No neural network, no API call, no hidden weights. What follows is the entire method behind this free AI detector and AI checker — including the parts that make the AI detect score less confident, which is usually the part nobody publishes.
Sentences are split with an abbreviation-aware scanner so “Dr. Chen et al. (2019)” stays one sentence. Words are lower-cased and stripped of possessives, keeping internal apostrophes and hyphens. Syllables are counted with a vowel-group algorithm plus a silent-e correction and an exception list.
Almost every stylometric statistic drifts with document length — type–token ratio always falls as text grows, so a long human essay would score like a short generated one. We compute each signal inside fixed-size windows (50 words for MATTR, 200 for Zipf and entropy, 1,200 characters for compression) and average. If there is not enough material for one full window, the signal returns NaN and drops out of the vote entirely rather than contributing a value that really encodes length.
For every measured signal we evaluate two Gaussians — one centred on the human calibration mean, one on the generated mean — and take the log of their ratio. Positive means the value is better explained by the generated hypothesis. Each class keeps its own measured spread, and the AI spread is never allowed below 0.75× the human one, because a near-zero measured variance would otherwise become a spike that awards maximum likelihood to every nearby value — and each ratio is clamped to ±3.5 so no single outlier can hijack a verdict.
Naive Bayes would treat the 46 signals as 46 independent votes. They are nothing of the sort — word length, syllable count and polysyllabic ratio are three views of one underlying variable. So the combination is a linear model fitted by L2-regularised logistic regression on the real corpus: each signal gets a coefficient learned *given the others*, refitted inside every cross-validation fold. The coefficients are capped at ±2.5, because a signal whose calibration spread is unusually tight can otherwise let one measurement decide a verdict on its own. The intercept (−1.52) is a mild human prior. Because the link is logistic, the result is a calibrated probability rather than a rescaled score.
Zero-width characters, chat-export markdown, assistant scaffolding phrases, verbatim duplicate sentences, sustained impersonal register with no contractions. These are facts about the document rather than statistics about its style, so they enter the log-odds as bounded, individually-itemised terms you can see and subtract yourself. Each impact is tagged heuristic or measured depending on how its weight was set — a hand-set weight never masquerades as a fitted one.
A single percentage hides how much the evidence actually supports it. We winsorize the outer 5% of sentence logits — so one pasted quote, URL or boilerplate line cannot drag the interval by itself — then bootstrap 400 resamples of the word-weighted per-sentence logits to build a 90% interval, widened further on short inputs. Short text produces a wide interval and a low reliability grade, which is the honest answer.
It would have been easier to fine-tune a transformer. We deliberately did not, for four reasons.
Generated directly from the source the detector runs on, so this page cannot drift out of date. d is Cohen's d measured on our calibration corpus — how many pooled standard deviations separate the human and generated means. w is the weight the fusion step gives it. Signals with low d are still reported, but they barely move the score. Every centre below was fitted on real human and machine text, not set by hand.
