AI detector FAQ: 50 questions, straight answers.

An AI detector detects the statistical fingerprints large language models leave behind: unusually predictable word choices, smooth uniform sentence lengths, recycled trigrams, stock transitions like “delve into” or “tapestry”, and flat punctuation rhythm. Text detectors blend these signals into an AI-likelihood score; image detectors check file provenance (C2PA, EXIF, generator tags) plus pixel physics; video detectors add frame-to-frame temporal grain; audio detectors measure spectral, bispectral and temporal voice physics.

This free AI detector measures 46 text signals and runs two-tier image forensics plus multi-domain video and audio forensics, then shows you every measurement — paste text or drop a photo, video or audio file above to see exactly what it detects in your own file.

“AI detect” means analyzing a piece of content to estimate how likely it was produced by an AI model such as ChatGPT, Claude, Gemini or Stable Diffusion. The result is always a probability, never proof: a high score means the writing strongly resembles machine output, while a low score means it looks human.

Because scores are probabilistic, always read the confidence interval and the underlying evidence — which is precisely what this free AI checker displays with every scan.

Paste at least 300 words of text (80+ at minimum), or upload a JPEG, PNG, WebP, AVIF or HEIC image, an MP4, WebM, MOV or MKV video, or a WAV, MP3, FLAC or OGG audio file, press Analyse, and read three things: the AI-likelihood percentage, its 90% confidence interval, and the sentence-by-sentence evidence map. Export the free JSON, Markdown or PDF report if you need a record.

Run your check here in seconds — no signup, no upload, no character limit — or practice first on the curated samples page.

An AI checker tokenizes your text, measures stylometric statistics (perplexity proxies, burstiness, lexical diversity, Zipf fit, n-gram repetition, punctuation rhythm, AI-phrase density), compares each against human and AI calibration baselines, and fuses them into one score. For images it inspects metadata first, then pixel-level capture physics. For video it samples frames for sensor physics plus temporal grain, flicker and motion coherence. For audio it measures spectral shape, vocoder bispectral artefacts and temporal voice dynamics.

Most checkers hide this math inside a black box. Ours prints all 46 signals with their fitted coefficients and contributions — read the full method on the how-it-works page.

To check whether a text was written by AI, paste it into a detector and look beyond the headline number: a high score with a narrow interval plus hard evidence (chat scaffolding phrases, zero-width characters, pasted markdown) is a strong signal, while a high score with a wide interval on short text means very little.

For important cases, cross-check with a second tool and compare where they agree — then run the free check here and inspect the sentence map yourself.

It depends on your job: universities overwhelmingly standardize on Turnitin, publishers and agencies favor Originality.ai, and individual teachers often use GPTZero. Independent 2026 tests put real-world accuracy for leading tools between roughly 60% and 95% depending on text type — far below the “99%” banners.

If you want a free AI detector that shows its entire working instead of asking for blind trust, this is it: 46 published signals, per-genre false-positive rates, confidence intervals, and unlimited free checks, all on-device.

The best AI detector is the one matching your stakes: Turnitin for institutions grading inside an LMS, Originality.ai or Pangram for publishers screening at scale, and a transparent free tool for everyday checks. No detector is reliable enough to be anyone’s sole evidence — treat every score as one signal among many.

Start with the free check on this page: you will see more diagnostic detail than most paid detectors show.

There is no undisputed winner. Vendor lab claims cluster near 99%, but independent 2026 testing lands very differently: Pangram led a 160-paper academic study, Originality.ai and GPTZero score in the 85–95% range on clean AI text, Turnitin reports under 1% document false positives, while free tools like ZeroGPT measure closer to 67–85%.

Accuracy also collapses on paraphrased, short, or non-native text for every tool. Demand published methodology and per-genre error rates — like the ones on our accuracy page — before trusting any “most accurate” crown.

Reliability means low false accusations plus transparent limits, not the biggest marketing number. Turnitin’s sub-1% document false-positive rate makes it the reliability benchmark in education, while tools that publish confidence intervals and say “inconclusive” — instead of guessing — are far safer for real decisions.

This detector is built for exactly that: a human-leaning prior, published miss rates by genre, and honest inconclusive verdicts. Try it free and compare its caution against any black-box score.

The best AI checker combines strong detection with evidence you can inspect: sentence-level highlights, named AI phrases, and exportable reports. Turnitin and GPTZero lead inside institutions; for everyone else, a free checker that includes all of that without a paywall wins on value.

