AI detection 101: a complete guide

Sometime in late 2022, a college professor in Texas failed an entire class because an AI detector flagged their essays. The students had written every word themselves. The detector was wrong.

That story, and dozens like it, captures the messy reality of AI text detection in 2026. The tools are better than they were three years ago. They're also still unreliable enough to ruin someone's day if you treat their output as gospel. We've spent a lot of time thinking about this at Bluffpedia, since our entire game is built around the boundary between human-written and AI-generated text. Here's what we've learned.

How detection tools actually work

Most AI detection tools rely on two core measurements, sometimes supplemented by a third newer technique. None of them are magic.

Perplexity scoring

Perplexity measures how "surprised" a language model would be by a given piece of text. The idea is straightforward: AI-generated text tends to be predictable, because the model that wrote it was essentially choosing the most likely next word at each step. Human writing is messier. We use unexpected word choices, switch topics abruptly, make grammatical decisions that are technically correct but statistically unusual.

A detection tool feeds your text through its own language model and calculates a perplexity score. Low perplexity means the text was predictable, which suggests AI. High perplexity means the text was surprising, which suggests a human wrote it.

The problem is obvious: a human writing clearly and simply about a well-understood topic will also produce low-perplexity text. A chemistry textbook chapter is highly predictable. That doesn't mean a robot wrote it.

Burstiness analysis

Burstiness looks at variation in sentence complexity. Humans tend to write with uneven rhythm. A short sentence. Then a longer one that packs in multiple clauses and subordinate ideas. Then maybe a fragment. AI text, especially from models like GPT-3.5 and early GPT-4, tends to be more uniform. Sentences hover around the same length, with similar structural complexity throughout.

This metric has gotten less reliable as newer models have improved at mimicking human writing patterns. Claude, GPT-4o, and DeepSeek V3 all produce noticeably more varied text than their predecessors. The burstiness gap between human and AI is narrowing.

Watermarking

This is the newer approach, and the one with the most theoretical promise. The idea is that AI providers embed an invisible statistical pattern in their model's output. Think of it like a digital watermark in a photograph, but for text. The watermark doesn't change the meaning or readability of the text, but a detection tool that knows the watermarking scheme can identify it.

Google DeepMind published their SynthID-Text watermarking system in late 2024, and OpenAI has been developing their own approach since at least 2023 (though they've been reluctant to deploy it, reportedly over concerns about user adoption). The advantage of watermarking is that it can theoretically be much more accurate than perplexity-based approaches. The disadvantage is that it only works on text produced by models that implement the watermark. If someone uses an open-source model without watermarking, or just paraphrases the output, the watermark disappears.

The tools: honest assessments

We tested the major detection tools extensively while building Bluffpedia. Here's where they stand in early 2026.

GPTZero

GPTZero launched in January 2023 and quickly became the most popular AI detector, largely because it was free and the founder, Edward Tian, was a Princeton student who made for a compelling story. The tool uses both perplexity and burstiness analysis, plus some proprietary techniques they've added since launch.

In our testing, GPTZero correctly identified AI-generated text about 78% of the time on straightforward cases. That number drops to around 60% when the AI text has been lightly edited by a human. It also flagged roughly 9% of purely human-written text as AI-generated, which is a problem if you're a teacher grading 200 essays.

GPTZero has improved significantly since its early days, and their enterprise product aimed at universities is more sophisticated than the free web tool. Still, a ~9% false positive rate means you should never use it as the sole basis for an accusation.

Originality.ai

Originality.ai positions itself as the premium option, primarily targeting content marketers and publishers. At $14.95/month for their base plan, it's the most expensive consumer-facing tool. The higher price buys you somewhat better accuracy: in our testing, it correctly identified AI text about 83% of the time, with a false positive rate around 5%.

Where Originality.ai struggles is with hybrid content, the increasingly common case where a human writes a draft and uses AI to polish it, or vice versa. The tool gives you a percentage score ("67% AI probability"), which sounds precise but often isn't. Is 67% a flag or not? That's on you to decide.

ZeroGPT

ZeroGPT is free and shows it. The accuracy in our testing was the lowest of any tool we tried, hovering around 65% correct identification with a false positive rate that spiked above 15% on certain types of content. Academic writing and technical documentation were particularly problematic. We wouldn't rely on ZeroGPT for anything where the stakes matter.

Turnitin's AI detection

Turnitin added AI detection to their existing plagiarism platform in April 2023. Because Turnitin already has deep institutional adoption (something like 16,000 institutions worldwide), this is the tool that affects the most students. Their accuracy is better than most standalone tools, around 80-85% in independent testing, with a false positive rate Turnitin claims is under 1%.

That sub-1% claim deserves scrutiny. Turnitin's own testing was conducted primarily on English text written by native speakers. When researchers at Stanford tested it on essays by non-native English speakers in 2024, the false positive rate jumped to over 60% for some language backgrounds. This is a pattern across all detection tools, and it's the single most important thing to understand about AI detection.