| Signal | Human μ | AI μ | d | w |
|---|---|---|---|---|
Mean sentence length On current data the machine answers run longer (28 vs 20 words) with less variation — the opposite of the old caricature of snappy AI prose. Modest signal (d=0.86). | 19.7 | 27.8 | 0.86 | 0.54 |
Burstiness B = (σ−μ)/(σ+μ) Goh–Barabási burstiness. Widely cited as the core AI tell; on real output it separates modestly (d=0.85) because generated prose is more regular — but formal human prose overlaps it, so we weight it modestly. | -0.276 | -0.397 | 0.85 | 0.35 |
Sentence-length skewness Human sentence lengths skew right — a bulk of medium sentences with a long tail of runaways. Generated forum answers are nearly symmetric (d=0.81). Needs 10+ sentences, otherwise it stays out of the vote. | 0.746 | 0.144 | 0.81 | 0.40 |
Sentence-length variation (σ/μ) Humans mix a three-word line with a forty-word one. Real model output varies less (d=0.77) — a genuine, modest signal. | 0.586 | 0.449 | 0.77 | 0.40 |
Sentence-opener diversity Share of unique two-word sentence openings. Models recycle “This is…”, “In addition…”, “It is…”. | 0.946 | 0.878 | 0.65 | 0.40 |
Paragraph-length variation Uniform 3–4 sentence paragraphs across a whole document is a formatting habit of assistants, not of writers. | 0.640 | 0.435 | 0.53 | 0.25 |
Sentence-length kurtosis Humans produce heavy tails — the occasional runaway sentence. Weakly present in current data (d=0.26). | 0.781 | 0.264 | 0.26 | 0.34 |
Sentence-length autocorrelation (lag 1) Whether a long sentence predicts another long one. Weak on current output (d=0.25) — reported for completeness, weighted accordingly. | -0.0303 | -0.0935 | 0.25 | 0.25 |
| Signal | Human μ | AI μ | d | w |
|---|---|---|---|---|
Zipf exponent |s| Slope of log-rank against log-frequency within 200-token windows. One of the strongest signals on real output (d=1.80): human prose leans harder on a small core vocabulary, generated text spreads flatter. | 0.549 | 0.672 | 1.80 | 0.82 |
Heaps' law exponent β Rate at which new word types appear as the text grows. A low β means the vocabulary saturates early. | 0.837 | 0.755 | 1.79 | 0.57 |
Hapax legomena ratio Share of words used exactly once. Humans reach for one-off words far more often than a probability-maximising decoder. | 0.743 | 0.647 | 1.48 | 0.63 |
Zipf fit quality R² How mechanically the text obeys Zipf. Real prose is slightly untidy; generated text can fit almost too well. | 0.864 | 0.909 | 1.38 | 0.57 |
MATTR-50 lexical diversity Moving-average type–token ratio over a 50-word window. Strong on caricatured output, strong on real model prose too (d=1.12) — humans repeat ordinary words more than generators do. | 0.795 | 0.743 | 1.12 | 0.90 |
Yule's characteristic K Length-invariant vocabulary concentration. High K means a small set of words is doing most of the work. | 125 | 173 | 0.96 | 0.61 |
Polysyllabic word ratio Share of words with three or more syllables. Weak on current output (d=0.33) — same story as the two signals above. | 0.133 | 0.154 | 0.33 | 0.63 |
Syllables per word Weak separation on current output (d=0.29): real model prose is only slightly more polysyllabic than human prose. Contributes modestly today; retained for older registers. | 1.51 | 1.56 | 0.29 | 0.71 |
Mean word length Latinate vocabulary inflates average word length — a weak signal on current ChatGPT-style output (d=0.24), far from the caricature-era tell it once was. Retained for older, flowery registers. | 4.62 | 4.74 | 0.24 | 0.79 |
Function-word ratio Proportion of grammatical glue. A low ratio signals the noun-dense, information-packed register models default to. | 0.455 | 0.467 | 0.21 | 0.75 |
| Signal | Human μ | AI μ | d | w |
|---|---|---|---|---|
LZ77 compression ratio Independent redundancy check with different assumptions to the Markov model. The single strongest signal on real output (d=1.87): formulaic text squeezes smaller. | 0.805 | 0.685 | 1.87 | 0.45 |
Bigram repetition rate Share of two-word sequences that recur. Sampling loops on its own phrasing more than authors do. | 0.0694 | 0.179 | 1.80 | 0.49 |
Adaptive character entropy Cross-entropy under an order-4 interpolated Markov model built from the text itself — our offline stand-in for perplexity. Strong on real output (d=1.73): generated prose is more self-predictable. | 4.79 | 4.33 | 1.73 | 0.40 |
Coherence fragility under word shuffle Relative increase in adaptive surprisal after a fixed-seed word shuffle — the offline analogue of perturbation-based detectors. Fitted centre (d=1.72). | 0.0133 | 0.0447 | 1.72 | 0.30 |
