The signs of AI writing, measured

August 9, 2026 · View on GitHub

The eight structural patterns below are what actually move an AI detector, ranked by how hard they hit. Missing contractions are the single strongest signal: "The software has been significantly improved" scores 92.6% AI, while "software's got a lot better" — the same fact, contracted — scores 0.03%. Formal verbs are next: "provides excellent stability" scores 99.6%, "doesn't flex" scores 0.01%. Every number here comes from a paired test, the same information written twice.

The finding underneath all of it: detectors have learned what clean, balanced, professional prose looks like, because that's what AI writes. Formality is the tell, not vocabulary. There's no word list to avoid.

Tested: 70+ paired sentence variations, February 2026 Model: desklib/ai-text-detector-v1.01 (DeBERTa-v3-large, 304M parameters, RAID benchmark #1) Previous model: dejanseo/ai-cop (DeBERTa-v3-small fine-tune) — replaced 2026-02-16 Tool: ai-detect (see the README) — run any of these yourself

How the tests were run

Each test was a paired comparison: a formal, AI-style sentence against a conversational rewrite carrying the same information. Pairing matters — it isolates structure from subject, so the gap can only come from how the sentence is built. Score one sentence on its own and you learn very little; write the same fact two ways and you can see what the model's reacting to.

The eight patterns

1. Contractions Are the #1 Signal

Without contractionAI probWith contractionAI prob
"The software has been significantly improved"92.6%"Software's got a lot better"0.03%
"Fanatec is not the cheapest option available"53.4%"Fanatec aren't cheap"0.05%
"They have partnered with several leading brands"98.8%"They've partnered with Porsche, BMW, Bentley"0.01%

2. "Provides/Offers/Delivers" Are Detector Magnets

With formal verbAI probWith colloquial verbAI prob
"provides excellent stability"99.6%"doesn't flex"0.01%
"offers comprehensive software support"99.9%"app's actually pretty good now"0.01%
"delivers impressive performance"92.5%"genuinely impressive for what you're paying"17.0%
"encompasses multiple price points"99.5%"ranges from basic stuff to high-end"0.02%

3. "For [group] who [condition]" Is a Strong AI Signal

AI frameAI probHuman frameAI prob
"For PC-only buyers who don't need console support, Moza offer..."99.1%"If you don't care about PlayStation, Moza will save you money"0.8%
Same, subject-first variant86.8%"No PlayStation? Moza are cheaper"1.2%

4. Formal Transitions Are Instant Flags

With transitionAI probWithoutAI prob
"Additionally, the QR2 provides..."99.6%"The QR2 doesn't flex and..."0.01%
"Furthermore, Fanatec offers..."99.9%"The Fanatec app replaced FanaLab and..."0.01%
"In conclusion, Fanatec remains..."99.9%"Look, Fanatec make good kit"0.05%
"It's important to note that..."99.8%"Prices vary by region, obviously"0.12%
"This represents a significant upgrade"99.8%"Massive upgrade over the old version"0.8%

5. Specific Names Beat Vague Abstractions

"several leading automotive brands" → 98.8% AI "Porsche, BMW, Bentley and the WRC" → 0.01% AI

6. Trailing Opinions Lower AI Probability

Removing a trailing personal clause ("and I haven't seen much disagreement on that") INCREASED AI probability from 1.6% to 18.6%.

7. Patterns That Always Pass as Human

  • Typos: 0.0% AI
  • Colloquial idioms ("how far down the rabbit hole"): 0.25% AI
  • "I reckon" + trailing "honestly": 0.3% AI
  • Fragment questions ("No PlayStation?"): 2.9% AI
  • Short blunt statements ("Fanatec aren't cheap"): 0.05% AI
  • Dash-interrupted sentences with opinion: 0.01-0.03% AI
  • Possessive contractions ("Software's got"): 0.03% AI

8. The Word "Noticeably" Is Flaggy

"The consensus has shifted noticeably" triggers AI regardless of framing (78-82% AI). "The mood has definitely changed" passes (5.4% AI).

The rule underneath all eight

Simplify, don't engineer. Every formal construction has a casual equivalent that reads as human. The detector has learned what clean, balanced, professional prose looks like — and that's what AI writes. Write like you're explaining something to a mate, not writing a press release.

What this is not

It isn't a guide to beating a detector. The patterns run one way: formal phrasing reads as machine-written because machines are trained on formal phrasing, so fixing a flagged sentence and fixing a badly written one are the same edit. Contract the verbs, name the actual brands, put the opinion back in.

It also isn't a verdict engine. A confident writer who contracts their verbs sails through, and a careful human writing a spec sheet will flag. Across a batch of forty articles the signal is real; on any one sentence it is a prompt to look closer, not a finding.

Checking your own text

pip install .
ai-detect --file draft.md

Every sentence is scored, and every flagged one comes back with the pattern that triggered it and a plain rewrite. The full tool, the model choices and the privacy position are in the README.