I have been dictating long-form thoughts into AI sessions and asking agents to help me turn them into blog posts. Sometimes my prompts, corrections, and source material are longer than the finished post. If you saw only the final page, though, you would not see any of that. You would see a polished artifact and have to guess where every phrase came from, what the agent changed, and whether I understood the argument or merely asked for something that sounded convincing.

That opacity is becoming a problem anywhere an artifact is used as evidence of ability. A graduate paper is supposed to show that a student can form a question, find relevant work, gather data, reason about it, and defend a conclusion. A software interview is supposed to show that a candidate can understand a problem, make trade-offs, and build a solution. AI can now produce both artifacts. Banning it does not restore the old signal, because the technology is already part of how people work. Accepting the artifact without asking how it was made does not solve the problem either.

The question I want to chase is not, "Did AI touch this work?" The more useful question is, "Can every phrase be traced to something the human already communicated, while the machine remains useful without receiving any authorship authority?"

From Correction to Collaboration

Spell-checkers and grammar tools removed mechanical friction without taking authorship away from the writer. We generally do not care whether a student remembered the spelling of every word. We care whether the words express the student's thinking. Word processors also made revision, formatting, citation management, and document assembly easier. Each tool moved some labor from the human into the machine while leaving the human responsible for the result.

AI continues that progression, but it crosses a more important boundary. It can reorganize an argument, discover a source, identify a gap, write a paragraph, generate code, or propose a conclusion. Those operations do not all represent the same kind of assistance. Fixing punctuation is not equivalent to introducing a new claim. Reordering ideas already present in a transcript is not equivalent to inventing the ideas. For the kind of academic authorship tool I am imagining, the boundary should be stricter: the agent must not generate new prose for the artifact at all.

The closest familiar analogy may be an executive working with an assistant. The assistant can arrange the material, prepare a memo, verify its format, and keep the process moving. The executive can still provide the direction, content, edits, and final approval. Responsibility remains with the executive. That relationship works because the people involved understand the roles. With AI, the assistant is invisible inside the document, so we need a way to make the roles visible again.

This blog is already one example of why the boundary needs to become visible. I provide long transcripts, lived examples, connections to earlier ideas, corrections, and the direction I want the argument to take. An agent helps organize that material into a cohesive story. I then steer, reject, revise, and approve it. That workflow is collaborative, but it is not yet the constrained system I am proposing. Clicking "accept" on agent-written prose does not make the prose human-authored. In the stricter mode, a suggestion can only become part of the document after the human says or types the substance of the change back into the system. The final prose should be traceable to human input, not merely approved by a human.

A File That Carries Its Own History

Imagine a local-first desktop application built around a research workspace rather than a blank page. It could hold the original dictation, notes, hypotheses, papers read, human-written source summaries, data, experiments, prompts, AI responses, edits, and final document in one place. Every transformation would become an append-only event in a timeline. More importantly, every phrase in the final document would carry lineage back to recorded human speech, human keystrokes, or an explicitly attributed external source.

The workspace would have an authorship mode with a hard capability boundary. The agent could delete, split, reorder, format, correct mechanical errors, or apply a declared dictionary mapping to words the human supplied. It could not invent a transition, complete a thought, add an argument, or silently strengthen a conclusion. Even a mechanically transformed sentence would retain links to its source spans and a record of the operation that changed it.

Other capabilities would operate outside the authored document:

• Review mode could identify repetition, unsupported claims, unclear reasoning, or missing connections. It could ask a question, but not answer it inside the paper.
• Research mode could search for sources and preserve what was retrieved. A machine summary would remain research material, not authored prose.
• Citation mode could insert a selected quotation with attribution and a link to the source passage.
• Organization mode could propose a new order, but every moved phrase would remain connected to its original human input.

There would be no one-click acceptance of generated content. If the review says an argument needs stronger evidence, the student must respond with the evidence and instruct the tool what to change. If research mode finds a new idea, the student must read it, cite it where appropriate, and explain the connection in their own recorded words. The agent can query the user for missing knowledge, much like an oral examiner, but it cannot answer on the user's behalf.

The capture layer matters as much as the editor. Dictation could preserve the original audio alongside its transcript. Typed input could preserve the text as it appeared, with optional keystroke timing in a proctored setting. Pasted material would be quarantined until the user identified it as an external source, linked it to its origin, or restated it as their own thought. These signals would not prove what happened inside a person's mind, but they would make the path into the document inspectable.

The result could be exported as a portable evidence package containing the final artifact, phrase-level lineage, its revision history, source references, research data or links to it, AI tool metadata, reproduction instructions, and a human-readable contribution report. Hashes and signatures could make later alterations detectable. Selective disclosure would also be essential, because raw research histories can contain private notes, confidential sources, unpublished data, or personal information. A verifier should be able to confirm the integrity of the package without automatically receiving every private thought inside it.

