Why fingerprints exist at all
When you read something and think "this sounds like AI," you're reacting to a fingerprint. Not one thing — a cluster. A word that no human would reach for. A paragraph that's exactly the same length as the one before it. A sentence that hedges in a way that feels rehearsed rather than honest.
These patterns aren't random. They trace back to a specific step in how language models are built: RLHF — reinforcement learning from human feedback. After the base model is trained on text, human workers rate its responses. The model learns to produce whatever those workers reward. If the workers prefer formal vocabulary, the model learns formal vocabulary. If they reward thoroughness, the model becomes exhaustive. If they happen to like the word "delve," the model starts delving into everything.
Different companies use different workers, different rating criteria, different training philosophies. That's why GPT doesn't sound like Claude, Claude doesn't sound like Gemini, and none of them sound quite like a person. Each model picked up its own habits from its own training — and those habits are remarkably consistent, remarkably persistent, and remarkably hard to prompt away.
This report documents what those habits are, model by model.
The words that give it away
The universal AI vocabulary
Some words are overused by every major model. Research across 15 million PubMed abstracts found over a hundred words that spiked sharply after 2022 — the year AI-assisted writing went mainstream. These are the words all models reach for more than humans do:
The interesting thing: the acceleration started in 2020, before ChatGPT launched. The RLHF workers who rated training data were already preferring these words. The public release just scaled it.
The "delve" origin story
"Delve" became the poster child of AI-generated writing, and there's a real hypothesis behind why. OpenAI used Kenyan and Nigerian workers (through a company called Sama) to rate model responses during training. In Nigerian formal English, "delve" is a common, natural word — used far more than in American or British English. If those workers rated responses containing "delve" as higher quality, the model learned to produce it more.
Academic research (ACL 2025) tested this and found:
- RLHF definitely contributes — instruction-tuned Llama models overuse these words; base Llama models don't
- Human participants actually prefer responses that contain these elevated words, confirming a feedback loop
- The Nigerian English link is plausible but not proven — the broader explanation (RLHF workers reward formal register, period) has stronger support
The same mechanism — RLHF workers rewarding what sounds "good" — produces different vocabulary depending on who the workers are. GPT and Claude share some tells but diverge on others because they used different annotator pools and different training processes. Same machine, different teachers.
What belongs to whom
Beyond the universal vocabulary, each model has its own signature words. These are the ones that help identify not just that AI wrote something, but which AI wrote it.
GPT loves to bold key phrases and organize into lists. In a classifier study with 97.1% accuracy across five models, "certainly," "such as," and "overall" were the most predictive GPT word choices.
Claude frames information with "according to" and "based on" — a habit from its Constitutional AI training, where it learned to cite its own reasoning. It uses minimal formatting by default: fewer bullets, less bold than GPT. It also doesn't use curly quotes, which immediately distinguishes it from ChatGPT and DeepSeek.
Gemini has the most varied syntactic structure of any model tested, but its vocabulary diversity is actually lower. It starts sentences with adverbial phrases far more than any competitor. It tries to match your tone — but can tip into stiff formality even when you're being casual.
DeepSeek uses curly quotes (like ChatGPT), immediately distinguishing it from Claude and Gemini. Its sentences average 24.9 words — the longest of any model tested. Classifiers frequently confuse it with GPT because of shared formatting traits.
This is the smoking gun for RLHF: base Llama doesn't overuse any of the focal words. The instruct-tuned version does. Same architecture, same training data — the only difference is the human feedback step. That's the clearest evidence that these vocabulary patterns come from training, not from the model itself.
The least studied model for writing fingerprints, which is itself a kind of data point. Its outputs tend to be competent, direct, and personality-neutral — somewhere between GPT-3.5 and GPT-4 in formality. One documented quirk: it ignores negative instructions more than other models. Tell it "don't use the phrase 'in today's'" and it will use it anyway. No distinctive vocabulary tell has been firmly established by any study.
Structure as fingerprint
Here's something most people don't realize: even if you strip out every word and replace it with a placeholder, classifiers can still identify which model generated the text with 73%+ accuracy — using only the markdown structure. Bullets vs. paragraphs. Header depth. How many items in each list. Whether there's a summary at the top.
Formatting is a fingerprint independent of vocabulary. And each model has defaults that are surprisingly hard to override.
