AI image, voice, video and code detectors ranked by independent studies
AI or Not and Hive are the two AI detectors in the index with a published study on a content type other than text, both scored on images in Ha and Passananti 2024: AI or Not, tested under its 2023 name Optic, at 90.67% accuracy, 24.47% false positives and 1.15% false negatives, and Hive at 98.03% accuracy, 0% false positives and 3.17% false negatives, on the same 280 human artworks and 350 AI-generated images (arXiv 2402.03214, read 9 November 2026). AI or Not ranks first on the tie rule alone, one record each and the earlier name; Hive was the only qualifying card until the AI or Not card was added on 9 November 2026. Twenty of the 72 cards state that they detect image, voice, video, code or music, and eighteen of those twenty have no study record on any of those types, by the count compiled from the index data file on 9 November 2026. On 10 November 2026 the tenth wave added four non-text cards at zero records, Deepfake-O-Meter (image, video, voice), ElevenLabs Classifier (voice, the vendor's own models), FakeCatcher (video) and Ircam Amplify (music), so 28 of 63 cards stated a non-text type at that version and 24 of them had no record on it. Later the same day the eleventh wave added two voice cards at zero records, Hiya Deepfake Detector and McAfee Deepfake Detector, and three records that score no indexed card, ASVspoof 2021 and ADD 2023 on voice and FaceForensics++ 2019 on video, so 30 of 72 cards state a non-text type and 26 of them have no record on it. The 35 study records in the index score text in sixteen cases and images in one; no indexed study scores voice, video, code or music.
Updated Studies dated 2023 to 2026
Short answer
Two records, four tools, one content type with an indexed card: AI or Not (as Optic), Hive and Illuminarty on images in Ha and Passananti 2024, and Hive and Sightengine on images in RAID image 2025, added 9 November 2026, so Hive leads with 2 records. The other fourteen cards that state image, voice, video, code or music detection rank at zero and are ordered by their text study records, so GPTZero (11 records, all on text) sits third and eight cards with no record at all, SynthID now among them, close the table alphabetically. The sort measures whether an outside paper has scored the non-text claim, not how well the tool scored; Hive's own record shows detection of AI images falling from 96.83% to 67.56% under high-intensity Glaze (arXiv 2402.03214, read 9 November 2026). Sources read 8 and 9 November 2026. No detector score should be the only evidence in a misconduct or hiring decision.
Which AI image, voice, video and code detectors have an independent study?
| Rank | Detector | Records on image, voice, video, code or music | Content types stated beyond text | Record on those types | Best for | Main limitation |
|---|---|---|---|---|---|---|
| 1 |  Hive | 2 | Image, video, voice (audio), music | Ha and Passananti 2024, images; RAID image 2025, images under adversarial perturbation | Image checks with an outside result | 67.56% of AI images detected under high Glaze, 3.23% false positives, Ha and Passananti 2024; 0.60 (accuracy) on 50 clean AI images and 0.44 after RAID's attack, Eddoubi et al. 2025 |
| 2 |  AI or Not | 1 | Image, video, voice (audio), music | Ha and Passananti 2024, images, as Optic | Image, voice and video checks with an outside image result | 24.47% of 280 human artworks flagged; AI image detection fell to 58.00% under high Glaze, Ha and Passananti 2024 |
| 3 |  Illuminarty | 1 | Image; text as a proof of concept | Ha and Passananti 2024, images | Image checks with an outside result and a stated false positive rate | 67.40% of 280 human artworks flagged as AI, 4.69% of 350 AI images missed, Ha and Passananti 2024 |
| 4 |  Sightengine | 1 | Image, video, deepfake, voice (audio), music | RAID image 2025, images under adversarial perturbation | API customers checking images and video at volume | 0.66 (accuracy) on 50 clean AI images and 0.00 after RAID's transferable adversarial perturbation, labelled preliminary, Eddoubi et al. 2025 Table D.13 |
| 5 |  GPTZero | 0 | Image (AI Vision, specs not published) | Text checks with 16 text records | No published specification or study for the image detector | |
| 6 |  Copyleaks | 0 | Code, image, video | Code plus text in one LMS tool | 9 records, all on text; no code or image result published | |
