Clarity
Try it free

How it works

Four checks.
One honest answer.

Every result comes from specialized detection models that we test on thousands of real examples they never saw in training. Here is exactly what happens to your content, and how often we get it right.

0.99image ranking accuracy (AUC) on 4,011 test images
94%of AI images caught, with 2% of real photos wrongly flagged
<1%of human writing wrongly flagged on 7 of 8 text benchmarks
0.4sto check an image on a standard server
01

Reading the score

Every check ends in a single likelihood from 0 to 100%. The scale is calibrated on benchmark data, so a higher score really does mean stronger evidence. We never force a yes or no when the evidence is mixed.

02

AI text detection

Language models leave statistical fingerprints: word choices that are a little too predictable, rhythm that is a little too even. Two independent detectors trained on millions of human and AI passages look for them, and we combine their opinions. In October 2026 we retrained both on writing from current models, including text deliberately written to imitate a real person's style.

  1. Your textPasted, or read from PDF, Word and 20+ formats
  2. CleanedHidden characters and look-alike letters used to fool detectors are removed
  3. Two detectorsBoth read each paragraph, trained on chatbots from 2023 through 2026
  4. Calibrated blendTheir scores are combined and mapped to a true likelihood
  5. VerdictOverall score plus paragraph highlights for mixed documents
Text scoring 60% or higher, by benchmarkBlue should be high, orange should be low. Each benchmark mixes human writing with text from many AI models. Ghostbuster's human side is polished essays, the kind most often mistaken for AI.

Where it struggles. Beemo contains human experts writing answers in the style of a chatbot, the hardest case for any detector, and 12% of those scored 60% or higher. Our October 2026 retraining also trades some recall on text from older AI models (2019 to 2023) for fewer false accusations: MAGE and COLING, which are mostly older models, are caught less often at our 60% threshold than before. That's why text in the 60 to 85% range is labeled "Possibly AI", and why a score should never be the only evidence against a person.

03

AI image detection

Image generators leave traces in the pixels that are invisible to the eye: in noise patterns, textures and the way detail is rendered. Our models are trained on images from thousands of generators, including 2025 and 2026 models like GPT Image, FLUX, Imagen and Seedream.

  1. Your imageJPG, PNG, WebP, HEIC, AVIF and more
  2. File evidenceWe read Content Credentials (C2PA), IPTC AI labels and generator settings hidden in the file
  3. Two forensic modelsVision transformers examine the image and its mirror
  4. Calibrated blendCombined into one likelihood; declared AI files are marked as such
  5. VerdictScore, file evidence and each model's reading
AI images caught, by generatorShare scoring 50% or higher on our held-out test set. Some test sets come from the same families as our training data; on an image collection we never trained on, ranking accuracy (AUC) was 0.94.
Real photos wrongly flaggedLower is better.
Compared with popular open-source detectorsRanking accuracy (AUC) on the same 1,051 test images. 1.0 is perfect; 0.5 is a coin flip.
04

Plagiarism check

Copied text shares long, identical runs of words with its source. We find candidate sources the way a careful editor would, then compare every word.

  1. Pick passagesDistinctive sentences from across your text
  2. SearchWikipedia, and the wider web
  3. Fetch sourcesThe matching pages are downloaded and cleaned
  4. FingerprintEvery six-word sequence is compared, ignoring filler phrases
  5. ReportSimilarity score, highlighted matches, links to sources
6words in a row must match before we count it, so common phrases don't trigger false matches
81%match found in our test when we pasted a Wikipedia article with a paragraph of our own
0%match for the original paragraph on its own
05

Hallucination check

Chatbots sometimes state things with confidence that simply aren't true, or that aren't in the document they were asked to summarize. We split the text into claims and test each one against evidence with a model trained specifically for fact verification.

  1. Find claimsSentences with names, numbers and dates
  2. Gather evidenceFrom your source document, or Wikipedia and the web
  3. Pick the best evidenceThe most relevant sentences for each claim
  4. VerifyA fact-verification model judges support or contradiction
  5. ReportEach claim marked, with the evidence quoted
A real test runWe asked for a check of a short summary against the company report it was based on. We had slipped in two details the report never said. Both were caught.
Claim in the summaryVerdictEvidence from the source
Acme's 2025 revenue rose 8% to $4.2 billion.Supported"Acme Corp reported revenue of $4.2 billion for 2025, up 8% from the prior year."
It opened 40 stores, mainly in Texas and Ohio.Supported"The company opened 40 new stores, mostly in Texas and Ohio."
The company announced a special dividend.Contradicted"CEO Dana Reyes said the company will not pay a dividend this year."
Profits doubled.Not in sourceNo mention of profits in the report.
Acme plans to hire about 1,200 people.Supported"...and plans to hire 1,200 people."