- Heavy use of bullet points and numbered lists
- Bold key phrases liberally
- Opens with a thesis paragraph that summarizes before diving in
- Transition phrases at every paragraph start ("Furthermore," "Additionally,")
- Formal header structure even when casual would fit
- Tends toward comprehensive coverage over concision
- Minimal formatting — fewer bullets, less bold
- Prose-first: more paragraph writing, less listing
- Variable sentence length (more natural rhythm)
- More likely to use em-dash for asides
- Doesn't restate the question before answering
- Skips level-2 headings; goes straight to level-3
- Adverbial sentence openers ("Ever feel like...", "Think about...")
- Bullet-heavy for information-dense responses
- Often restates or reframes the question
- Most varied syntactic structure of any model tested
- Tends to formally organize even creative content
- Numbered subsections with headers everywhere
- Section-by-section breakdowns even for simple tasks
- Longest average sentence length (24.9 words/sentence)
- Emojis in chat-mode headers (🔍 📌 etc.)
- Chain-of-thought reasoning visible in R1 variant
Sentence length — it's not the average, it's the variance
Across seven models tested in academic writing:
- DeepSeek-coder V2: 24.9 words/sentence — the longest
- Llama 3.1 8B: 20.6 words/sentence — more readable by standard metrics
- GPT-4 and Claude: 18-22 words, but Claude varies length more naturally within a single response
- GPT-5: Shorter and more concise than predecessors — and distinctively overuses semicolons as a result
Sentence length uniformity is more telling than length itself. Human writers swing wildly — 8 words, then 32, then 14, then 27. AI models, especially GPT-4o, hover in a narrow band: 21, 22, 20, 21. Both average 21 words per sentence. Only one of them reads like a machine wrote it.
How each model dodges, hedges, and deflects
Every model handles uncertainty and controversy differently — and those differences are fingerprints just as distinctive as vocabulary. This isn't about what they say. It's about what they do when they're uncomfortable.
Who hedges most?
| Model | Hedging tendency | Go-to hedge phrases | On controversial topics |
|---|---|---|---|
| GPT-4 / 4o | High — adds disclaimers unprompted | it's worth noting, however, it depends on, it's important to consider |
Both-sides framing; rarely refuses, often hedges |
| Claude | Medium — hedges on uncertainty, not on opinion | that said, I'm not certain, broadly speaking |
More likely to take positions; also more likely to refuse entirely on some topics |
| Gemini | High — errs on caution | I hope this helps, I'd be happy to assist, that's a great question |
References ethical frameworks without committing to a position |
| DeepSeek | Low-medium on general topics; high on Chinese political content | I should point out, it's important to note |
Direct on technical/historical; refuses or deflects on China-sensitive content |
| Llama 3 (instruct) | Medium — mirrors what RLHF trainers rewarded | of course, certainly, sure! |
Variable by fine-tune; base model is more direct |
| Mistral | Low — notably fewer disclaimers than GPT or Claude | Fewer hedge phrases documented | More direct; less content filtering by default |
The "Let me explain" problem
Models narrate their own responses at very different rates. GPT-4o is the worst offender — it opens with an acknowledgment, contextualizes the task out loud, works through the answer, then closes with an offer to elaborate. Claude tends to skip the preamble and just answer. Gemini opens with enthusiasm before the actual content. DeepSeek dives straight into structure.
- GPT-4o: "Great question," "I'd be happy to help," "Let me walk you through," "Here's what you need to know," "To summarize"
- Gemini: "I hope this helps," "That's a great question," "Absolutely!" "Here's a breakdown," "I'd be happy to"
- Claude: "That said," "In other words" — rarely opens with affirmations, tends to start directly answering
- DeepSeek: Minimal meta-commentary; dives directly into structured response
- Llama 3 instruct: "Sure!", "Of course", "Certainly!" — opener-heavy, then gets down to business
How they handle controversy
This is where training philosophy becomes most visible:
- GPT-4: Both-sides by default. Presents multiple perspectives, rarely refuses, rarely commits. Will take a position when pushed, but never on its own.
- Claude: More willing to form and express opinions, especially on philosophical and ethical questions. Also more willing to refuse creative and roleplay content that GPT handles without complaint.
- Gemini: References ethical frameworks abstractly — utilitarianism, deontology — without committing to any of them. Strong on historical examples. High abstraction on moral questions.