That is what this page is: a free AI checker with sentence heatmaps, hard-evidence detection, and free JSON, Markdown and PDF export. Run your text and judge for yourself.

For teachers and editors, the best AI checker is one that explains itself — sentence highlights plus the reasons behind them — and respects privacy by analyzing on-device. Server-side checkers that silently upload student essays or client manuscripts fail both tests.

Paste a sample into the free checker above and see the difference evidence makes.

Independent tests, not marketing pages, decide this — and they disagree by design, because every study uses different texts. Recent results cluster Originality.ai, Pangram, GPTZero and Turnitin at the top on clean AI prose (mid-80s to high-90s percent), with every tool degrading sharply on edited or paraphrased writing.

Rather than chasing one percentage, use a checker that quantifies its own uncertainty. Ours reports a 90% interval and reliability grade with every score — free, right here.

Somewhat — and only in the right conditions. AI detectors are good at flagging raw, unedited model output (typically 80–95% for leading tools) but unreliable on short passages under ~80 words, heavily edited or paraphrased text, and non-native English writing, where false positives cluster.

Use detectors as screening tools, never as proof, and prefer ones that publish their miss rates the way this site does on its accuracy page.

As accurate as the detector behind it — same technology, same limits. Expect solid screening on long, unedited AI text and skepticism-worthy output on short, mixed, or polished drafts. Any checker promising 100% certainty is selling something the science does not support.

Verify any important result with a second checker and a human conversation — starting with a free, fully transparent check here.

Turnitin uses its own in-house AI writing detector — not a third-party tool. Its current system (models AIW-2 and AIR-1) is a transformer classifier that slides few-hundred-word windows across a submission, scores each sentence, and only displays a percentage when over 20% of sentences look AI-generated. It needs at least 300 words and is tuned primarily for English academic prose.

Because it lives inside institutional Turnitin licenses, students and freelancers cannot run it themselves — which is exactly why a free pre-check here, with sentence-level detail, is the smart move before you submit.

It is among the most careful detectors available: Turnitin reports under 1% document-level false positives (0.5% in a 700,000-paper stress test) with roughly 84% document recall — and independent studies have repeatedly placed it at or near the top. But “careful” is not “infallible”: paraphrased AI text, short documents, and non-native writing still defeat or confuse it, which is why some universities (Vanderbilt famously) disabled it.

Treat a Turnitin flag as the start of a conversation, never a verdict — and pre-screen your own drafts with the free checker on this page first.

Turnitin’s published numbers: 0.5% document false-positive rate, ~0.33% sentence false-positive rate, 84.2% document recall and 92.3% sentence recall on mixed test sets. Independent 2026 research found it far weaker on paraphrased text (down to ~46% on some paraphrased sets) and on deliberately humanized academic papers.

Bottom line: excellent at avoiding false accusations on standard work, beatable by editing — so no Turnitin score alone should decide an academic case.

Turnitin does not license an outside AI checker — the AI writing report inside Feedback Studio and sim-check workflows is Turnitin’s proprietary detection stack, bundled with its similarity (plagiarism) report. Schools access it through their institutional license, not as a standalone consumer product.

If your school runs Turnitin, assume every submission gets both a similarity and an AI score — and self-check your drafts here first, free and privately.

Yes, within its design envelope: long English academic submissions, where it keeps false accusations under 1% at the document level. Outside that envelope — short text, heavy paraphrasing, AI-human hybrids, non-English prose — its accuracy falls like every other checker’s.

The practical takeaway: draft in your own voice, keep your revision history, and verify with a free transparent check before submitting.

Nothing outside a Turnitin license truly replicates it — its model, thresholds, and 20% display rule are proprietary and LMS-embedded. The closest experience comes from institutional tools with sentence-level highlighting and low false-positive tuning, such as GPTZero’s education plans; for publishers, Pangram led the head-to-head academic study that included Turnitin.

For a free approximation with the same spirit — sentence maps, evidence lists, cautious thresholds — run your text through this detector and read the interval, not just the number.

No consumer checker is truly similar under the hood, since Turnitin’s stack is closed and trained on decades of student writing. Tools most often compared to it are GPTZero (education focus, LMS integrations), Copyleaks (institutional bundling), and Pangram (strongest recent academic benchmark). Each differs in thresholds and behavior, so scores will not match Turnitin’s.

