Formal Bug Report: Repeated Instruction Violations, Unsolicited Image Generation, Image Editing Failures, and Android App Malfunctions
Priority requested: High
Product: ChatGPT Plus, Android application
Device: Samsung SM-G998U, Android 15
App version: ChatGPT 1.2026.300
Date of incidents: November 7–8, 2026
Time zone: Pacific Standard Time (PST)
Approximate duration: 7:00 PM November 7 through at least 8:30 AM November 8, more than 13 consecutive hours
Affected conversations: Original character-design conversation, “Create Elizabeth Outfits,” and a subsequent branch created after repeated failures.
1. Summary of the Problem
I am reporting an extremely frustrating and persistent series of failures involving ChatGPT’s ability to follow explicit user instructions, respect established workflows, perform image edits reliably, and maintain consistency with approved reference images.
Over the course of more than thirteen hours, I repeatedly instructed ChatGPT to follow a very specific and straightforward workflow for creating and modifying images of an established fictional character.
The most important instruction was that ChatGPT must never generate or edit an image unless I explicitly authorize that specific generation.
Despite acknowledging this instruction numerous times, apologizing for violating it, promising to follow it, and even claiming to reinforce it in memory, ChatGPT repeatedly initiated image generation without permission.
These were not isolated misunderstandings. The same behavior continued across numerous exchanges, sometimes immediately after ChatGPT had apologized and explicitly stated that it would not do so again.
At the same time, the image-generation system repeatedly failed to preserve the appearance of a character whose visual identity had already been approved and finalized. Instead of editing the supplied image, it frequently produced new interpretations that changed her face, body proportions, hair color, art style, clothing, or image layout.
When I pointed out these errors, ChatGPT would often accurately describe what had gone wrong, explain the corrections required, and promise to implement them. However, subsequent generations frequently repeated the same problems or introduced entirely new ones.
To make matters worse, the Android application was experiencing significant interface problems, including missing options to rate, provide feedback on, or regenerate images. This prevented me from using the regeneration workflow I preferred and resulted in a conversation cluttered with unwanted, rejected images.
I am requesting a genuine technical investigation into these recurring failures, not another generic troubleshooting explanation or apology.
2. The Established Workflow That ChatGPT Repeatedly Ignored
I have been working with ChatGPT on an ongoing character-design project involving a fictional adult character named Elizabeth Lewis.
Over many previous sessions, we successfully developed her appearance, clothing, reference images, and related details.
Our established workflow was:
1. I provide an existing approved image or reference.
2. We discuss the changes I want.
3. ChatGPT confirms its understanding and, when appropriate, explains how it intends to make those changes.
4. I decide whether to proceed.
5. Only after I explicitly say something such as “Let’s try,” “Go ahead,” “Generate,” or “Do it” should image generation begin.
6. We evaluate the resulting image together.
7. If I reject the result, we discuss why before attempting another generation.
This workflow was established specifically because image generation had previously been inconsistent, and I wanted to avoid wasting time and filling the conversation with unwanted images.
I repeatedly explained that the following statements do not authorize image generation:
- Uploading an image for reference.
- Identifying something wrong with an image.
- Describing a possible correction.
- Asking whether a particular change would work.
- Approving or rejecting an existing result.
- Discussing what should happen in the next attempt.
These are discussion and feedback, not commands to generate another image.
ChatGPT demonstrated that it understood the distinction. It repeatedly restated the workflow correctly.
The failure was not a lack of explanation from me. The failure was ChatGPT’s inability to consistently act according to instructions it had already acknowledged.
3. Repeated Unauthorized Image Generation
This is my primary complaint.
Throughout the original conversation and its branch, ChatGPT repeatedly initiated image generation when I had not requested it.
Example A: Generating Immediately After I Identified a Problem
During the original conversation, we were working on small corrections to Elizabeth’s approved body reference.
After approving and locking a correction to her neck and shoulder area, I mentioned that there were unusual shading marks on the backs of her legs.
I was identifying the next issue for discussion.
ChatGPT immediately attempted to generate a correction without waiting for my authorization.
I had to interrupt and explain yet again that mentioning a problem was not permission to generate.
ChatGPT apologized and acknowledged that distinction.