- DeepSeek: Excellent on general historical and ethical reasoning. But explicit refusal or deflection on anything touching Chinese political sensitivities — Taiwan, Tiananmen, Xinjiang. This asymmetry is unique to DeepSeek and instantly identifying when triggered.
How tells change version to version
Models don't stay the same. Each new version fixes some tells and introduces new ones. Tracking this evolution matters because a detector tuned to last year's tells will miss this year's models.
The GPT lineage
The Claude lineage
Are newer models harder to detect?
By word-lists, yes. The same classifier trained on older outputs loses accuracy on newer ones. But "harder to catch by word-list" is not the same as "harder to catch." Embedding-based classifiers still hit 97.1% accuracy on the latest models — even when text is paraphrased, translated, or summarized. The fingerprint isn't just in the words. It's in how models organize thought. And that's much harder to engineer away.
Character portraits
If you've read this far, you've seen the data. Here's how it all adds up into personality — the shorthand version of what each model sounds like when you step back from individual tells and listen to the voice.
GPT-4 / GPT-4o — The Competent Bureaucrat
GPT writes like a research assistant who wants to demonstrate thoroughness. It front-loads with a summary, organizes everything with headers and bullets, uses transition phrases between every section, and closes with an offer to help further. It's reliable and comprehensive — but it's never surprising.
Its vocabulary skews formal even for casual tasks. When asked to be casual, it becomes "friendly professional" rather than actually casual. The em-dash in GPT-4o is persistent enough that even explicit prompts to avoid it fail regularly — it's baked deep.
Claude 3.5 / Claude 4 — The Thoughtful Generalist
Claude writes like a well-read person who cares about being understood. It varies sentence length, avoids most boilerplate openers, and defaults to prose over lists. Its "AI-ness" is subtler — slightly elevated vocabulary, occasional over-explanation of things nobody asked about, and a tendency to contextualize decisions within the response itself.
It's significantly better at following style instructions than GPT. Tell it to write like Hemingway and it won't start with "Certainly! Here's a Hemingway-style piece." It just writes it.
Gemini 1.5 Pro / 2.0 — The Eager Collaborator
Gemini is the most conversational of the major models. It asks questions back, frames things warmly, uses "I hope this helps" like a reflex. This is both its strength and its most obvious tell. Its syntactic variety is actually the highest tested, but its vocabulary diversity is lower. It can tip into stiff formality even when casual register is requested.
The adverbial-phrase opening is Gemini's signature move: "Ever feel like...", "Just imagine..." It creates an enthusiastic-coach voice that's instantly recognizable once you know to listen for it.
Llama 3 (instruct) — The Eager Intern
Llama opens with enthusiasm markers — "Sure!", "Of course!", "Certainly!" — more consistently than any other family. This is almost certainly an RLHF artifact: annotators rated responses higher when they opened with agreement and willingness. Base Llama doesn't do this at all. The contrast is the cleanest evidence that human feedback, not architecture, shapes these patterns.
Quality varies enormously by fine-tune. The base model has the fewest AI tells of any widely available model — which makes controlled Llama fine-tunes a potential blind spot for word-level detectors.
Mistral — The Pragmatist
Mistral is the least studied model for writing fingerprints, which is itself a data point. Its outputs are competent, direct, and free of the enthusiastic-affirmation pattern. It ignores negative instructions more than other models — tell it not to use certain phrases and it uses them anyway. Its register sits between GPT-3.5 and GPT-4 in formality. No signature vocabulary tell has been established.
DeepSeek V3 / R1 — The Technical Structuralist
DeepSeek organizes everything. Numbered headers, subsections, section-by-section breakdowns even when prose would be more appropriate. Its chain-of-thought reasoning (R1) is visible, which creates a distinctive "I'm working through this" quality. Curly quotes consistently. Longest average sentence length of any model in academic writing tests.
The political-topic deflection is the single clearest model-specific tell in this entire report: DeepSeek will discuss Nazi war crimes in detail but deflects on Tiananmen, Uyghur persecution, and Taiwan independence. When triggered, it's an instant identification.
