Analyze images and photos

ChatGPT Vision: Understanding images

ChatGPT isn’t limited to text: every model accepts images. Photos, screenshots, charts, photographed documents, diagrams—it describes them, reads them, and comments on them. The value is immediate because this feature eliminates the transcription step: you show instead of describing. It’s available on Free and paid plans, subject to each plan’s quotas.

Sending an image to ChatGPT

On the web and in the desktop app, three paths lead to the same result: the + button in the input area, then Add photos & files; dragging and dropping an image into the input area; or pasting a copied image from the clipboard (Ctrl+V or Cmd+V). On mobile, the same + button adds a photo, either one you just took or an older one; on iPhone, a long press on + opens your recent photos directly. ChatGPT accepts PNG, JPEG, and non-animated GIF formats, up to 20 MB per image.

You don’t have to send just one image at a time. Including multiple visuals in a single message lets you ask for a comparison, and you can add new images as the conversation progresses to go deeper or shift angles.

What vision enables in daily work

The first and most valuable use case involves charts. Capture a curve from an annual report, a dashboard, or a presentation, and have ChatGPT comment on it instead of deciphering it yourself.

[image of chart] “Analyze this chart. What are the main trends? Are there any anomalies? Summarize the conclusions in 3 points.”

Next comes document reading: ChatGPT reads text present in an image. “Transcribe the text from this document, then summarize it in 3 sentences.” The same approach works for invoices, with structured extraction that saves you manual entry—“Extract the information from this invoice: vendor, date, amount excluding tax, VAT, amount including tax.”—and for translation when you photograph a sign or menu in a foreign language.

Identification forms a third category: an error message displayed on screen whose meaning and resolution you need, an object whose name you’re looking for, a dish for which you want a similar recipe. Finally, vision serves as a critical eye for professional visuals.

[screenshot of a website] “Analyze this homepage. Strengths and weaknesses in terms of user experience. Propose 3 concrete improvements.”

What separates a good analysis from a bad one

Result quality depends first on image quality. OpenAI recommends enlarging text to be read without cutting off important details, and avoiding rotated or upside-down images, as well as panoramic or “fish-eye” views, which are often misinterpreted. To draw attention to a specific area, annotate the image before sending it, using your phone’s editing tool, for example.

The second variable is the text accompanying the image. “What is this?” invites a general description, often useless. “Here is our Q1 sales chart: analyze the trend and identify problematic months” invites an actionable response. The principle also applies when combining multiple visuals: “Here is my competitor’s logo [image]. Compare it with ours [image]. What is each one’s visual positioning?”

Limits to know

OpenAI itself publishes a list of limits. The model isn’t designed to interpret specialized medical images, such as CT scans, and must not serve as medical advice. It reads text in non-Latin alphabets, such as Japanese or Korean, less reliably. It can make mistakes on charts where only color or line style (solid, dashed, dotted) distinguishes the curves, on precise localization tasks like reading a chess position, and provides only approximate counts. It doesn’t read the file name or its metadata, and the image is resized before analysis.

One caveat deserves emphasis: for numbers read from an image, such as a photographed table or scanned PDF, exact values aren’t extracted reliably. When precision matters, send the original spreadsheet or text file instead, and manually double-check anything driving a decision. Videos can finally be attached as files, even on Free, but their analysis may be partial.

Your turn

Three tries. Photograph or capture a report page or an invoice, and ask for extraction of key information into a table; verify each amount. Next, capture a chart from an article or dashboard, and ask for a trend analysis. Finish with a photo of any object, sent twice: once without instructions, once with a specific question. Compare the depth of the two responses.

Official sources: ChatGPT Image Inputs FAQ (OpenAI Help Center: plans, formats, adding images, limits) · Data analysis with ChatGPT (scanned or image-based tables) · ChatGPT capabilities overview (image analysis)

Key takeaways

• All ChatGPT models analyze photos, screenshots, charts, and photographed documents, across all plans
• Send an image with the + button, drag-and-drop, paste, or on mobile, from your photos
• A specific question and a sharp image beat a vague “what is this?”
• Official limits: medical images, non-Latin alphabets, curves distinguished by style only, counts
• Numbers read from an image aren’t reliably exact: prefer the original file