| 7 |  Originality.ai | 0 | Code | Publishers who also check code | 7 records, all on text | |
| 8 |  ZeroGPT | 0 | Image, video | Free monthly text checks | 7 records, all on text | |
| 9 |  Pangram | 0 | Image, video | Text checks with 2025 and 2026 records | 4 records, all on text | |
| 10 |  QuillBot | 0 | Image | Students inside the QuillBot suite | 2 records, both on text, scores not rates | |
| 11 |  Winston AI | 0 | Image, handwriting (OCR) | Text and image in one account | 2 records, both on text | |
| 12 |  Codeleaks | 0 | Code | Source code checks on the Copyleaks platform | No study record; the AI code verdict was removed from the API on 16 December 2025 per the release notes | |
| 13 |  Codequiry | 0 | Code (no text) | Source code only | No study record; vendor states 1.3% false flags with no sample size | |
| 14 |  Deepfake-O-Meter | 0 | image, video, voice (audio) | None: the platform hosts 41 research detectors and no indexed record scored the platform | Researchers who want several published methods run on one file | Each detector's score is its own; the paper reports usage, not accuracy |
| 15 |  Deepware | 0 | Video (face manipulation only) | Free video deepfake checks by link or upload | No study record; the scanner host was unreachable on 9 November 2026 | |
| 16 |  ElevenLabs Classifier | 0 | voice (audio): ElevenLabs-generated speech only | None; SAFE 2025 names ElevenLabs only as a generator of test clips | Readers checking whether a clip came from ElevenLabs | A self-detector for one vendor's models; 99% precision and 80% recall are the vendor's own figures |
| 17 |  FakeCatcher | 0 | video (portrait videos, through blood-flow signals) | None: the 2019 paper's datasets are the authors' own evaluation | Readers checking what Intel announced in 2022 | No product, price, sign-up or API exists; the 96% is the newsroom figure |
| 18 |  Hiya Deepfake Detector | 0 | voice (audio): the voice in a video or audio played in a Chrome tab | None: no indexed record scored the extension or Hiya's platform; ASVspoof 2021 and ADD 2023 scored anonymised teams | Readers who want a free one-second voice check in the browser | The 99% is stated for the platform, not the extension; no study record |
| 19 |  Ircam Amplify | 0 | music (AI-generated tracks) | None: the 99.7% and 0.14% figures are on the vendor's internal benchmark datasets | Labels and distributors checking catalogues for AI-generated tracks | No indexed study tested AIMD; access is by application |
| 20 |  McAfee Deepfake Detector | 0 | voice (audio): AI-generated audio in a video playing in the browser | None: no indexed record scored the detector; FaceForensics++ 2019 scored retrained research methods on face manipulations | Readers who want an on-device alert on a Windows AI PC | The 96% is a 2024 press release figure with no dataset; English audio only; no study record |
| 21 |  Phrasly | 0 | Image | Writers checking text and images | No study record | |
| 22 |  Pindrop Pulse | 0 | Voice (audio), video (Pulse for Meetings) | Live call and meeting deepfake checks in a contact centre | No study record; the vendor's 99% figure is computed on ASVspoof5 by the vendor; no public price | |
| 23 |  Proofademic | 0 | Image (API), voice (API) | API users wanting image and voice | No study record; image and voice through the API only | |
| 24 |  Reality Defender | 0 | Video, image, voice (audio) | Enterprise deepfake screening | No study record; no numeric accuracy published | |
| 25 |  Resemble Detect | 0 | Voice (audio), image, video, music | Audio, image and video checks through one API with per-second pricing | No study record; the 99.5% figure is the vendor's on the Podonos Audio DFD Bench | |
| 26 |  Sensity | 0 | Video, image, voice (audio), file forensics | Forensic teams with a sales contract | No study record; no prices or trial terms published | |
| 27 |  SynthID | 0 | Image, video, audio (watermarked media only) | Checking whether a file carries a Google watermark | No study record; detects only SynthID-watermarked media; the Nature paper is vendor-authored | |
| 28 |  TruthScan | 0 | Image, PDF documents, video (deepfake), voice (audio) | Image and PDF checks with a results-based free plan | No study record; the vendor's image accuracy figures are its own | |
