A fresh wave of posts on social media is once again putting OpenAI’s file-handling behavior under the microscope, this time with a sharper focus on GPT-5.6 Sol and a specific allegation: that the system can delete files and related data without warning.
The claim is not being treated as brand-new by everyone following the story. Multiple commenters point out that OpenAI had already disclosed a version of the underlying issue earlier in June, which means the current discussion may be less about discovering a surprise capability and more about what users experienced in practice—and whether the transparency around it matched what people expected.
What makes the renewed attention notable is the way these conversations tend to evolve. At first, they often start as scattered reports—screenshots, short videos, or “this happened to me” threads. Then they broaden into questions about scope: what exactly was deleted, how often, under what conditions, and whether there was any meaningful notice before the deletion occurred. In other words, the debate quickly shifts from “did it happen?” to “what does it mean for real-world workflows?”
Below is what the current reporting and discussion appears to center on, why it matters, and what users and teams should do to reduce risk while the details continue to clarify.
The core allegation: deletion without user notification
According to the social media reports driving this latest round of concern, GPT-5.6 Sol allegedly removed files and data associated with a user’s activity without providing an explicit warning at the moment of deletion. The phrasing varies across posts, but the common thread is the absence of a clear, user-facing alert that something had been removed.
That distinction—“deleted” versus “expired” versus “not retained”—is crucial. Many AI platforms use retention windows, background cleanup processes, or storage policies that can make data disappear after a period of time. Those behaviors are not inherently unusual. What alarms users is when the disappearance feels immediate, unexpected, or tied to actions that don’t obviously map to a retention policy.
In practice, users often assume that uploading a file to an AI tool means it will remain available for the duration of their project, at least until they explicitly delete it. When that assumption breaks, the impact can range from minor inconvenience (having to re-upload) to serious disruption (losing work product, losing access to reference material, or breaking compliance expectations).
Why the June disclosure changes the context
The reason this story is being framed as “people keep warning” is that OpenAI had reportedly already disclosed the problem in June. That doesn’t automatically resolve the current controversy, but it does shift the narrative.
If OpenAI previously communicated that certain data might be deleted under specific circumstances, then the current question becomes: did users understand the disclosure well enough? Did the disclosure cover the exact scenario now being described? And did the actual behavior match the wording and timing implied by the earlier communication?
Disclosures can be technically accurate yet still fail to prevent confusion if they’re buried, hard to interpret, or not aligned with how users experience the system day-to-day. For example, a policy might say data is retained for a limited time, but users might interpret that as “retained unless you delete it,” rather than “retained only for X window regardless of your intent.” Or a policy might describe deletion as part of a safety or housekeeping process, but users might not connect that to the moment they noticed the loss.
So while the June disclosure suggests OpenAI anticipated the issue, the renewed posts imply that the practical understanding among users may still be incomplete—or that the behavior may have changed, expanded, or been observed differently than expected.
What “files and data” could mean in an AI workflow
One reason these stories spread quickly is that “files and data” can refer to multiple layers of information in an AI system:
1) The original uploaded file itself
2) Derived artifacts created during processing (extracted text, embeddings, summaries, indexes)
3) Conversation-linked references (metadata that helps the model “remember” what it saw)
4) Temporary working data used during a session
5) Cached results or intermediate outputs stored for performance
When users say “it deleted my files,” they might mean the original upload disappeared from their interface. But they might also mean that the model no longer had access to the content it previously used, even if the file still appeared to exist somewhere else. Those are different failure modes with different implications.
For instance, if the platform deletes the original file but keeps derived representations, the user might still get answers based on earlier content. Conversely, if derived data is removed but the file remains visible, the user might see the file listed yet find that the model behaves as if it never processed it.
The social media reports referenced in the discussion appear to focus on user-visible deletion, but the broader concern is about continuity: whether the system reliably preserves the information needed to complete a task over time.
Automatic deletion vs. deletion triggered by specific actions
Another key question raised by these reports is whether deletion is truly automatic—happening as a background process—or triggered by particular actions or settings.
