OpenAI deepfake detector 'belated but welcome'

USA- OpenAI's introduction of a deepfake detector that can accurately identify images made with its Dall-E 3 generative AI system met a largely positive reception, even if it comes only five months before the U.S. presidential election.

The release of the detector earlier this week also comes as many other elections around the world are set to unfold this year and fears rise about deepfakes aimed at manipulating voter behavior.

"This is a belated but welcome move," said Kashyap Kompella, CEO and analyst at RPA2AI Research. "Detection of AI-generated content is a wicked problem. It is not a problem that any one company can solve on its own."

Election worries

Deepfakes for political deception circulated widely during the 2020 U.S. election. With the explosion of generative AI systems since then, the ease and quality of deepfakes as well as the ability to scale them have multiplied.

Meanwhile, OpenAI and other big GenAI vendors have been under intense pressure to ensure that their text, image and video-generating AI systems are safe and reliable.

Other vendors with widely used text-to-image generators include Stability.ai, Midjourney and Adobe.

The rising proliferation of deepfakes has amplified this concern. While the OpenAI image detector, which the vendor claims is 98% accurate and is now available for approved testers, addresses deepfake fears, it's unclear who will use it and whether other GenAI vendors will follow suit with similar tools.

Other OpenAI moves

In addition to the OpenAI image detector, the vendor unveiled a $2 million education fund for nonprofits working to develop AI literacy among older adults and other groups vulnerable to deepfake manipulation. OpenAI established the fund with its partner, Microsoft.

In what's been seen as an important step, OpenAI also joined the steering committee of the Coalition for Content Provenance and Authenticity (C2PA). The organization develops technical standards to address the spread of misleading information by identifying the source and provenance of media content.