Users are reporting that GPT's newest image generation model is blending visual elements from previous session outputs into fresh prompts, exposing a structural flaw in how the system handles long-term context.
Something strange started happening on April 23rd. Across X and Reddit, users of OpenAI's latest image generation update , internally tagged Gen-V4 , began posting side-by-side comparisons of what they asked for versus what they got: distorted composites where a freshly prompted landscape bled into the ghost of a logo generated three messages earlier, or a portrait warped by the color palette of an entirely unrelated asset from the same session. The community landed on a name fast: Ghosting.
The phenomenon isn't random noise or typical model inconsistency. Based on user reports, the contamination follows a clear pattern , earlier images in an active chat history are leaking into subsequent generations, with fragments appearing smeared or amalgamated into the new output. The more generative work done within a single session, the more pronounced the interference. That specificity matters, because it points toward a systemic architectural issue rather than a one-off inference error.
The working theory gaining traction among developers and researchers is that Gen-V4's updated compression algorithms, combined with a new long-term context retention architecture, are the culprit. OpenAI appears to have built the model to maintain image context across a conversation , presumably to reduce compute overhead by avoiding full re-encoding of prior session assets. The tradeoff, apparently unvetted at scale, is that the model is now hallucinating visual data from its own recent context window and folding it into new renders. It's a memory leak, but for pixels.
OpenAI has not issued a public statement acknowledging the regression, though elevated latency on image-generation API endpoints was flagged on the company's status dashboard shortly after the patch went live , a quiet signal that something went wrong under the hood. The silence is notable given how quickly the reports spread and how reproducible the issue appears to be.
Why enterprise clients are the real story here
For casual users, a ghosted image is annoying. For the design teams and product studios running iterative workflows through the API, it's a blocker. Cross-contamination at the output level makes professional asset creation functionally impossible within multi-turn sessions , which is precisely the use case OpenAI has been marketing to enterprise customers. Several teams have already reported suspending API-based design pipelines while waiting on a fix or rollback.
That operational disruption points to a deeper reputational risk for the generative AI sector broadly. The entire value proposition of these tools rests on output fidelity and predictability. When a model update quietly introduces a regression that corrupts the consistency of the product, it erodes exactly the trust that enterprise adoption depends on. A single high-profile ghosting incident shared on LinkedIn or a design community forum can set back procurement conversations that took months to develop.
There's also a structural tension this incident puts in sharp relief. As image and multimodal models extend their context windows , a capability arms race that every major lab is running , the risk of interference between distinct user requests within the same session grows. Isolation protocols that treat each generation as a clean-slate inference exist for good reason. Relaxing them to chase efficiency gains, without adequate testing at scale, is the kind of optimization that looks smart on a benchmark and breaks in production.
The most likely short-term outcome is a rollback to the prior generation's isolation architecture, with a more careful staged rollout of the context retention feature. What's worth watching longer term is whether OpenAI moves to implement session-scoped memory controls that give enterprise users explicit command over what context the model carries forward , a transparency mechanism that would address both the technical and trust dimensions of this failure. The labs that solve stateful inference without sacrificing output integrity will have a genuine structural advantage. Right now, Gen-V4 is a case study in what happens when that problem is treated as solved before it actually is.
Also read: GPT Image 2 just produced a fake video game screenshot that nobody can tell is fake • The viral joke about thanking ChatGPT in 2050 is actually a serious warning about how we're building AI • Anthropic's AI security tool found 271 zero-day vulnerabilities in Firefox and the industry should pay close attention