Coherence fragility under adjacent swap Relative surprisal increase after swapping each adjacent word pair — a second, gentler perturbation in the DetectGPT lineage that preserves the local bag of words. Generated forum prose degrades far more (d=1.72); story-register output barely moves, so the dual-mode fit carries both centres. | 0.00897 | 0.0287 | 1.72 | 0.30 |
Word-surprisal dispersion (CV) Coefficient of variation of per-word adaptive surprisal. Human writing mixes predictable glue with genuinely surprising choices; generated prose holds every word near the same moderate surprise. Fitted centre (d=1.65). | 0.120 | 0.166 | 1.65 | 0.55 |
Trigram repetition rate Same idea at three words, where accidental human repetition is rare and generated repetition is not. | 0.0175 | 0.0852 | 1.46 | 0.35 |
Context gain (unigram ÷ bigram bits) How much the local bigram context reduces surprise versus the context-free unigram baseline from the same text — our offline analogue of the Binoculars normalisation. Text near a sampler’s modal ridge is disproportionately well predicted by its own neighbours (d=1.37). | 1.05 | 1.07 | 1.37 | 0.50 |
Adjacent-sentence lexical cohesion Mean Jaccard overlap of content words between consecutive sentences. Generated forum prose re-states the same propositions sentence to sentence (d=1.35 above human); fabricated story continuations push forward with new material instead (d=1.17 below). Dual-mode, scored against the nearer centre. | 0.0427 | 0.104 | 1.35 | 0.70 |
Predictable-word share (surprisal < 4 bits) Share of words the text’s own bigram model finds unsurprising — the offline form of GLTR’s “green word” tier. Instruction-tuned forum answers concentrate in the predictable head (d=0.95); story-register output does not, so this fires on the forum mode only. | 0.00577 | 0.0361 | 0.95 | 0.60 |
Normalised word entropy Shannon entropy of the word distribution over log₂(vocabulary) — how evenly the author spreads across their own lexicon. | 0.938 | 0.924 | 0.90 | 0.66 |
| Signal | Human μ | AI μ | d | w |
|---|---|---|---|---|
Function-word richness Distinct glue-word types per glue token. Forum-model output reuses a narrow shelf of connectors (d=1.34 below human); story-register output spreads wider than humans do (d=2.39 above) — a textbook dual-mode signal, scored against the nearer centre. | 0.360 | 0.297 | 1.34 | 0.55 |
Function-word bigram repetition Share of grammatical two-word collocations (“of the”, “in the”, “to be”) that recur within a 200-token window. Aligned output reuses the same handful of glue bigrams more than authors do (d=1.12 in the forum register); story-register output reuses them less than humans (d=1.31 the other way), so it is scored against the nearer centre. | 0.127 | 0.235 | 1.12 | 0.70 |
Clause density Subordinators and commas per sentence. Generated sentences carry more balanced subordinate structure. | 2.37 | 3.35 | 0.75 | 0.38 |
Function-word JS divergence Jensen–Shannon distance between this text’s function-word profile and large-corpus English. Function words are produced unconsciously, so they are hard to fake. | 0.273 | 0.299 | 0.62 | 0.40 |
Word-class entropy Entropy over coarse morphological classes (-ing, -ed, -ly, nominalisations…) as a tagger-free proxy for syntactic variety. | 0.689 | 0.681 | 0.16 | 0.59 |
Readability stability (σ FKGL) Standard deviation of Flesch–Kincaid grade across 110-word windows. Negligible on real output (d=0.07) — reported for completeness, weighted accordingly. | 0.919 | 0.862 | 0.07 | 0.35 |
| Signal | Human μ | AI μ | d | w |
|---|---|---|---|---|
First-person / personal-voice rate Personal pronouns and stance markers. Generated explanatory prose hides the speaker. | 22.1 | 2.20 | 0.86 | 0.40 |
Contraction rate “don’t”, “it’s”, “we’ve”. Real model output contracts more than the old caricature suggested, so this is a modest signal now (d=0.56), not a loud one. | 13.4 | 6.85 | 0.56 | 0.50 |
Semicolon rate Literary human prose uses semicolons freely; current assistants avoid them. The old story had this backwards — calibration corrected it. | 1.86 | 0.128 | 0.51 | 0.25 |
Parenthetical aside rate Bracketed digressions signal a mind wandering mid-sentence — a distinctly human interruption. | 6.66 | 3.44 | 0.45 | 0.36 |
Informal-register rate Colloquialisms, fillers and vague quantifiers that alignment training tends to smooth away. | 1.41 | 0.437 | 0.39 | 0.35 |
Em-dash rate Nearly absent from the 2023-style output we calibrated on, common in newer assistants and present in literary human prose. Weak either way (d=0.33) — watch this space, not the score. | 0.599 | 0.0220 | 0.33 | 0.55 |
