OpenAI Releases 'Deepfake' Detector

OpenAI releases a new 'deepfake' detector that identifies 98.8 percent of images created by DALL-E 3. The tool is being shared with researchers to test in real-world situations, as the company also works on other methods to combat AI-generated misleading content.

Share
OpenAI Releases 'Deepfake' Detector

OpenAI, a prominent AI company, is taking part in a cross-industry initiative to identify content created with artificial intelligence. This comes amid warnings that AI-generated images, audio, and video could influence this fall's elections.

A New Tool for Detecting AI-Generated Images

OpenAI has launched a new detector that can identify 98.8 percent of images created by DALL-E 3, its own image generator. This tool is now being shared with a small group of disinformation researchers to test it in real-world situations and contribute to further development.

"This is to kickstart new research," says Sandhini Agarwal, a researcher at OpenAI focused on safety and policy. The detector, however, isn't designed to detect images created by other popular generators like Midjourney and Stability.

More Measures to Combat Deepfakes

Since the detector works with probabilities, it can never be entirely perfect. That's why OpenAI, like many other companies, nonprofits, and academic labs, is working to combat the problem on multiple fronts.

The company is joining the steering committee of the Coalition for Content Provenance and Authenticity (C2PA), an initiative to develop digital content credentials. The C2PA standard functions as a kind of "nutrition label" for digital files, showing when and how they were produced or modified, including with AI.

Watermarking AI-Generated Content

OpenAI is also developing methods to "watermark" AI-generated audio so it can be easily identified in real time. These watermarks are designed to be difficult to remove.

Growing Pressure on the AI Industry

With major companies like OpenAI, Google, and Meta leading the way, the AI industry faces increasing demands to account for the content its products create. Experts urge the industry to prevent users from creating misleading and harmful material, and to offer ways to trace the origin and spread of such content.

Greater Demand for Monitoring AI Content

In a year filled with major elections around the world, demand is growing for methods to monitor the provenance of AI content. In recent months, audio and images have already influenced political campaigns and elections in countries like Slovakia, Taiwan, and India.

OpenAI's new deepfake detector could help reduce the problem, but it's no silver bullet. As Sandhini Agarwal put it: "There's no silver bullet" in the fight against deepfakes.