OCR: Extract Text from Photos

Of all ChatGPT’s visual capabilities, extracting text from images delivers value the fastest. You don’t need a dedicated OCR app: simply snap a photo of a document, sign, label, or screen. ChatGPT transcribes the text and processes it immediately.

An OCR That Understands What It Reads

The difference from traditional tools comes down to one word: context. A classic OCR engine recognizes characters and hands you a raw string of text, leaving you to make sense of it. ChatGPT understands what it reads, letting you extract, structure, summarize, or translate in a single step.

A receipt illustrates this gap perfectly. Send a photo with the prompt, “Extract all items and their prices, then calculate the total.” The model reads each line, distinguishes labels from amounts, organizes the data, and performs the calculation. Always verify the total; a single misread digit skews the entire result. A classic OCR tool would have stopped at transcription.

Four Professional Use Cases

A business card picked up at a trade show is the textbook example, mostly because these cards usually end up buried in a drawer. Photograph it and ask, “Extract the contact information and format it for a CRM import (first name, last name, company, email, phone).” The output comes out structured and ready to paste.

Administrative documents follow the same logic. Faced with a partially filled paper form, the prompt “Transcribe the filled fields from this form into a table format” digitizes in one pass what manual entry would take much longer to achieve. You can handle handwritten meeting notes the same way: “Transcribe these handwritten notes, correct any errors, and organize them into action items” turns scribbles into a distributable summary, provided the handwriting remains legible. Accounting benefits too—on a photo of an invoice, “Extract the invoice number, date, pre-tax amount, VAT, and total amount” returns the five fields you need for your tracking.

Extract and Act in One Move

You don’t have to separate extraction from processing. A sign in a foreign language can be read and translated with a single prompt: “Read this sign and translate it into English.” However, the OpenAI Help Center notes that text in non-Latin scripts, such as Japanese or Korean, is read less accurately. For these documents, verify the transcription before trusting the translation. A photographed newspaper article can be read and summarized into three key points. A menu can be read and filtered—“identify vegetarian options with their prices.” A printed table converts directly to CSV.

Make it a habit to state the final goal rather than an intermediate step. Asking for transcription in one message and translation in a second costs you an extra round trip for the same result.

The Photo Decides Everything

Extraction accuracy is determined at the moment you take the picture. No prompt can recover information lost there.

Lighting comes first. A well-lit document without shadows or glare reads easily; diffuse natural light yields the best results. Counterintuitively, flash often degrades readability by creating glare that wipes out entire sections of text. Angle matters just as much—shoot flat and perpendicular, as an oblique angle distorts characters. Use the highest resolution available and resist the temptation to resize before sending; well-intentioned compression destroys precisely the useful information. Finally, framing should include all text without excessive empty space around it, as the model analyzes images better when content occupies a significant portion of the frame.

Difficult Cases and How to Handle Them

Some text resists extraction. In these cases, the goal is to make uncertainty visible rather than ignore it. Facing messy handwriting, ask, “Transcribe this text. When you aren’t sure about a word, indicate it in brackets with a question mark.” This tells you exactly where to focus your review. For text on a complex background—such as a poster where letters stand out against a photograph—crop to the text-only area. For old or damaged documents, a bit of context guides the reading: “This document dates from the 19th century; some words may use archaic spelling.”

Two checks remain your responsibility. First, numbers: confusions between 0 and O, 1 and l, or 5 and S are common and silent, so double-check amounts, ID numbers, and dates. Second, languages: on a multilingual document, specify which languages are present. Otherwise, the model may normalize a passage to the dominant language.

Your Turn

Test extraction reliability. Put this to the test with a real document—a receipt, letter, business card, book page, or handwritten notes. Use the prompt, “Extract all text from this image. Then organize the extracted information structurally and identify key data points.” Compare the output line by line with the original. This check is the only way to gauge real extraction reliability.

Official sources: ChatGPT Image Inputs FAQ · Does ChatGPT tell the truth? (OpenAI Help Center)

Key Takeaways

• ChatGPT acts as advanced OCR: it extracts, structures, and processes text in a single step
• Combine extraction with action (translation, summarization, reformatting) to save time
• Photo quality is critical: good lighting, perpendicular angle, high resolution
• Always verify numerical data and proper nouns in transcriptions
• Specify context and language to improve accuracy on complex texts