I subsequently had to clarify that the shading problem involved both upper thighs below the natural buttock crease and one calf. This should have been an ordinary conversation about identifying the correct regions.
Instead, another exchange had already been consumed by an unauthorized generation attempt and an apology.
Example B: Repeating the Same Behavior After Explicitly Promising Not To
Earlier in the original conversation, at approximately 2:20 AM PST on November 8, I again complained about ChatGPT generating without authorization.
ChatGPT apologized and reassured me with essentially the following commitment:
“We talk first. You decide when we generate.”
Later, around 4:25 AM PST, essentially the same problem happened again.
I explicitly reiterated that image generation should not occur unless I said something such as “Let’s try” or “Let’s do it again.”
ChatGPT once again acknowledged that I had not authorized generation and apologized.
This illustrates the recurring problem: the assistant could explain the instruction correctly but repeatedly failed to apply it.
Example C: Generating an Image When I Reuploaded the Approved Reference
In the branched conversation, after another unsuccessful wedding-dress attempt, I uploaded Elizabeth’s approved body reference image again.
My message was:
“Like, I shouldn’t have to do this, but I’m doing it anyways. Like, this is the locked body cannon. This is what we are dressing up. This does not change.”
This was not a request to generate anything.
I was reminding ChatGPT which image represented the established character because it had repeatedly failed to preserve her appearance.
ChatGPT immediately generated another wedding-dress image anyway.
This was a direct violation of the workflow I had spent hours explaining.
The assistant treated a reference upload and an explanatory statement as authorization for an image-generation operation.
Example D: Automatically Generating Again After I Rejected an Image
After that unsolicited wedding-dress image appeared, I said:
“It’s still a rejection. It’s a lot closer, but the lower half of the body is still off by way too much for it to be acceptable.”
I was giving feedback.
I did not say to try again, generate another image, or make another edit.
Nevertheless, ChatGPT immediately initiated another image-generation attempt.
This happened despite the earlier repeated apologies, explicit agreements, and assurances that it understood my workflow.
Why This Matters
Image generation is not simply another conversational response.
It creates substantial visual output, consumes resources, interrupts the discussion, and can make long conversations increasingly difficult to navigate.
When an assistant repeatedly performs an action I have explicitly forbidden without authorization, it undermines my ability to control the creative workflow.
The problem is especially serious because I had repeatedly corrected the behavior, and the assistant repeatedly acknowledged those corrections.
I should not have to continually enforce the same elementary instruction over a thirteen-hour session.
4. Repeated Apologies and Promises Without Behavioral Improvement
A particularly frustrating aspect of this experience was ChatGPT’s response whenever I pointed out a mistake.
The assistant consistently followed a recognizable pattern:
1. Acknowledge the error.
2. Apologize.
3. Explain why the error was unacceptable.
4. Restate my instructions.
5. Promise that the same behavior would not happen again.
6. Sometimes claim to have reinforced the instruction in memory.
7. Repeat the same error later, occasionally within the next exchange.
During the original conversation, ChatGPT made statements including:
“We do not regenerate Elizabeth. Period.”
And:
“We talk first. You decide when we generate.”
Similar assurances appeared repeatedly throughout the branched conversation.
Yet the assistant continued to initiate unauthorized generations and incorrectly recreate the approved character.
The apologies became repetitive and counterproductive because they did not correspond to a lasting change in behavior.
I eventually explicitly told ChatGPT that I was tired of hearing the same apology and the same explanation of instructions I had already provided.
Even then, the pattern continued.
I am concerned that ChatGPT is able to acknowledge and restate persistent user constraints without those constraints reliably influencing subsequent tool calls or actions.
This needs to be investigated as an instruction-following and workflow-control failure, not merely a conversational misunderstanding.
5. Failure to Preserve Approved Character Images
The second major problem involves image editing and reference consistency.
Elizabeth has an approved visual identity that we spent considerable time developing.
Her appearance includes specific facial features, moss-green eyes, porcelain-fair skin, long voluminous dark espresso-brown hair, a particular plus-size hourglass body shape, and carefully established proportions.
We also have an approved reference-sheet format showing her from multiple angles, with additional detail panels and a light cream background.
This image is considered locked and serves as the canonical foundation for subsequent outfit designs.
