Researchers have introduced SpinDoctor, a method for discovering innocuous prompts for text-to-image models. The approach is designed to produce prompts that lead to images exhibiting a chosen target bias, while keeping both the prompt and the resulting image high in quality. This means the biased outputs are not obviously distorted or low-fidelity, which could make them more difficult to detect.
The authors frame this as a potential risk. Because the prompts appear innocuous and the images retain quality, such a method could be used to generate biased content in a way that evades simple screening. The work points to a need for better safeguards in text-to-image systems, though the abstract does not detail specific countermeasures.