Treat All Work as AI-Assisted. Then What Happens to Voice?
A mathematician’s proposal on Terence Tao’s blog, read by a translation student: when disclosure stops being the test, explaining your choices becomes one.
The essay arguing that we should evaluate all mathematical work as if it were AI-assisted opens with an editor’s note from Terence Tao: “This blog post was initially written in a different file format and converted using AI. — T.” It’s the most ordinary disclosure imaginable. It also shows exactly why Frank Calegari thinks disclosure is the wrong test. Nobody reading that note learns anything about whether the argument is good. If that’s true for proofs, translation students should ask what replaces it for sentences.
A useful starting point is a choice you can explain: what your sentence preserves, what it changes, and what would make you revise it. That still leaves a separate question about permission to use the tool. A good explanation doesn’t cancel that obligation.
What Calegari actually proposed
In his guest essay on Tao’s blog, Calegari makes six proposals for mathematics in an AI-assisted world. Two matter especially to a graduate student reading across disciplines: evaluate work on the assumption that AI may have contributed, and broaden graduate education while retaining serious technical training.
His immediate context is mathematical research. Calegari says OpenAI released more than 700 papers on October 6, and criticizes their writing and attribution. That’s his account of the release, not an assessment of those papers here. The practical question he raises is how to judge work when a finished proof tells you less about the route its author took.
He wants authors to explain the interest of their mathematics: why a question matters, where a result belongs, what it connects to. He also preserves a place for learning to do difficult mathematics without AI. Breadth doesn’t mean outsourcing the foundations.
Translation isn’t a proof with prettier punctuation. A sentence can be accurate and still sound wrong for the person saying it. I’m borrowing Calegari’s question about evaluation; the answer has to survive contact with a different kind of work.
Why disclosure stopped doing its job
Calegari identifies a problem with the incentives around disclosure: if using AI carries a stigma, people have a reason to describe its contribution as narrowly as possible. The declaration starts doing reputational work before anyone has evaluated the result.
In the blog’s comments, jesusvilelajato adds that paraphrasing can circumvent most AI detectors. That’s a commenter’s claim, not a detector benchmark. It does, however, expose a weakness in treating a text’s apparent origin as the answer to a question about its quality. An origin label doesn’t tell us whether a particular translation choice holds up.
A literary-translation study by Ferstler and colleagues makes that distinction unusually tangible. Fifteen avid readers compared human and machine translations of excerpts from fifteen novels, translated into English from French, Polish and Japanese. Across thirty whole-excerpt judgments, readers preferred the human translation nineteen times, but correctly identified the human version seventeen times. These are different questions, with different results. The small study doesn’t establish a rule for Spanish literary translation; it gives a reason to keep preference and origin separate.
Disclosure still has work to do. The Authors Guild’s model translation clause distinguishes consent to AI translation from a human translator’s controlled use of AI assistance. These are proposed contract terms, not a universal rule for every publisher or university. They make the distinction concrete: a process can require permission even when the resulting prose reads well.
Disclosure can describe the process. It can’t, by itself, justify the sentence. We still need to read what happened to the words.
From “did you use it” to “can you explain it”
Consider this invented Spanish passage, written for illustration rather than taken from a book or a student’s work:
— ¿Vas a volver? — preguntó Ana. Él dejó la llave sobre la mesa. — Ya veremos.
One English rendering is:
“Are you coming back?” Ana asked. He put the key on the table. “We’ll see.”
Another is:
“Will you return?” Ana asked. He placed the key on the table. “That remains to be seen.”
Both renderings are illustrative choices made for this article. They aren’t outputs from a human-versus-AI experiment. That missing label is useful: we have to discuss the actual differences.
Get Maya Bennett’s stories in your inbox
Join Medium for free to get updates from this writer.
At the level of meaning, both keep the question about returning, the key placed on the table, and an answer that leaves the future unresolved. Neither tells us whether he intends to return. An AI assistant can help propose versions at this level; the translator still has to check what each one says.
For voice, listen to the reply. “We’ll see” leaves room for weariness, evasion, reassurance or a refusal to decide. “That remains to be seen” gives the speaker a more detached, measured response. It might suit a character who habitually keeps people at a distance. We don’t have that character history in this fragment.
Register is the more local question of formality. “Will you return?” and “That remains to be seen” raise it together. “Placed” also makes the small action feel more deliberate than “put.” None is a dictionary error. Together, though, they change the temperature of the exchange. The second speaker sounds as if he’s had time to prepare his statement.
Then there’s silence: what the narration doesn’t explain. The key could belong to a flat, a car or something else. The fragment doesn’t say. Adding “reluctantly,” or making the key a sign that the relationship is over, supplies an interpretation the Spanish hasn’t settled. A fluent translation can close that space without looking conspicuously inaccurate.
With only this fragment, I’d use the first rendering. It keeps the dialogue conversational and leaves his intention open. I’d revisit it if the surrounding pages established a deliberately formal speaker. That condition matters more than a declaration that the translation “preserves the author’s unique voice.” The declaration sounds reassuring. It doesn’t name a decision.
A short translator’s note could say: “I used ‘We’ll see’ to keep the reply conversational and noncommittal. I left the meaning of the key unexplained because this passage doesn’t supply it. I’d check the speaker’s earlier dialogue before settling the register.” That gives another reader something specific to challenge.
The commenter jesusvilelajato describes asking a model to edit “in the manifold of my words.” It’s an appealing request: improve the English without making the speaker disappear. But a request still needs a reading of the result. Which habits survived? Which awkwardness carried meaning? Which smoother phrase moved the speaker a little further away?
An AI system can generate explanations of those choices too. The translator’s responsibility is to check and stand behind the explanation they adopt, including its limits. A plausible rationale written after the fact isn’t proof that the choice fits the passage.
Where the proposal makes me uneasy
There is an easy way for this standard to become unfair: reward the person who performs the most polished defense. The student who hears something wrong but needs time to locate it loses to the student with a fluent account of a mediocre decision.
Calegari explicitly distinguishes deep knowledge from fluent explanation, and gives priority to depth. That qualification belongs inside any translation of his proposal into another discipline. Otherwise, we’ve merely added an oral presentation to the workload.
The unease is already in the responses. An anonymous commenter reads the proposal as leaving ordinary mathematicians without a place. In the Hacker News discussion, ethanlipson warns against building a career around the expectation that AI won’t become good at explaining mathematics. Meanwhile, Nick describes AI as allowing him to learn at his own pace. His experience points toward access to learning, rather than another performance to pass.
I don’t think these concerns disappear when we ask people to explain their choices. They change how the question should be asked. Give the translator the passage, the alternatives and time to annotate them. Let a written note count. Ask what would change their decision, and check the answer against the text. A confident delivery shouldn’t outweigh a well-supported revision.
Return to the key on the table. A student who initially prefers “That remains to be seen,” then changes it after noticing the character’s informal speech elsewhere, has shown something worth evaluating. So has a student who keeps it and can point to the pattern that supports the choice. The valuable part is the connection between the words and the reason, including the willingness to reconsider.
That leaves voice somewhere more demanding than a checkbox. It lives in the particular phrase, the neighboring sentences, and the space a translator decides to leave alone.
If you had to defend one sentence of your work without mentioning whether a model touched it, which sentence would you pick?
