Can AI Detectors Detect GPT-5 and the Newest Models? Yes — After a Lag, and the Lag Is the Story

Mostly yes, but with a catch that matters: every detector is trained on yesterday's models. When GPT-5 launched in August 2025, some major detectors visibly stumbled for days to weeks, then retrained and recovered. The same cycle repeated with GPT-5.5 in 2026. Detection works — on a delay. If you're relying on a score the month after a big release, that delay is your problem.

This article is about that delay: where it comes from, how big it actually gets, what vendors claim about it (with receipts), and why it keeps reopening no matter how many times the industry declares it closed. A note on scope before we start — this is a different question from "can detectors catch text from my model brand." Detectors don't target brands at all, and if that's your question, our companion piece on whether AI detectors detect Claude answers it. This piece is about time: the gap between a model shipping and detectors catching up.

One disclosure, since we'll be quoting detection vendors throughout: we build a detector and humanizer ourselves. Here's our editorial policy — judge accordingly.

Why a lag exists at all

An AI detector is a classifier, and classifiers learn from examples. A vendor assembles a large corpus of human writing and a large corpus of machine writing, trains a model to separate them, and validates it against held-out samples. The machine half of that corpus can only contain output from models that exist. That's not a design flaw anyone can engineer away; it's an arrow-of-time problem.

So when a genuinely new frontier model ships, the detector meets its output cold. Whether detection holds up depends on one question: does the new model's prose land inside the statistical territory the classifier already maps? Usually, mostly, yes — new models are still transformers choosing high-probability words, so their output stays low-perplexity and low-burstiness in the ways detectors key on (the mechanics are in our how detectors work explainer). That shared statistical core is why detection survives model launches at all.

But "mostly inside the territory" is not "identical to the old models." Each generation shifts the distribution: different fine-tuning, different default rhythms, sometimes deliberate work by the model's maker to reduce recognizable house style. The best independent evidence that these shifts hurt detectors comes from the RAID benchmark (ACL 2024), which tested twelve detectors against six million generations and found them consistently vulnerable to exactly this: "unseen generative models," along with unfamiliar sampling settings — a repetition penalty alone cut accuracy by up to 38 points. Detectors generalize to models they've never seen imperfectly. The imperfection is the lag.

GPT-5, August 2025: a detection dip in real time

The GPT-5 launch on August 7, 2025 is the best-documented case study we have, because for once the vendors published their numbers in public, week by week, and the numbers disagree with each other in instructive ways.

Originality.ai tested GPT-5 shortly after launch and said the quiet part out loud: GPT-5 was initially "less detectable than GPT-4o." Its own detector scored well by its own account — 94% on newly generated GPT-5 content and 99% on GPT-5 rewrites of human text, each on a quick 100-sample test — but its competitor spot-check was brutal: on 10 GPT-5 samples, it reported GPTZero catching 1 of 10, QuillBot 2 of 10, Grammarly 2 of 10, and Copyleaks 5 of 10. Ten samples is a spot-check, not a study, and it was run by a competitor — both caveats matter. But the direction of the finding was corroborated elsewhere.

Pangram, publishing on launch day itself, claimed 99.8–100% detection across GPT-5 variants on roughly 3,700 samples per variant without any GPT-5-specific training — and reported that in its own 10-prompt test, competing detectors including GPTZero, ZeroGPT, and Originality.ai misclassified GPT-5 output as human. Notice the shape of the disagreement: Originality says it caught everything and GPTZero missed; Pangram says Originality missed too. Both are self-published, same week, tiny adversarial samples. When vendor tests contradict each other this directly, the honest reading is not "one of them is lying" but "10-sample launch-week tests measure noise as much as signal."

GPTZero is the most useful witness, because it documented its own catch-up in public. On August 14 — one week post-launch — it published benchmarks showing 95% recall on GPT-5 (92.2% on GPT-5-mini) at a 1% false-positive rate, stating plainly that its model had "NOT been trained to detect GPT-5 text" and had been trained on GPT-4.1, o3, Gemini 2.5, and Claude Sonnet 4. Ten days later, on August 24, it shipped an update trained on actual GPT-5 output and reported recall rising to 97.6% on GPT-5 and 97.1% on GPT-5-mini. That's the retraining cycle compressed into seventeen days, with before-and-after numbers attached: real but recoverable degradation, closed by feeding the new model's text back into training.

Put the three accounts together and you get the honest composite. Unedited GPT-5 text was never invisible — even untrained detectors caught most of it, because GPT-5 still writes like a language model. But the launch window produced measurably worse performance on some tools, wildly inconsistent results across tools, and a flurry of contradictory vendor claims. If your institution ran a batch of submissions through a detector in mid-August 2025, the scores were less reliable than the same tool's scores in October. Nobody flagged that on the report.

