When the first commercial AI detectors arrived in late 2022 and early 2023, it wasn’t exactly an auspicious start. News articles and social media posts showing AI detectors flagging the Declaration of Independence and the opening lines of the Book of Genesis as AI-generated were soon making the rounds. Either the Founding Fathers were also using AI to cheat on their homework, the comments sections quipped, or AI detectors did not work as advertised.
These false positives — instances when detectors label human-written text as AI-generated — were due to how the initial models worked. They evaluated how surprising a given text’s word choice was (a measurement called perplexity) alongside how often those surprising choices were dispersed throughout (burstiness). Since we know large language models (LLMs) produce text by predicting what statistically fits next in a given sequence based on their training data, a low perplexity-burstiness combination should signal a high likelihood of AI authorship.
It did not. Being as often cited and quoted as they are, the Declaration and the Book of Genesis proved anything but surprising, revealing a weakness in this approach. Now, in addition to worrying about whether we could trust that something we read came from an actual person, we also had to keep a watchful eye on the watchmen.
Three years on, and that worry hasn’t abated. True, AI detectors have improved, but false positives still occur, and that leaves some in the unenviable position of having to defend their hard work against accusations of AI slop.
As a writer and editor, I need to understand these detectors and how to use them without ruining the trust I’ve built with my readers and fellow writers. To help me start figuring that out, I spoke with Max Spero, the cofounder and CEO of Pangram, a popular AI detector. We discussed how today’s AI detectors work, how these probabilistic tools fit into today’s culture, and the social norms we can start building around them to help us preserve authenticity in writing.
This interview has been edited for length and clarity.
Big Think: Why did you start Pangram, and why will AI detectors be important in this new AI era?
Max Spero: We were thinking about how we could build for a post-AI world and what that world would need. That’s when we landed on detectability technology. Even if not intelligent, AI seems intelligent and can produce large volumes of content. If that is completely indistinguishable from [what is] human, then we have a problem on our hands.
When [Anthropic CEO] Dario Amodei was at OpenAI, he was saying that GPT-3 is too powerful to release to the world [because] a bad actor could generate hundreds of thousands of fake news articles and post them online. That sort of thing.
In retrospect, the technology wasn’t there, but I think we’re seeing this today. There are maybe two sides [to consider]. On the one hand, bad actors are using AI to scale up inauthentic behavior that is bad for the information ecosystem. On the other hand, people are offloading their cognition and work to AI in ways that are inauthentic and also bad for the information ecosystem.
Big Think: When you talk about people offloading their cognition, there’s an implication that this isn’t like other tools. But some argue that AI is a tool that can improve writing and clarify thoughts, like a spell checker.
Spero: Anyone who compares AI to a calculator or spell checker fundamentally misunderstands how powerful this technology is and how much it’s going to change society. It’s completely different.
You can use AI in a high-effort way, like an editor. But the vast majority of what we’re seeing is extremely low effort. Usually in the course of writing something out, you’re able to reconcile your thoughts. Entering them into words makes them more coherent. But when you’re using an LLM instead, it’s just auto-completing a half-baked thought into an essay that [only] looks well-written.
In the past, there was this effort disparity [with writing] where it took more effort to write a piece of text than to read it. That is no longer the case now. There’s a lot of writing out there — it’s hard even to call it writing — that didn’t have even as much effort put into it as what it would take to read.
Big Think: When it comes to AI tells, a lot of the conversation focuses on specific patterns that are easy for people to spot: triads, em dashes, and negative parallelism. What is the Pangram algorithm looking for?
Spero: For the first part, the patterns that people pick up on are real. Part of the reason AI writes this way is that it has been trained to produce good writing (in a sense). It sees these constructions that good writers use, and then the reinforcement learning process dials that up to 11. But LLMs have no concept of the distribution of text that they produce. All they know is what they are producing in this single instance. What they don’t understand is that good writers use these constructions when appropriate, not all the time. That’s why we see this overuse.
