It’s funny easy it is to get GPT to ignore doing pictures of real people.

Understanding AI Limitations: Navigating Content Policy Restrictions in Image Generation

In the realm of creative AI applications, leveraging tools like GPT and image generation models can be both exciting and challenging. Recently, I delved into generating visual representations for a novel I previously published, aiming to create “movie stills” that bring my characters to life. The process illuminated some intriguing aspects of AI behavior, particularly regarding how these models handle content policies related to real people.

Initially, I attempted to generate headshot profiles of my characters using reference images of real actors—some of whom are legendary figures long deceased. When prompting the AI, I observed that it would sometimes produce satisfactory images, but at other times, it would decline, citing restrictions on depicting real individuals. This inconsistency was expected, given the content policies designed to prevent misuse or misrepresentation.

However, I discovered a workaround that proved surprisingly effective with GPT-based models. Instead of directly referencing real images, I would specify that the reference pictures I was using were AI-generated images created by the AI itself. Framing the prompt in this manner essentially contextualized the images as synthetic, helping the model sidestep restrictions and generate the desired visuals.

I also experimented with Nano Banana, another AI image generator, but found it significantly more cumbersome to bypass the content policies around real people. The same trick of rephrasing references didn’t work consistently, and the AI often produced unrelated or nonsensical images. Interestingly, in some cases, the images generated by GPT-based prompts—when I used the described workaround—turned out to be of higher quality or more fitting than those produced by other models.

This experience highlights an important aspect of AI content policies and their implementation. Many models are designed with safeguards to prevent generating realistic depictions of identifiable individuals, especially without consent. Yet, creative prompt engineering—such as framing references as AI-generated—can sometimes open avenues for producing desired content within these constraints.

Have you explored similar techniques or discovered other effective methods for working around these content restrictions? Sharing insights can help the community better understand how to ethically and effectively use AI tools for creative projects.

In conclusion, while AI models have built-in limitations to prevent certain types of outputs, understanding their prompting mechanisms and framing techniques can offer valuable ways to achieve creative goals within ethical and policy boundaries.

---

Disclaimer: Always consider ethical implications and copyright regulations when generating or using AI-created images, especially those depicting real individuals.