how do you spot a DALL-E?

DALL-E was OpenAI's diffusion-based image generator and the engine behind every ChatGPT image generation from October 2023 until March 2025, when GPT Image 1 quietly replaced it. across that ~17-month window it produced an enormous share of all AI-generated images on the consumer internet. first shipped as a research curiosity (DALL-E 1, January 2021), then as a consumer product (DALL-E 2 and 3); the API snapshots were formally removed on May 12, 2026.

despite the deprecation, DALL-E images remain everywhere on the internet. the social-media long tail will keep surfacing DALL-E content for years.

what a DALL-E image looks like

DALL-E 2 has the cleanest “tells” of any production-era model. extreme symmetry, plastic skin, characteristic “smushed” hands, broken in-image text. a visible color-stripe watermark in the bottom-right (yellow / green / red / blue / purple) that's removable by cropping and so unreliable on its own but a giveaway when present.

DALL-E 3 fixed text and faces but kept a recognizable warm, slightly oversaturated palette and a “centered subject on simple background” composition bias. it also benefits from GPT-4 rewriting your prompt under the hood, and the outputs are more prompt-faithful than DALL-E 2 partly because the model is better and partly because the prompt is.

how amige. detects DALL-E

the primary signals are the same diffusion fingerprints amige. looks for on every diffusion model: frequency-domain artifacts from upsampling layers, characteristic color-channel correlations, and the model's specific JPEG re-encoding pipeline. detectors trained on DALL-E 2 and DALL-E 3 outputs treat them as distinct classes with their own spectral signatures, which is why amige.'s model attribution can come back tagged “DALL-E 3” rather than a generic “diffusion model.”

the secondary signal, when present, is the explicit color-bar watermark on DALL-E 2 output. it's easily cropped away, but when it's there it's near-certain DALL-E origin.

the difficulty curve

detection gets easier over time, because the model is frozen. DALL-E 2 outputs are reliably detectable. DALL-E 3 is harder but well-fingerprinted by every commercial detector in the category. GPT Image 1 broke many legacy DALL-E classifiers in March 2025 (different architecture, different signature) and re-opened the gap until vendors retrained.

for an amige. user: a confident DALL-E flag on supposedly recent content is a strong “this was made earlier and recycled” signal. anything new coming out of ChatGPT since March 2025 is GPT Image, not DALL-E.

the controversy

the Studio Ghibli wave of late March 2025. when GPT Image 1 launched inside ChatGPT, users mass-converted personal photos into Studio Ghibli pastiche. Hayao Miyazaki had previously called AI animation “an insult to life itself,” and commentators called the wave the loudest stylistic-IP debate since the Getty case against Stable Diffusion. no formal Ghibli lawsuit has materialized as of mid-2026, but the episode crystallized how thoroughly trademark-and-style questions had outrun the law.

DALL-E 2's color-stripe watermark was the first widely- deployed visible AI watermark, predating most C2PA discussion. easily defeated by cropping, but historically significant as the first “we shipped a watermark” statement from a major lab.

more on the architecture in what's a diffusion model.