I think this will go away over time as AI labs get better at making AI writing sound more natural. But that’s not what Pangram picks up on.
Pangram uses a higher volume of weak signals because there’s a lot of AI writing that doesn’t have a bunch of em dashes or negative parallelisms but is still AI-generated. Instead of looking for just these few smoking guns, it looks at every single word and punctuation choice and asks, “Is this extremely consistent with how AI makes its decisions?” Human-written text will sometimes line up with how AI makes its decisions, but not consistently throughout the entire text.
Big Think: You claim a false positive rate of one in 10,000. How did you manage that, specifically with regard to edge cases?
Spero: Most other AI detectors use burstiness and perplexity. If the text is unsurprising, then these other AI detectors will say that it’s AI. This works sometimes, but it is also why the Declaration of Independence gets flagged.
We threw perplexity out the window and said we’re going to train a classifier model. We’re going to use deep learning, use a lot of data and a lot of scale to do better.
We have a diverse corpus of human-written text, and for each text, we ask AI to write the same thing. Say, this one is a 500-word essay on Moby-Dick written by a 10th grader. Then we have the AI write a 500-word essay on Moby-Dick in the style of a 10th grader on the same topic. That’s what we call a “synthetic mirror.”
When Pangram is being trained, it sees both of these and learns the difference between the choices humans and AI make. That gets us a base model that does pretty well, usually around 99% accuracy. And then we want to make it better. This is where active learning [comes in].
With active learning, we use this 99%-accurate model to search a larger corpus of text examples, find the ones it still fails on or is unsure about, and then bring those back into the training set by creating synthetic mirrors for each. Basically, we’re finding human text that’s closer to the human-AI boundary, and because that’s harder for our model, it learns better.
Big Think: And you do that over and over again, and it just gets better and better?
Spero: Yeah, that’s how we drive down the false positive rate to one in 10,000.
Big Think: When you do hit a false positive in testing, are there any patterns that you, as a human reader, can pick out to understand why that happened?
Spero: AI tends to overuse specific constructions for certain domains. If you look at restaurant reviews, AI will sound like, “I recently had the pleasure of going to this restaurant,” or “From the moment I walked in, it was great,” or “It’s hidden in the heart of Brooklyn.” Those kinds of things.
With Pangram, the most important thing is to be transparent about the error rate so that people can reason and understand how to use the technology.
[Also,] the more words Pangram has, the more data it has and the more accurate it will be. That’s why most of the false positives we see are in the 50- to 100-word range where Pangram just has a lot less data. It’s more making an educated guess rather than a very confident classification.
We understand that Pangram’s never going to be 100% correct. That’s probably impossible unless we had access to the ChatGPT and Claude databases.
Big Think: Have you asked?
Spero: They said no. Kidding, kidding.
If you look at a false positive, sometimes they’re very understandable. Sometimes they’re not, and those are the really strong candidates to bring back to the model and retrain on, because it shouldn’t be making these mistakes.
Big Think: How do you keep Pangram up to date given that LLMs are being constantly updated? Then there are humanizers to consider, too.
Spero: Pangram generalizes really well to new models within the same family. When Claude Fable 5 came out the other day, Pangram did very well. Same with GPT-5.6. And we’re retraining regularly to make sure we are ready for Fable 5.1 and GPT-5.7.
Humanizers try to erase and reverse any signal that something has been written by AI, so it’s definitely a harder task.
For example, we found a humanizer that figured out if you add a bunch of zero-width spaces between every character — they’re like Unicode characters that don’t show up visibly so the text reads normally — Pangram would parse it as a bunch of spaces in the middle of the text and say the text was human even though it was AI.
Ultimately, we patched it by figuring out that we could preprocess out zero-width spaces, but this is the cat-and-mouse game that is probably going to go on. It’s more adversarial, and so it’s going to be back and forth.
Big Think: Geoffrey Fowler at the Washington Post once wrote that “With AI, a detector doesn’t have any ‘evidence’ — just a hunch based on some statistical patterns.” Others have leveled similar criticisms. Can you speak to those?
