Quantitative Signals
AI vs Human Writing in Numbers

The AI Writing Fingerprints report names the patterns. This report puts numbers on them.

We used a public research dataset called HC3 — the Human ChatGPT Comparison Corpus — where real humans and ChatGPT answered the exact same questions across four domains: open Q&A, finance, medicine, and encyclopedia-style topics. That pairing is what makes the comparison fair: same prompts, different authors.

The dataset contains 7,210 human answers (about a million words) and 10,243 AI answers (about 1.8 million words). We ran every measurement on both halves and compared. Where GPTZero has published their own ratios from a much larger corpus, we include those too — so you can see whether the numbers agree.

They usually do.

You’ve probably seen lists of "AI words" — delve, tapestry, vibrant, landscape. But how overused are they, really? GPTZero measured it across millions of documents. The top result is striking:

AI uses the phrase "play a significant role in shaping" 182 times more often than human writers do. That’s not a subtle difference. That’s a fingerprint.

We verified these patterns independently on the HC3 corpus. The ratios are lower — HC3 is a Q&A dataset with shorter answers, less room for elaborate phrasing — but the direction is the same:

This is probably the most reliable non-vocabulary signal. When you read AI text and feel like something is off even though you can’t point to a specific word — this is often why. AI sentences are longer, and they’re more uniform. Human writing has a rhythm that speeds up and slows down. AI writing hums at a steady frequency.

But here’s the twist that matters. This signal doesn’t work the same way across all subjects:

If you’ve ever looked at a page of AI writing and thought "this looks too organized" — it’s not your imagination. AI organizes text into paragraphs more aggressively than humans, and those paragraphs are eerily uniform in size.

And there’s an interesting surprise in the last row. Humans use topic sentences more than twice as often as AI. That goes against the intuition that AI writing is "more structured." AI looks structured because the paragraphs are uniform, but it actually does less of the work of orienting the reader at the start of each paragraph. It substitutes neatness for clarity.

AI doesn’t just overuse specific words. It uses fewer different words overall. Three standard measures all point the same direction — and the gap gets worse the longer the text runs.

Notice that the MATTR metric (a length-independent measure of vocabulary diversity) holds perfectly steady at 0.703 vs 0.631 across all chunk sizes. That stability is the point — it confirms the 10% diversity gap is real and consistent, not an artifact of how we sliced the data.

Beyond words and structure: how often does AI use specific rhetorical moves compared to humans? Two patterns stood out sharply. One was expected. The other was interesting for what it didn’t show.

The hedging premium is a direct consequence of how AI models are trained. During reinforcement learning, models are rewarded for epistemic caution — "may," "might," "could," "appears to," "suggests that" — and penalized for overconfident claims. The result: AI hedges 59% more than humans, even in formal encyclopedia text where both sides hedge less overall.

The summary-conclusion signal is the one that matters most for detection. Only 0.9% of human answers in this corpus end with an "In conclusion" or "Overall" summary paragraph. AI does it 9.7 times more often. If you’ve read the Good AI benchmark, you already know this pattern: formulaic conclusions were the single most common detection signal across the entire library.

There’s one more signal we should mention, even though WROITER doesn’t use it: perplexity.

Perplexity measures how "surprising" text is to a language model. If you feed a sentence to GPT-2 and it finds each word highly predictable, the perplexity is low. AI-generated text has consistently lower perplexity than human text — because it was generated by a process that maximizes token probability. It’s predictable because a prediction machine wrote it.

The HC3 research shows this clearly. Human text has a broad perplexity distribution, with significant mass at high values — humans write things that genuinely surprise language models. AI text clusters at the low end, concentrated in a narrow band of high predictability.

GPTZero was one of the first detectors to use perplexity as a primary signal, combined with "burstiness" (how much perplexity varies from sentence to sentence). It works. So why doesn’t WROITER use it?

Every study has limitations. Ours are worth being explicit about.

The corpus is from early ChatGPT (2023). HC3 captures a specific moment in AI writing — before GPT-4o, before Claude 3.5, before Gemini 2. The patterns measured here are real, but newer models may exhibit different ratios. We’re working with the best publicly available paired corpus, not the most current one.

The format is Q&A, not essays. These are short answers to specific questions, not blog posts, marketing copy, or academic papers. Some patterns that are strong AI signals in long-form writing — like meta-introductions and elaborate "both sides" framing — simply don’t appear in this format. The numbers here are a floor, not a ceiling.

GPTZero’s full data is gated. We used their public Top 10 ratios. Their full Top 50 list and monthly vocabulary updates appear to be account-gated or dynamically rendered. Adding those would strengthen the word-level analysis significantly.

Multi-corpus validation is still to do. Adding numeric tables from Ma et al., StyloAI, and CHEAT analyses would provide robustness beyond a single dataset. We haven’t done that yet.

Keep reading

This report provides the numbers. Rule-Based Signals turns them into 50 named, implementable detection patterns. AI Writing Fingerprints maps which patterns belong to which model. And the Good AI Benchmark shows what happens when you point the detector at someone else’s library.