This week, OpenAI introduced a new image generation feature in ChatGPT, which gained widespread attention for its capability to produce images in the style of Studio Ghibli. Beyond these pastel-themed illustrations, GPT-4o’s built-in image generator significantly enhances ChatGPT’s functionality, offering improved picture editing, text rendering, and spatial representation abilities.
A significant development from OpenAI this week is the revision of its content moderation policies, which now permit ChatGPT to generate, upon request, images of public figures, symbols associated with hate, and racial features. Previously, OpenAI had declined such prompts due to their controversial or harmful nature. However, the company has since adjusted its stance. According to a blog post by Joanne Jang, OpenAI’s model behavior lead, the company is transitioning from absolute refusals in sensitive domains to a more targeted approach aimed at preventing real-world harm. Jang emphasized the importance of humility and the need to adapt as the company gains greater understanding.
These changes appear to be part of OpenAI’s broader strategy to “uncensor” ChatGPT. In February, OpenAI shared its plans to modify how its AI models are trained, aiming to enable ChatGPT to handle a wider range of requests, present diverse perspectives, and increase the number of topics it can engage with.
Under the revised policy, ChatGPT can now generate and modify images of public personalities like Donald Trump and Elon Musk, which were previously restricted. OpenAI aims to avoid acting as a gatekeeper in deciding which figures should or should not be depicted by ChatGPT. Instead, users are given the option to opt-out if they prefer not to have their likeness represented.
A white paper released by OpenAI on Tuesday noted that ChatGPT will also allow the generation of hateful symbols, such as swastikas, as long as they are used in educational or neutral contexts and do not promote or endorse extremist agendas.
Additionally, OpenAI is redefining what constitutes “offensive” content. Previously, ChatGPT would reject requests related to physical characteristics, such as altering someone’s eye shape to appear more Asian or modifying someone’s weight. However, testing by TechCrunch revealed that ChatGPT’s updated image generator can now accommodate these types of requests.
While ChatGPT can emulate the styles of creative studios like Pixar or Studio Ghibli, it still limits replicating the styles of living individual artists. This issue ties back to an ongoing debate about the fair use of copyrighted materials in AI training datasets.
OpenAI is cautious about widespread misuse; GPT-4o’s image generator continues to decline many sensitive queries and includes more robust safeguards around generating images of minors compared to its predecessor, DALL-E 3, according to the GPT-4o white paper.
OpenAI’s relaxation of certain content moderation constraints comes amid critiques of alleged AI “censorship” from conservatives and tech companies in Silicon Valley. Google faced criticism for its Gemini AI image generator, which inaccurately depicted historical figures in multiracial contexts.
The contentious issue of AI content moderation continues to escalate. Earlier this month, Republican Congressman Jim Jordan questioned OpenAI, Google, and other technology giants about potential coordination with the Biden administration to censor AI-generated content.
OpenAI refuted claims that the content moderation updates are driven by political motives in a prior statement to TechCrunch. Instead, the company states these changes reflect a long-standing commitment to increase user control, with advancements in technology now enabling more nuanced handling of sensitive matters.
Regardless of the reasoning behind the changes, OpenAI’s timing aligns with potential regulatory scrutiny during the Trump administration. Other Silicon Valley companies like Meta and X have similarly adjusted their policies to permit more controversial topics on their platforms.
While the new image generator from OpenAI has resulted in some viral Studio Ghibli-themed memes, the broader impact of these policy changes remains uncertain. Though the adjustments might sit well with the Trump administration, allowing an AI bot to engage with sensitive inquiries could introduce challenges for OpenAI in the future.