If you are about to be graded by Turnitin, the winning strategy is not score-matching — it is original drafting plus a free pre-check here to catch accidental AI-like patterns.

Most colleges use whatever their LMS license bundles — overwhelmingly Turnitin, whose AI writing report appears automatically alongside similarity reports in systems like Canvas and Blackboard. Individual instructors without institutional access most often reach for GPTZero’s free tier, while some departments add Copyleaks.

Because you cannot see your Turnitin score before submitting, smart students pre-check with a free detector here — privately, with zero upload — and fix flagged phrasing before the grade depends on it.

Colleges standardize on Turnitin’s AI checker inside their learning platform, occasionally supplemented by Copyleaks or GPTZero institutional plans with admin dashboards. Free consumer checkers are rarely “official” but are widely used by students for self-screening.

Self-screen here first: it is free, unlimited, and — unlike server-side checkers — never stores or trains on your essay.

Teachers use Turnitin Feedback Studio when their school provides it; otherwise GPTZero is the most common classroom choice thanks to its free tier, sentence highlighting, and Canvas integration. Privacy-conscious teachers increasingly prefer on-device detectors that never upload student work to third-party servers.

This detector was built for that classroom: paste an essay, get sentence-level evidence plus a reliability grade, export a free report — with student data never leaving the browser.

The same shortlist: Turnitin inside the LMS, GPTZero for quick individual checks, and Copyleaks where districts bundle it. The responsible workflow is identical everywhere — treat the score as a prompt to talk to the student, compare against their known writing, and never penalize on a number alone.

Try that workflow here free: the ethics checklist on our accuracy page walks through every step.

Canvas has no built-in AI detector — it provides the plumbing (Document Processor / LTI integrations) and each school plugs in its own tool. Turnitin was the first partner to deliver AI reports into SpeedGrader, GPTZero offers a Canvas integration with sentence highlights, and Copyleaks plugs in the same way.

So “the Canvas AI detector” is really your school’s choice of Turnitin, GPTZero, or Copyleaks — ask your admin which one grades you, then pre-check drafts with this free detector.

Canvas itself checks nothing — it forwards submissions to the school’s chosen checker (usually Turnitin’s AI writing report, optionally GPTZero or Copyleaks) and displays the returned report in SpeedGrader. Availability therefore varies by institution, not by Canvas.

Whatever your school plugged in, a free private pre-check here catches risky passages before they reach SpeedGrader.

If you are asking “will this get flagged as AI?” — paste the text into the free checker above and read the full result, not just the percentage. A low score with a narrow interval means your writing looks human; a high score on a short draft may just mean there is too little text to judge.

Students: self-check before submitting, keep your drafts as proof of process, and rewrite flagged passages in your own voice rather than gaming the score.

Mid-tier. QuillBot cites ~99% on the RAID academic benchmark (at a 5%-false-positive operating point), but independent tests of 3–160 texts land between roughly 76% and 80%, with one 160-sample study finding 0% false positives but a 38% miss rate on AI text. The free tier caps at ~1,200 words and 6 scans daily.

Useful as a free second opinion — QuillBot itself warns never to rely on it alone for academic or career decisions. Cross-check important texts here, where every signal is shown.

Overstated in marketing, weak in independent testing. Grammarly claims 99% and a #1 RAID ranking, yet a 2026 peer-reviewed study measured just ~71% baseline accuracy collapsing to ~40% on paraphrased text — and it is deliberately tuned to miss AI rather than risk false accusations. One 160-sample test found it let half of all AI samples pass as human.

Fine as a writing assistant, unreliable as a verdict. Verify anything important with a dedicated, transparent check — like the free one on this page.

By the numbers: ~99% claimed (RAID benchmark, 2023-era models, no Claude/Gemini in corpus) versus ~58–71% in independent 2026 tests, with paraphrased-text detection falling as low as ~19% and one study logging a 34% false-positive rate. Grammarly’s own docs admit no detector is conclusive.

Treat its score as a gentle hint, then confirm with a second detector that publishes its methodology — start free here.

Not accurate enough to trust alone: independent tests consistently rank it below dedicated detectors like GPTZero, Originality.ai, and Turnitin, especially on edited or non-GPT text. Its strength is low-friction self-checking inside an editor you already use.

For decisions that matter, pair it with a transparent free check here and compare the evidence, not just the percentages.

