I’ve discovered a fundamental limitation when using ChatGPT (even GPT-4o) for visual editing support in Photoshop or any image-based creative workflow.
Whenever the model offers help like:
- “Place an object in square G5,”
- “Draw a shadow under the subject’s arm,” or
- Mark a specific place in an Image
…it often places the element in the wrong location. For example, when I asked it to place a marker in G5, it placed it in G3 or J5 instead. This happens because the AI doesn’t understand visual layout or grid overlays the way humans do — it calculates placement based on image dimensions, not based on what we actually see on screen.
Why this matters:
This makes spatial guidance unreliable for:
- Painting/masking instructions
- Shadow/highlight placement
- Grid or guide-based compositing
- Lighting or retouching workflows
The time spent correcting or guessing AI suggestions defeats the purpose of using it to save time — and breaks trust in the tool during precision-based tasks.
Why offer a service that can’t be done reliably yet?
ChatGPT often offers to place an item on a marker for me as a guide - such as pointing to where to place shading or a shadow on a subject and always get’s it incorrect. It confidently gives spatial instructions, but without true visual anchoring or spatial awareness, those instructions end up being off — and that becomes frustrating fast for creative professionals.
I believe this is a key usability gap that’s being overlooked, and it needs urgent attention as AI tools move further into the creative space.
Suggested Solutions:
- Allow users to define a visual anchor point (e.g., “This is G5”) before offering placement-based instructions.
- Provide a way for AI to calibrate a reference grid based on user input or image metadata.
- Implement a feedback loop so the AI can ask, “Did this land in the right place?” and adjust its logic over time.
- Don’t offer precise coordinate-based placement unless calibration is confirmed — it wastes time and causes misalignment.
Supporting Documentation:
I’ve put together a short PDF with real screenshots and examples showing how this issue plays out and how frustrating it can be in practice.
Download: AI_Visual_Placement_Feedback_Richie_Final.pdf
Calling other creatives:
If you’ve also experienced this kind of misalignment in image tasks, please comment, share your examples, or upvote this post. The more of us that highlight it, the more likely it is to be taken seriously in future model iterations.
Final Thought:
AI is evolving fast — but if it’s going to become a true assistant in graphic design, compositing, and Photoshop-style editing, this kind of visual misplacement issue needs to be addressed head-on. Until then, it limits the real creative potential of what these tools can offer.
Thanks for reading — and thanks to the OpenAI team if you’re listening.
— Richie 