OpenAI's classifier (2023, defunct)

OpenAI launched their own AI text classifier in January 2023 and shut it down six months later, in July 2023. The reason? It only correctly identified 26% of AI-written text while incorrectly labeling 9% of human text as AI-generated. Those numbers are barely better than a coin flip.

The fact that the company that built GPT couldn't build a reliable detector for their own model's output tells you something about the fundamental difficulty of this problem. OpenAI has since focused on watermarking approaches instead, though nothing has shipped publicly as of this writing.

Tool	AI detection rate	False positive rate	Price	Best for
GPTZero	~78%	~9%	Free / $10+ mo	Quick checks, education
Originality.ai	~83%	~5%	$14.95/mo	Content publishing
ZeroGPT	~65%	~15%	Free	Nothing, honestly
Turnitin AI	~80-85%	<1% (disputed)	Institutional	Universities (with caveats)
OpenAI Classifier	~26%	~9%	Shut down	Historical curiosity

The accuracy problem

The numbers above might make some of these tools seem reasonable. 83% accuracy sounds pretty good, right? But accuracy rates hide important nuances.

False positives disproportionately hit certain groups

The Stanford study we mentioned found that AI detectors consistently flagged writing by non-native English speakers at much higher rates. The reason makes sense once you understand how the tools work: non-native writers often use simpler vocabulary, more predictable sentence structures, and fewer idiomatic expressions. To a perplexity-based detector, this looks like AI text.

This isn't a minor edge case. In a university with international students, it means the students who are already navigating language barriers are the ones most likely to be falsely accused of cheating. Several lawsuits have been filed over exactly this scenario.

Paraphrasing tools break detection easily

On the other side, anyone actively trying to evade detection can do so with minimal effort. Paraphrasing tools like QuillBot, or simply asking a different AI model to rewrite a passage in a more casual style, drops detection rates dramatically. In one test, running GPT-4 output through a paraphrasing tool reduced GPTZero's detection rate from 78% to 31%.

This creates a perverse incentive structure: the students who are already inclined to cheat invest five extra minutes in a paraphrasing step, while honest students face false accusations.

The fundamental arms race

AI detection is stuck in a losing position. Detection tools are essentially trying to identify statistical patterns in AI output. But AI models are explicitly trained to produce text that is indistinguishable from human writing. Every improvement in language models makes detection harder. Every improvement in detection gets incorporated into the next round of model training data.

Watermarking could break this cycle, but only if every AI provider implements it and it can't be easily removed. Neither condition is true today. Open-source models have no watermarking, and text-based watermarks can be destroyed by paraphrasing.

What actually works

If automated tools aren't reliable enough, what should you do?

For educators: The best approach is a mix of process-based assessment and skeptical tool use. Require drafts and revision histories. Ask students to explain their reasoning verbally. Use detection tools as a flag for further investigation, not as evidence. If a tool flags a paper, have a conversation with the student before making any judgment.

For publishers and content teams: Originality.ai and GPTZero are useful for catching the laziest AI-generated content. But your editorial team's judgment is more reliable than any tool. Editors who know their writers' voices will spot AI-generated submissions faster than any algorithm.

For everyone else: Develop your own instincts. Read critically. Notice when text is suspiciously smooth, when it avoids taking positions, when it provides vague attributions instead of specific sources. These are the same instincts you build playing Bluffpedia, actually.

Where Bluffpedia fits in

We built Bluffpedia as a game, not a detection tool. But the skills it trains are directly relevant to this problem.

Every round of Bluffpedia is a micro-exercise in exactly the kind of judgment that automated tools can't replicate. You're reading four pieces of text, looking for the subtle signals that distinguish a real Wikipedia summary from an AI-generated fake. After a few hundred rounds, you develop an intuition for how AI writes differently from humans. That intuition doesn't give you a percentage confidence score. It gives you something better: a sense of when something is off, even if you can't immediately articulate why.

In our data, experienced Bluffpedia players (those with 500+ rounds played) correctly identify AI-generated text at a rate of around 85%. That's on par with or better than most automated tools, and it comes with zero false positives on non-native English writing, because human judgment accounts for context in ways that statistical analysis can't.

We're not suggesting you should play Bluffpedia instead of using detection tools. But we are suggesting that the human skill of reading critically is the piece that most organizations are underinvesting in. The tools are supplements. Your brain is the detector.

Where this is all heading

Three years into the AI detection arms race, a few things seem clear.

Watermarking will eventually become standard for the major commercial models. Google is already doing it. OpenAI is working on it. Regulation may force it. But watermarking won't cover open-source models, and it won't catch text that's been heavily paraphrased.

Perplexity and burstiness-based detection will continue to degrade as language models improve. The statistical gap between AI and human text is shrinking. In two or three years, the current generation of detection tools may be no better than random chance for well-crafted AI text.

The organizations that invest in human judgment, critical reading skills, and process-based verification will be better positioned than those who outsource the problem to an algorithm. The tools have a role. But they're not the answer.

And if you want to get better at telling real from fake, you know where to find us.