Then it happened again: GPT-5.5

The cycle is not a one-time event; it's a loop. When OpenAI released GPT-5.5 on April 23, 2026, Originality.ai published a fresh evaluation on May 5 — reporting 94% accuracy for its Lite model and 96% for Turbo on a 5,000-sample-pair test, while noting explicitly that "GPT-5.5 didn't yet exist" when its detectors were trained. Read that fine print the way it deserves: mid-90s accuracy on an unseen model is a genuinely good generalization result, and it's several points below the ~99% the same vendor claims on models it has trained on. A few points of accuracy sounds small until you translate it into essays: on 10,000 documents, the difference between 99% and 94% detection is hundreds of additional wrong calls — misses, or worse.

Every future frontier release re-runs this loop. New model ships; detectors run on generalization alone for days or weeks; vendors scramble to collect the new model's output and retrain; accuracy recovers; blog posts announce victory; next model ships. The gap keeps reopening because it structurally must — the only way to train on tomorrow's model today is to be the company building it, and even that didn't help OpenAI, whose own classifier managed just 26% detection of AI text before being retired in July 2023 for low accuracy.

The launch-window timeline, with receipts

One table, one lesson: every strong number comes with a condition, and the conditions cluster around the same variable — whether the detector had seen the model before. When you read any vendor's accuracy claim, the first question to ask is which models, tested when, trained on what. (That skill generalizes; our accuracy hub exists to teach it.)

Why vendors can close the gap so fast — and why that's cold comfort

Seventeen days from "not trained on GPT-5" to "retrained, +2.6 points" is genuinely fast, and it's worth being fair about why. Collecting training data from a new model is easy — you buy API access and generate a corpus overnight. The statistical core detectors rely on doesn't move much between generations, so retraining is refinement, not reinvention. And vendors are commercially desperate to close launch gaps, because "detector misses new model" headlines are existential for them. The competence is real. Turnitin, GPTZero, Originality, Copyleaks — these teams have now survived multiple frontier launches, and detection of unedited output keeps climbing back to wherever that tool sat before the launch. Where that is depends entirely on the tool: on RAID's threshold sweep, Originality bottomed out at a 0.62% false-positive rate and ZeroGPT at 16.9%, on the same corpus, on the same day.

Here's why that doesn't settle the question. The retraining cycle fixes tomorrow's scores. It does nothing for the scores already generated during the gap — and nothing in a detection report tells you whether it was produced during one. A percentage looks equally confident on August 10 and October 10. The August number was quietly worse, and neither the instructor reading it nor the student accused by it had any way to know.

There's also an asymmetry worth naming. During a launch window, detectors mostly fail in one direction on the new model — false negatives, AI text passing as human. That failure hurts the detection business but doesn't directly hurt innocent people. The direct human damage comes from the other error, false positives, and those don't spike because a new model shipped — they're a permanent baseline (Liang et al. in Patterns found an average 61.22% false-flag rate on non-native English speakers' essays across seven detectors, on no new-model launch at all; see our false positives explainer). But launch windows corrode something subtler: trust calibration. A tool that was 99% last month and 94% this month, with the same confident interface, teaches its users a precision it doesn't currently have. Decisions get made on last month's accuracy with this month's tool.

What this means if you rely on a score

For instructors and editors: treat the month after any major model release as a known low-confidence period. A "human" verdict on a suspiciously polished submission means less in that window than usual — and a borderline AI score deserves even more skepticism than we'd normally urge. If the stakes are real (a grade, a job, a byline), a detector score should be the start of a conversation in any month, but especially in a launch month. Check when your vendor last shipped a model update; the good ones publish changelogs, and the changelog date versus the model's release date tells you exactly where in the cycle you're standing.

For writers being screened: the lag is not a loophole worth building on. Any advantage a brand-new model gives you expires in weeks, silently, when the vendor retrains — and text you submitted during the gap can be re-scanned after it closes. Betting on the gap means betting that nobody ever looks again. Meanwhile the thing that actually determines how text scores — how much of it a human genuinely wrote and rewrote — doesn't expire. That's the subject of our piece on edited and mixed AI text, and it's a far sturdier place to stand than a countdown you can't see.

For what it's worth, this lag applies to us too. Our AI detector faces the same arrow-of-time problem as everyone else's — a model released tomorrow is a model we haven't trained on — and it doesn't promise to catch any particular model or to beat any benchmark, because nobody can honestly promise that. We publish no accuracy percentage for it without published methodology. The whole premise of this pillar is that you should demand the same from every vendor quoted above.

The question that stays open

Will detectors eventually get ahead of the cycle — catching models before seeing them? Pangram's launch-day GPT-5 numbers, if they hold up to independent scrutiny, suggest strong generalization is possible. Originality's unseen-model GPT-5.5 results (94–96%) point the same direction. Watermarking, where the model's maker embeds a signal at generation time, would end the lag entirely for cooperating models — but it requires the generator's cooperation and degrades under heavy paraphrase, which is a different bargain altogether. The realistic near-term future is the one we already live in: a cycle of gap and catch-up, with unedited AI text detectable most of the time, and a few weeks of fog after every launch.

