Argument architecture

How this editorial can be challenged

Core question

How can a customer check a promise that depends on a company they never chose?

Proposed mechanism

In a chain of contractors, each participant can know its own settings while missing the actual path of a request. An independent observation can reveal the gap between the documented route and what happened.

Strongest counterargument

Detailed logs can create precisely the extra copies that zero retention is meant to prevent. A provider also needs room to suspend a route on suspicion without calling it a confirmed breach.

Our response

The relevant evidence need not contain the prompt. Record the identity and policy of the processor, restrict access to those records, and use synthetic probes to test selected claims.

Evidence limits

A tested decoy establishes that a particular credential traveled somewhere unexpected; it neither counts every exposed customer nor certifies all routes where no decoy fired.

What would change our mind

If independent tests routinely confirmed actual routing and timely incident notice through less intrusive means, requiring more detailed processing receipts would be hard to justify.

The wrong person knew the secret

A decoy credential has almost no economic value. Its usefulness lies in who knows it. If a stranger tries it, the person who planted it has learned something a privacy policy could never reveal on its own. In Together AI’s account of its prompt-sample leak, five such decoys connected a real download to Dragos’s separate experiment. Five is not a victim count. It is the difference between suspecting an unexpected journey and having evidence that it happened.

A buyer can inspect the wrong company

When an employer buys access to a language model, the company named on the invoice need not be the one running the computation. Another operator may handle a request, and still another system may hold an authorized testing copy. A careful buyer could ask the first company all the right questions about its own storage and come away knowing remarkably little about where the words went.

The list describes tomorrow

OpenRouter’s response illustrates both the value and the limit of a fast intervention. It removed providers and withdrew zero-retention eligibility from routes, some as a precaution. That changes the choices available to an application now. A person whose information appeared in a request months ago has a different question. Knowing that a train no longer stops at a station does not establish where yesterday’s passenger got off.

A mathematician offers a useful contrast

Yesterday we described an AI-assisted Riemann proof as a proposal awaiting independent checks. Today the Lean and Mathlib communities say it proves the intended statement. Andrew Sutherland also reports that a separate rebuild produced an identical proof export and independent checkers accepted it. Readers can inspect what was tested; a mathematical explanation that people can learn from is a different task. The evidence changed, so our wording should change with it.

Privacy cannot be tested like a theorem: one cannot enumerate all the things that never happened to a prompt. But the distinction between checking a claim and trusting its author travels well. A test that never leaves the team that made the promise gives the public little reason to revise its view either way.

More logging can be the wrong answer

Keeping a copy of every sensitive conversation to prove that conversations are not being kept would be a peculiar victory. A better request from a buyer is narrower: identify the processor and the applicable policy without storing the text, and explain how someone tested the exception path. Even that metadata deserves access and retention limits. A receipt should be a way to reduce the amount of trust required, not an excuse to collect more.

Put one old request on the table

The next time a vendor presents a retention badge, ask its team to choose a past test request containing nothing personal. Can it show which company handled that request, which rule applied, and what the customer would have learned if the route changed? A clear account of what cannot be reconstructed is more useful than a confident answer about the default.

There is a commercial opportunity for a provider that can answer without keeping the conversation. Its strongest sales pitch would be modest: here is what we checked, here is who checked it, and here is where the evidence stops.

Evidence behind the argument

Read the reporting

Opinion is ours. The factual record is linked below.

Together AI — November 21 incident notice Dragos — instrumented prompt investigation OpenRouter — actions on inference-provider incidents Lean and Mathlib — verification of the Riemann hypothesis Andrew Sutherland — independent verification log