Mark the Machine's Paragraphs
Meta AI published six mathematics papers in early October, co-authored with human mathematicians and drafted with help from its Muse Spark 1.1 and 1.2 models running in Thinking Mode. Five of the six answer previously open problems, including a 384-element group that disproves a 2024 conjecture by Kida. The publication format caught my attention more than the results, because it addresses a problem any organization using AI in documents runs into.
Each paper marks which passages the model drafted and which a human drafted. A second group of mathematicians reviewed the papers independently. A reader can see, line by line, where machine output enters the argument, and the reviewers who checked the work did not generate it.
A second case from the same fortnight pairs well with it. An independent researcher ran OpenAI's Dots as a swarm of six subagents over three days and raised a 1979 combinatorics bound, C(24,14,4) at least 20. The result arrived with a proof verified in Lean, a proof assistant that checks each step mechanically. The trust sits in the checker, a small program that accepts or rejects each step.
Together, those two cases show two practices that I think form a pattern. The first is provenance marking: a durable record of which parts of a document came from a model. The second is an independent check whose verdict the generating system cannot influence. The Lean proof achieves it through mechanical verification, and the Meta papers through a separate group of human reviewers. Marking tells a reader where to look. Independent checking tells a reader whether to believe what they find there.
A contrasting case appeared in the same week's coverage. OpenAI published 722 machine-generated mathematical manuscripts from an unreleased model, and only 162 had their main results formalized in Lean. The mathematics community's objection, voiced through the Association for Human Mathematics, centered on the absence of the second practice, with publication by blog post standing in for review. Those numbers show what happens when generation scales without the checking infrastructure to match.
Mathematics has an unusual property that makes both practices easy to apply, which is that proofs can be machine-checked. Most documents that organizations produce lack that luxury. A policy, a clinical summary, a financial narrative, or the evidence package for an audit cannot be run through a theorem checker. Yet the same questions press on them. A regulator reading an AI-assisted clinical summary wants to know which sentences a model wrote. An auditor reading a control narrative wants to know that a person with accountability reviewed it.
The pattern transfers with a few adaptations, and the scheme I have settled on has three parts. Each section of a document carries a provenance tag, a short label stating whether a human drafted it, a model drafted it and a human edited it, or a model drafted it and a human only approved it. Each document names a specific reviewer who is accountable for it, a person whose name appears on the sign-off and who is distinct from whoever prompted the model. And each document has a verification step whose pass condition the model cannot influence, which in most business settings means a check against a source of record: numbers matched against the ledger, citations matched against the cited documents, dates matched against the system that holds them.
The third part is where regulated documents differ most from mathematics. In a proof, the checker is formal. In a financial summary, the check is a reconciliation against data the model had no ability to alter. In a clinical summary, it might be a comparison against the underlying chart. The property that matters carries over: the verifier gets its information from somewhere the generator cannot touch.
Provenance tags also have a second-order effect that I did not expect to find valuable. They change how reviewers read. A reviewer who knows which paragraphs were model-drafted applies a different kind of scrutiny than one who assumes everything is human, and the difference shows up in what gets caught. Unmarked AI text invites the reader's usual trust in a colleague's prose, which is the wrong trust for a system that produces fluent text whether or not it is correct.
What Meta's papers show is a disclosure norm available now, built from a few ordinary habits. The cost is a tagging convention and a named reviewer. The benefit arrives later, when someone asks how a document came to say what it says and the answer exists on the page.
Sources: Meta AI, "Solving open research problems together" (Oct 2026); "I used OpenAI Dots as an agent swarm" write-up (Oct 2026); Mixture of Experts Ep. 128 (Oct 9, 2026); All-In Ep. 292 (Oct 10, 2026).