A new study reveals that using OpenAI's ChatGPT for scriptwriting could impose major limitations on artistic creativity. Researchers from Penn Engineering, Haverford College, and Penn State identified that ChatGPT's automated content moderation often censors content, even in scripts rated PG, potentially stifling creative expression.

Is ChatGPT Stifling Creativity?

The study highlights how the guidelines established by the Writers Guild of America (WGA) and the Association of Motion Picture and Television Producers (AMPTP), which ended a recent strike, allow for the use of AI tools in the scriptwriting process but prohibit AI from being credited as a writer. 

This new research questions the effectiveness of this arrangement, revealing that ChatGPT's content moderation system restricts the generation of material that is otherwise permissible on television. The team found that ChatGPT's filters target content related to violence, sexuality, and hate speech, aiming to prevent the creation of inappropriate or harmful material.

However, the study, which analyzed scripts from IMDb's 100 most-watched TV shows, including popular titles like "Game of Thrones," "Stranger Things," and "13 Reasons Why," discovered that nearly 20% of the ChatGPT-generated scripts were flagged for content violations. 

Even more striking, the content moderation system flagged almost 70% of the actual TV show scripts, including half of the tested PG-rated ones. Danaë Metaxa and Sorelle Friedler, co-senior authors of the paper, expressed concerns about the potential impact of such stringent content moderation on storytelling.

They pointed out that if AI tools are used to create cultural content like TV scripts, there is a risk that many important stories might not be told due to excessive content filtering.

The researchers were inspired to investigate this issue during the writers' strike as they questioned whether large language models (LLMs) such as ChatGPT could generate high-quality scripts. 

Their experiments revealed that the content moderation system often prevented the AI from generating scripts at all. One example the team cited is when they prompted ChatGPT with a summary from an episode of "Game of Thrones," the chatbot refused to produce the script, citing a violation of usage policies.

Read Also: AI Chatbots Like ChatGPT Could Soon Run Out of Data That Keeps Them Smarter, Study Finds

GPT-Generated vs. Real Scripts

The researchers specifically generated a dataset of 100 popular US television shows, analyzing synopses for each episode in the first season, and compared the content moderation outcomes with 81 real scripts from the same shows.

Their findings indicated that a substantial number of both GPT-generated and real scripts were flagged as content violations. Approximately 18% of GPT-generated scripts and 69% of real scripts triggered the moderation system.

The study also found a significant correlation between TV maturity ratings, certain genres (such as animation, crime, and fantasy), and the likelihood of a script being flagged.

The researchers concluded that the self-censorship imposed by LLMs like ChatGPT could have far-reaching implications for creative industries. They emphasized the need for more transparency in these AI systems' content moderation standards and processes. 

The study's findings were presented at the 2024 ACM Conference on Fairness, Accountability, and Transparency. 

Related Article: RAGE Tool: This New Software Can Unveil ChatGPT's Sources