Moderately: roughly 76–80% overall in independent tests, very safe on human text (near-zero false positives in several studies) but leaky on AI text (missing over a third in one 160-sample trial). Good for casual screening within its free limits.

For high-stakes checks, supplement it with a detector that shows its full working — run the same text here free and compare signal by signal.

Contested. JustDone advertises ~80–94% accuracy with low error rates, but a 2026 independent retest measured only ~59–61% overall (70–79% on long documents, coin-toss territory under 100 words), and reviewers found the free tier buggy while the paid tier behaved far better. Its plagiarism checker consistently tests stronger than its AI detector — and note it also sells a “humanizer” to beat detectors.

Usable as one weak signal on long documents; never as a sole judge. Confirm anything serious with a free second opinion here.

Only for casual, low-stakes checks. ZeroGPT is free and fast with a generous 15,000-character limit, but independent 2026 testing puts real-world accuracy at roughly 67–85% with false-positive rates of 15–26% — including infamous flags on classic literature and pre-AI academic papers. It publishes no methodology.

Fine for curiosity; dangerous for accusations. For anything consequential, use a detector that publishes its math — like this free one — and cross-check.

Strong vendor claims, shaky independent record. Copyleaks advertises 99%+ accuracy with a 0.03% false-positive rate, yet peer-reviewed tests found 74% dropping to 59% under simple evasion with the highest false-accusation rate of its group — and a 2026 university study saw it fully flag zero of 40 AI-generated papers. Its real strengths are enterprise APIs and LMS bundling, not infallibility.

If Copyleaks screens your work, pre-check it here free and arrive with drafts and evidence, not just hope.

Yes — this page includes a full AI checker for images and photos. It reads provenance first (C2PA credentials, EXIF, PNG text chunks, generator tags for Midjourney, DALL·E, Stable Diffusion and more), then pixel physics. On original files it cleared 29 of 29 photographs and caught 14 of 14 AI images in our validation.

Drop any photo into the Image tab above for a free, on-device verdict — and remember recompressed social-media screenshots honestly deserve an “inconclusive”.

Yes — the Video tab is a production-grade AI video detector, not a demo. It parses the MP4, WebM, MOV or MKV container for Sora, Veo, Runway, Pika, Kling and C2PA provenance, then decodes sampled frames locally and measures per-frame sensor physics (cross-channel residual agreement, spectral shape) plus temporal physics (frame-to-frame grain fluctuation, flicker coherence, optical-flow motion consistency).

Every frame measurement, LLR, weight and contribution is printed in the report with the exact score arithmetic, and the full JSON, Markdown, CSV, HTML and PDF export matches the on-screen numbers. Heavily recompressed or very short clips honestly return low reliability or inconclusive instead of a confident fake number.

Yes — the Audio tab is a production-grade AI audio detector for speech and voice. It parses WAV, MP3, FLAC, OGG and M4A containers locally, decodes the waveform in your browser with no upload, then runs multi-domain forensics: spectral shape (centroid, rolloff, flux, flatness, entropy), bispectral vocoder-artefact tests (quadratic phase coupling, phase coherence, peak regularity) and temporal voice dynamics (zero-crossing rate, energy variation, pause patterns).

The report shows every measured value, both calibration centres, the LLR, the weight and the signed contribution, plus the exact logit arithmetic that produces the percentage. Short clips under a second or heavily compressed files return low reliability by design — a screening signal, never proof.

Honestly: weaker than text detection, and any tool claiming otherwise is overselling. There is no large public corpus of verified camera-original versus AI-generated video, so the temporal baselines are physics-based engineering priors, not corpus fits — the per-frame pixel tests reuse the image engine centres measured on 29 camera originals versus 14 generated images (d = 1.35 on the cross-channel test).

That is why this video detector publishes every number, damps the score toward the prior when frames are few or compressed, and says inconclusive when the evidence is gone. Use a high score as a reason to inspect provenance and frames, never as proof of generation.

Useful as screening, not proof. There is no large public corpus of verified human versus AI-cloned voice with ground truth, so the spectral, bispectral and temporal baselines are acoustic engineering priors (natural voice has richer spectral entropy, more variable zero-crossing rate and irregular phase; vocoded speech is brighter, flatter and phase-locked), not fitted corpus statistics.

This detector therefore shows every measurement with its centre, spread, LLR and contribution, requires at least a second of voiced audio for confidence, and widens the interval on short or noisy clips. Treat a high score as a reason to listen and verify provenance, never as speaker identification.

