Did GPT-4 to GPT-5 Make AI Writing Harder to Detect? A Look at the Evidence

Every time a new flagship model drops, the same question lights up in forums and faculty Slack channels: did this one finally outsmart the detectors? GPT-4 to GPT-5 was no exception. Plenty of people noticed their AI-assisted drafts started scoring lower and concluded the writing had simply gotten too good to catch.

The real story is more interesting, and a lot less magical.

Key takeaways

• Newer models often do score lower on some detectors, but the main reason is that detectors lag behind new releases, not that the text became truly “human.”
• Detection is a moving target. Each big model launch opens a short window where older detectors are shakier until they’re retrained.
• The measurable shift isn’t raw eloquence. It’s that newer models default to more varied, less repetitive phrasing, which nudges the statistics detectors watch.
• Detectors report a probability, not proof. A low score on new-model text doesn’t mean undetectable, and a high score never proves authorship.
• The skill that lasts is editing for genuine voice and specifics, not chasing whichever model happens to score low this quarter.

What “harder to detect” usually means

When people say a newer model is harder to detect, they’re almost always describing one observation: they fed it a prompt, ran the output through a detector, and got a lower AI score than they used to.

That’s a real result. But it blends together two very different explanations.

The first is that the model genuinely writes more like a person. The second is that the detector just hasn’t caught up to the new model’s output yet. In the short term these produce identical scores, and people mix them up constantly.

Most “newer models beat detectors” claims are really the second thing. A detector is itself a model, trained on examples of human text and AI text. If almost all its AI examples came from GPT-3.5 and GPT-4, it learned what *those* models sound like. A new model with a slightly different default style is, briefly, unfamiliar territory. Lower scores follow, until vendors retrain on fresh samples and the gap narrows again.

A quick history check: detectors have always lagged

This isn’t behavior unique to GPT-5. It’s the pattern every model generation has followed.

Worth remembering: OpenAI launched its own AI Text Classifier in January 2023 and quietly pulled it that July, citing a low rate of accuracy. Detection was hard even when the company building the detector also built the model it was trying to catch. The arms-race framing is fair, but it’s a slow, lopsided race where the detectors are usually a step behind by design.

So when GPT-5-era output scored lower at launch, that fit a years-long rhythm. New model ships, scores dip, detectors update, scores partly recover. Calling that “the model won” is like calling halftime the final score.

The thing that genuinely changed: phrasing, not genius

Here’s the part with actual substance. Newer models didn’t just get “smarter.” Their default output got statistically less robotic in a specific, measurable way.

Older models had a tell: they were too predictable. Word after word, they picked the safe, expected option. That predictability is exactly what many detectors measure through perplexity and burstiness, the two statistics most detectors lean on. Low perplexity means the text is easy to predict. Low burstiness means the rhythm from sentence to sentence stays flat. Human writing spikes and dips. Older AI writing tended to hum along at one level.

Newer models, partly through different training and tuning, default to more varied word choices and a bit more rhythmic variation. That pushes their statistical fingerprint closer to the human range, not because they understand anything more deeply, but because their probability distributions over the next word flattened out and widened.

The catch: this is a difference in *degree*, not a clean escape. The signals shifted toward human territory. They didn’t disappear. And the moment detectors retrain on new-model samples, that shifted fingerprint becomes the new thing they look for.

An illustration

Picture a student, Maya, drafting a short literature response. (The numbers below are made up to show the shape of the problem, not measurements from any real test.) She pastes older-model output into a detector and gets a high AI score. A year later, same assignment, newer model, same kind of prompt, and the score comes back far lower.

It’s tempting to read that as “the AI got undetectable.” But two other things could easily be true at once. The newer model used less formulaic phrasing out of the box, *and* the detector she used hadn’t been updated in months. Run the same passage through a second, recently retrained detector and the number could jump right back up.

Same text, very different numbers depending on the tool and the month. That spread is the real lesson, not any single low score.

Why you shouldn’t trust a single low score

If the newer-model bump taught us anything, it’s that detector outputs are noisy and time-sensitive.

