How accurate is it?
On a clean screenshot or a flat scan of printed text, high enough to paste and skim. On a photograph taken at an angle or in poor light, low enough that you should read it against the picture before using it. Rather than ask you to guess which you have, the tool reports its confidence for every word and marks the doubtful ones, so the checking has somewhere to start.
Can it read handwriting?
No. Recognising handwriting is a genuinely different problem — the shapes vary per person rather than per font — and needs a model trained for it. This engine is trained on printed type and will produce confident-looking nonsense from handwriting rather than an error, which is the reason to say so plainly here.
Which languages does it handle?
English only, in this version. The language model is about two megabytes on its own, and shipping several so that most visitors download models they will never use is the trade this site keeps refusing. Other languages are on the list for the image tools and will arrive as a choice you make rather than a cost everybody pays.
Why is it slow, and why the download?
Because a real OCR engine and its language model are being fetched and then run in your browser, rather than a photograph being sent to a machine in a data centre that does the work. That is the trade this whole site makes: about five megabytes once, some seconds per image, and your document stays on your device.
Can it read the text in a scanned PDF?
Yes, and it has its own tool now: OCR a PDF draws every page and recognises it in one step, so you no longer have to export the pages as images and bring them here. Use that one for a document; this one is for a picture you already have.
Is my picture uploaded?
No. The engine, the language model and the recognition all live in this tab. This is the tool where that matters most on the whole site: the pictures people run OCR on are receipts, letters, prescriptions, payslips and passport pages, and every free online OCR service receives all of them.