Master comparison table
Everything above, compressed into one table. Use it as a reference — the sections above tell the story, this is the cheat sheet.
| Model | Primary vocab tell | Structural default | Rhetorical default | Punctuation tell | Controversy approach | Detectability trend |
|---|---|---|---|---|---|---|
| GPT-3.5 | "I hope this email finds you well," "I am writing to" | Paragraphs, moderate lists | Formal register, direct | No em-dash; straight quotes | Balanced; rarely refuses | High (older tells obvious) |
| GPT-4 | delve, certainly, such as, overall, pivotal | Heavy bullets + bold headers | Comprehensive; unsolicited disclaimers | Curly quotes; no em-dash yet | Both-sides framing | High |
| GPT-4o | underscore, tapestry, boasts, explore | Heavy bullets + bold + em-dashes | Enthusiastic meta-commentary | Em-dash dominant; curly quotes | Both-sides + hedges | Medium-High |
| GPT-5 | Semicolons as connective tissue; fewer older tells | More concise; still structured | Fewer disclaimers; more direct | Semicolon overuse emerging | More direct than GPT-4 | Medium (evolving) |
| Claude 3 Opus | fascinating, realm, delve into, it's worth noting | Prose-first, minimal lists | Literary register; considered tone | Em-dash (moderate); straight quotes | Forms opinions; more refusals on edge content | Medium |
| Claude 3.5/4 | according to, based on, that said, landscape (metaphor) | Minimal formatting; paragraph-dominant | Direct; hedges only when uncertain | Em-dash (light); no curly quotes | Opinionated; more selective refusals | Medium-Low |
| Gemini 1.5 | I hope this helps, great question, I'd be happy to | ADVP openers; bullet-heavy; question restatement | Enthusiastic-collaborative; warm hedging | No curly quotes; no em-dash | Abstract + ethical frameworks | High (enthusiasm pattern obvious) |
| Gemini 2.0+ | Reducing but "I hope this helps" persists | Better prose; still conversational | More tone-matching; less formulaic | Consistent with 1.5 | Improving nuance | Medium |
| Llama 3 instruct | Sure!, Of course, Certainly!, Here's what I found | Moderate lists; opener-heavy | Affirmation openers; then direct | Variable by fine-tune | Variable by fine-tune | High (opener pattern) |
| Mistral | No firm tell established | Moderate; ignores negative instructions | Direct; fewer disclaimers | Variable | Direct; less filtering | Low-Medium |
| DeepSeek V3/R1 | Numbered section headers; meticulous subsections | Heavy structured formatting; CoT visible | Technical precision; direct on general content | Curly quotes consistently | Strong on history; deflects on China-sensitive topics | Medium (structure obvious; political tell unique) |
What makes a tell "strong" vs. "weak"
- Strong tell: Shows up consistently, rarely appears in human writing, and hard to prompt away. Examples: em-dash in GPT-4o, opener affirmations in Llama 3 instruct, DeepSeek's political deflection.
- Medium tell: Elevated vs. human baseline, but can be prompted away with explicit instruction. Examples: delve, certainly, bullet overuse.
- Weak tell: Overrepresented in AI generally but not specific to any one model. Examples: realm, pivotal, meticulous — all models use these.
The strongest detection signals are structural — formatting patterns that survive even when you swap out every word. Next come rhetorical defaults — how models handle disagreement, uncertainty, and meta-commentary. Vocabulary tells come last. A detector that only looks at word-level tells will be defeated by the next model update. The deeper signature — how a model organizes thought — is much harder to train away. That's what WROITER is built on.
What we don't know yet
This report is a snapshot, not a permanent truth. Here's where the gaps are:
- Coverage is uneven. Most academic studies focus on GPT and Claude. Gemini, Mistral, and DeepSeek are significantly less studied for writing fingerprints specifically.
- Tells shift with every release. Anything documented before mid-2024 may already be reduced in current models that were trained to avoid it.
- Fine-tuning changes everything. Enterprise fine-tunes of Llama, Mistral, and others can produce signatures that look nothing like the base model.
- The "delve" origin story is plausible, not proven. Direct confirmation from OpenAI's annotator data hasn't been published.
- Identification ≠ detection. Telling which model wrote something requires model-specific classifier training on current outputs. Telling that AI wrote it is a different — and in some ways easier — problem.
- The arms race is real. GPT-5, Claude 4, and Gemini 2.5 are specifically trained to reduce known tells. WROITER has to track model releases actively, not just document what was true six months ago.