| 29 |  Walter Writes | 0 | Image (API), voice (API) | API users wanting image and voice | No study record; image and voice through the API only | |
| 30 |  WasItAI | 0 | Image (no text) | Free image checks, 10 per month | No study record; the detector is labelled BETA |
Records by content type
| Content type | Cards stating it | Cards with a study record on it | Record |
|---|---|---|---|
| Text | 58 | 34 | 25 records at the count of 9 November 2026, 29 cards with a record on 10 November 2026 and 34 after the eleventh wave the same day (Rashidi 2023 added as a 26th text record), listed on the ranking by independent testing |
| Image | 20 | 4 (AI or Not, Hive, Illuminarty, Sightengine) | Ha and Passananti 2024, 280 human artworks and 350 AI images; RAID image 2025, 50 clean and 50 adversarial AI images for Hive and Sightengine |
| Video | 15 | 0 | DFDC 2020, 2,114 Kaggle teams on 10,000 black-box videos, and FaceForensics++ 2019, six retrained research methods on 1.8 million manipulated frames; no indexed card scored in either |
| Voice | 22 | 0 | ASVspoof 5, 53 anonymised teams, SAFE 2025, 12 teams on 13 TTS models, ASVspoof 2021, 54 teams with valid scores, and ADD 2023, 49 Track 1.2 submissions on Mandarin; no indexed card scored in any of the four (Hiya and McAfee joined at zero on 10 November 2026) |
| Code | 4 | 0 | |
| Music | 5 | 0 | Ircam Amplify joined on 10 November 2026 with eleven named music generators and no record |
Counts from the modalities field of each card, read 8 November 2026 for the first sixteen cards and 9 November 2026 for the twelve added that day in two waves; Winston AI's handwriting (OCR) input counts under text and image, as on the content type matrix. The voice and video rows list records added on 9 November 2026 that scored no indexed card (ASVspoof 5, SAFE 2025, DFDC 2020), so they count toward the records column of no card, and the zero-record rows for Pindrop Pulse, Resemble Detect, Reality Defender, Sensity, Deepware and the rest stand; the twenty-six cards added on 9 November 2026 are counted in the first column.
The sixth wave on 9 November 2026 added the second non-text record with an indexed card and the first voice and video records without one. RAID image 2025 scored Hive and Sightengine on 50 AI images before and after a transferable adversarial perturbation (0.60 to 0.44 and 0.66 to 0.00 accuracy, labelled preliminary), lifting Hive to 2 records and Sightengine to 1. ASVspoof 5 and SAFE 2025 on speech and DFDC 2020 on video score anonymised challenge entrants and Kaggle teams, so they appear in the records-by-content-type table below with zero indexed cards and lift no card's count; Pindrop Pulse, Resemble Detect, Reality Defender, Sensity and Deepware stay at zero.
One of ten rankings in the rankings hub, the content-type twin of the ranking by independent testing, which counts every record a card has. The content type matrix shows which card states which type, and the home board shows every AI detector in the index on the same fields.
Which AI detectors should you check first?
- Ifyou need an image detector that an outside paper has scored: Hive, 1 record, 98.03% accuracy on 630 images, or AI or Not, 1 record, 90.67% on the same 630 images, both in Ha and Passananti 2024.
- Ifyou need video, voice or music detection with a published study: no card qualifies on 9 November 2026; read the vendor statement on the card and run your own test.
- Ifyou need code detection with a published study: no card qualifies; Codequiry, Copyleaks and Originality.ai state the capability on the code view.
What did each study find, tool by tool?
AI or Not
- Verdict
- Fits readers who want an image detector with an outside result and a free monthly text allowance.
- Value
- 1 record on images, under the name Optic: accuracy 90.67%, false positive rate 24.47%, false negative rate 1.15% on 280 human artworks and 350 AI-generated images from five generators, with 4.29% of decisions in the dropped 40% to 60% band (arXiv 2402.03214, Table 1, read 9 November 2026).
- Best for
- Catching AI images, where the record shows 1.15% missed; the false positive rate on human artworks is the cost.
- Main limitation
- Detection of AI images fell from 98.85% to 62.62% under medium Glaze and 58.00% under high (Table 3); the record says nothing about the tool's text, voice, video or music detection, and the vendor's 98.9% accuracy claim covers all four (aiornot.com, read 9 November 2026).
