ChatGPT Human Reviewers and the Trust Boundary | Elacity
Contractors were paid to read real ChatGPT chats. A privacy toggle can't fix that; moving the original home and gating every use with a revocable grant can.
ChatGPT Human Reviewers Read Your Chats. A Settings Toggle Won't Move the Trust Boundary.
You typed it late at night because a machine felt safer than a person. Some of those conversations were read by ChatGPT human reviewers anyway: OpenAI hired hundreds of contractors to read ChatGPT prompts and rate the replies, 404 Media reported.
That is the cost, stated plainly. Your most candid writing passes through hands you never met, under terms you skimmed once. This piece is not about whether one company behaved badly; it is about where your words actually live, and who stands on the near side of that line.
What Project Lily Revealed About ChatGPT Human Reviewers
Reviewers in the program summarised what a user wanted and scored up to four candidate answers on a 1-to-7 scale, according to a proposed class action described by Decrypt. The complaint alleges the automated filter meant to strip personal details does not always catch them.
The case, Vredenburgh v. OpenAI OpCo in the Northern District of California, argues users were never clearly told a human might read their chats. It seeks opt-in consent before any outside reviewer sees a conversation, and model training switched off by default.
None of this is new. In 2019, Amazon workers were found listening to Alexa recordings to improve the service. Each wave of computing rediscovers that the system includes people.
The Concession: Review Is Not the Villain
Human feedback is how models learn to stop flattering you; OpenAI has described the reviews as aimed at reducing sycophancy and human-like behaviour. That is a legitimate goal, and banning human review would not make anyone safer.
A hosted model also has to read your prompt in plaintext to answer it. No cryptography makes a remote datacenter respond to words it cannot see, and any honest account of chatbot privacy starts there.
So the failure is narrower: the stored copy. Once your history sits in someone else's storage, every later use (a reviewer, a vendor, a training run, a subpoena) is decided by their policy. OpenAI's data controls let you opt new conversations out of training, but a toggle is a setting inside their boundary, not a wall around yours.
The Lens: Where the Trust Boundary Sits
Every era moved the boundary. The machine room trusted its operators. The personal computer pulled trust into a box on your desk. The cloud pushed it back onto someone else's servers, and every decision about who reads what went with it.
AI forces the question back to you, because what you now feed machines is not files but intent: plans, fears, drafts, symptoms. It is the most valuable data you produce, and today it is archived on the far side of the line.
The Elacity Approach: Move the Plaintext Home
Elacity treats that data as property you hold, with terms you set, rather than exhaust a service keeps. The engine beneath is ElastOS, the open-source runtime for Personal Cloud Compute, where your own machine is the source of truth and the cloud is a guest. The mechanism that matters here is the capability rail: zero ambient authority.
1. Nothing reads by default
No app, script, or AI can touch your files, network, or money until you grant a specific, narrow, expiring permission. Revoke it and the action stops mid-flight. A third party reading your data stops being a policy footnote and becomes a permission you either issued or did not.
2. The original stays with you
Your history lives on hardware you control. A remote service receives what a task needs, not a standing archive it can later mine, review, or lose in a breach.
3. Sharing happens at a sealed moment
When you do want your data used by a model, a lab, or a buyer, Elacity dDRM can package it as a Wealth Capsule: encrypted, with rights and royalties written in. It stays sealed everywhere except a brief moment of use inside a sandbox, and its key is split across an owned 2-of-3 quorum where each node re-checks the buyer's on-chain rights. Consent becomes a gate that runs every time, not a sentence in a privacy policy.
The honest edge: this protects data you hold and choose to license. It cannot make a third-party chatbot private once you paste into it, and the agent tooling that would let your own assistant work this way end to end is still being built.
What Changes for You
Treat every hosted chat window as a room with an audience you cannot see, and keep your originals where you can. The commercial side of the same problem is in ChatGPT Ads Turn Your Chats Into Inventory; how narrow grants work is in Capability-Based Security, Explained.
"The people who create the value should own it. That is the entire reason Elacity exists." (Sasha Mitchell, Founder)
You were the product. Now you own the asset class. More from this line of thinking lives in The Manifesto; Follow Elacity on X for the next one.