—saving time or money—the detection score matters less than the quality of the underlying source material.

Katie Robbert – 29:00

Since I developed the 5P Framework, using AI to organize my previous writing makes sense, but people might dismiss it as “AI-generated.”

Christopher Penn – 29:45

You can break your corpus into a database by topic—purpose, people, process—and have AI select and assemble your original sentences into a coherent structure. This preserves your exact syntax while organizing the content.

Alternatively, you can benchmark text against 40+ stylistic metrics from your known human writing and instruct the AI to match those scores, effectively neutralizing AI detection algorithms.

Katie Robbert – 31:15

AI detectors are problematic, but how people misuse them is the real red flag. What do you think, John?

John – 31:30

Unless you are on a crusade, no. It comes down to whether the content is useful. AI detection could serve as a screening filter to deprioritize automated slop, but because scores represent probabilities, you can never say definitively whether something is AI-generated.

Christopher Penn – 32:15

If your audience values 100 percent human-crafted content, double down on transparency. At Trust Insights, our RAFT framework emphasizes transparency—disclosing if and how AI was used.

The real issue is misrepresentation: claiming work is yours when it isn’t.

Katie Robbert – 33:45

That is not a new problem; AI just amplifies it. People often go on autopilot, creating reports without understanding the underlying data. AI removes the manual effort of building the deck, making it easier for people to skip critical thinking.

Christopher Penn – 35:15

AI detectors contribute to the problem by focusing on output volume rather than encouraging metacognition and critical thinking. AI didn’t create the problem of people not using their brains; it amplified it.

John – 36:00

In smaller organizations, a report or dashboard isn’t the final product—the goal is understanding what is happening in the real world and what the data means.

Katie Robbert – 36:30

AI handles repetitive tasks, freeing up time for human analysis and critical thinking. Using AI detectors to dismiss content blindly misses the point.

Christopher Penn – 37:15

To prevent AI from sounding like AI, give it explicit linguistic negative constraints: forbid negative parallelism, triadic rhythm, uniform staccato, rhetorical Q&A, and meaningless filler sentences. Have the AI score and remove these patterns from its output.

John – 38:15

Today’s human statement: take heed.

Katie Robbert – 38:30

Thanks for tuning in. Subscribe to the show, check out the Trust Insights podcast at TrustInsights.ai/tipodcast, subscribe to our newsletter at TrustInsights.ai/newsletter, and join our free Slack community at TrustInsights.ai/analyticsformarketers. See you next time.