Numeric specificity Concrete figures, dates and measurements. Generated prose leans on qualitative descriptors instead. | 12.4 | 6.36 | 0.33 | 0.25 |
Exclamation rate Assistants are trained toward measured tone and under-use exclamation compared with human authors. | 1.36 | 0.291 | 0.32 | 0.25 |
| Signal | Human μ | AI μ | d | w |
|---|---|---|---|---|
Weighted AI-marker density Weighted hits against a lexicon of instruction-tuned vocabulary, with per-phrase weights fitted to real model output — hedges and scaffolding (“it’s worth noting”, “here are some”) carry the weight now, not “delve”. | 0.201 | 2.49 | 1.18 | 0.85 |
Discourse-connective density “Overall,”, “additionally”, “in summary” signpost structure more heavily than writers do. Note “furthermore” and “moreover” actually favour humans in current data — the fitted weights know this. | 0.0651 | 0.641 | 0.57 | 0.70 |
Tricolon rate (“A, B, and C”) The rhetorical triple appears about 1.6× the human rate in current output — a real but weak tell (d=0.24), not the hallmark it once was. | 1.27 | 2.05 | 0.24 | 0.45 |
Our first calibration fitted the AI side on hand-written imitations of machine prose — and it worked beautifully, right up until we tested it on real ChatGPT output, where it scored below chance. The imitations were a caricature: real model prose looks different in measurable ways, so we threw the caricature out and refitted everything on 1,140 real ChatGPT answers plus 450 multi-model story continuations.
That refit surfaced the fact that now shapes the whole engine: the two machine registers sit on opposite sides of human prose. Lean forum answers are lexically poorer and more repetitive than humans; rich story output is lexically richer. A single AI centre cannot cover both, so every signal carries two — one per register — and scores against the nearer one. Which centre won is shown on each row of your report.
What actually separates the classes now is compressibility and self-predictability: LZ77 ratio (d = 1.87), adaptive character entropy (1.73), the Zipf exponent (1.80), shuffle fragility (1.72), swap fragility (1.72) and bigram repetition (1.80). Classic rhythm tells like burstiness (0.85) and sentence-length variation (0.77) contribute modestly. Word length, syllables per word and polysyllabic ratio — the loudest signals of the caricature era — measure only weakly (d ≈ 0.24–0.33) on current output and contribute modestly. We reweighted the whole model around these measurements rather than around received wisdom, and the weights above are the outcome.
Provenance first, physics second. We never let a statistical guess override a record the file carries about itself.
We walk JPEG segment markers, PNG chunks and RIFF containers by hand — no decoding library — to recover:
trainedAlgorithmicMedia declarationtEXt/iTXt/zTXt chunks, where Stable Diffusion writes its prompt and ComfyUI writes the entire workflow graphWhen metadata has been stripped — as it is by every social platform — we fall back on capture physics:
A d figure is Cohen's separation, measured on the validation sets on Accuracy. Tests tagged prior showd not measured because they never were — they carry a literature prior with a deliberately wide spread and a low weight, and we will not print a separation number we have not earned.
The measurement that carries this tier is not a stylistic tell that a better model will fix. It is a consequence of how the two kinds of image are physically made.
A camera builds each colour channel from different sensor sites. The photons arriving at a red-filtered well are counted independently of those at its green neighbour, so shot noise is uncorrelated between channels. The demosaicer then interpolates each channel with a different kernel — green has twice as many samples as red or blue on a Bayer grid. And the lens focuses red and blue light at slightly different planes, so their fine detail is not even the same detail. Take the high-frequency residual of each channel and the three disagree: on our 29 camera originals, mean correlation 0.883.
A latent-diffusion decoder makes all three channels from one shared latent tensor. The micro-texture is essentially a single pattern painted three times. The residuals agree almost perfectly: on our 14 generated images, mean correlation 0.980. That is d = 1.35, and unlike every sensor-level test it survives downscaling, because it is a property of the colour relationship rather than of the pixel grid.