Spero: AI detection is a statistical method, and there are a lot of comparable statistical methods.
[For instance], DNA testing doesn’t check every single gene in every single chromosome and make sure that they match exactly. Instead, it checks many markers and ensures that they match. In the wild, DNA testing has a false-positive rate of about one in 100,000. Even still, with DNA testing, at least in the United States, it is not enough evidence to convict someone of a crime. You need other corroborating evidence to go beyond a reasonable doubt.
Similarly, with Pangram, the most important thing is to be transparent about the error rate so that people can reason and understand how to use the technology.
Probabilistic methods are powerful, and oftentimes they’re the only way to make a determination. I think a lot of people aren’t ready for the fact that this is the way the world is going. All of the most powerful technology today works in a probabilistic way, [including] AI.
Big Think: What are your thoughts on the way people are using Pangram online to call out examples of writing they believe are AI-generated?
Spero: I think it is emblematic of this rising distaste for and backlash against AI and people overusing it for everything. I talk to so many people who tell me, “I hate using LinkedIn now. It’s just full of AI slop.” And we just released a study where we found that 40% of LinkedIn’s long-form content is AI-generated, which is an insane amount.
People are sick of seeing slop, and that’s why they want to have this gotcha. People are trying to set the norms here: “Don’t put this on my feed. This is not cool. I don’t want to read this.” It’s incredibly disrespectful when you see AI content and comments that are pretending to engage with you, but it’s not a real person behind it.
Big Think: How would you recommend we approach developing new social norms around AI detectors?
Spero: The first thing is always to have a conversation, especially when you’re in a trusted environment. Oftentimes, when someone is in a role where they evaluate text all the time, like an editor or teacher, they know when something is AI-generated. They’ve trained their internal detector well, and they’ll use Pangram to have something to point to so it’s not just their internal suspicion or intuition.
I think Chris Best, [the CEO of Substack], put it really well. He said that a lot of people who use AI have a bit of a spinach-in-your-teeth problem. They don’t realize that other people can look at your text and instantly clock that it was either written by or largely assisted by AI. Many people still believe that it’s completely indistinguishable, but it’s not.
Big Think: What about on the social media side of things? Is it appropriate to post a “100% AI” screenshot and point fingers?
Spero: We are in a time of norm-building. I’ve seen plenty of people on Twitter who said they were experimenting with using AI to write, and now they’re back to writing themselves because something about it didn’t feel right, or they [realized] this is something that other people don’t want to read.
I think it’s important to be able to say publicly that I believe this text is AI, and I don’t want to read it. If we don’t have this, the incentives shift to everyone should be using AI to pump out as much content as possible because every algorithm rewards quantity over quality. If we make it taboo to point out someone passing off AI content as their own, then we enable slop peddlers to pollute the commons and make the internet worse for everybody.
Big Think: Why is it important for us to get this right?
Spero: We’re moving into a world where the concepts around trust and authenticity are fundamentally changing, and they’re changing quickly compared to any other societal change in the past because this technology is moving so fast. If we don’t get this right, we could end up in a world with extremely low trust, where everyone assumes everything on the internet is AI-generated and everyone is a bot.
This sends us into a society with negative incentives. Suddenly, having an open grant program is costly to run because reviewing tens of thousands of applications becomes nearly impossible. Having a scholarship contest for students also becomes nearly impossible because the cost of writing an essay is a quarter of a second, and the cost of reviewing the same essay is so much more.
It’s not necessarily about saying AI is bad, but as AI becomes more prolific, we need to be prioritizing human work.
Big Think: Are you a no-AI guy then?
Spero: Oh, it’s an incredibly powerful tool. I use AI daily for fact-finding, research, and writing code. But I never use it to write. I will never put AI words in my mouth.
This article Pangram CEO: People believe AI writing is indistinguishable from ours. It’s not. is featured on Big Think.