DALL·E 3 lands in ChatGPT with a 99% accurate AI detective

DALL·E 3 is no longer a research paper. It’s a product, and as of today, it’s woven directly into ChatGPT for Plus and Enterprise subscribers. Describe a scene in plain English, and the model spits out a set of images you can refine through follow-up chats. The underlying leap isn’t just aesthetic—though OpenAI says the images are more visually striking and handle text, hands, and faces far better than DALL·E 2 ever could. The real trick was building a state-of-the-art image captioner to generate richer descriptions of training data, then training DALL·E 3 on those improved captions. The result is a model that actually pays attention to what you ask for, including specific aspect ratios.

OpenAI is acutely aware that a model this capable can be a liability. They layered on a multi-tiered safety system that checks both prompts and the resulting images for violent, adult, or hateful content before anything reaches the user. Red-teamers and early users stress-tested edge cases—sexual imagery and convincingly misleading images were flagged as particular risks. The company also took steps to block the generation of content in the style of living artists, limit images of public figures, and improve demographic representation. Creators can opt their work out of future training datasets.

Perhaps the most interesting bit of the announcement is the provenance classifier. It’s an internal tool that can identify whether an image came from DALL·E 3 with over 99% accuracy on unmodified images. Even after cropping, resizing, or JPEG compression, that figure stays above 95%. Those are impressive numbers, but OpenAI is quick to frame it as probabilistic—not definitive proof. The classifier is part of a broader exploration into how we understand what’s real when audio and visual content can be generated by anyone with a subscription.

The company is positioning this as a learning exercise. They’re leaning on user feedback via a flag icon to surface unsafe outputs or missed prompts, and they’re explicit that solving AI-generated content detection will require collaboration across the entire value chain, including the platforms that distribute images. That’s a quiet acknowledgment that a 99% accurate tool inside OpenAI’s walls doesn’t mean much if the images are already circulating on social media with no provenance information attached.