Hive
- Verdict
- Fits readers who want an image detector with a published, outside result.
- Value
- 1 record on images: accuracy 98.03%, false positive rate 0.00%, false negative rate 3.17% on 280 human artworks and 350 AI-generated images from five generators, scores under 40% read as human and over 60% as AI (arXiv 2402.03214, Table 1, read 9 November 2026); a second record joined on 9 November 2026: RAID image 2025, 0.60 (accuracy) on 50 clean AI images, 0.44 after a transferable adversarial perturbation and 1.00 on 50 real images (Table D.13, labelled preliminary, read 9 November 2026).
- Best for
- Checking artwork and photographs; the paper's authors called Hive "the clear winner among all five detectors".
- Main limitation
- Under high-intensity Glaze, detection of AI images fell to 67.56% and false positives rose to 3.23% (Table 4); 6 of 70 Adobe Firefly images were marked human (Table 2). The record says nothing about Hive's text, video, voice or music detection.
GPTZero
- Verdict
- Fits text checks; the image detector has no record.
- Value
- 0 records on image; 11 on text, from Weber-Wulff 2023 to Van Vlasselaer 2026.
- Best for
- Readers weighing the largest text record in the index.
- Main limitation
- The card records "image (AI Vision, specs not published)" from the vendor page read 8 November 2026 (gptzero.me); no sample, rate or study exists for it.
Copyleaks
- Verdict
- Fits schools that want code and text in one LMS tool, on the vendor's word for code.
- Value
- 0 records on code, image or video; 6 on text.
- Best for
- Canvas, Moodle, Blackboard or D2L deployments.
- Main limitation
- The code, image and video capabilities are stated on the product page read 8 November 2026 (copyleaks.com) and no indexed study tested them; Orenstrakh 2023 excluded code submissions because they were unusable without the code (arXiv 2307.07411, read 9 November 2026).
Originality.ai
- Verdict
- Fits publishers who also scan code, on the vendor's word for code.
- Value
- 0 records on code; 5 on text.
- Best for
- Content teams scanning long-form articles.
- Main limitation
- Code detection is stated in the 5.0 launch post read 8 November 2026 (originality.ai); no indexed study scored it.
ZeroGPT
- Verdict
- Fits free text checks; image and video have no record.
- Value
- 0 records on image or video; 5 on text.
- Best for
- Students on the free monthly allowance.
- Main limitation
- Image and video detection are stated on the home page read 8 November 2026 (zerogpt.com) with no published figure.
Pangram
- Verdict
- Fits text checks with 2025 and 2026 records; image and video have none.
- Value
- 0 records on image or video; 2 on text.
- Best for
- Readers who want recent text results on long papers and large samples.
- Main limitation
- Image and video detection are stated on the home page read 8 November 2026 (pangram.com) with no published figure.
Winston AI
- Verdict
- Fits users checking text and images in one account; the image detector has no record.
- Value
- 0 records on image; 2 on text.
- Best for
- Writers on a credit plan who also upload images.
- Main limitation
- Image detection and handwriting OCR are stated on the pricing page read 8 November 2026 (gowinston.ai) with no published figure.
Codequiry
- Verdict
- Fits source code checks, on the vendor's own figures.
- Value
- 0 records; the card states code only.
- Best for
- Instructors and teams checking code for AI authorship and peer similarity.
- Main limitation
- The vendor states that model v3e "catches 90% of AI/GPT-written code and falsely flags 1.3% of human-written code" with no sample size (codequiry.com, read 9 November 2026); no outside study exists.
SynthID
- Verdict
- Fits readers checking for a Google watermark, not for AI generation in general.
- Value
- 0 records; the card states image, video and audio, each only where the file carries a SynthID watermark.
- Best for
- Verifying media from Google, NVIDIA and partner models through the Detector portal or Gemini.
- Main limitation
- "This is not a general AI detector" (synthid.com, read 9 November 2026); the Nature paper (Dathathri et al., 2024) is vendor-authored and reports detectability at a fixed 1% false positive rate on the authors' own outputs, so it is cited on the card and not filed as a study record.
Reality Defender
- Verdict
- Fits enterprise deepfake screening, on the vendor's statements.
- Value
- 0 records; the card states video, image, voice and text.
