Real vs fake: how to spot AI-generated text
In January 2026, a viral LinkedIn post about a supposed breakthrough in quantum battery technology got shared over 200,000 times before anyone noticed the author didn't exist. The profile photo was generated by StyleGAN. The text was written by GPT-5. And the "peer-reviewed journal" it cited was a hallucinated URL that returned a 404.
Nobody caught it for eleven days.
We're in an era where AI can produce text that passes casual inspection by most readers. That's not a hypothetical concern. It's Tuesday. So how do you actually get better at telling the difference between something a human wrote and something a language model produced?
I've spent the last year studying this question, partly because we build a game around it at Bluffpedia, but mostly because I think it's one of the most useful skills a person can develop right now. Here's what I've learned.
The linguistic tells
AI text has quirks. They're not always obvious, and they change as models improve, but certain patterns have persisted across multiple generations of language models.
Hedging and softening. AI models are trained to be helpful and avoid controversy. This produces a distinctive verbal tic: excessive hedging. Phrases like "it could be argued that," "there are various perspectives on," and "while opinions differ" appear far more frequently in AI output than in human writing. Humans with opinions just state them. AI tries to cover all angles.
Uniform sentence rhythm. Read a paragraph of AI text aloud. You'll often notice the sentences have similar lengths and follow similar grammatical structures. Human writing is messier. We write a short sentence. Then a long one that goes on for a while and maybe loses its thread before circling back to the point we were trying to make. Then another short one. AI tends toward a medium-length default.
Vocabulary distribution. This one is subtle but measurable. AI models have preferences for certain words that don't match normal human usage patterns. A 2024 study by researchers at Stanford found that AI-generated academic papers used the word "intricate" 117% more often than human-written ones. "Commendable" appeared 200% more. These aren't wrong words, exactly. They're just oddly popular with language models.
The 'swap test'
Pick any adjective in a suspicious paragraph and ask: would a normal person use this word in conversation? If the answer is no for more than two adjectives in a single paragraph, that's a signal. Words like "multifaceted," "nuanced," and "comprehensive" appear in human writing, sure, but rarely three of them in the same paragraph.
Absence of error. This sounds backwards, but real human writing usually contains minor imperfections. Not typos necessarily (editors exist), but small deviations from perfect grammar, colloquial turns of phrase, or sentence structures that technically break a rule but feel natural. AI text, especially from larger models, tends to be grammatically flawless in a way that reads as sterile.
Structural patterns
Beyond individual word choices, AI text follows recognizable structural patterns that become obvious once you know what to look for.
The five-paragraph essay. AI loves symmetry. Ask it to write about any topic and it will often produce an introduction, three body sections, and a conclusion. Each section will be roughly the same length. Each paragraph will make one main point. This structure isn't wrong (it's what we all learned in school), but real published writing rarely follows it this neatly.
The "to be sure" pivot. AI text frequently follows a pattern of claim, then immediate qualification. "Social media has changed communication. To be sure, there are drawbacks." or "This technology is powerful. However, it also raises concerns." This balancing act happens in human writing too, but AI does it almost reflexively, paragraph after paragraph. It's like watching someone who can't stop both-sidesing every statement.
Predictable topic sentences. In AI-generated articles, the first sentence of each paragraph almost always states the paragraph's main idea clearly and directly. Human writers sometimes bury the point in the middle, or build up to it, or start with an anecdote that seems unrelated until the connection clicks three sentences later. AI rarely takes that kind of structural risk.
| Trait | Human writing | AI writing |
|---|---|---|
| Sentence length | Varies a lot. Short bursts, long tangents, fragments. | Mostly medium. Consistent and predictable. |
| Vocabulary | Personal favorites, slang, field-specific jargon | Slightly formal, favors certain 'safe' words |
| Structure | Messy, nonlinear, sometimes surprising | Clean, symmetrical, follows templates |
| Hedging | States opinions, takes sides | Covers all angles, qualifies constantly |
| Errors | Occasional grammar quirks, colloquialisms | Near-perfect grammar, reads as polished |
| Specificity | Concrete examples, personal anecdotes, exact numbers | General claims, round numbers, vague references |
| Emotional range | Sarcasm, frustration, humor, boredom | Consistently measured and pleasant |
| Transitions | Sometimes abrupt or implicit | Smooth, signposted, often formulaic |
Factual verification methods
Stylistic analysis is useful, but the strongest signal is often factual. AI generates plausible-sounding claims that may be partially or entirely wrong.
Check specific numbers. AI loves to include statistics because they make text feel authoritative. But the numbers are often wrong. If a text claims a bridge was built in 1847 or that a country has a population of 12.3 million, take thirty seconds to verify. AI frequently generates dates and figures that are close but not quite right, which is actually worse than being wildly wrong because it's harder to catch.
