OpenAI, known for its advancements in generative AI, has revealed a new deepfake detection tool designed to tackle the surge of misleading content on social media. Presented at the Wall Street Journal’s Tech Live conference in Laguna Beach, California, OpenAI’s Chief Technology Officer, Mira Murati, unveiled the tool claiming it can reliably identify AI-produced images with 99% accuracy.
The ubiquity of AI-generated imagery ranges from harmless caricatures like Pope Francis donning a trendy Balenciaga coat to malicious content capable of causing significant financial harm. Consequently, distinguishing authentic content from AI-generated manipulations has become increasingly challenging as AI tools grow more sophisticated.
While the specific release date of OpenAI’s new tool remains undisclosed, its announcement has sparked considerable interest, particularly considering OpenAI’s previous setbacks.
In January 2022, the company introduced a text classifier intended to discern human-generated writing from machine-generated text, like that produced by models such as ChatGPT. However, by July of the same year, OpenAI discreetly terminated the tool, citing an unacceptably high error rate. The classifier mistakenly identified genuine human writing as AI-generated in 9% of cases.
Acknowledging the limitations of current automated methods in detecting AI-generated images, enthusiasts have primarily relied on intuition, emphasizing common challenges that impede generative AI, such as rendering accurate depictions of hands, teeth, and intricate patterns. Moreover, distinguishing between AI-generated and AI-edited images remains challenging, particularly when attempting to use AI to detect AI.
OpenAI not only focuses on detecting harmful AI images but also implements measures to regulate its own models, surpassing publicly stated content guidelines. Decrypt’s findings revealed that OpenAI’s Dall-E tool seemingly alters prompts without notification and intermittently generates errors when tasked with creating specific outputs, even if they adhere to published guidelines and avoid generating sensitive content related to particular names, artistic styles, and ethnicities.
While OpenAI is actively involved in deepfake detection, other entities like DeepMedia are also developing similar capabilities tailored to governmental clients. Established tech giants such as Microsoft and Adobe have joined the effort, introducing an ‘AI watermarking’ system through the Coalition for Content Provenance and Authenticity (C2PA). This system embeds a distinctive “cr” symbol within a speech bubble, indicating AI-generated content, and aims to promote transparency, enabling users to discern the content’s origin.
Despite these advancements, loopholes exist where metadata containing the symbol can be removed. To address this, Adobe has developed a cloud service capable of recovering lost metadata, ensuring the symbol remains intact. However, even these measures are not entirely foolproof.
As the conversation around criminalizing deepfakes gains traction, these innovations underscore the need for both technological progress and societal responsibility. Implementing effective strategies to ensure authenticity in the digital sphere necessitates the active involvement of tech corporations, lawmakers, content creators, social media platforms, and end-users. In the interim, human judgment and vigilance remain crucial in combating AI misuse, although humans themselves are not immune to errors. Long-term solutions will demand collaborative efforts from various stakeholders to navigate this complex terrain presented by evolving generative AI technology.
