Why does ChatGPT keep generating images instead of text when I ask it to write?

ChatGPT generates images instead of text because its multimodal routing sends your request to the image module when it detects image-associated wording — or when earlier image turns are still in the session context. Fix it by opening a fresh chat and starting your prompt with: "Respond in plain text only. Do not generate any images."

To stop it immediately, do two things: start a new chat, and put an explicit output instruction at the very top of your prompt. Both attack the actual cause, which is routing, not comprehension.

Why it happens: ChatGPT (and Gemini — the behavior is identical) decides per request which output module to use. Two signals bias that decision toward images. First, trigger verbs: "create," "generate," "make," and "design" are statistically associated with image requests, so the router treats them as image intent even when you meant writing. Second, session carry-over: once one image has been generated in a thread, that context bleeds into every subsequent turn — the model keeps assuming visual output is what the conversation is about. This carry-over compounds the longer the thread runs, which is why the problem gets worse mid-session, not better.

Fix 1 — lead with a plain-text instruction: Begin the prompt itself (not the middle, not the end) with: "RESPOND IN PLAIN TEXT ONLY. Do not create, generate, or attach any images." Placing the constraint first means the router reads it before it reads any image-flavored language in your actual request. If the model still slips, stop it and correct it explicitly rather than re-asking the same way. "I had to stop it because it started to generate the actual video. It got too excited, got ahead of itself. So I'm going to simply correct it," as one creator documenting an AI agent production workflow put it — the stop-and-correct discipline is the same for text.

Fix 2 — reset the session: A fresh chat clears the image context entirely, and it is the reliable fix once a thread has already produced images. Keep writing threads and image threads separate from the start: documented AI production workflows report losing around 20 minutes per session re-establishing context when a tool drops or misroutes it, so isolating the two intents up front is cheaper than resetting repeatedly.

Fix 3 — swap the trigger verbs: Rephrase requests to use unambiguous writing language: "Write a paragraph describing...", "Draft the text of...", "Give me prose for..." instead of "Create a..." or "Make me a...". Asking for a named text format — "a bulleted outline," "a 200-word description," "dialogue only" — removes the ambiguity the router is reacting to.

The same three fixes work in Gemini, where users report the identical complaint on Google's own support forums: reset the session, then open with the plain-text instruction. This matters in practice if you draft scripts in ChatGPT before production — one documented workflow develops a story outline and a treatment outline in ChatGPT, then uploads both to the invideo agent for video production. Prepending the plain-text instruction to every drafting prompt keeps those documents coming back as text instead of unwanted concept art.

I had to stop it because it started to generate the actual video. It got too excited, got ahead of itself. So I'm going to simply correct it.

— a creator documenting an AI agent production workflow