Verify named sources. When AI text attributes a quote or finding to a specific person or study, look that up. AI commonly attributes real-sounding quotes to real people who never said them. It can also invent entire studies, complete with plausible journal names and publication years. A quick search usually reveals whether the citation exists.
Look for anachronisms. AI models have training data cutoffs and sometimes mix up temporal details. A text about events in 2025 might include information that was only known in 2023, or it might project trends forward incorrectly. These temporal inconsistencies are hard for AI to avoid because it doesn't have a clear model of "when things happened relative to each other."
AI detectors aren't reliable enough to trust blindly
Tools like GPTZero, Originality.ai, and Turnitin's AI detection feature can be helpful as one signal among many, but none of them are accurate enough to serve as a sole judge. False positive rates (flagging human text as AI) range from 5% to 20% depending on the tool and the type of text. False negatives (missing AI text) are even more common, especially with paraphrased or edited AI output.
How accurate are the detection tools?
Since everyone asks about automated detection tools, let's look at how they actually perform. These numbers come from a combination of published benchmarks and independent testing by researchers at the University of Maryland and ETH Zurich in late 2025.
The "Human avg" bar at the end is the one that should worry you. In controlled studies, average humans correctly identify AI-generated text only about half the time, which is barely better than a coin flip. Trained humans do much better (around 70-75%), and automated tools score higher still on unedited text.
But here's the catch: all these numbers drop when the AI text has been lightly edited by a human. Even basic paraphrasing and restructuring can cut tool accuracy by 15-25 percentage points. The arms race between generators and detectors is ongoing, and the generators are currently winning.
Why this actually matters
This isn't just an academic exercise. AI-generated text is already being used to create fake product reviews on Amazon (an estimated 42% of reviews on some product categories, per a 2025 Fakespot analysis), fabricate news articles that spread on social media, generate phishing emails that are more convincing than anything a human scammer could write, and produce academic papers submitted to real journals.
The World Economic Forum's 2025 Global Risks Report ranked AI-generated misinformation as a top-five global risk. Not because any single piece of fake text is catastrophic, but because the sheer volume is overwhelming our collective ability to verify what we read.
Every person who gets better at spotting AI text is a small but real countermeasure against this trend.
How to practice
Reading about detection techniques is a start, but the skill is mostly developed through practice. Here are some concrete ways to build your AI-text radar:
Read AI output regularly. Go to ChatGPT or Claude, ask them to write a paragraph about something you know well, and study the output. Notice the patterns. Then ask for a revision and see how the patterns change.
Play detection games. There are several online tools that test your ability to spot AI text. Bluffpedia focuses specifically on Wikipedia-style content, with AI-generated fake summaries that you have to distinguish from real ones across eight different game modes. It's a useful way to calibrate your instincts because you get immediate feedback on whether you were right.
Verify before sharing. This is the simplest and most impactful habit. Before you share an article, statistic, or quote on social media, spend sixty seconds checking whether the source is real and the claims check out. This single habit, practiced consistently, would reduce misinformation spread more than any detection tool.
Cross-reference with primary sources. When something seems suspicious, go to the original source. If a news article claims a study showed something, find the actual study. If a social media post quotes a public figure, find the original statement. This takes more effort, but it's the most reliable verification method available.
Your detection skills improve faster than you'd expect
Research from the University of Waterloo (published in Cognitive Research, 2025) found that participants who spent just two hours practicing AI text detection improved their accuracy from 52% to 68%. Ten hours of practice pushed accuracy above 75%. The skill curve is steep at the beginning, which means even a small investment of time pays off significantly.
The asymmetry problem
Here's something that bothers me about this whole situation. Generating convincing AI text takes seconds and costs fractions of a cent. Verifying whether text is AI-generated or factually accurate takes minutes to hours of human effort. That's a fundamental asymmetry that favors the generators.
We can't solve this through individual vigilance alone. We also need better tools, clearer disclosure requirements, and platform-level interventions. But individual skill still matters. A population that's better at spotting fakes is harder to fool at scale.
Think of it like hand-washing. Individual hygiene didn't eliminate infectious disease. Public health infrastructure did that. But individual hygiene still helps, and a population that practices it is more resilient. The same logic applies to media literacy in the age of AI.
Where we are now
AI detection is an imperfect, evolving discipline. The tools are getting better, but so are the generators. Stylistic analysis catches a lot of current AI text, but models are already being fine-tuned to avoid the tells I described above. Factual verification remains the most robust approach because it targets the fundamental limitation of language models: they don't know what's true.
The best defense is a layered one. Use stylistic analysis as a first filter. Apply factual verification when something seems off. Use automated detection tools as a supplementary signal, not a verdict. And maintain a healthy baseline of skepticism about text you encounter online, regardless of how polished it looks.
The fact that you're reading an article about this means you're already ahead of most people. The next step is practice.