The models will keep shipping. The gap will keep reopening. The scores will keep looking equally confident on both sides of it. Knowing which side you're on is, for now, your job — the report won't tell you.

FAQ

Can AI detectors detect GPT-5? Yes — current versions of major detectors catch unedited GPT-5 text at high rates, and vendors have published post-retraining figures in the 96–99% range. During the first weeks after the August 2025 launch, though, performance was measurably worse and inconsistent across tools.

Can AI detectors detect GPT-5.5? Largely yes. Originality.ai reported 94–96% accuracy on GPT-5.5 shortly after its April 2026 release, while noting the model "didn't yet exist" when its detector was trained. Expect those numbers to have risen since as vendors retrain on GPT-5.5 output.

How long does the detection lag last after a new model launches? The documented GPT-5 case suggests days to a few weeks for major vendors: GPTZero shipped a GPT-5-trained update 17 days after launch, raising its reported recall from 95% to 97.6%. Smaller or slower vendors may lag longer, and no tool tells you mid-scan whether it has been updated.

Does a new model ever fully evade detectors? Not in any documented case among major releases. New frontier models still write statistically machine-typical prose, so even untrained detectors catch most unedited output — GPTZero reported 95% on GPT-5 before any GPT-5 training. The gap shows up as a dip of several points and tool-to-tool inconsistency, not invisibility.

Is using the newest model a way to avoid detection? It's a bad bet. Any launch-window advantage expires silently within weeks when vendors retrain, and previously submitted text can be re-scanned afterward. If AI use is prohibited in your context, using a newer model to hide it is a violation either way.

Why do vendor tests of the same model disagree so sharply? Launch-week tests are small (often 10–100 samples), self-published, and run by competitors on each other. In August 2025, Originality reported GPTZero catching 1 of 10 GPT-5 samples while GPTZero's own benchmark showed 95% recall — a gap explained by sample size, sample construction, and incentive, not by one party lying.

Do detectors detect Claude's newest models the same way? The same dynamic applies to every maker's releases — the lag is about model freshness, not brand. For the separate question of whether detectors handle Claude differently from other brands at all, see our dedicated article on Claude detection.

Should schools pause AI detection after a major model release? Pausing is a policy call, but at minimum scores from launch windows deserve wider error bars: more human review, less weight on borderline results, and awareness of the vendor's update timeline. A score is evidence, never a verdict — in launch months it's weaker evidence than usual.

Key facts

• GPT-5 launched August 7, 2025; within the first weeks, Originality.ai stated GPT-5 was initially "less detectable than GPT-4o," and its 10-sample spot-check reported GPTZero catching 1/10 and Copyleaks 5/10 (Originality.ai blog, Aug 2025).
• GPTZero, August 14, 2025: 95% recall on GPT-5 at a 1% false-positive rate without GPT-5 training data; its model had been trained on GPT-4.1, o3, Gemini 2.5, and Claude Sonnet 4.
• GPTZero, August 24, 2025: after retraining on GPT-5 output, reported recall rose to 97.6% — a documented 17-day gap-to-close cycle.
• Pangram, August 7, 2025: claimed 99.8–100% GPT-5 detection on ~3,700 samples per variant with no GPT-5-specific training, and reported competitors misclassifying GPT-5 as human on a 10-prompt test.
• GPT-5.5 released April 23, 2026; Originality.ai reported 94–96% accuracy on May 5, 2026, noting the model "didn't yet exist" when its detectors were trained.
• RAID benchmark (ACL 2024): 12 detectors were consistently vulnerable to "unseen generative models" and sampling changes — a repetition penalty alone cut accuracy by up to 38 points.
• OpenAI retired its own classifier in July 2023 at 26% true-positive / 9% false-positive — the founding proof that even a model's maker can't take detection for granted.

Sources

1. GPTZero, "GPTZero's Massive AI Detector Update for Summer 2025" (GPT-5 benchmark, Aug 14, 2025) and "GPTZero detects GPT-5 even better in our August update" (Aug 24, 2025), gptzero.me/news.
2. Originality.ai, "Is GPT-5 Content Detectable?" (Aug 2025) and "Is GPT-5.5 Content Detectable?" (May 5, 2026), originality.ai/blog.
3. Pangram Labs, "Do AI Detectors Work Against GPT-5?" (Aug 7, 2025), pangram.com/blog/gpt-5.
4. Dugan et al., "RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors," ACL 2024 (aclanthology.org/2024.acl-long.674).
5. OpenAI, AI text classifier retirement announcement, July 2023.
6. Liang et al., "GPT detectors are biased against non-native English writers," Patterns (Cell Press), 2023.