- Best for
- Banks and platforms screening calls, meetings and uploads.
- Main limitation
- No numeric accuracy or false positive figure is published; the FAQ describes a 1 to 99% probability rating and the home page cites a bank case study of 2.4M+ calls analysed in 2025 (realitydefender.com, read 9 November 2026).
WasItAI
- Verdict
- Fits free image checks with no stored copy; no record.
- Value
- 0 records; the card states image only.
- Best for
- One-off image checks, 10 free per month.
- Main limitation
- The detector is labelled BETA and no accuracy figure is published (wasitai.com, read 9 November 2026).
Deepware
- Verdict
- Fits readers who want a free face-swap video check, with the host's availability noted.
- Value
- 0 records; the card states video only, "deepfakes that use a real video of a person and modify or swap their face", and the FAQ says the scanner does not detect voice, images or text (deepware.ai, read 9 November 2026).
- Best for
- One-off checks of a suspect video up to 10 minutes.
- Main limitation
- No indexed study tested Deepware Scanner, and scanner.deepware.ai did not answer from two networks on 9 November 2026.
Sensity
- Verdict
- Fits forensic and trust-and-safety teams buying through sales.
- Value
- 0 records; the card states video, image, voice (audio) and file forensics, each submission running four services and returning a suspicious or valid verdict (docs.sensity.ai, read 9 November 2026).
- Best for
- Investigators who need on-premise or air-gapped deployment.
- Main limitation
- No indexed study tested Sensity; no prices, quotas or trial terms are published, and new cloud accounts are held for manual review.
Sightengine
- Verdict
- Fits platforms that moderate images and video through one API call.
- Value
- 0 records; the card states image, video, deepfake, voice (audio) and music, called as models=genai on the check.json endpoint (sightengine.com, read 9 November 2026); the count rose from 0 on 9 November 2026 when RAID image 2025 was added: 0.66 (accuracy) on 50 clean AI images, 0.00 after a transferable adversarial perturbation and 1.00 on 50 real images (Table D.13, labelled preliminary, read 9 November 2026).
- Best for
- Developers who want generator attribution scores with the verdict.
- Main limitation
- No indexed study tested Sightengine; the vendor's page cites an 80,000-image benchmark (arXiv 2404.14581) without printing a figure.
TruthScan
- Verdict
- Fits teams checking images and PDF documents on a metered plan.
- Value
- 0 records; the card states text, image, PDF documents, video (deepfake) and voice (audio), with 25 free results a month (truthscan.com, read 9 November 2026).
- Best for
- Finance and operations teams screening receipts and invoices.
- Main limitation
- No indexed study tested TruthScan; the 99.3% image accuracy and the "<1%" image false positive rate are the vendor's own figures.
Illuminarty
- Verdict
- Fits readers who want the image detector with the highest measured false positive rate in the one 2024 study.
- Value
- 1 record: Ha and Passananti 2024, 72.65% accuracy, 67.40% false positive rate on 280 human artworks and 4.69% false negative rate on 350 AI images (Table 1), read 9 November 2026.
- Best for
- Readers comparing Hive, AI or Not (as Optic) and Illuminarty on the same artworks.
- Main limitation
- The authors note the model had not been updated for six months when tested; the text detector is a proof of concept in the vendor's words.
Codeleaks
- Verdict
- Fits teams on the Copyleaks platform who want the code product's stated coverage read beside its withdrawal.
- Value
- 0 records; the card states source code in Ruby, JavaScript, CSS, Markdown, TypeScript, Python, C++, C, Java, C# and PHP, and the release notes record the AI code verdict's removal from the API submit endpoints on 16 December 2025 (copyleaks.com/codeleaks, read 9 November 2026).
- Best for
- Readers tracing what Codeleaks still does after the API change.
- Main limitation
- No indexed study tested Codeleaks; no free allowance is published and code is metered in tokens.
Pindrop Pulse
- Verdict
- Fits contact centres that want live synthetic voice scoring and can buy through sales.
- Value
- 0 records; the card states voice (audio) in the IVR and with a live agent, and video, synthetic avatars and face swaps in Pulse for Meetings (pindrop.com, read 9 November 2026).
- Best for
- Enterprises on Genesys, Five9, Amazon Connect, Cisco or NICE platforms.