A few reasons one low score means very little:

• Scores drift with updates. Text that reads as low-AI today can read as much higher after a detector retrains. Nothing about your writing changed; the measuring tool did.
• Detectors disagree with each other. The same paragraph can score low on one tool and high on another, because they were trained on different data with different thresholds.
• Short text is unreliable in both directions. A handful of sentences doesn’t give these statistics enough to work with, so scores swing.
• Probability isn’t proof. Reputable vendors say plainly that their output is an estimate. A number near zero doesn’t certify “human,” and a number near 100 doesn’t prove anyone used a specific tool.

There’s also a fairness angle that should make anyone cautious about treating scores as truth. A 2023 study from a Stanford team led by Weixin Liang, published in *Patterns*, found that GPT detectors were biased against non-native English writers, flagging genuinely human writing as AI at high rates. Across the essays they tested, the detectors misclassified a majority of the non-native samples while rarely misflagging native-English ones. The likely culprit: second-language prose tends to have lower perplexity, the same plainness the statistic mistakes for machine output. The very machinery that makes detectors lag on new models also makes them misfire on real people.

So, are newer models harder to detect? An honest answer

Yes, temporarily and partially, and mostly for reasons that have little to do with the writing being “better.”

Stack it up:

• Right after launch, newer models often score lower, because detectors were trained on older output. Real, but short-lived.
• After detectors retrain, the gap shrinks. Scores on new-model text climb back toward where older models sat.
• Over several generations, default phrasing has gotten less formulaic, which moves the statistical fingerprint but doesn’t erase it.
• Throughout all of it, detection stays probabilistic, lagging, and imperfect, no matter the model. That hasn’t changed and isn’t about to.

If your mental model is “find the newest model and you’re invisible,” you’re optimizing for a window that closes. The text that scored low at launch may score much higher after the next detector update, with zero edits from you.

What actually holds up over time

The traits that make writing read as a person’s are the same ones detectors *and* human readers respond to, and they don’t depend on which model you used.

Concrete details. A real example from your own experience. An opinion you’d actually defend. Sentence rhythm that breaks up, with short punches between longer thoughts. The specific over the generic. Even good models tend to default to smooth, safe, slightly hollow prose. Editing that back toward your own voice is the part that survives every detector update.

That’s also why “which model is hardest to detect” is the wrong question to obsess over. The model hands you a draft. What you do with the draft is what lasts. If you’re curious how detection keeps evolving, you can browse the rest of our writing on detection for the deeper mechanics, and you can compare what each plan includes if you want to put a real workflow around your editing.

Frequently asked questions

Are newer AI models actually harder to detect than older ones? Often, yes, but with a caveat. When a new model ships, detectors trained mostly on older output score the new text lower for a while. That gap shrinks once detectors retrain on the newer model. “Harder to detect” is frequently a timing effect, not a permanent property.

Did GPT-5 specifically beat AI detectors? There’s no credible blanket evidence that any single model permanently beats detectors. Newer models can score lower on some tools, especially right after release, but results vary a lot by detector, prompt, and length. The same passage can get very different scores from different tools.

Why would a newer model score lower at all? Most detectors lean on statistical patterns, like how predictable each word is. Newer models default to more varied, less repetitive phrasing, which nudges those patterns toward the human range. Detectors trained on older, flatter output haven’t learned the new distribution yet.

Does a low AI score mean my text is safe? No. Detectors give probabilities, not verdicts. A low score today can rise after a detector update, and short text produces unreliable scores both ways. Treat any single number as one weak signal.

What’s the reliable way to write with AI without sounding obviously AI? Edit for what models skip: specific details, your own examples, a real opinion, and natural rhythm. Those traits make writing read as yours regardless of the model, and they hold up better than gaming one detector.

The bottom line

Newer GPT models did make AI writing *temporarily* harder to detect, mostly by shifting their default phrasing and by outrunning detectors that hadn’t caught up. Neither of those is a permanent shield. The scores move, the tools update, and the fairness problems persist on all sides.

Chasing the lowest-scoring model is a treadmill. Writing with real specificity and a voice that’s yours is the thing that actually lasts. Want to see where your own draft stands and how editing changes it? Paste a passage and see how our checker scores it.