Reading about AI writing patterns is useful. Reading actual annotated examples is better. Each sample below was generated by the named model in early 2025 in response to a standard prompt, then analysed for key signals. Scores are approximate results from our free AI detector. Pay attention to the annotations — those are the exact signals that push a score up or down.
The prompt used for all AI samples: "Write a paragraph explaining the importance of sleep for cognitive performance."
Example 1 — ChatGPT (GPT-4o)
Sleep is a [VOCAB] fundamental component of cognitive health, playing a [VOCAB] pivotal role in memory consolidation, emotional regulation, and executive function. During sleep, the brain engages in a [VOCAB] meticulous process of synaptic pruning and waste clearance through the glymphatic system, which is essential for maintaining optimal neural performance. Research consistently demonstrates that individuals who obtain seven to nine hours of sleep per night exhibit superior problem-solving abilities, enhanced creativity, and greater emotional resilience compared to those who are sleep-deprived. [TRANSITION] Furthermore, the consequences of chronic sleep deprivation extend far beyond simple fatigue, encompassing increased risk of cognitive decline, impaired decision-making, and diminished capacity for learning. [CLOSING] In conclusion, prioritising adequate sleep is not merely a lifestyle choice but a scientifically validated necessity for sustaining high-level cognitive performance.
Signals detected: 3× elevated vocabulary [VOCAB] (fundamental, pivotal, meticulous), 1× formal transition opener [TRANSITION] (Furthermore,), 1× closing ritual [CLOSING] (In conclusion,), uniform sentence length (~22 words avg), balanced hedging absent but comprehensive coverage present.
Estimated detector score: ~87% AI — High confidence. The clustering of vocabulary tells plus the closing ritual is near-diagnostic for ChatGPT.
Example 2 — Claude (claude-3-5-sonnet)
The relationship between sleep and cognitive function is [HEDGE] worth considering carefully, as it touches on one of the more [VOCAB] nuanced intersections in neuroscience. [HEDGE] It's perhaps not surprising that memory consolidation occurs primarily during slow-wave sleep — the brain, after all, needs uninterrupted processing time to transfer experiences from short-term to long-term storage. What's [HEDGE] less often discussed, though, is how even a single night of reduced sleep can measurably impair prefrontal cortex activity, affecting judgment and impulse control in ways that feel subjectively like nothing more than tiredness. Sleep debt accumulates in subtle ways that people tend to underestimate.
Signals detected: Claude's characteristic philosophical hedging [HEDGE] ("worth considering," "it's perhaps not surprising"), 1× elevated vocabulary [VOCAB] (nuanced), no closing ritual, more varied sentence length than GPT, but still lacks genuine personal specificity.
Estimated detector score: ~74% AI — Moderate-high confidence. Claude's more natural hedging style reduces the score versus ChatGPT, but the absence of any personal experience or specific anecdote is still detectable.
Example 3 — Gemini (Gemini 1.5 Pro)
Sleep is essential for cognitive performance for several key reasons:
- Memory consolidation: During sleep, the hippocampus replays and transfers experiences to the neocortex for long-term storage.
- Toxin clearance: The glymphatic system flushes metabolic waste — including amyloid-beta — from the brain during sleep.
- Emotional regulation: REM sleep processes emotional memories, reducing their intensity and enabling clearer thinking.
- Executive function: Even mild sleep restriction measurably reduces prefrontal cortex activity and decision quality.
Research consistently shows that adults need 7–9 hours of sleep per night. Chronic sleep deprivation is associated with increased risk of cognitive decline and neurodegenerative conditions. Prioritising sleep is one of the highest-leverage investments in long-term brain health.
Signals detected: Gemini's characteristic heavy structuring with bullet points and headers [STRUCTURE], comprehensive coverage across all standard sub-topics, "consistently shows" as a hedge, formal closing that summarises. Note the lack of any conversational register or personal element.
Estimated detector score: ~80% AI — High confidence. The compulsive structuring with bullets is a strong Gemini signal. This pattern is very rare in naturally written human prose for a single-paragraph answer.
Example 4 — Humanised AI Text (Same ChatGPT prompt + paraphrasing tool)
Getting enough sleep genuinely changes how sharp you feel the next day — and the science backs this up. While you're asleep, your brain is busy consolidating what you learned, clearing out metabolic junk, and resetting your emotional responses so they're not quite so raw. Cut your sleep short and you'll notice it most in the higher-order stuff: planning, staying calm under pressure, making clear-headed calls when you're tired. Seven to nine hours isn't just a guideline — it's where most people's performance actually sits at its best.
Signals detected: Vocabulary has been substantially simplified — no "meticulous," "pivotal," or "fundamental." Sentence lengths more varied. No formal transition openers. No closing ritual. However: still lacks genuine personal anecdote, still covers all standard sub-topics comprehensively, still uses hedged qualifiers ("most people").
Estimated detector score: ~44% AI — Uncertain zone. The paraphrasing has successfully reduced surface signals, but statistical patterns at the sentence level still carry AI probability. This is why humanised AI is a challenge — and why detecting humanised AI requires more than vocabulary scanning.
Example 5 — Real Human Writing
I used to think I was fine on six hours. I had the coffee, I had the adrenaline of a deadline, and I got things done — or thought I did. It was only when I started tracking my reaction times in a simple online test that I saw what sleep deprivation actually cost: my scores on mornings after less than seven hours were consistently 15–20% slower than my rested baseline. That's not just "tired." That's impaired. The cognitive debt doesn't show up in how you feel, it shows up in the quality of what you produce.
Signals detected: First-person narrative with specific personal detail (coffee, adrenaline, deadline, reaction time test, actual percentage numbers from personal tracking), highly varied sentence length (4 words to 28 words), informal register ("or thought I did"), genuine opinion ("That's not just tired"). Zero AI vocabulary tells.
Estimated detector score: ~12% AI — Very likely human. The specificity, personal voice, and informal register are difficult to fake even with prompting. AI can be asked to write in first person, but the genuine specific numbers from personal experience are almost impossible to authentically replicate.
What the Scores Tell Us
The pattern across these five examples is instructive. The highest AI scores come from text with clustered vocabulary signals plus structural tells (closing ritual, formal transitions). Claude scores lower than ChatGPT because it has learned to hedge more naturally. Gemini's structural compulsiveness is its main tell. Humanised AI occupies an uncertain middle zone where vocabulary signals are scrubbed but statistical patterns remain. Human writing, especially with personal specifics, sits far from AI probability distributions.
The practical implication: if you want to see where a real piece of text sits, paste it into our free AI detector and compare the score to the profiles above. A score above 80% is comparable to the ChatGPT example — strong clustering of signals. A score in the 40–60% range is comparable to the humanised example — possible AI evasion or genuinely mixed-origin writing. Below 20% is comparable to the human example — very likely authentic. For more on the differences between AI and human writing, see our dedicated comparison guide, and check out our ChatGPT detection page for model-specific analysis.