When you ask an AI about something on your screen, there are two ways to hand it over. You can turn the pixels into text first (OCR) and send the text, or you can send the picture itself to a model that reads images. They fail in different ways, so it helps to know which to use when.
The short rule
- If you could read it aloud without losing anything, send text.
- If the layout or a drawing carries meaning, send the image.
A paragraph, a multiple-choice question or an error message is fine as text. A bar chart, a circuit, a geometry figure or a table with merged cells is not.
What OCR does well, and where it slips
OCR recognizes characters on your own computer and produces plain text. Only that text is sent.
- Small and cheap. A text question is a fraction of the size of an image, so it uses less of a provider’s allowance and comes back faster.
- More private. Nothing outside the recognized words leaves your machine: no window chrome, no other tabs, no faces in a corner of the screen.
- Works with any model, including fast text-only ones.
Where it slips:
- Small marks that change the meaning. A minus sign, a “not”, a decimal point, a subscript or an exponent can be dropped or misread.
x²may arrive asx2. - Structure. Columns can be read in the wrong order, and a table becomes a run of words.
- Anything that is not text. OCR sees the labels of a chart but not the bars.
What an image does well, and where it costs
Sending the image lets the model see exactly what you see.
- Charts, diagrams and figures stay intact.
- Formulas keep their layout: fractions, roots and matrices are not flattened.
- Handwriting and unusual fonts often read better than OCR manages.
The costs:
- Larger requests. Images use more of a free allowance and take longer.
- Not every model accepts images. A text-only model will refuse or ignore it.
- Everything in the selection is sent. Select tightly.
A quick way to decide
- Is there a figure, chart or formula that matters? Image.
- Is it plain text, clearly rendered? Text.
- Is it text, but the answer depends on a symbol or a number? Text, then read what was recognized before you trust the answer. If it looks wrong, ask again as an image.
How LuciQA handles both
LuciQA has both modes in its tray menu: Ask from text (OCR) and Ask from image / diagram. OCR runs on your device in nine languages. The answer panel shows exactly what was sent, the recognized text or the image, so you can spot a misread minus sign before it becomes a wrong answer. If your chosen model reads text only and you send an image, LuciQA uses that provider’s image model for that one request.
In Multi, which handles a whole page of questions at once, you choose text or image for each question separately and can correct the recognized text before anything is sent.