--- id: more-than-text title: Using files, images and voice level: 2 modes: [day-pack, rucksack] moment: "They upload a long file and the answer misses something in it, send a photo or screenshot, ask whether the AI can see or hear things, or try talking to it by voice." keywords: [upload a file, attachment, photo, screenshot, image, video, audio, voice mode, read the whole file, PDF] source: "research/2026-10-23-concepts/more-than-text.md; mapping (product/learning-system/MAPPING-1-2.md): FAQ 2.04 (home), 2.05 (home), 2.06 (home); MISTAKES 2.11 (home); tips none in the register; strand C" --- ## What it is Besides typing, many AI apps can take files, photos, screenshots and your voice. With a long file, such as a 60-page report, an app may read only parts of it, so point it to the page or section that matters. It can also misread a photo of small, tilted or handwritten text, so check any numbers it takes from one. A voice chat usually saves a written transcript, so treat what you say as you would what you type. ## Tips - When you upload a long file, name the page or section you need, or paste that part in, because the app may have read only parts of it. - Crop and straighten a photo of writing before you send it, because the AI can misread tilted or tiny text. - Check numbers it reads from a photo of handwriting or a table against the original, because it can fill in what it expects instead of what is there. ## Questions people ask - Q: Why did it ignore the file I uploaded? A: It may have read only part of a long file, and some apps read only the text of a document and skip pictures and charts in it. Name the page or section you need, or paste that part into the chat. - Q: Can AI see images, watch videos and hear audio? A: Images, usually yes, though it can misread small, tilted or handwritten text. Video and audio depend on the app and your plan, since some take clips or recordings with a length limit and others may not take either. - Q: What's voice mode, and when should I use it? A: Voice mode is a spoken, two-way chat, where you talk and your AI talks back. It suits busy hands or talking an idea through, but check any exact figures in writing, and remember a transcript is usually saved. - Q: Are my photos and voice kept? A: Assume so. Voice chats usually save a transcript, uploaded files and photos may be stored, and each app has its own settings for how long they are kept and whether they can help train future AI. ## For your AI Facts (checked 23 October 2026; research note research/2026-10-23-concepts/more-than-text.md). - What each takes: ChatGPT, documents, spreadsheets, slides, text and images on every plan, audio on paid plans, with Free limited to 3 uploads a day (https://help.openai.com/en/articles/8555545-file-uploads-faq); Claude, files and images, PDFs to 1,000 pages but charts and images in them read only to 100 pages (https://support.claude.com/en/articles/8241126-upload-files-to-claude); Gemini, files, images, video and audio, with more on paid plans (https://support.google.com/gemini/answer/14903178); Copilot, documents and images, kept up to 18 months (https://support.microsoft.com/en-us/microsoft-365-copilot/file-upload-in-microsoft-copilot). - Long files: OpenAI's work guide says only the first part of a very long document sits in front of the AI and the rest is searched in pieces (https://help.openai.com/en/articles/10029836-optimizing-file-uploads-in-chatgpt-enterprise); Gemini says past its limit replies can miss details; facts in the middle of a long input are often found least well (Liu et al., 2023). An answer can miss a part because the app read only some of the file, read only its text and not its pictures, or summarised past it; you cannot always tell which. Never say a file was read in full. - Images: the companies list trouble with rotated, tiny or low-quality images, counts and charts; on handwriting and tables the best of 18 models scored about 72 percent against people's 77, and most of the models' errors came from what they expected rather than what was in the image, while printed text read well (WildHandBench, preprint, 24 August 2026, https://arxiv.org/abs/2608.22959). - Voice: ChatGPT keeps voice clips 30 days with the transcript and does not train on them unless shared (https://help.openai.com/en/articles/8400625-voice-mode-faq); Claude's voice mode saves a transcript and counts toward usage; Gemini Live audio and video are not used to improve Google's services by default; transcripts are kept like chats. Photos and recordings of people can be sensitive information (openrucksack.com/map/other-peoples-information.md). Teach it like this: when an answer misses something in a long file, ask them which page or section it is on and work from that part, or have them paste it. For a photo or screenshot of figures, read back the key numbers for them to check. Voice suits talking something through or practising, hands-free.