Checking against Wikipedia is more cautious: on 180 held-out FEVER claims, fewer than 2% of verdicts were confidently wrong, and claims without clear evidence are reported as "couldn't verify" rather than guessed.

06

Tested on someone else's rules

Our own benchmarks could flatter us, so we also run tests built by others. In September 2026, TextSight, another AI detection company, published a benchmark designed to "not flatter detectors": human writing provably from before ChatGPT (Wikipedia revisions dated before 2022 and research papers by non-native English authors) against text from five current AI models writing in the same styles. We never trained or tuned on its texts.

0.999ranking accuracy (AUC). The strongest open detector TextSight tested scored 0.87.
100%of AI text caught while wrongly flagging 1 in 100 human texts. TextSight's best open detector: 50%.
100%of Claude Opus 5 text caught at that same strictness, where TextSight reported 0%.
AI text caught at a 1% false-positive rate, by modelClarity, measured with TextSight's published data and protocol, next to the strongest open detector in TextSight's own published results.

Source: TextSight AI text detection benchmark 2026 (700 passages; Wikipedia and PubMed Central human sets, 56 AI passages; the per-model samples are small, 8 to 12 passages each). Competitor figures are as published by TextSight for Oxidane/tmr-ai-text-detector. Human texts wrongly flagged at our 60% threshold: 0% of Wikipedia, 1.1% of non-native research papers. Our training data includes other text from some of the same model families, but none of TextSight's passages.

AI that imitates a real writer

The hardest test we know of comes from Epoch AI, an independent research group. In 2026 they had three frontier models (Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro) write blog posts, fiction and scientific prose, once with a plain request and once told to imitate a specific human author's style, and scored the results with Pangram, GPTZero and Originality.ai. We ran Clarity on the same 1,089 texts.

Epoch AI 2026 benchmarkShare of each group scoring as AI. Clarity at its standard 60% "Possibly AI" threshold; the other detectors at their own verdicts, as published by Epoch AI.
92%of style-imitating AI caught, ahead of Pangram (90%) and GPTZero (89%). Before our October 2026 retraining: 6%.
0%of human writing wrongly flagged (0 of 495), the same as Pangram and GPTZero. Plain AI text: 99.7% caught.
100%of "humanized" AI text caught in a separate 2025 study, where Pangram caught 90 to 93% and GPTZero 47%.

Where we still fall short. Scientific writing that imitates a specific author is still the case we miss most: about 1 in 5 of those texts score below 60%, and Gemini 3.1 Pro is the hardest model to catch in that mode. Zero false flags on 495 texts is a strong result, not a guarantee: on a separate benchmark of 150 human-written articles we flagged one. We report our weak spots because a detector that hides them can't be trusted on its strong ones.

Sources: Epoch AI data insight on AI text detectors (495 human and 594 AI texts; Pangram, GPTZero and Originality.ai results as published by Epoch AI, not tested by us) and Russell et al., ACL 2025 (150 human and 150 AI articles, including paraphrased and "humanized" text; at our 60% threshold Clarity caught 144 of 150 AI articles and flagged 1 of 150 human ones; Pangram and GPTZero results as published by the authors). Product names are trademarks of their owners. We never trained on either set. Our training data does include other text from the same model families, including GPT-5.5, written for different prompts.

07

What we promise, and what we don't

  • Your content is never stored. It's analyzed in memory and discarded the moment your result is ready. It is never used to train anything. Plagiarism and fact checks send short excerpts as search queries to find sources.
  • We publish our misses. Every number on this page includes the cases we get wrong.
  • No detector is proof. Edited, translated or very short content is harder to judge. Use a result as one piece of evidence, never the only one.
  • New models are a moving target. New AI generators appear constantly, and accuracy on a brand-new one can be lower until detection models catch up.

See it on your own content.

AI text checks are free, with no account needed. Image, plagiarism and fact checks are coming soon.

Check text free