Deciphering OpenAI's Plans: The DALL-E 3 Image Classifier Contemplation

• OpenAI
• DALL·E

OpenAI, a leading player in artificial intelligence (AI) research, has been engrossed in extensive discussions and debates about when to release its AI-generated image detector that can discern whether an image was created using DALL-E 3, its generative AI art model. The decision, however, seems to be on hold for the time being.

"There is no AI without IA (Information Architecture)" - Peter Morville

Sandhini Agarwal, an OpenAI researcher specializing in safety and policy, shared these insights in a recent phone interview. According to Agarwal, the classifier tool's accuracy is commendable but has yet to reach OpenAI's high-quality standards.

What's holding back OpenAI?

The hesitance in releasing the tool may be due to the potential ramifications of its decisions on photographs, such as categorizing a piece of work as either artistically painted or deceptively inauthentic. Agarwal underlines the importance of the tool's reliability in making these significant determinations.

The company's targeted accuracy for the tool is extraordinarily high. Mira Murati, OpenAI's chief technology officer, disclosed that the classifier is nearly 99% reliable at discerning if an unmodified photo was generated using DALL-E 3. Whether the ultimate goal is 100% accuracy, Agarwal chose not to divulge.

A draft blog post from OpenAI revealed that the classifier maintains over 95% accuracy when an image has been subjected to common types of modifications, such as cropping, resizing, JPEG compression, or when text or cutouts from authentic images are superimposed onto small portions of the generated image.

Previous Controversies and Philosophical Questions

OpenAI's reluctance could also be attributed to the controversy surrounding its previous public classifier tool designed to detect AI-generated text not only from OpenAI's models but also from third-party text-generating models. The device was eventually withdrawn due to its low accuracy rate, which attracted widespread criticism.

Apart from technical considerations, OpenAI is also grappling with philosophical questions about what entails an AI-generated image. Artwork generated from scratch by DALL-E 3 is undeniably AI-generated. Still, the classification of an idea that has undergone multiple edits, combined with other images, and subjected to post-processing filters must be clarified.

Agarwal articulated this dilemma, "At that point, should that image be considered AI-generated or not?" OpenAI strives to navigate this complex question and is keen to hear from artists and others whom such classifier tools would significantly impact.

Deepfakes and the Search for Watermarking and Detection Techniques

Many organizations, including OpenAI, are exploring watermarking and detection techniques for generative media in response to the proliferation of AI deepfakes. DeepMind, for instance, recently proposed a specification, SynthID, to mark AI-generated images in a way invisible to the human eye but can be spotted by a specialized detector.

Other companies are developing similar tools. French startup Imatag, launched in 2020, offers a watermarking tool that it claims isn't affected by resizing, cropping, editing, or compressing images, much like SynthID. Another firm, Steg.AI, uses an AI model to apply watermarks that withstand resizing and other edits.

However, the AI industry has not yet unified around a single watermarking or detection standard. Further, there needs to be a guarantee that the watermarking techniques and detectors will be foolproof.

OpenAI's Future Plans for the Image Classifier

When asked whether the OpenAI image classifier would ever support detecting images created with other non-OpenAI generative tools, Agarwal remained non-committal. However, she acknowledged that depending on the reception of the image classifier tool as it currently exists, it's a possibility OpenAI would consider exploring.

Agarwal explained, "One of the reasons right now DALL-E 3-specific is because that's, technically, a much more tractable problem." She added that while a general detector isn't something they're working on at the moment, depending on where the classifier tool goes, she isn't ruling out the possibility.