Dall-E 2 created 3M images since April and OpenAI is adding up to 1,000 users per week, applying a content policy to try to avoid sensitive or biased images
Alex Kantrowitz / Big Technology : Tweets: @chrisbrogan , @martinsfp , @jon_gluck , @_lamaahmad , @cebsilver , and @kantrowitz Tweets: @chrisbrogan : I want access to this. THIS is cool. If THIS was on the other half of an NFT, I'd give a rat's ass. Face to Face With Dall-E, The AI Artist That Might Change The World https://bigtechnology.substack.com/ ... @martinsfp : “The demand for quality art exceeds illustrators' ability to deliver it, and Dall-E can fill the gap” https://twitter.com/... Jon Gluck / @jon_gluck : .@Kantrowitz: “If Dall-E-style images become ubiquitous, whoever controls the technology will be in a pretty influential position.” https://kantrowitz.medium.com/ ... @_lamaahmad : Chatted with @Kantrowitz yesterday about the complicated considerations around bias and representation in DALL-E 2 https://twitter.com/... @cebsilver : I can't wait til someone starts making rule 34 type requests https://twitter.com/... Alex Kantrowitz / @kantrowitz : My reflections: Dall-E is powerful communication technology. It can draw just about anything in the world. But should it? New on @BigTechnology: https://bigtechnology.substack.com/ ...
Context & Ripple Effects
OpenAI is running Dall-E 2 as a deliberately throttled rollout — 3 million images since April, with access expanding by up to 1,000 users per week rather than opening the gates — and pairing it with a content policy aimed at sensitive or biased output. The moderation question is not hypothetical: days after this piece, a hacktivist documented that [[a:980143|Craiyon, the unmoderated DALL-E mini clone, defaulted to stereotyped portraits of brown-skinned women in saris]] on blank requests.
The strategy also sets up the business model. Stratechery's earlier analysis of zero-marginal-cost generation framed image models as infrastructure for social networks and the metaverse, and within weeks OpenAI converted the waitlist into revenue with a credit-based beta charging $15 for additional generations. Kantrowitz's interviewees see the stakes clearly — if Dall-E-style images become ubiquitous, control of the generator becomes control of the imagery.
First-order effects
- Roughly 1,000 new users per week get access to a system that can produce commercial-grade images on demand, while everyone else stays on a waitlist that OpenAI — not market demand — paces.
- The content policy gives OpenAI a live moderation layer over every generated image, making it the de facto editor of what its user base can depict.
Second-order effects
- Unmoderated clones like Hugging Face's DALL-E Mini absorb the spillover demand, and their visible bias failures hand OpenAI both a differentiation argument and a demonstration of why gating matters.
- Once the waitlist converts to paid credits, per-image pricing becomes the template competitors must match, turning image generation from a research demo into a metered product line.
Third-order effects
- If the pattern holds, content policy plus staged access becomes the industry-standard release mechanism for generative models — with the lab that owns the policy effectively setting the boundaries of synthetic visual culture.
- The alternative is a split market: moderated first-party models alongside open clones whose bias failures, like Craiyon's, keep surfacing — pushing regulators and platforms toward demanding provenance and moderation standards for AI imagery.
The trend: Consumer AI image generation is consolidating around gated, policy-moderated first-party models whose owners pace access and set the limits of what can be depicted.