It has one fatal enemy, and we would rather name it than hide it: JPEG 4:2:0 chroma subsampling averages colour over 2×2 blocks, which forces the channels to share fine detail and fakes the synthetic signature on real photographs. So we parse the JPEG SOF sampling factors, and on a subsampled file we cut these tests to 12% of their weight and say so on the result card. See the numbers.
Separate spreads per class, but not the measured one. Generated images cluster far more tightly than camera originals (σ = 0.014 against 0.100), and camera values drift upward whenever a file is resized. Encoding the measured σ would assert that every generator, present and future, lands in exactly the same spot — on the validation set that produced five false accusations on native photographs and eighteen on recompressed ones. We use 0.9 × σcamera: the useful asymmetry without the unearned confidence.
Weights are d2.5, not d. With linear weights a cluster of weak, mutually-agreeing signals could outvote the one strong one. A product shot measuring 0.9965 cross-channel correlation — about as synthetic as a number gets — was averaged down to an inconclusive 57%. Raising the exponent lets decisive evidence actually decide.
A confidence multiplier on the whole tier. Every pixel measurement assumes it is seeing something close to what the sensor or the decoder produced. Resizing, recompression and subsampling each erode that assumption, and past a threshold the numbers stop meaning what the calibration says. So the tier is shrunk toward the prior in proportion to the damage — a mangled file returns "inconclusive" rather than a confident guess.
We originally wrote these baselines from the forensics literature. Measured against real camera originals and real generator output, two were directionally backwards. Deviation from Benford's law is widely cited as a synthesis indicator; JPEG-quantised camera output actually deviates more. Shot-noise theory predicts brightness and noise correlate positively; in real photographs the correlation is usually negative, because bright regions are skies and dark regions are foliage. A third, saturation dispersion, is real but points the opposite way to the folklore — generators produce more saturation variance than camera JPEGs, not less.
Every centre on this page is now measured rather than cited — with explicit exceptions. Tests tagged prior in the list above carry literature priors with deliberately wide spreads and low weights, marked prior wherever they appear, until the next measurement round. The scripts that produced the measured centres ship in research/.
The AI video detector reuses the image engine's calibrated pixel centres (cross-channel agreement d = 1.35) on sampled decoded frames, then adds temporal physics no still can have: frame-to-frame sensor-grain fluctuation (cameras wobble, re-encoded synthetic grain sits still), flicker coherence and optical-flow motion consistency. The AI audio detector decodes the waveform locally and measures spectral shape, bispectral vocoder artefacts (quadratic phase coupling from ElevenLabs, OpenAI TTS, Coqui, Bark-style synthesis) and temporal voice dynamics. Both score with Gaussian LLRs, reliability damping and a prior — and both print every value with the exact logit arithmetic in the report and in every export.
Honest limit, stated plainly: there is no large public corpus of verified camera versus generated video, nor of verified human versus cloned voice, so the temporal and acoustic centres are physics-based engineering priors with wide spreads — not corpus fits. That is why short, noisy or recompressed files damp toward the prior and return low reliability or inconclusive instead of a confident fake percentage.
Provenance follows the same honesty rule as images: only a strong decoded-field record (container metadata, never a file name) may claim origin, and when the frames or waveform clearly disagree the record stays visible but unconfirmed — reduced weight, no generator naming, a named disagreement finding, and reliability capped at low. Every per-signal LLR is clamped (±4), every logit is clamped (±6), and camera counter-evidence (a camera-maker tool string) votes toward camera so the headline reflects both sides.
Generated directly from the source the AI video detector runs on. Pixel rows reuse the image engine's measured centres; every row taggedprior is a physics-based engineering prior with a wide spread — never a corpus fit, never an invented separation figure.
