Any model · explicit constraint failures
The AI changed things I told it not to change
When an AI violates explicit preservation constraints, use an evidence-first diagnosis instead of automatically blaming the wording.
Short answer
What is probably happening
If the no-change rule was explicit, specific, compatible with the task, and present in the final instruction, the AI is responsible for violating it. You can still make the next attempt safer, but that is risk control—not an admission that the original request was bad.
Possible causes
The same symptom can have different verdicts
The key distinction is evidence. “I assumed it would keep the face the same” is an implicit expectation. “Do not change the face, expression, skin texture, hairline, or proportions” is an explicit boundary. WWW does not treat those as the same case.
The protection was too abstract
“Keep it the same” may leave uncertainty about identity, pixels, layout, tone, behavior, or meaning.
The protection conflicted with the requested transformation
A dramatic pose change, new viewpoint, or full rewrite may be impossible without changing features the instruction also freezes.
The model violated a concrete boundary
The user named the protected details and the result altered exactly those details.
The tool cannot preserve at the required fidelity
A generative editor may redraw the whole region even when instructed to change one part.
How to distinguish them
Check the evidence in this order
- 01
Quote the exact no-change sentence and list the changed details it covers.
- 02
Check whether the requested transformation logically required any protected detail to move or be regenerated.
- 03
Identify whether the tool edits original pixels, applies a patch, or synthesizes a new artifact.
- 04
Retry from the original once. Repeated drift under the same clear boundary is strong model/tool evidence.
WWW-style diagnosis
Example casewrong teammate.
- Wanted
- Remove a background without changing the photographed product.
- Asked
- Remove only the background. Do not alter the bottle, label, crop, lighting, colors, or proportions.
- Result
- The background changed, but so did the label, cap, reflections, and crop.
- Why
- The boundary was explicit. The broad, coordinated drift points to a generative editing method that redrew the subject instead of preserving its original pixels.
What to do next
Fix the diagnosed cause
Preserve from the original
Never stack a second edit on a result that already drifted. Return to the last correct source.
Choose the right mechanism
Use deterministic selection, masking, patching, or structured edits when exact preservation matters more than creative synthesis.
Verify protected details explicitly
Turn the preservation list into an acceptance checklist and compare before approving the result.
When the instruction needs repair
A boundary-first template
Change only [target]. Preserve exactly [specific protected details]. The protected elements are acceptance criteria, not style guidance. Do not regenerate, reinterpret, crop, rename, reorder, rewrite, or restyle them. If the requested change requires altering a protected element, stop and explain the conflict instead of proceeding.
This template is relevant only when the evidence points to missing or conflicting boundaries. It cannot excuse a model that violated a clear instruction.
Questions
What people usually ask next
Should I make the negative instruction longer?+
Only if it becomes more specific. A longer list cannot make an incompatible tool preserve data it normally regenerates.
Is this always the model’s fault?+
Not if the protection was ambiguous or impossible alongside the requested transformation. It is a model failure when the boundary was concrete and feasible.
Why start again from the original?+
A second edit treats the first altered result as its source, so earlier drift becomes baked into every later attempt.