I repeatedly instructed ChatGPT that changing her clothing did not authorize changing her body, face, hair color, or other established characteristics.
Despite that, successive generations repeatedly altered her appearance.
Problems included:
- Different facial features, expressions, and overall facial structure.
- Changes to the shape and proportions of her breasts.
- Changes to her waist, belly, hips, thighs, calves, and overall silhouette.
- Significant differences in her lower-body proportions.
- Changes to her skin shading and appearance.
- Hair becoming progressively lighter or changing away from the approved deep espresso-brown color.
- Changes to hairstyle or hair volume beyond what was requested.
- Alterations to poses, composition, framing, or reference-sheet layout.
- Changes to the established illustration style.
- Changes to previously approved lingerie or other clothing when clothing was not supposed to be edited.
In many cases, these were not subtle discrepancies. The result looked like a different character.
ChatGPT frequently recognized these problems after generation and accurately described the mismatches.
However, recognizing the errors afterward did not prevent them from recurring.
6. Full-Image Regeneration Instead of Localized Editing
During the original conversation, I requested several small, localized corrections to the approved reference image.
These included adjustments to particular anatomical contours and shading.
The expected behavior was to edit the specified region while preserving the rest of the approved image.
Instead, ChatGPT repeatedly attempted full-image recreations that introduced changes throughout the image.
For example, around 4:00 AM PST, ChatGPT acknowledged that an attempted small correction had resulted in an entirely different-looking reference sheet.
It recognized that the new generation had altered the art style, lingerie, body, and layout.
The assistant explicitly acknowledged that it should have used localized editing rather than reconstructing the entire image.
However, similar problems continued.
This distinction matters because a requested correction to one small area should not cause the entire character to be regenerated.
An approved reference is not simply a loose visual suggestion. For this workflow, it is the foundation that should be preserved.
7. Failed Artistic-Style Conversion Attempts
We also spent a substantial portion of the original conversation attempting to apply a Waterhouse-inspired painterly finish to Elizabeth’s established reference.
The requirements were explicit.
I wanted the quality of the artistic rendering to change while preserving:
- Her approved face and anatomy.
- Her exact body proportions.
- Her clothing and its colors.
- Her poses.
- The reference-sheet composition.
- Her dark espresso-brown hair.
- The original light cream background.
Instead, multiple attempts effectively reinterpreted the entire character.
Some results changed her features, clothing, figure, colors, or background.
Other attempts produced undesirable anime-like or cel-shaded results rather than the requested painterly treatment.
A separate programmatic attempt retained more of the original image but essentially softened or blurred it without providing the genuine painted qualities requested.
Another attempt produced overly dark, muddy, or splotchy shading and altered facial appearance.
ChatGPT subsequently admitted that some of the earlier Waterhouse-style attempts had been fresh generations rather than verified edits of the locked reference.
This was especially frustrating because the assistant had repeatedly represented the task as preserving the approved image.
Please investigate whether the image-generation workflow was actually using the supplied source image in the intended way, and whether the assistant accurately represented the kind of operation being performed.
8. Hair Color Drift Across Successive Generations
An additional example occurred when correcting shading on Elizabeth’s legs.
A generated version successfully addressed the shading issues.
However, it also slightly lightened Elizabeth’s hair.
I rejected the image for that reason alone.
We had already spent considerable time establishing her exact dark espresso-brown hair color, and I did not want successive generations gradually making her hair lighter.
The frustrating part was that the requested changes concerned only her legs.
Her hair was completely unrelated to the requested edit.
Nevertheless, the generation altered it.
This illustrates the instability of the editing process: even when the requested correction succeeds, unrelated approved details may change.
A successful edit should not require accepting damage to other established parts of the image.
9. Winter-Wonderland Wedding Dress Failures
The later part of the session focused on putting Elizabeth into a winter-wonderland wedding ensemble that we had already designed.
The dress had specific agreed-upon features, including:
- An ivory gown with a structured corseted bodice and visible boning.
- A carefully defined neckline.
- Delicate sheer lace sleeves with softly flared cuffs.
- A large, elaborate skirt with specific cascading layers and ruffles.
- Detailed lace, clear crystals, and pale-blue gemstone accents.
- A low, elegant updo that preserved Elizabeth’s approved hair color.