- Main limitation
- No indexed study tested Pindrop Pulse; the 99% (accuracy) against known engines is the vendor's own computation on the ASVspoof5 dataset.
Resemble Detect
- Verdict
- Fits developers who want audio, image and video deepfake checks from one API.
- Value
- 0 records; the card states voice (audio), image, video and music, with 17 upload formats and per-second or per-image credits (resemble.ai/products/detect, read 9 November 2026).
- Best for
- Teams that want a pay-as-you-go plan with no subscription fee.
- Main limitation
- No indexed study tested Resemble Detect; the 99.5% pooled accuracy and the sub-1% rates are the vendor's own benchmarks.
Deepfake-O-Meter
- Verdict
- Fits readers who want to see that an open research platform has no outside record of its own.
- Value
- 0 records on image, video or voice: the platform's own paper describes the service and its users and prints no accuracy for the platform (arXiv 2404.13146, read 10 November 2026); the detectors endpoint lists 41 methods (detectors endpoint, read 10 November 2026).
- Best for
- Comparing several published detectors on the same upload.
- Main limitation
- Each hosted detector carries its own authors' results and scope line; the index holds no study of the platform.
ElevenLabs Classifier
- Verdict
- Fits readers who want to see that the one voice detector built by a generator has no outside record.
- Value
- 0 records: no indexed study scored the classifier; SAFE 2025 names ElevenLabs among the generators of its test clips, not among the detectors (elevenlabs.io/docs, read 10 November 2026).
- Best for
- Checking an ElevenLabs clip with the vendor's own tool.
- Main limitation
- The tool scores ElevenLabs audio only and does not reliably classify Eleven v3; the precision and recall figures are the vendor's (elevenlabs.io/ai-speech-classifier, read 10 November 2026).
FakeCatcher
- Verdict
- Fits readers who want the status of a detector that was announced and never sold.
- Value
- 0 records: the FakeCatcher paper reports 96%, 94.65%, 91.50% and 91.07% on four datasets, the authors' own evaluation (arXiv 1901.02212), and no indexed study scored the product (intel.com newsroom, 14 November 2022, read 10 November 2026).
- Best for
- Readers tracing the 96% figure to its source.
- Main limitation
- Intel publishes no product page, price, sign-up, download or API for FakeCatcher on 10 November 2026; the DFDC 2020 record scored Kaggle teams, not it.
Ircam Amplify
- Verdict
- Fits readers who want the one music detector with a named generator list and no outside record.
- Value
- 0 records: no indexed study scored AIMD; the V2.3 release note's 99.7% (accuracy) and 0.14% false positive rate at the 50% threshold are measured on Ircam Amplify's internal benchmark datasets (docs.ircamamplify.io, read 10 November 2026).
- Best for
- Catalogue scans against eleven named music generators.
- Main limitation
- Every figure is the vendor's own; access is by application and no price per track is published (ircamamplify.com, read 10 November 2026).
Hiya Deepfake Detector
- Verdict
- Fits readers who want to see that a free voice extension has no outside record.
- Value
- 0 records: no indexed study scored the extension or the Hiya AI Voice Detection platform it is "Powered by"; the two voice challenges added on 10 November 2026, ASVspoof 2021 and ADD 2023, scored anonymised teams and name no vendor (Chrome Web Store listing, read 10 November 2026).
- Best for
- Checking the voice in a playing video from one second of audio.
- Main limitation
- The "Over 99%" is stated for the platform with no dataset named, and the Chrome Web Store disclaimer says the extension's outputs "may not always be 100% accurate" (hiya.com AI Voice Detection, Chrome Web Store listing, read 10 November 2026).
McAfee Deepfake Detector
- Verdict
- Fits readers who want to see that a bundled on-device voice tool has no outside record.
- Value
- 0 records: no indexed study scored the detector; the video benchmark added on 10 November 2026, FaceForensics++ 2019, scored six retrained research methods and no product (mcafee.com/ai/deepfake-detector, read 10 November 2026).
- Best for
- An alert while watching English-language video on a supported PC.
- Main limitation
- The 96% rests on the 21 August 2024 press release with no dataset, the live page prints no figure, and the tool runs only on named Lenovo, HP and Intel Core Ultra PCs (mcafee.com newsroom, 21 August 2024, read 10 November 2026).
