ChatGPT can incorrectly route
a purely textual request
into image generation.
This can happen even when the user
clearly asks for text only.
Observed example:
The user requested a short,
concise sales text
and a price assessment.
Instead of responding with text,
ChatGPT started a full image-generation
workflow.
The UI showed image-generation progress,
and the final result was an image
rather than the requested text output.
This is not simply a misunderstanding
of wording.
It is an intent-routing problem.
The system should distinguish reliably
between:
- writing text
- analyzing text
- estimating price
- researching information
- editing an image
- generating a new image
A request involving an object,
product or existing image
does not automatically mean
that a new image should be generated.
The most recent explicit user intent
should have priority.
Expected behavior
If the user asks for:
- sales copy
- description
- price assessment
- wording
- analysis
- comparison
the system should remain
in a text workflow
unless image generation
is explicitly requested.
Why this matters
Incorrect image generation:
- interrupts the workflow
- wastes processing time
- creates unnecessary output
- may alter context
- forces the user to repeat
an already clear instruction
This becomes especially disruptive
in professional workflows
where text and image tasks
are closely interleaved.
Acceptance criteria
-
Explicit text requests
remain text requests.
-
Mentioning or uploading an image
does not automatically trigger
image generation.
-
Image generation starts only
when the user requests
image creation or editing.
-
The system prioritizes
the latest explicit task intent.
-
A text-only instruction
cannot silently switch
into an image-generation workflow.
Hi @ralph_apfelbaum,
Could you please share the exact prompt that triggered image generation? It would be easier for others to try to reproduce it to see if it is a broader issue.
Thanks for asking.
The exact prompt I documented was:
“Please write a short, concise sales text
with a headline at the top
and a price assessment.”
The request was made inside an existing
product / sales workflow
where image material was already present.
Instead of returning text,
ChatGPT started an image-generation process
and showed:
“A more detailed image is being created…”
So the issue appears to be
that existing image context
can incorrectly influence tool routing,
even when the latest user instruction
is explicitly text-only.
I believe this happened in the same
used-product workflow around the sneaker,
but I want to separate recollection
from what I can verify directly.
What I can verify is:
- the exact text-only prompt above
- image context was already present
- image generation started anyway
I can also provide the screenshot
showing the incorrect image-generation state
if useful for reproduction.
Thanks for sharing. Was this in Chat or Work and do you have memory enabled?
If you had requested image generation earlier in that same conversation, the previous context could have influenced how the next prompt was interpreted, even though you were asking for text. Have you tried the same prompt in a new chat to see whether it happens there too?
You could also try adding “Text-only output. Do not generate an image” to make the intended output explicit and see whether that helps.
Could you please share the original screenshot showing the sales-text prompt and the response immediately after it? I can’t really see where the text-only request triggered image generation.
Thanks, Larisa.
I checked the original material again,
and I can now clarify the context more precisely.
This was not the sneaker workflow.
It happened in a Rosenthal SUOMI porcelain
sales workflow.
Earlier in the same conversation,
image generation had been used,
including a request to create
a complete-set image
from the existing Rosenthal images.
After that, I asked:
“Please write a short, concise sales text
with a headline at the top
and a price assessment.”
Instead of returning text,
ChatGPT started image generation again.
So yes:
there was previous image-generation context
in the same chat immediately before
the text-only request.
I will attach the original screenshot directly
to this reply rather than linking to it.
I have not yet repeated the same prompt
in a completely fresh chat,
so I cannot say yet
whether it reproduces without
the previous image context.
This screenshot shows the issue directly:
the immediate request was text-only, but image generation was triggered anyway.
