Drop a photo, screenshot or scan, and download the words in it as a .txt file. The reading happens on your device, which matters, because the images people need text from are letters, receipts, forms and screenshots of things they were sent.
What reads well. Screenshots read best of all: the text is already crisp pixels. Scans of printed pages read well. Phone photos of documents read reasonably when the page is flat, the light is even and the text is in focus.
What reads badly, said plainly. Handwriting mostly fails, because the recogniser is built for print. Curved book spines, low light, glare and stylised lettering all cost accuracy. If a photo reads poorly, a flatter and brighter retake usually beats any setting this tool could offer.
A crop first often helps. Reading only the paragraph you need, rather than the whole cluttered page, is one stop at the crop tool away and usually sharpens the result.
First use fetches the reading engine - about 17MB, kept by your browser afterwards, and the second image starts immediately. It is the accurate model rather than the quick one, the right trade when the alternative is retyping. English print is what it reads.
Nothing you drop here is uploaded. The network tab will show the engine arriving once and your images going nowhere.