Beyond text, images, voice and files

Through the whole previous stretch I was fine-tuning one single thing: how to talk to the chat well in writing. But the day I discovered I didn't have to type, the tool suddenly became much bigger. I want you to reach that point now, without taking as long as I did to realise the text box was only one of the doors.

This step closes Level 1. It assumes you already know what a chat is and that you already know how to ask it for things; what it adds is that you can ask in several ways.

The chat stopped being a text box

I had a very specific image lodged in my head: the chat was a box where I wrote and it wrote back. Letters on one side, letters on the other. And for a while it was exactly that. The misunderstanding wasn't in what I saw, but in believing that was all there was.

Most of today's tools are multimodal. The word sounds technical, but it means something simple: they work with more than one kind of input and output, not just text. You can give it an image, you can speak to it out loud, you can drop it a document. And depending on the tool, it answers you in several ways too.

The underlying shift arrived in a big way in May 2024, when OpenAI unveiled GPT-4o, a model trained to process text, voice and image within the same network, without passing them through separate parts. It wasn't the first or the only one, but it marked the moment this stopped being an experiment and became part of the chat anyone opens. From there on, taking for granted that a chat is text only fell short. GPT-4o is no longer in ChatGPT today —OpenAI retired it from the chat in February 2026, replaced by the GPT-5 family—, but the door it opened is still open.

A photo the machine looks at

The first time I uploaded an image to the chat I could hardly believe what it did with it. I passed it a photo of a sign in another language and asked it to translate it for me. It read it and translated it. I hadn't written the sign's text anywhere: the machine pulled it out of the photo itself.

Since then that's part of how I use it. A screenshot with an error message and the question "what does this mean and how do I fix it?". A photo of a dish and "how do you make this?". The image of a document and "extract for me the data that appears here." In every case I do the same: I put the image down and, alongside it, I tell it in text what I want from it. The image is the material; my instruction is still what steers.

It's worth remembering what you saw in the first steps: the machine doesn't understand the photo the way you do. It processes it and estimates what's in it, with the same hits and the same firm misses you already know. A blurry image, slanted text or a small detail can throw it off just like an ambiguous text. It's not an infallible scanner; it's the same probabilistic engine, now looking at pixels instead of letters.

Talking to it and getting an answer

The second door is voice. You can talk to the AI instead of writing: you dictate your message out loud and, if the tool allows it, it answers you out loud too. For me it changed the use in the stretches when my hands are busy or I'm just tired of typing on the phone.

This is where GPT-4o really made itself felt. Before, talking to a chat was slow: the voice took several seconds to turn into an answer, and you noticed the lag. That model brought the reaction time to voice down to about 320 milliseconds on average, similar to a conversation between people. The difference between waiting three seconds and barely waiting at all is the difference between dictating to a machine and talking with something that answers at the pace of a chat.

Inside there's no new magic. In some tools your voice turns into text and the answer turns back into voice; in others, as GPT-4o did, the model processes the sound directly. Either way it does its usual thing: chaining the next likely chunk. Knowing it saves you surprises. If you speak somewhere noisy or it swallows a word, it'll answer what it thought it heard, not what you said. It's the same old care, moved to the ear.

Dropping a file on top

The third door was the one that saved me the most work: attaching files. A long PDF I didn't feel like reading whole and a "sum it up for me in ten lines." A spreadsheet and "what does this table tell me?". An endless text and "find where it talks about such-and-such." Instead of copying and pasting chunks by hand, I give it the whole document and work on it.

The gesture is identical to the image one: you attach the file and, in the same message, tell it what you want done with it. Summarise, search, compare, explain a hard paragraph. The file is the raw material; your question is still the helm.

And back comes the usual warning, which weighs especially here. The machine hasn't "verified" the document: it processes it and answers about what's in it with its usual poise, right or wrong. If you ask it to summarise a fifty-page PDF, it's worth going over it to check it hasn't invented a detail along the way. The file gives it more to work with, it doesn't give it infallibility.

The input changes; the way you talk to it doesn't

If there's one single idea I want you to take from this step, it's this: everything you learned in the previous stretch about asking for things well still holds exactly the same. You change the door you come in by —photo, voice, file instead of text—, but you don't change the craft of explaining clearly what you want.

You upload the image and tell it what for; you speak to it and ask for something specific; you attach the document and pin down what you're looking for in it. A vague message still gives a vague answer, wherever the material comes from. That's why this step doesn't make you learn anything new about how to steer the chat: it only teaches you that the chat accepts more things than you thought.

What you upload is your data

There's one caution left that I don't feel like dressing up, because it's the one that really matters. When you upload a photo, dictate something or attach a document, you're handing that content to a service that isn't yours. An invoice with your address, a screenshot with someone else's data, a confidential work document: all of that leaves your screen and enters the machine. It's worth thinking twice before dropping certain things.

There's also a more prosaic detail: not all tools have all the doors open, and some features —smoother voice, large files— tend to come with tighter limits on the free tiers and more room on the paid ones. So what you can do will depend on the one you use.

With this you close Level 1. You now know what you're talking to, you know how to talk to it, and now you know how many ways it can listen to you. What comes from here is to stop treating each conversation as a blank page and start setting yourself up a place where the chat keeps what you've been giving it. But that's already another floor of the building.

Definitions

- Multimodal: said of a chat or model that works with more than one kind of content, not just text. It understands and, depending on the case, also produces images, voice or files. - GPT-4o: model unveiled by OpenAI in May 2024, trained to process text, voice and image within the same network. It popularised fast voice conversation in the chat. OpenAI retired it from ChatGPT in February 2026. - Uploading an image: giving the chat a photo or screenshot so it works with it —reading a sign, describing a scene, pulling the data out of a photographed document—. - Voice (input and output): dictating to the chat instead of typing and, if the tool allows it, hearing its spoken answer instead of reading it. - Attaching a file: handing it a document, a spreadsheet or a long text to summarise, search or explain, without having to copy and paste its content.

Further reading

- Xataka, OpenAI presenta GPT-4o: un sorprendente modelo de IA gratuito que trabaja con voz, texto e imagen al mismo tiempo — an explanation in Spanish of what it means for a chat to be multimodal. https://www.xataka.com/robotica-e-ia/openai-quiere-volver-a-revolucionar-ia-presenta-nuevo-modelo-multimodal-que-trabaja-voz-imagenes-video - Genbeta, ChatGPT se hace más inteligente: tendrá voz para hablar con nosotros y analizar las imágenes que subimos — a plain-language round-up of voice and image inputs. https://www.genbeta.com/actualidad/chatgpt-se-hace-inteligente-tendra-voz-para-hablar-nosotros-analizar-imagenes-que-subimos - DigitalDot, Las capacidades de ChatGPT en el reconocimiento de voz e imagen — a practical description of uploading images and talking to the chat. https://www.digitaldot.es/las-capacidades-de-chatgpt-en-el-reconocimiento-de-voz-e-imagen/ - OpenAI, Retiring GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini in ChatGPT — the announcement of GPT-4o's retirement from the chat. https://openai.com/index/retiring-gpt-4o-and-older-models/