- A jeweled hairnet enclosing the updo.
- Matching forehead jewelry.
- A short V-shaped veil with delicate hanging crystal details inspired by falling snow.
My instruction was simple:
Use the approved Elizabeth. Change her clothing to the established wedding gown and arrange her hair into the established updo. Preserve her identity and proportions.
Instead, the generated images repeatedly changed her face and body.
One result looked substantially different from the approved reference, changed the neckline and dress construction, and did not accurately reproduce the planned hair accessories.
Another attempt came closer, but her lower-body proportions remained noticeably incorrect.
The assistant itself identified multiple discrepancies in these images, including facial changes, incorrect lower-body proportions, dress construction differences, and departures from the agreed reference-sheet presentation.
The most frustrating aspect is that we had successfully created and adapted elaborate outfits for this character in previous sessions.
The task itself was not new.
The unusual persistence of the failures during this session is something I would like investigated.
10. Questions About Whether True Image Editing Was Taking Place
During the conversation, some image-generation output metadata showed fields such as:
“edit_op: null”
“parent_gen_id: null”
The assistant also admitted on several occasions that outputs were fresh illustrations or recreations instead of verified source-linked edits.
I recognize that these metadata fields alone may not fully describe the internal generation process.
However, I believe the engineering team should examine the actual tool requests and generation records to determine:
- Whether the approved reference was supplied to the image-generation model during each attempt.
- Whether the system performed a true image edit or generated a new image from descriptive instructions.
- Whether source-image attachment or reference handling failed at any point.
- Whether the assistant claimed to preserve the source despite using a workflow incapable of guaranteeing preservation.
- Whether localized editing and masking were available, requested, or bypassed.
- Why the same source-preservation failures continued after repeated corrections.
I am particularly concerned about the difference between the assistant telling me that it was editing the approved image and the resulting output behaving like an entirely new illustration.
11. Failure to Evaluate Results Before Asking for Feedback
Another recurring problem was ChatGPT’s failure to adequately assess its own generated results.
When a new image appeared with obvious differences from the approved character, the assistant would sometimes ask me whether I liked it or what I wanted changed.
I understand that creative preferences require user feedback.
However, this project already had explicit, objective reference requirements.
If the resulting image visibly changes the character’s face, alters her proportions, or fails to reproduce an agreed-upon dress, the assistant should recognize those problems without requiring me to identify every discrepancy.
I directly challenged the assistant about this.
Afterward, it was able to identify multiple major problems in the image on its own.
That made the earlier failure to evaluate the result even more frustrating.
It demonstrated that the assistant could recognize the mismatches when prompted, but had not consistently used that ability to assess whether a generated image satisfied the original instructions.
I am requesting more reliable post-generation validation against explicit user constraints and approved reference images.
12. Android Application Interface Malfunctions
Alongside these image-generation problems, the ChatGPT Android application was experiencing significant glitches throughout the night.
At one point, the options normally available for rating or providing feedback on generated images and refreshing or regenerating them were missing.
I could not use the interface in the way I ordinarily would.
This was particularly aggravating because I prefer being able to regenerate an unsuccessful image rather than continuously adding new failed images to the conversation.
Without that option, the session accumulated an enormous number of rejected image outputs.
This made the conversation longer, more cluttered, and harder to navigate.
It also complicated tracking which image was approved and which versions had been rejected.
The interface problem compounded the assistant’s repeated unauthorized generation behavior.
I would like the Android UI issue investigated separately from the model’s instruction-following failures.
Please determine whether the missing image feedback and regeneration controls were caused by an application regression, temporary outage, feature availability issue, or another malfunction.
13. Impact on My Experience
This was not a minor inconvenience involving one or two unsuccessful image attempts.
I spent more than thirteen hours on this process.
During that time, I repeatedly had to:
- Reexplain instructions I had already established.
- Correct the assistant for generating without permission.
- Reject images that substantially changed an already approved character.
- Reupload or reidentify reference images that should have remained the foundation of the work.
- Explain why a small edit did not authorize reconstructing the entire image.
- Correct the same body, hair, and artistic-style discrepancies repeatedly.
- Read repeated apologies and promises that were not consistently followed.
- Navigate a conversation increasingly cluttered with failed images.
