Hallucination in image AI: what it is and why it matters
Quick facts
- Definition
- A detail the AI invents in the generated image, with no match in the original photo.
- Examples
- Change in skin texture, a different tooth, a new asymmetry or a feature that did not exist.
- Cause
- The model fills gaps with learned patterns, not with facts about the patient.
- Mitigation
- Limit what the model can change and check the image before showing it.
- In HOF
- It is a reason to call the projection an illustrative simulation.
Hallucination, in AI vocabulary, is when the model produces something that looks plausible but has no basis in the input. In images, it shows up as an invented detail: a texture that changed, a different tooth, a new asymmetry, a feature the original face did not have.
Why does image AI hallucinate?
Generative AI models learn patterns from many images and fill gaps with what is statistically likely. When the task is to alter only one area, the model has to decide what to do with the rest, and it does not always get it right. Techniques such as inpainting try to restrict the change to the chosen area.
How does this affect a facial simulation?
The risk is that the projection shows a change that nobody asked for or that is not realistic. This misleads the patient and puts the professional in a difficult position. The most common signs are:
- skin that looks too smooth or different from the original;
- a subtle change of expression or gaze;
- a neighboring area altered without having been marked.
How can the problem be reduced?
Clearly delimit what may change, keep restrictive instructions and check the result before showing it. This is why the conversation with the patient treats the image as an illustrative simulation, and why the acceptance in informed consent includes the limit of the tool.
If something looks off, generate again or adjust the marking before showing it to the patient.
Also known as: generation error, AI artifact.
Frequently asked questions
Can Harmy's projection hallucinate?
Any AI-generated image can contain details that do not match the patient. That is why the result is presented as an illustrative simulation and should be checked by the professional before it is shown.