Max Spero

The through-line

Max Spero's consistent argument is that AI detection is not primarily about catching cheaters. It is infrastructure that lets a teacher, an editor or an institution draw a line and know whether it was respected. He is blunt that "using AI in and of itself is not plagiarism. It's not cheating" [1], and that the problem arises only when an instructor has said, in effect, "I want this in your own words. I want to hear what you have to say about this, not what chat GPT has to say about this" [1], and the student ignores that. Everything else follows from this framing: accuracy matters because a false accusation destroys the trust the boundary depends on, and adoption matters because a tool nobody uses sets no boundaries at all.

The second, quieter through-line is that this is a credibility problem before it is a technology problem. He believes the detection works and that the field's reputation was damaged by earlier tools; his job, as he describes it, is getting people to try Pangram at all. The same concern extends outward from classrooms to journalism, verification and the general question of what pervasive AI text does to the internet [5][8][10].

On what a detector is actually looking at

Pangram works stylistically rather than by any watermark or metadata. The system examines "the style of text, say an essay or any other written content, and we're looking what are the stylistic ticks and choices that AI makes more consistently than humans," and then surfaces that as evidence for why a passage is or is not machine-written [1]. The output for a faculty user is a report that either says the document is clean or points to where in the submission the AI text sits [1]. He has gone further into the underlying techniques, and into the twin risks of false positives and false negatives, in conversations aimed at general and financial audiences as well as educational ones [10].

On accuracy, false positives, and why the number has to be extreme

He is willing to make a hard quantitative claim and to show his working. Three academic studies from computer science departments, some peer-reviewed, have found Pangram to be over 99% accurate, measuring both directions of error: human text wrongly flagged as AI, and AI text that slips past [1]. Against that he sets a rival tool's own published benchmark, which put its accuracy on GPT-4 output at about 74 percent, a gap he considers decisive [1]. The false positive figure is the one he leads with: "our false positive rate is 1 in 10,000," derived by running the detector over "hundreds of millions of documents" that are known to be human because they predate 2022 and the arrival of ChatGPT, and finding fewer than 0.01% flagged [1].

The reason he dwells on this is that he understands the institutional cost of being wrong. Older detection systems, he acknowledges, carry bias against non-native English speakers and higher false positive rates, and many faculty already know this [1]. A single wrongly accused student pulls in grievance procedures, administrative time and money, which is why he treats the false positive rate as the number that determines whether the product is usable at all rather than as a marketing statistic [1].

On boundaries, and preparing students for workplace AI governance

The most developed idea in his thinking is that classroom rules about AI are a rehearsal for professional ones. He accepts that students entering companies will be expected to use AI to be more productive, but adds that "they're going to need to learn the norms of using AI which is not trying to pass off something that came out of AI as your own but instead being transparent and using AI to make yourself more productive learn faster not to kind of like cheat your way and like pretend you had greater output than you did" [1]. His example is drawn from his own former world: "If I work at Google and I am pasting my code into chatgpt and having it help me write my code, then like that's a huge red flag," because internal code leaks to OpenAI, and "You don't want something you upload to go into OpenAI's training set" [1]. Setting guard rails early, and expecting students to be capable of following them, is therefore "a completely reasonable thing to do" [1]. He endorses the hybrid assignment design where a student is permitted to use AI for one stage and required to write unaided in another, with detection used to confirm the instruction was followed rather than as an alarm [1]. He has also taken part in conversations laying his approach alongside other teacher responses to AI in the classroom, presenting detection as one of several viable strategies rather than the only one [3][4], and in back-to-school discussions of the pressures schools face as the technology moves faster than institutional policy [2].

On how it gets deployed in institutions

He is deliberately flexible about the commercial shape. An individual faculty member can buy a license and get a dashboard for uploading student submissions and generating PDF reports; there is a free tier allowing up to five checks a day for anyone who wants to try it; and universities can buy institutional licenses covering all faculty, which he says brings the price down substantially and, more to the point, "honestly get more people using Pangram, which is always our goal" [1]. Reactions across the sector span the full range, from faculty who follow the AI-in-education literature closely and already know the limitations of older detectors, to people who did not know accurate detection was possible; in both cases he finds that the reaction comes from trying the tool rather than from being told about it [1]. He is candid that higher education is a difficult market for a startup, with procurement, committees and shared governance slowing purchases even when there is genuine interest [1].

On the parts of founding a company that come easily and the parts that don't

His background as a machine learning engineer means the technical work is the comfortable part. Building the product felt "fun and natural" even during 100 hour weeks, and he recalls the early days fondly: the two co-founders renting a house in San Diego for a week with no distractions, "eating fish tacos and writing code" [1]. The hard part is commercial. "The main thing for us is like most people just don't know we exist," he says, naming obscurity as the real barrier [1]. Marketing runs against his temperament, and he catches himself hedging in real time: he describes needing to "get out of my shell and get out of my head and, you know, talk about how good the product is accurately," noting that he had just said people were "fairly impressed" when the truthful version is that they are really impressed [1]. He has separately walked through the mechanics of raising capital for the company [6].

On AI text beyond the classroom

Detection, in his account, is a general verification problem rather than an education one. He has extended the argument to what pervasive machine-written text does to the wider internet, including deceptive chatbots and the degradation of user-generated platforms such as Reddit [10], and to journalism specifically, where the questions he raises go past whether AI text can be detected to what over-reliance on the technology does to the work itself [5][8].

Takeaways

• Detection is framed as boundary enforcement, not accusation: "using AI in and of itself is not plagiarism. It's not cheating," the problem is doing it against an explicit instruction [1].
• Pangram identifies "the stylistic ticks and choices that AI makes more consistently than humans" and shows where in a document the AI text appears, rather than returning a bare score [1].
• Three academic studies put the tool above 99% accuracy, against a competitor's self-reported 74% on GPT-4 output [1].
• The claimed false positive rate is 1 in 10,000, validated against hundreds of millions of documents written before 2022 [1].
• Classroom AI rules are training for corporate AI governance: pasting internal code or company financials into a chatbot is the professional failure mode students should learn to avoid [1].
• Distribution is layered deliberately: five free checks a day, individual faculty licenses, and cheaper institutional licenses because the goal is maximum usage [1].
• Building the product was the easy part; the hard part is that "most people just don't know we exist," and self-promotion runs against his instincts [1].
• The same verification problem extends to journalism and to the open internet, including chatbots and platforms like Reddit [5][8][10].