How we ranked
- Field sorted
- Count of indexed studies that scored the tool on image, voice, video, code or music
- Direction
- Higher first
- Tie rule
- More study records in all first, then name
- Qualification
- A card that states at least one content type other than text
- Qualified
- 30 of 72 detectors
Field sorted: the count of indexed study records that scored the tool on image, voice, video, code or music, compiled on 9 November 2026. Direction: more records rank higher. Tie rule: cards with the same count are ordered by their total study records, most first, then by product name alphabetically, which is why the eighteen cards at zero run from GPTZero (11 text records) to the twelve cards with none, and why AI or Not (1 record) sits above Hive (1 record) by name. Qualification rule: the card states at least one content type other than text in its modalities field; the eighteen cards that state text alone are not listed (Compilatio's handwriting OCR reads handwritten text, so it counts as text). Only one record in the index, Ha and Passananti 2024, scored a non-text type, so the field is 1 for AI or Not and Hive and 0 for every other qualifying card. 14 of 22 qualified on the morning of 9 November 2026; 16 of 28 qualified after the six cards of the third wave, AI or Not with a record and SynthID without; 20 of 38 qualified after the ten cards of the fourth wave, TruthScan, Sightengine, Deepware and Sensity joining at zero records. 24 of 72 qualify after the four cards of the fifth wave on 9 November 2026: Illuminarty joins with 1 record, the third card scored on images in Ha and Passananti 2024 (72.65% accuracy, 67.40% false positives, 4.69% false negatives), and Codeleaks, Pindrop Pulse and Resemble Detect join at 0 (Illuminarty card). Full rules: methodology page; every record: studies index.
Why numbers differ between sources
Two pages added 9 November 2026 read the same modality fields: the text and image question lists the fifteen cards that state both text and image with their text and image record counts, and the code languages ranking counts the programming languages the three code cards name, the code row here standing at zero records.
The 35 indexed studies were built for text. Weber-Wulff 2023, Walters 2023, Perkins 2024, RAID 2024, Orenstrakh 2023, Krishna 2023, Sadasivan 2023, Jabarian and Imas 2025 and Van Vlasselaer 2026 all fed written passages or papers to the tools; Gao 2023 fed medical abstracts; Chaka 2023 fed chatbot responses; the NIST pilot scored summaries. Ha and Passananti 2024 is the one paper that fed images, and it tested five image detectors of which two, Hive and Optic (now AI or Not), have a card (arXiv 2402.03214, read 9 November 2026). Vendor pages describe image, video and voice detection in general terms; Hive's documentation gives chunk sizes for audio (10 seconds) and music (3 seconds) but no error rate (docs.thehive.ai, read 8 November 2026). A content type stated without a study is a capability claim, and the index records it as one.
Public leaderboards are a separate kind of record from the studies counted here: the leaderboard question, added 9 November 2026, lists the 6 cards with a RAID entry and notes that no indexed card is named in the voice and video challenge tables. The voice and video question, added the same day, lists the 9 cards that cover both media and finds 0 records on either.
Music is the one content type in the index with cards and no record at all: five cards state it and the music detectors question lists them with their generators and limits, read 10 November 2026; the two voice vendors with no record are set side by side in the Pindrop vs Resemble Detect comparison.
Questions about independently tested AI image, voice and video detectors
Which AI image detectors have been independently tested?
Two cards in the index: Hive, scored by Ha and Passananti 2024 at 98.03% accuracy and 0% false positives on 280 human artworks and 350 AI images, and AI or Not, tested as Optic in the same paper at 90.67% accuracy and 24.47% false positives (arXiv 2402.03214, read 9 November 2026). The same paper tested Illuminarty, DIRE and DE-FAKE, none of which has a card.
Has any AI voice or video detector in the index been independently tested?
No. Seven cards state voice detection and nine state video detection, read 8 and 9 November 2026 (Sightengine, Deepware, Sensity and TruthScan added the video claims of the fourth wave), and none of the 35 indexed studies scored either type. The claims rest on vendor pages until a paper that names the tools, the sample and the results is indexed.
Has any AI code detector been independently tested?