This should not become an "authenticity score" or a percentage claiming that a paper is 73 percent human. A score would hide uncertainty while creating a new target to game. The claim should be narrower and verifiable: every phrase in the submitted artifact has a recorded path to human input or an attributed source, and the constrained editor had no capability to insert untraceable prose. The tool still cannot prove that an idea is original or that the human understands it. It can show when the idea entered the record, what the author had read at that point, how it changed, and which questions caused the author to add more material. The outcome is an audit trail, not a verdict.

There could also be a stronger proctored mode for specific certifications. A student might complete a designated research exercise in an environment that records source intake, tool access, notes, and transformations. That could be useful when the process itself is under examination, but it should not become the default. Constant recording introduces privacy, accessibility, surveillance, and power concerns. Ordinary provenance and high-assurance proctoring solve different problems and should remain separate.

The Artifact, the Process, and the Defense

An audit trail alone is not enough. A person could carefully stage a transcript, repeat suggestions they do not understand, or learn how to produce a convincing history. The final artifact is not enough either, because an unconstrained agent can generate one. Validation becomes stronger when three forms of evidence agree:

• The artifact shows what was produced.
• The process shows how it was produced.
• The defense shows that the person understands and can extend it.

For graduate work, a teacher could inspect whether the research question emerged from the student's notes, whether claims trace back to evidence, whether sources were summarized and connected rather than merely cited, and whether experiments can be reproduced. The recorded history could then generate specific oral-defense questions. Why did you reject this alternative? What changed your interpretation of this result? Reproduce this analysis with a different assumption. Explain the strongest source that disagrees with you. Those questions test the student's relationship to the work rather than their ability to recite the final paper.

The same model applies to hiring. A candidate could bring an AI-assisted implementation together with its work history, then explain the architecture, identify weaknesses, debug a failure, or extend the system under a new constraint. CoderPad describes Meta's pilot of AI-enabled coding interviews as a way to give candidates tools resembling those used in real engineering work. That is one sign that interviews are moving from pretending the multiplier does not exist toward evaluating how a person uses it.

The appropriate level of autonomy depends on the context. In day-to-day software engineering, I may want an agent to implement, test, and iterate with broad autonomy while I guide the outcome. In a qualification interview, the system may need to expose what the candidate knows and how they direct the tool. In a graduate paper, the constrained editor should have no authority to write new prose. These are different modes with different trust boundaries, not one universal policy for AI use.

The point would not be to prove that the candidate can work without AI any more than we ask an engineer to work without an editor, compiler, or search engine. The point would be to show judgment: knowing what to delegate, recognizing when the machine is wrong, understanding trade-offs, and remaining accountable for the result. A defense remains necessary because traceable words demonstrate provenance, not understanding.

That changes what we measure. Remembering syntax becomes less important when syntax is readily available. Forming useful questions, evaluating evidence, designing experiments, connecting ideas, and defending decisions become more important. AI does not remove the need for mastery. It makes weak proxies for mastery easier to see.

Showing the Work Without Rejecting the Tool

The first version of this system does not need to decide who is qualified or whether a paper is valid. A minimum useful product could capture voice and typed input, quarantine pasted text, preserve sources, version a document, enforce a mechanical transformation policy, display phrase-level lineage and diffs, and export a self-contained evidence package. A read-only reviewer could explore the history and create questions for a defense. Judgment would remain with the teacher, reviewer, or interviewer, while responsibility would remain with the author.

This is closer to a scientist's lab notebook than an AI detector. Detection guesses whether a machine generated the surface of the artifact. A constrained environment prevents the machine from inserting unattributed prose and records what happened instead. Provenance shows where the language and evidence entered. Reproducibility asks whether another person can follow the method. Defense asks whether the author understands the decisions. Together they offer more confidence than a prohibition that cannot be enforced or a detector that can be wrong.

The larger shift is from submitting artifacts to submitting accountable work. We should expect people to use capable tools for mechanical and organizational labor. We should also expect them to disclose material assistance, preserve the path from evidence to conclusion, and stand behind what they submit. The technology should make that discipline easier, not make every student or candidate reconstruct it after the fact.

This post began as dictated exploration and became an argument through collaboration with an agent, making it an example of both the opportunity and the unresolved problem. Its current history can show extensive human direction, but it was not produced inside the phrase-level constrained environment proposed here. That distinction matters. The next step beyond the persistent context in Why I Switched to Worklogs and the deliberate handoff in The Bicycle of the Mind is a tool whose limits provide the evidence: the machine can organize, question, and correct, but the human must supply every idea and every piece of prose they claim as their own.