OpenAI has released an enhanced version of its AI system, GPT-4o, which is capable of generating more realistic photographs. The upgrade is the result of a year-long collaboration with human trainers.
GPT-4o has replaced DALL-E 3 as the default image generation model powering OpenAI’s ChatGPT chatbot, and users of ChatGPT Free, Plus, Team, and Pro can now access it, according to the company.
Billed as a more affordable version of OpenAI’s most advanced AI model at the time, GPT-4o was first launched last year as a multimodal system capable of producing and analysing text, video, audio, and images.
OpenAI claims that the improved GPT-4o model enables both consumers and businesses to generate more realistic images, coherent paragraphs of text, commercial logos, and PowerPoint presentations with greater ease.
According to Gabriel Goh, the project’s principal researcher, the advancements in GPT-4o were made possible by a team of human trainers who annotated training data, identifying AI-generated errors such as typos, misplaced hands, and distorted faces.
This approach, known as “reinforcement learning from human feedback” (RLHF), is a widely used technique by AI companies to refine their models after initial training. Goh noted that this method allowed GPT-4o to follow human instructions more accurately, producing visuals that are both more useful and more precise.
Given the scale of OpenAI’s AI systems, the impact of these human trainers is significant. The company reports that ChatGPT has over 400 million weekly users. OpenAI says that around 100 human workers collaborated on the RLHF process for GPT-4o.
As a result of this research, OpenAI states that ChatGPT’s image generation capabilities are now much more beneficial to both individual users and businesses. For instance, GPT-4o can now generate paragraphs of understandable text alongside images—something earlier iterations of OpenAI’s models struggled to achieve.
However, AI image generators remain controversial. Some artists argue that these tools jeopardise their livelihoods by replicating elements of their original work.
OpenAI says that GPT-4o was trained using both confidential data from its collaborations with companies such as Shutterstock and “publicly available data”.
