OpenAI is providing wider access to its newest text-to-image generator DALL-E 3, making it available within ChatGPT for Plus and Enterprise customers. This expands upon Microsoft’s recent integration into Bing.
OpenAI says it has implemented more rigorous safety mitigations to address issues that have plagued generative AI image models. These include outputting copyrighted content, problematic social biases, and misrepresentation.
The company cites research on its website into DALL-E 3‘s enhanced safety capabilities. This includes limiting mimicry of living artists’ styles, reducing inappropriate depictions of public figures, and improving representation across generated images.
Provenance Detection Tool Can Identify DALL-E 3 Images
Additionally, OpenAI has developed an internal “provenance classifier” tool claimed to be 99% accurate at detecting images created by DALL-E 3. This could help identify its content being misused online.
Past text-to-image models like DALL-E 2, Midjourney and Stable Diffusion have faced backlash over harmful imagery slip-ups despite attempted safeguards. Users have found ways to evade blocking on words and phrases.
Cautious Wider Release Follows Microsoft Integration
This wider but limited release for select users follows Microsoft rapidly incorporating DALL-E 3 into Bing search and image generation. Despite advertised protections, some concerning images slipped through.
OpenAI appears to be taking a more cautious approach by restricting access initially. However, the true test will come from monitoring how safely the technology performs at larger scale.
Generative AI’s risks remain highly challenging to eliminate. But OpenAI’s commitments and transparency around safety research are welcome steps amidst deployment. Ongoing vigilance, not complacency, is key for this powerful technology.