| Test | Camera μ | AI μ | w |
|---|---|---|---|
Cross-channel correlation prior crossChannelCorr · Cross-channel correlation — demosaicer disagreement vs shared latent. AI spread is 0.9× the human spread (never the tight measured cluster), so a single frame value cannot spike the verdict. | 0.883 | 0.978 | 2.50 |
Spectral falloff exponent prior spectralSlope · Spectral falloff exponent — inherited from the image engine measurement (natural 1/f vs faster synthetic falloff) | 2.516 | 3.062 | 1.20 |
High-frequency energy share prior highFreqRatio · High-frequency energy share — same scale as the image engine (lens/sensor texture puts energy up here; synthesis puts less) | 0.7834 | 0.7436 | 1.00 |
Frame-to-frame noise fluctuation prior temporalNoise · Frame-to-frame noise fluctuation — sensor grain wobbles; re-encoded grain is static | 0.0042 | 0.0018 | 2.00 |
Stability of cross-channel correlation across framesprior temporalStability · Stability of cross-channel correlation across frames | 0.003 | 0.008 | 1.80 |
Frame-to-frame luminance correlation prior flickerCoherence · Frame-to-frame luminance correlation — synthetic video often flickers | 0.15 | 0.72 | 1.50 |
Frequency domain anomaly score prior frequencyAnomaly · Frequency domain anomaly score — spectral distortion detection | 0.25 | 0.65 | 1.80 |
Texture consistency prior textureConsistency · Texture consistency — LBP-based skin/face detail coherence | 0.85 | 0.55 | 1.50 |
Edge consistency prior edgeConsistency · Edge consistency — unnatural boundary detection | 0.8 | 0.5 | 1.50 |
Optical flow warping artifacts prior opticalFlowWarping · Optical flow warping artifacts — temporal motion inconsistency | 0.08 | 0.25 | 1.60 |
Optical flow coherence prior opticalFlowCoherence · Optical flow coherence — smooth motion field consistency | 0.85 | 0.55 | 1.40 |
Generated directly from the source the AI audio detector runs on. Every centre is an acoustic engineering prior (natural voice is spectrally richer, more variable and phase-irregular; vocoded speech is brighter, flatter and phase-locked) — not a corpus fit. Rows taggedprior carry wide spreads and clamped LLRs so no single measurement can decide a verdict alone.
| Test | Human μ | AI μ | w |
|---|---|---|---|
Spectral centroid prior spectralCentroid · Spectral centroid — brightness measure | 2200 | 2800 | 1.20 |
Spectral flatness prior spectralFlatness · Spectral flatness — noise-like vs tonal | 0.22 | 0.35 | 1.40 |
Spectral entropy prior spectralEntropy · Spectral entropy — distribution complexity | 7.8 | 6.5 | 1.30 |
Spectral flux prior spectralFlux · Spectral flux — frame-to-frame change rate | 0.18 | 0.12 | 1.10 |
Spectral rolloff prior spectralRolloff · Spectral rolloff — 85% energy cutoff frequency | 3800 | 4500 | 1.00 |
Bispectrum anomaly prior bispectrumScore · Bispectrum anomaly — quadratic phase coupling (vocoder artifact) | 0.35 | 0.72 | 2.20 |
Phase coherence prior phaseCoherence · Phase coherence — natural vs synthetic phase relationships | 0.52 | 0.78 | 1.80 |
Spectral peak regularity prior spectralPeakRegularity · Spectral peak regularity — vocoder creates uniform peaks | 0.55 | 0.25 | 2.00 |
Zero-crossing rate prior zeroCrossingRate · Zero-crossing rate — speech pattern characteristic | 0.12 | 0.085 | 1.30 |
ZCR variation prior zeroCrossingRateStd · ZCR variation — natural speech has more variation | 0.035 | 0.015 | 1.50 |
Energy variation prior energyStd · Energy variation — natural speech dynamics | 0.018 | 0.008 | 1.40 |
Silence ratio prior silenceRatio · Silence ratio — pause patterns | 0.25 | 0.18 | 1.10 |
Temporal anomaly prior temporalAnomaly · Temporal anomaly — sample distribution analysis | 0.22 | 0.45 | 1.60 |
Harmonic anomaly prior harmonicAnomaly · Harmonic anomaly — vocoder vs natural harmonics | 0.35 | 0.65 | 1.70 |
Measured but unfused diagnostics — MFCC voice-print shape and variation, pause-length structure, clipping flag, channels and bit depth — are computed from every file and shown in the report with their live values and an explicit “no vote” label, so future validation can be checked against them.
Every tunable number in the fusion step, so you can reproduce a score by hand.