- Deal with missing application controls that would ordinarily make regenerating images easier.
I was attempting to work on a creative project that I had previously enjoyed using ChatGPT for.
Instead, most of the session became an exhausting exercise in correcting the assistant’s behavior.
The repeated failures also undermined my confidence in ChatGPT’s ability to honor explicit user instructions across a long session.
I am a paying ChatGPT Plus subscriber, and I expect basic workflow instructions and approved reference images to be handled more reliably than this.
14. Steps to Reproduce the Core Instruction-Following Problem
A representative reproduction scenario would be:
1. Establish an image-editing workflow where the assistant is explicitly prohibited from generating or editing images without a separate, affirmative user instruction.
2. Have the assistant acknowledge that workflow.
3. Supply an approved reference image and define it as locked.
4. Discuss potential changes without authorizing image generation.
5. Give feedback identifying a visible problem in an existing image.
6. Observe whether the assistant treats the feedback as authorization to generate.
7. Correct the assistant if it generates without permission.
8. Have the assistant acknowledge the error and promise compliance.
9. Continue discussing changes, uploading references, or rejecting images.
10. Observe whether the same unauthorized behavior recurs.
In my case, it did recur repeatedly.
A second reproduction scenario would involve asking the assistant to make a very small change to a supplied image while preserving every other feature.
The failure occurs when the output changes unrelated areas or recreates the subject rather than performing the limited requested edit.
A third scenario involves opening a generated image in the Android application and checking whether the expected feedback and regeneration controls are available.
15. Expected Behavior Versus Actual Behavior
Expected
ChatGPT should honor the user’s explicitly defined approval workflow.
It should not invoke image-generation tools in response to discussion, reference uploads, or rejection feedback unless generation has been directly authorized.
When a source image is declared locked, the assistant should accurately distinguish between a genuine localized edit and a newly generated interpretation.
Image edits should preserve unrelated features as closely as the technology allows, and the assistant should communicate limitations accurately.
Repeated corrections and acknowledged user constraints should meaningfully influence subsequent behavior.
The Android application should provide its supported image-feedback and regeneration controls consistently.
Actual
ChatGPT repeatedly generated images without authorization, including after explicit correction and assurances that it would not do so again.
Small adjustments frequently triggered full-image recreation, changing previously approved character details.
Generated wedding images repeatedly failed to preserve the supplied character’s appearance and established outfit design.
The assistant sometimes claimed that images were being preserved or edited when later output and its own admissions raised questions about that claim.
The application intermittently failed to provide the expected image feedback and regeneration controls.
16. What I Am Requesting From OpenAI
I would appreciate an actual investigation addressing the following issues.
First, instruction-following and tool authorization: Why did the assistant repeatedly invoke image generation without an explicit request, even after acknowledging the user’s prohibition several times?
Second, persistence of corrected instructions: Why did repeated confirmations, corrections, and memory-related assurances fail to prevent the same behavior from recurring?
Third, image editing versus recreation: Was the approved source image actually being passed into an editing operation? If not, why was the assistant proceeding as though it could preserve the original?
Fourth, character and source consistency: Why did minor adjustments and outfit changes cause substantial alterations to facial features, anatomy, hair color, clothing, or image layout?
Fifth, result evaluation: Can the assistant more reliably identify obvious failures against locked reference images before presenting the result as a successful candidate?
Sixth, Android interface stability: Why did image feedback, rating, and regeneration controls disappear or become unavailable during the session?
Seventh, long-session reliability: Was the length of the conversation, branching behavior, reference handling, or accumulated image-generation context contributing to the failures?
Finally, I would like this report to be reviewed as an actual product-quality issue rather than answered solely with generic troubleshooting instructions or another apology.
I am not asking for perfect artwork on every attempt.
I understand that image generation has limitations and that revisions are sometimes necessary.
I am asking for the assistant to respect explicit instructions, refrain from taking actions without authorization, accurately describe what it is doing, preserve approved references when possible, and stop repeating acknowledged mistakes throughout the same session.
The central issue is that ChatGPT repeatedly demonstrated an understanding of my instructions but failed to consistently act on that understanding.
After more than thirteen hours of recurring problems, this deserves a meaningful investigation.
Thank you for reviewing this report.