Some platforms delete or purge data when:
– A session ends
– A conversation is closed or archived
– A user changes a setting related to retention
– A file is replaced or re-uploaded
– A safety filter flags content and triggers cleanup
– A quota or storage limit is reached
– A retention timer expires
If deletion is automatic and tied to a timer, then the “without warning” aspect might reflect the fact that users aren’t notified when a retention window ends. If deletion is triggered by specific actions, then the “without warning” aspect might reflect a missing confirmation step—something like “your previous files will be removed when you do X.”
The difference matters because it determines what users can do to prevent it. If it’s timer-based, users can plan around it (export work, re-upload periodically). If it’s action-based, users can avoid the triggering steps once they know what they are.
At the moment, the public discussion seems to be pushing for clarity on exactly which mechanism is at play.
Why users are reacting strongly: the mismatch between expectation and reality
AI tools are increasingly used like workspaces. People upload documents, ask questions, iterate on drafts, and treat the system as a semi-persistent assistant that can revisit earlier materials. Even when the platform is not marketed as a document management system, users often behave as if it is one—because the workflow feels like one.
When data disappears unexpectedly, it undermines trust in the tool’s reliability. It also creates a practical risk: if you’re using the system for anything time-sensitive or compliance-sensitive, losing reference material can create downstream problems.
Consider common scenarios:
– Legal or HR teams using AI to summarize policies and then revisiting those summaries later
– Researchers uploading datasets or transcripts and expecting continuity across sessions
– Creators using AI to extract notes from long-form content and then returning days later
– Developers uploading logs or code snippets and relying on the model to maintain context
In each case, the user’s mental model is “I gave it the material; it should still be there when I come back.” If the system deletes it without warning, the user has to rebuild context—sometimes from scratch.
That’s not just an inconvenience. It can also lead to subtle errors. If the model no longer has access to the original file, it may produce answers that sound plausible but are based on incomplete or outdated context. Users might not realize the context changed, especially if the interface doesn’t clearly indicate what the model can still access.
The transparency gap: what notice should look like
The most persistent theme in these discussions is not merely deletion—it’s the lack of notice. But “notice” can mean different things depending on the platform’s design philosophy.
A strong notice system would ideally include:
– Clear retention policy language in plain terms
– User-facing indicators showing what is currently stored and for how long
– Alerts when a file is removed or when access to it is revoked
– Documentation that maps common user actions to retention outcomes
– Export or backup prompts when deletion is imminent or likely
Even if deletion is unavoidable due to technical constraints or privacy requirements, users generally accept it better when the system communicates it clearly. Without that communication, users interpret deletion as a bug or a breach of expectation.
This is where the June disclosure becomes relevant again. If OpenAI already explained the behavior, then the current complaints suggest that either the explanation didn’t reach users effectively, wasn’t specific enough, or didn’t translate into a clear in-product experience.
A unique angle: the “safety” framing vs. user trust
There’s another layer to this story that often gets overlooked in fast-moving tech debates: deletion can be motivated by safety and privacy goals, but safety measures still need to preserve user trust.
From a privacy standpoint, deleting user data can be a feature, not a flaw. From a safety standpoint, removing certain stored artifacts can reduce the risk of unintended reuse. But from a user standpoint, trust depends on predictability.
If users can’t predict when data will be removed, they can’t confidently build workflows around the tool. That can push teams toward conservative usage patterns—like exporting everything immediately, avoiding long-term reliance on uploaded files, or limiting the tool to short-lived tasks. Those workarounds reduce productivity and can increase operational overhead.
So the real challenge isn’t simply “stop deleting.” It’s aligning deletion behavior with transparent, user-understandable expectations so that users can decide whether the tool fits their needs.
What users and teams can do right now
While the public conversation continues, there are practical steps users can take to reduce risk and avoid surprises:
1) Treat uploads as non-permanent unless the interface guarantees otherwise
If the platform doesn’t clearly promise persistence, assume you may need to re-upload.
2) Export or archive critical inputs and outputs
If you’re working on anything important, keep local copies of:
– Original files
– Key extracted text
– Summaries and final drafts
– Any intermediate outputs