Source: raw/How_to_Use_AI_on_Files_You_re_Not_Allowed_to_Upload.md
Creator: Nate B Jones (AI News & Strategy Daily) URL: https://www.youtube.com/watch?v=EuVvLwWZ5wc Platform: YouTube
A walkthrough of running a frontier model against a document too sensitive to upload, by redacting it first — and the useful part is not the tool but the decision rule. The framing is a complaint about standard privacy guidance: “don’t paste your sensitive information into AI” stops too soon, because the contract still needs its risk review and the performance review still needs writing. If the safe path is manual cleanup or abandoning the tool, the warning has handed the problem back rather than solving it. The creator demonstrates a tool he built called Airlock against a synthetic pricing plan.
Key Takeaways
- Start with the job, not the file. This is the load-bearing idea. Whether a field must be redacted depends on the question being asked — the same price is essential when the model is evaluating the pricing, and unnecessary when the question is “what assumptions could break the operating plan.” Redaction policy is task-relative, so a fixed detector cannot decide it for you.
- Keep what the work actually needs. In the demo, operational facts stayed because the analysis depended on them: all three warehouses move in the same week, the plan assumes an ERP integration lands before the first billing cycle, training happens during normal shifts. The model produced the three assumptions most likely to break the launch from that alone.
- Confidential ≠ detectable. Protected terms are entered by the operator up front — a customer name, an internal codename like “Project Lantern,” a product code. Such a string “doesn’t look private to a machine,” and the context that makes it sensitive exists only in the operator’s head. A pattern detector alone will miss exactly the terms that matter most.
- Default to hide; retention is an explicit act. Where there is uncertainty, the starting choice is to redact. Keeping an item requires a deliberate decision — the safe direction is the default rather than the exception.
- Credentials are bundled, not needed. The API key in the file had “no conceivable reason to be inside the task” — it was simply bundled in. The creator names this as the ordinary mechanism by which confidential data reaches frontier models: not deliberate pasting, but incidental inclusion.
- The original never leaves the machine. The model saw only the redacted document; the source file stayed local, and the work still got done.
Why not just rely on zero-data-retention? (2026-09-29 addition)
Source: raw/AI_Kills_Everybody_or_Doomer_Psyop_OpenAI_s_Math_Breakthrough_Nike_s_200B_Collapse.md — All-In Podcast, youtube.com/watch?v=cvxjqbfLVk0. Auto-caption transcript with no speaker labels, so names below are inferred from context. Everything here is attributed opinion from investors with stakes in the outcome.
- What started it. OpenAI said it had solved a 200-year-old maths problem on the Navier-Stokes equations. Researchers who had been working on the same problem with OpenAI’s products suggested their work may have been used. The hosts read OpenAI’s statement: “While unlikely, we cannot rule out that deidentified data derived from their usage of our products helped improve our models.”
- Chamath Palihapitiya: ZDR is “a commercially best efforts basis… They can’t guarantee it.” He names one leak path: even with retention off, “if anybody that’s using it clicks the like button inside of the chat window, that’s not necessarily guaranteed.” Later he calls ZDR “flimsy and it’s Swiss cheese.” His fix is a “sovereign” deployment: your own hardware or VPC through a vendor you trust, running models provisioned to you.
- De-identification doesn’t protect ideas (Chamath). Stripping names and company details does not remove the method: “the approach is the IP.” An iterative chat can be de-identified and still teach the model how to solve the problem.
- The local-hardware counterpoint. A co-host (by context Jason Calacanis) “just ordered two Mac Studio M5s” to run local models and is moving sensitive data off Claude (caption: “claw”). Chamath replies that local boxes lack the “multiplayer experience,” knowledge base and memory that make the cloud tools useful.
- The legal angle (by context David Sacks). He finds data leakage the less plausible explanation for the OpenAI dispute. His concern is legal: AI chat data can be obtained “with a subpoena or court order,” not the search warrant and probable cause that email requires. He wants AI chats protected at least to the email standard.
- Disclosures. Chamath says his company 8090 partnered with EY and Deloitte on this problem. Jason’s LAUNCH accelerator incubated an on-prem AI box company that he promoted in the same segment.
What this adds to the method above.^[inferred] ZDR is a contract term you cannot inspect. Task-relative redaction is the layer you control, so do it regardless of retention settings. Until you have checked your provider’s terms, also avoid feedback clicks (thumbs, likes) on chats that contain sensitive material. Chamath’s claim that a like click can override retention has not been checked against any provider’s terms.
Try It
- Take your next genuinely sensitive task and write down the job first. Before touching the file, state the question in one sentence. Then decide field-by-field whether the model needs that field to answer that question — you will find fields you would have reflexively kept.
- List your organization’s protected terms once. Codenames, internal project names, customer names, unreleased product names. This list is the part no detector can infer, and it is reusable across every document.
- Audit one file you have already uploaded for incidentally-bundled credentials. The API-key-in-the-planning-doc case is the failure mode worth checking for retroactively.
Open Questions
- Airlock is the creator’s own tool and this wiki has not verified it. No pricing, availability, licensing, or repository was captured from the transcript — treat the tool as illustrative and the method as the transferable part.
- No verification step is shown. The demo does not cover how you confirm redaction succeeded before sending, which is the obvious failure point in any such workflow.
- Effectiveness is asserted, not measured. The claim that the model still found “the three assumptions most likely to break my launch” is the creator’s own assessment of a synthetic document, with no comparison against the unredacted result.
- Does a feedback click override zero-data-retention? This is Chamath’s claim on All-In, with no provider named and no documentation cited. Check the specific provider’s terms for how rated or flagged conversations are handled under ZDR before relying on it either way.