AI detectors that state OCR of handwriting or scanned pages: the three cards that say so, and the eight that take an image without reading it
An AI detector states that it reads handwriting or an image of text by OCR on 3 of the 72 cards in the index, by the modalities, input limits and plans fields read 8 and 9 November 2026: Winston AI ("handwriting (OCR)" among its modalities and "OCR, PDF reports" in the Essential plan), Compilatio ("OCR on handwritten documents" as a Magister feature) and Copyleaks ("OCR files: gif, png, bmp, jpg, jpeg" in its input limits) (gowinston.ai/pricing, compilatio.net/magister, copyleaks.com/ai-content-detector). None of the three names a scanned PDF, and 0 of the 72 cards state that a scanned PDF is read by OCR and scored, the finding of the scanned PDF question read 10 November 2026; Turnitin's capabilities FAQ, re-read the same day, says the vendor is "focusing on typed text" and that Gradescope, built for handwritten work, will not gain AI detection. Eight more cards accept an image file for a purpose other than reading text, scoring it as a possible AI image or as a page to be judged by forgery, and sit in the second table so a reader does not mistake an image upload for OCR. This page is one of the views under AI detectors by feature; the content type matrix shows every modality each card states, the file types ranking counts upload formats, and the full index shows all 72 cards.
- Field 1OCR statement as the vendor makes it
- Field 2What is read
- Field 3Where it is stated
- Field 4Study records
- Field 5Source, read date
Three fields, any one of which can qualify a card. The modalities field qualifies a card when the vendor lists handwriting or OCR as something the detector takes (Winston AI, Compilatio); the plans field qualifies a card when a tier names OCR as an inclusion (Winston AI's Essential); the input limits field qualifies a card when it names OCR files or a scanned-page route (Copyleaks). The statement has to be the vendor's: a format list that includes png or jpg is not an OCR statement, because every image detector in the index takes png and jpg to judge whether the picture was generated, and GPTKit's upload control accepts .jpeg, .png and .gif with no sentence about what happens to the pixels. The second table holds those cards with the vendor's stated purpose for the image. A third group, cards that accept a PDF, is the scanned PDF question's first table (18 cards), of which four state a position on scans, all requiring a text layer or typed text, and none states OCR. Field rules are in the methodology.
Which fields decide whether a detector states OCR?
| Detector | OCR statement as the vendor makes it | What is read | Where it is stated | Study records | Source, read date |
|---|---|---|---|---|---|
 Copyleaks | "OCR files: gif, png, bmp, jpg, jpeg", listed apart from "Image files: JPG, JPEG, PNG, WEBP, BMP, HEIC, HEIF" | Text inside an image file; handwriting not named | Input limits field, from the detector page | 9 | copyleaks.com/ai-content-detector, 8 November 2026 |
 Winston AI | "handwriting (OCR)" among the modalities; Essential includes "OCR, PDF reports"; uploads in .docx, .png and .jpg | Handwriting in an image; "PDF reports" is an output, not an input | Modalities and plans fields, from the pricing page | 2 | gowinston.ai/pricing, gowinston.ai/ai-content-detector, 8 November 2026 |
 Compilatio | "OCR on handwritten documents" as a Magister+ feature beside AI detection | Handwritten documents uploaded by teachers through Magister | Modalities field, from the Magister page | 1 | compilatio.net/magister, 9 November 2026 |
Statements as each vendor makes them, read 8 and 9 November 2026 and ordered by study records then name; no record in the studies index tested any of the three on handwriting or on a scanned page.
Eight cards whose image or PDF input is for something other than reading the text, read 8 to 10 November 2026; Detecting-AI.com is kept because it is the one text card that states in words that scanned pages are not read. The other 61 cards state no OCR, no handwriting and no image route for text.
Why each card is in this view
- AIDI-D-0003Copyleaks
the detector page keeps two format lists, one headed "OCR files" and one headed "Image files", so the vendor distinguishes reading text out of a picture from judging a picture, and the OCR list is for text (copyleaks.com/ai-content-detector, read 8 November 2026); 9 study records, none on an image of text.
- AIDI-D-0005Winston AI
the only card that lists handwriting as a modality in its own right and sells OCR as a plan inclusion from the lowest paid tier; the upload formats are .docx, .png and .jpg, so a handwritten page arrives as a photograph (gowinston.ai/pricing, read 8 November 2026); 2 records, neither on handwriting.
- AIDI-D-0032Compilatio
OCR on handwritten documents is a listed Magister+ feature beside the AI detection the Standard tier lacks, so the two sit in the same institutional licence (compilatio.net/magister, read 9 November 2026); 1 record, the 2023 free tool in Weber-Wulff 2023, on typed text.