Yes — Sora, Veo, Runway, Pika, Kling, Hailuo, Luma and C2PA declarations are all parsed from the MP4, WebM, MOV or MKV container first, then corroborated against sampled frames. Only a strong decoded-field record may claim origin: a file merely named “sora.mp4” is recorded as a weak filename hint and can never trigger a conclusive verdict, and when the frames clearly disagree the record stays visible but unconfirmed with dampened scoring.

Because feeds strip metadata, always test the original export: drop it into the Video tab above and read the provenance plus frame-temporal tables with their exact arithmetic.

MP4, WebM, MOV and MKV up to 512 MB, decoded entirely in your browser tab — nothing is uploaded. Seven frames are sampled across the duration and measured for per-frame sensor physics plus temporal grain fluctuation, flicker coherence and optical-flow motion consistency, with per-frame observations printed alongside the fused score.

If decoding fails on an exotic codec, the report says so and shows container measurements only instead of inventing a percentage. Convert to MP4 (H.264) and retry for the full analysis.

It screens for them: ElevenLabs, OpenAI TTS, Coqui, Bark, Tortoise, Speechify and similar synthesis leave two traces this detector reads — a metadata/software string when the exporter writes one, and a bispectral vocoder fingerprint (quadratic phase coupling, locked phase coherence, over-regular spectral peaks) that natural voice does not share. A metadata claim the waveform contradicts stays unconfirmed with dampened scoring rather than deciding the verdict.

Treat a high score as a reason to verify provenance and listen — never as speaker identification or proof. Run a clear multi-second voice clip through the Audio tab and inspect the acoustic table.

WAV, MP3, FLAC, OGG and M4A, decoded locally with no upload: WAV goes through a strict local parser with exact container facts, while compressed formats decode via Web Audio in-tab (capped at the first 90 seconds to stay responsive). Every clip needs at least ~0.3s of decodable audio, and confidence requires seconds of actual voiced speech.

Pure tones, silence, music beds and sub-second clips honestly return low reliability — the tests judge voice, and with no voice to judge the score damps toward the prior instead of faking certainty.

Because short inputs carry almost no statistical evidence: temporal grain needs multiple frames across seconds, and voice forensics needs seconds of voiced speech. Rather than guessing, the engines shrink short, noisy or recompressed inputs toward the prior, widen the interval and drop the reliability grade to low.

That “inconclusive” is the correct answer, not a failure — re-test with a longer original export (several seconds, direct from camera or DAW rather than a recompressed feed) for a decisive reading.

No — and distrust any tool that claims otherwise. Text detection here is cross-validated on 2,743 real documents (ROC AUC 0.966); video temporal centres and audio acoustic centres are physics-based engineering priors with wide spreads, because no large public corpus of verified camera-versus-generated video or human-versus-cloned voice exists to fit against.

What makes them production-grade is not a marketing percentage but the discipline around the number: clamped LLRs, reliability damping, disagreement findings, every value printed, and inconclusive verdicts where the evidence is gone. Use high scores as screening evidence, never as proof.

Honestly: there is no reliable trick. Paraphrasers and “humanizers” lower scores on some detectors but leave statistical traces, fail on others, and improve nothing about the underlying dishonesty — detectors keep adapting, and several now flag paraphrased text specifically. The only method that always passes is writing the piece yourself, in your own voice, from your own research.

Use this free detector the right way instead: check your own draft, find passages that read as generic or machine-like, and rewrite them with concrete detail and personal reasoning. That improves your writing and your score together.

Run a documented, humane workflow: collect original files, screen with a sentence-level detector, and review the evidence and confidence interval — never the headline score alone. Compare against the employee’s known writing, check draft history and sources, then discuss findings privately before concluding anything. Get policy and consent right first, especially for confidential material.

This on-device detector fits audits perfectly: confidential reports never upload anywhere, and free PDF/JSON exports create the paper trail. Start with a trial screen above.

An “AI” label from any checker means “statistically resembles machine writing,” not “proven machine-written.” High scores with narrow intervals and hard evidence (chat scaffolding, pasted formatting, machine-like uniformity across dozens of signals) are strong; bare percentages on short or formal text are weak.

When a label surprises you, run the text here free and inspect which of the 46 signals actually fired — the evidence, not the badge, is the answer.