In the 1998 poker classic Rounders, a young Matt Damon realizes that his nemesis Teddy KGB slowly eats an Oreo cookie when he has good cards. A ‘tell’ in poker is a giveaway that subconsciously leaks how a player is feeling about their hand. Someone who is slouching might suddenly sit upright after looking at their cards, or a player who had been distracted by the crowd is now staring intently at their opponents.
Just like human poker players, it turns out that Large Language Models (LLMs) like OpenAI ChatGPT, Anthropic Claude, and Google Gemini also have tells that giveaway their AI-generated writing. When a student uses AI to generate an essay, the machine doesn’t just produce text. Each chatbot has a distinct pattern of linguistic choices and structure.
But just like an old poker pro reading the table, a professor who knows what to look for can easily spot the synthetic cadence of a student essay that’s presented confidently to look like a winning hand. But spotting tells is about instincts, not exact science. Confronting students (or coworkers) about AI isn’t about running every paper through inaccurate and unprovable detection software that doesn’t meet the burden of proof for accusing them of a major academic violation. Identifying AI is about developing an eye for the patterns and having a feeling for the right time to call it out.
If you want to call the bluff on AI writing by students, coworkers, news reporters, LinkedIn posters, or even friends who can’t write a party invite themselves, here are ten tells I look for.
Tell #1: Perfect grammar with soulless tone
This is the most common way I can spot AI. Most AI writing reads like a corporate white paper that’s been wordsmithed by a dozen people into the perfect statement.
My prompt: “Write a perfect two sentence summary of this article”
Gemini: Just as Mike McDermott exposed Teddy KGB’s hollow hand by reading a subtle poker tell in Rounders, educators can unmask synthetic student essays by spotting the unmistakable linguistic tics, rigid structures, and contextual vacuums left behind by Large Language Models. By training themselves to recognize these ten algorithmic giveaways, teachers can move past flawed software detectors, reclaim classroom authenticity, and design more human-centric assignments that a machine cannot replicate.
ChatGPT: AI-generated writing isn’t hard to spot once you learn its predictable “tells,” including sterile tone, repetitive structure, hallucinated citations, and overly polished but emotionally empty prose. The article argues that instead of relying on flawed AI detectors, educators should redesign assignments to require genuine human thinking, discussion, and lived experience that chatbots still struggle to fake.
AI’s writing style is grammatically flawless but devoid of human emotion or personality. Each model’s opening sentence has stacked adjectives and a second sentence with exactly three points. There is not a single typo and it sounds overly polished. When human students are writing real thoughts, there are a mix of short, punchy sentences with some complex or overly run-on ideas too.
I try to carefully proofread these Substack articles, but I still spot grammatical errors that I missed every time I look back at them. The paradigm has flipped with my students’ papers because some poorly written sentences would mean they are actually writing and that should get a higher grade than ‘perfectly written’ AI slop.
Tell #2: Over-Engineered APA Citations
AI is a prediction engine that is constantly guessing what the right output is. When a LLM is missing a piece of data, it will guess what should be there based on a pattern. This leads to my second big tell which is hyper-formatted citations. While a human might forget a date or a comma, a LLM will hallucinate a “perfect” citation by filling every possible slot in the APA template including adding things that shouldn’t be there. An APA citation should look like this:
A big red flag is seeing a citation like (Doe, January n.d., 2024) or (Smith, Month Day, Year) with “Month Day” literally written as words in the citation because ChatGPT couldn’t find the date.
Human student mistake: (2021 Smith)
AI-generated mistake: (Smith, January n.d., 2021, pp. XX–XX)
The LLM knows a date belongs there but since it can’t find that piece of information in the real source, it provides a text placeholder or a weirdly specific (yet incorrect) hybrid that no human student would create.
Tell #3: “Hallucinated” Citations
AI models are not a library of facts because the models turn every word and pixel of an image into numbers and vectors to make a probabilistic map of language. When asked for a source, LLM will often invent a plausible-sounding book or article that doesn’t exist.
Earlier this spring, I needed to modify my resume into a one-page bio for a conference and even though I have tons of publications and uploaded a word doc with my resume…AI made up a paper that I didn’t write!
Read my article about this specific issue: AI is making mistakes on purpose
Tell #4. “In conclusion, it is important to note...”
AI is obsessed with structural signposting (the cues and transitions that guide the reader through a text) and LLMs love to wrap things up with a neat bow by using repetitive transition phrases.
In conclusion, critical infrastructure is the foundation of modern society, supporting everything from basic daily needs to national security and economic stability. However, these systems are increasingly exposed to serious risks, including cyberattacks, natural disasters, aging equipment, and human error. Because energy grids, transportation networks, water systems, healthcare facilities, and financial institutions are so interconnected, even a single disruption can have widespread consequences. As technology continues to advance and societies become more dependent on digital systems, the importance of protecting and modernizing critical infrastructure will only grow. Ensuring its resilience through investment, planning, and security measures is essential for maintaining stability and safety in the future.
Some of the most common AI phrases you will see are “Ultimately, the balance between X and Y remains...” and “Furthermore, it is essential to consider...”
Tell #5: Excessive Hedging
LLMs consistently frame their arguments with soft language and artificial balance. The result is writing that feels terrified to commit to a real opinion as AI written papers try to present every possible side in the same paragraph.
You’ll see phrases like “on one hand”, “while there are valid arguments on both sides,” “it is important to consider multiple perspectives,” or “the issue is complex and multifaceted” repeated without the writer actually saying anything meaningful. Pre-LLM writing, my human students were the opposite and usually lean into their opinions too hard. Students would have a half-baked concept and write pages about it without enough citations or clear arguments. AI is the complete opposite and will present 4 different sides of an argument with references (some of them fake) for each one.