Not in the index. Codequiry, Copyleaks and Originality.ai state code detection; Orenstrakh 2023 mined over 2,000 code submissions and excluded them as unusable without the code (arXiv 2307.07411, read 9 November 2026). Codequiry's 90% and 1.3% are the vendor's own figures with no sample size.
How well did Hive do on images?
Accuracy 98.03%, false positive rate 0.00% and false negative rate 3.17% on 630 unperturbed images; detection by generator ranged from 91.04% on Adobe Firefly to 100% on CIVITAI and SDXL images; under high-intensity Glaze, detection fell to 67.56% and the false positive rate rose to 3.23% (arXiv 2402.03214, Tables 1, 2 and 4, read 9 November 2026).
Why does GPTZero rank second with no image study?
Because the tie rule orders cards at zero by their total study records, and GPTZero has 11, all on text. The rank says an outside paper has examined the tool often; it says nothing about the image detector, whose specifications the vendor has not published (gptzero.me, read 8 November 2026).
Why are text-only detectors left out?
The ranking asks whether a non-text claim has been tested, so a card that states text alone has no claim to test here; the eighteen text-only cards, Grammarly, Scribbr, Sapling, JustDone, Turnitin, GLTR, AIDetectPlus, the GPT-2 Output Detector, Writer, Undetectable AI, Grammarly Authorship, Turnitin Clarity, Smodin, BrandWell, Surfer, Compilatio, Quetext and DetectGPT, appear on the ranking by independent testing where their text records count.
Does a 98% image accuracy transfer to photographs or video?
The paper's sample was artwork in seven styles plus 70 upscaled photographs and 40 hybrid images; the authors state the fixed set of styles as a limit (arXiv 2402.03214, Section 10, read 9 November 2026). The result applies to that sample, those generators and the versions tested between late 2023 and June 2024.
How often is this ranking updated?
The ranking was created on 9 November 2026 and changes whenever a study record that scored image, voice, video, code or music is added, or a card adds or drops a content type, with the Updated line matching. Candidate papers on voice and video detection are read in the primary source before any is indexed. Figures from the index, compiled 9 November 2026.
Sources and revisions
Sources
- Ha, Passananti et al. 2024, Organic or Diffusedarxiv.org/abs/2402.03214
- Orenstrakh et al. 2023arxiv.org/abs/2307.07411
- Hive AI-generated content detection docsdocs.thehive.ai/docs/ai-generated-content...
- GPTZerogptzero.me
- Copyleaks AI content detectorcopyleaks.com/ai-content-detector
- Originality.ai 5.0 launch postoriginality.ai/blog/introducing-originali...
- ZeroGPTzerogpt.com
- Pangrampangram.com
- Winston AI pricinggowinston.ai/pricing
- Codequiry AI code detectioncodequiry.com/detect-ai-written-code
- Reality Defender FAQrealitydefender.com/faq
- WasItAIwasitai.com
- AI or Notaiornot.com
- SynthID Detectorsynthid.com
- TruthScan pricing (read 9 November 2026)truthscan.com/pricing
- Sightengine AI image detection documentation (read 9 November 2026)sightengine.com/docs/ai-generated-image-d...
- Deepware FAQ (read 9 November 2026)deepware.ai/us/faq/index.html
- Sensity documentation (read 9 November 2026)docs.sensity.ai
Revision log
- Ranking created, 14 of 22 qualify, 1 record on a non-text content type
- Six wave 3 cards added, ai or not (1 record, as optic) ranks first on the tie rule and synthid joins at zero, 16 of 28 qualify
- Ten wave 4 cards added, deepware, sensity, sightengine and truthscan join at 0 records, 20 of 38 qualify
- Four wave 5 cards added, illuminarty joins at 1 record (ha and passananti 2024), codeleaks, pindrop pulse and resemble detect at 0, 24 of 42 qualify
- Eleven wave 6 records added
- Text and image question and code languages ranking linked (seventh wave)
- Eighth-wave pages linked
- Twelve wave 10 cards added
- Nine wave 11 cards and five records added
- Twelfth-wave pages linked
). Illuminarty, one of four cards added in the same wave, joined with its one image record from Ha and Passananti 2024, so 18 of 72 cards now qualify.: register plate](/img/independently-tested-480.webp)