What the evidence says about handwriting and scans
Nothing in the index tests it. The 35 records in the studies index were run on typed or generated text, on images, on audio or on video; no record scored a detector on a photographed handwritten page or on a scanned PDF, and the three cards above carry no figure for OCR accuracy from their own pages either. What a reader can take from the records is the chain of steps a scanned page adds: an OCR pass produces text that may differ from what was written, and the detector then scores that text. The short-text and non-native findings apply to whatever the OCR pass yields: Liang 2023's seven detectors flagged an average of 61.22% of 91 human TOEFL essays (Liang 2023 record), and Sapling's own page states that the shorter and more general a text is, the more likely a false positive (sapling.ai/ai-content-detector, read 8 November 2026). A one-page handwritten exam answer is short, and a page the OCR pass reads badly is shorter still.
The clearest vendor statement in the index is a refusal. Turnitin's capabilities FAQ, edited 3 November 2026 and read 10 November 2026, says "We do not currently have plans to add these capabilities to Gradescope, since the primary use case for Gradescope is handwritten text while for AI detection we're focusing on typed text" (guides.turnitin.com), which the Gradescope question and the Turnitin card both record; Detecting-AI.com states the same limit for its own tool in one sentence. The scanned PDF question lists the 18 cards that accept a PDF and the four that say the file needs a text layer, and the content type matrix shows the two handwriting cells. No detector score should be the only evidence in a misconduct, hiring or publishing decision, and a score on text that an OCR pass produced from a handwritten page carries that pass's errors as well as the detector's.
Sources
- https://copyleaks.com/ai-content-detector (read 8 November 2026)
- https://gowinston.ai/pricing/ (read 8 November 2026)
- https://gowinston.ai/ai-content-detector/ (read 8 November 2026)
- https://www.compilatio.net/en/magister (read 9 November 2026)
- https://gptkit.ai/ (read 9 November 2026)
- https://truthscan.com/truthscan-ai-text-detection-api-documentation (read 9 November 2026)
- https://www.pangram.com/ (read 8 November 2026)
- https://docs.thehive.ai/docs/supported-file-formats (read 8 November 2026)
- https://docs.realitydefender.com/sdks/quickstart (read 9 November 2026)
- https://docs.aiornot.com/api-reference/reports-by-modality/image (read 9 November 2026)
- https://walterwrites.ai/pricing/ (read 8 November 2026)
- https://detecting-ai.com/ (read 10 November 2026)
- https://guides.turnitin.com/hc/en-us/articles/28477544839821-Turnitin-s-AI-writing-detection-capabilities-FAQs (read 10 November 2026)
- https://sapling.ai/ai-content-detector (read 8 November 2026)
- https://arxiv.org/pdf/2304.02819v3 (Liang et al. 2023, read 8 November 2026)
Revision log: 2026-11-10: view created at index version 2026.11.13 (twelfth wave)
FAQ: AI detectors and handwriting
Which AI detectors can read handwriting?
Two of 72 state it on 10 November 2026: Winston AI lists "handwriting (OCR)" among its modalities and includes OCR in its Essential plan, and Compilatio lists "OCR on handwritten documents" as a Magister+ feature; Copyleaks states OCR of image files without naming handwriting, by the fields read 8 and 9 November 2026.
Do any AI detectors OCR a scanned PDF?
None of the 72 cards states it, by the input limits and file types fields read 10 November 2026: 18 cards accept a PDF upload, four of them (Sapling, ZeroGPT, Detecting-AI.com and Turnitin) say the file needs a text layer or typed text, and 14 say nothing; the scanned PDF question lists all 18.
Does Turnitin detect AI in handwritten work?
No. The capabilities FAQ, read 10 November 2026, says AI detection focuses on typed text and that Gradescope, whose primary use case is handwritten text, will not gain it; the Gradescope question records the statement and its date.
Which detectors accept an image but do not read the text in it?
Eight on 10 November 2026: GPTKit (image formats in the text tool's upload control, purpose not stated), TruthScan (a PDF page scored as an image), Pangram, Hive, Reality Defender, AI or Not and Walter Writes (AI image detection), and Detecting-AI.com, which states that scanned PDFs are not supported, by the input limits fields read 8 to 10 November 2026.
Has any study tested an AI detector on handwriting?
No record in the studies index has, as of 10 November 2026; the 35 records cover typed text, images, audio and video, and none of the three OCR cards publishes an OCR accuracy figure on its own pages.
). 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)
