Burstiness Over Time: How AI Writing Patterns Have Shifted Across Model Generations

Tracking AI writing burstiness over time shows a steady climb: the flat, metronomic rhythm that made GPT-2 and GPT-3 easy to spot has given way, generation by generation, to deliberately varied prose — and detectors have had to retreat from their once-best signal. Understanding that shift explains a lot of otherwise confusing detection behavior, including why a 2023-era rule of thumb (“AI text sounds even”) now misfires in both directions.

Key takeaways

• Burstiness — the variance in sentence length and complexity — was one of the first practical AI-detection signals because early models genuinely lacked it.
• The flatness wasn’t an accident of intelligence; it fell out of how models decode text, favoring safe, high-probability continuations.
• Each model generation since has gotten rhythmically more human: better decoding strategies, RLHF-shaped style, and training that rewards engaging prose.
• No public study tracks one burstiness metric across every model generation — the trend is visible in the research record and in how often detectors retrain, not in a single tidy chart.
• Burstiness survives today as one feature among many, still useful against cheap and templated output, no longer decisive against frontier models.

What burstiness measures, in one minute

Take any paragraph and plot the length of each sentence. Human writing produces a jagged line. Seventeen words, then thirty-one, then four. Writers speed up and slow down without thinking about it — that variance is burstiness. The concept borrows from an older statistical observation that words in human text arrive in bursts rather than at uniform rates.

Detectors care because variance is hard to fake and easy to measure. Perplexity asks “how predictable is each word?”; burstiness asks “how much does the rhythm swing?” They’re related but distinct, and we’ve unpacked the pair properly in our piece on burstiness and perplexity, the two signals detectors measure.

Generation one: the flatline era

Early generative models earned their reputation for evenness honestly. Research from the period documents it from several angles. The GLTR tool (2019) showed that GPT-2 text was dominated by high-probability word choices — visualized, it lit up green with predictability. Holtzman and colleagues’ “Curious Case of Neural Text Degeneration” (2020) explained why: decoding strategies that chase likely words produce text that is repetitive and weirdly uniform compared to the spiky probability profile of human prose. And Ippolito’s team found the paradox that defined the era — the decoding tricks that made text *sound* better to humans made it easier for machines to catch, and vice versa.

The practical upshot: a GPT-2 or GPT-3 essay read like cruise control. Sentences of similar length, ideas advancing at a constant speed, no risk-taking. Burstiness barely needed measuring; you could feel it.

Generation two: polished, but still even

The chat-tuned models that arrived with ChatGPT changed the texture without fixing the rhythm. RLHF — training models against human preferences — produced the now-famous house style: helpful, hedged, structured, allergic to fragments. Polished in a way GPT-2 never was, but still rhythmically narrow. The tell shifted from “robotic repetition” to “relentless mid-length competence”: sentence after sentence in the same 15-25 word band, every paragraph shaped the same way.

This is the era that built the detection industry. GPTZero launched in early 2023 with perplexity and burstiness as its two headline signals, and for that generation of models, the pair worked well enough to go viral. It’s also the era that produced most people’s mental model of “what AI writing sounds like” — a model that’s aging badly.

Generation three: variance on purpose

Newer frontier models write with noticeably more rhythmic range, for converging reasons. Training data and preference tuning began rewarding engaging, magazine-style prose rather than encyclopedic evenness. Users hammered every model with “don’t sound like AI,” and labs noticed. Style controls got better, so a one-line prompt now reliably produces fragments, asides, and long-short contrast that 2022 models couldn’t sustain.

The result shows up wherever people measure current-generation output: sentence-length variance closer to the human range, fewer of the old uniformity tells. We should be precise about the evidence here — there is no definitive longitudinal study charting one burstiness metric across every model release, and vendors don’t publish that data. But the indirect evidence is consistent: detector companies retrain constantly, the burstiness-only heuristics of 2023 demonstrably misfire on frontier output, and the research community has largely moved from rhythm statistics toward learned classifiers and paired-model methods. We looked at the adjacent question — whether newer GPT models became harder to detect — and found the same pattern: each new generation opens a gap that retraining then partially closes.

What this did to detectors

The shift forced a quiet architectural change across the industry. Burstiness didn’t disappear from detectors; it got demoted. Where 2023 tools could nearly lead with it, modern systems treat rhythm variance as one feature inside a larger learned model, weighted alongside perplexity profiles and classifier features trained on fresh output.

Two consequences matter for anyone reading scores today:

• Old intuitions over-flag distinctive humans. Writers with naturally even cadence — technical writers, second-language writers taught to keep sentences short and safe — look “AI-flat” to rhythm-based heuristics. That failure mode existed in 2023; it’s relatively worse now that the signal separates the classes less cleanly. It’s part of why OpenAI’s own classifier retired within six months for low accuracy.
• Rhythm tricks help evaders less than advertised. Prompting a model to “vary sentence length” shifts one measured feature while leaving others — word-choice predictability, semantic evenness — largely intact. Mechanical variation is itself a pattern.

AI writing burstiness over time: the honest read on where this goes

Burstiness is following the path every surface-level signal follows in an arms race: strong while the gap is real, weaker as the generator learns the evaluator’s test. What tends to survive is depth — specificity, genuine opinion, information the model didn’t have. A draft revised to say something concrete in your own cadence moves a detector more than any rhythm hack, because it changes many signals at once.

If you want to see how your own writing scores on these measures, you can try it on your own text — the heatmap shows exactly which sentences read as flat. For heavier workloads, see what each plan handles, and for the wider context, browse the rest of our writing on detection.

Frequently asked questions

What is burstiness in AI writing? Burstiness is the variation in a text’s rhythm — how much sentence length and complexity swing from one sentence to the next. Human writing is bursty: a long winding thought, then a three-word punch. Early AI writing was famously even, with sentences clustered around the same length and shape, which made low burstiness one of the first reliable detection signals.

Has burstiness in AI text actually changed across model generations? Directionally, yes. GPT-2 and GPT-3-era output was measurably flat and repetitive. Post-RLHF chat models developed a house style that was polished but still rhythmically even. The newest generation varies sentence length far more, partly because training rewarded engaging prose and partly because users kept asking for writing that doesn’t sound like AI. There’s no single longitudinal study tracking one burstiness metric across every model, but detector vendors’ repeated retraining tells the same story indirectly.

Do AI detectors still use burstiness? Most still include it, but few lean on it alone. As models learned to vary rhythm, burstiness weakened from a headline signal into one feature among many, alongside perplexity and learned classifier features. A modern detector that relied on burstiness by itself would miss most current-generation output.

Can I just ask a model to write with high burstiness and beat detectors? Prompting for varied sentence length does shift the statistics, and it’s one of the oldest evasion tricks for that reason. It helps less than people expect: detectors read many signals at once, and prompted variation tends to be mechanical — alternating long-short-long on a metronome, which is its own kind of pattern. Genuine revision that adds specifics and opinion moves scores more than rhythm tricks alone.

If new models write with human-like rhythm, is burstiness dead as a signal? It’s diminished, not dead. On short text it was always noisy, and on current models it separates human from machine far less cleanly than in 2023. But it still contributes in combination with other features, and flat rhythm remains a useful tell for older models, cheap models, and templated bulk content — which is still a huge share of AI text in the wild.

The bottom line

Burstiness over time is the story of the whole detection field in miniature. A real, measurable gap between human and machine writing existed; it powered the first useful detectors; and then the models closed it, generation by generation, until the signal became one modest feature in a bigger ensemble. The rhythm gap is mostly gone. What still separates human writing from the machine kind isn’t the shape of the sentences — it’s whether they contain anything only you could have written.