Dear Editors: If You Want My Voice, Give Me Back My Language

Thu, 27 Aug 2026 00:00:00 +0000

Another day, another journal trumpeting its AI policy. This time, it is Progress in Human Geography, whose editors have published a policy statement warning that using AI as a substitute for one’s own reading, thinking and writing is ‘a breach of academic ethics akin to plagiarism’. Papers found in breach ‘may be withdrawn’; at best, they will be corrected, with a notice printed in the journal. Authors, the editors add, should ‘carefully ponder whether they wish their names to be associated with withdrawals and correction notices’.

That sentence deserves a moment. Retraction sits at the top of academic publishing’s sanctions, the penalty normally reserved for fabrication and fraud. Here it is attached to an offence the same editorial cannot define. There are, the editors concede, ‘genuine grey areas’ and ‘no clear lines demarcating the zone of “irresponsibility”’. Enforcement will rest on ‘judgement calls that might upset some authors, who will claim no undue AI use’: one’s own declaration, in other words, may be overruled. The advice that follows is to remain ‘safely in the zone of “responsibility”’, and one only counsels safety where there is danger. Sanctioning people for crossing a line you admit you cannot draw makes sense only if the aim is to rule by fear. Hold that thought.

What does responsible use look like? Authors are told: ‘great care must be taken … to retain their own voice’. Misuse is to resort to AI ‘to write in a language authors would never themselves normally employ’. This makes sense for a native speaker: do not pretend to be what you are not. But for foreigners who must write in a language not their own, the injunction contradicts itself. The voice the policy protects is the one thing the journal’s own rules deny us.

Take me as the test case. English is my fourth language, exactly the kind I would never employ ‘normally’. Anglophone universities trained me to write it short, signposted, almost mechanically: topic sentences up front, hedges in the approved places, transitions from the approved list. The training worked: there has never been a voice in my English to protect; it was the first thing that had to go. On the editors’ own definition, everything I have written in English would be inappropriate use, with or without a machine.

The editors say they have thought of us: the policy recognises non-native speakers benefit from AI checking their English, even from its ‘translational, writing and editing services’. Yet AI involvement in writing ‘should amount, at most, to editing and proofreading assistance’, and must stop short of ‘large chunks of AI written text’. A machine-translated paper is AI-written from first word to last; it is nothing but such chunks. So the concession stands only if translating is not writing. Where does that line run? They have already told us: ‘no clear lines’.

There is worse, because the suspicion is not only unfalsifiable. It is loaded. In 2023, a Stanford team tested seven widely used AI detectors on essays written by humans: the detectors flagged, on average, 61% of TOEFL essays by non-native speakers as machine-generated, while classifying essays by native speakers almost perfectly; one detector flagged 97.8% of the TOEFL set (Liang et al., 2023). The reason is instructive. Detectors measure predictability, and second-language English, trained towards the mean, is predictable by design: the most probable word in the most probable order. The same study found the signal in real scholarship, with lower linguistic variability in conference papers by non-native first authors. The stylistic cues that read as ‘AI’ are the cues of competent second-language English, and a human judgement call reads the same cues as the machine. The double bind is complete: write my own English and look like a machine; use the machine and break their rule.

The paradox would be benign if it were not accompanied by threats. Yet the threats reveal a deeper, older problem that the policy never names: monolingualism. The submission guidelines ask for ‘the widest possible international coverage’ while, on the same page, ‘the language of the journal is English.’

The cost of that arrangement is not hypothetical. Surveying 908 environmental scientists, Amano and colleagues found that non-native English speakers spend up to 51% more time writing a paper, see it rejected 2.5 times more often and are asked to revise it 12.5 times more often, on language grounds alone (Amano et al., 2023). That tax is levied before any AI policy arrives; the declaration form and its threats are an audit added on top of it.

This is the axiom no one discusses. It is an imperial inheritance that acquired, with time, a practical justification: translation was slow and costly, and no board could assess what it could not read. Yet none of this survives advances in machine translation. State-of-the-art machine translation can be used as a reliable layer between authors and editors.

The ideal way out is human: scholars should read more than one language. Instead, anglophone universities prefer closing modern languages schools. Note that the closures are an Anglo-American phenomenon: scholars elsewhere never had the luxury of monolingualism. And note the timing. To close a language school is to place a bet, that English has won for good and that nothing worth reading will ever again be written in anything else. That bet is being made at precisely the wrong moment, twice over. First because machine translation is making it possible, for the first time, to do without a lingua franca altogether; second because the order that made English the default is receding: the anglosphere is turning inward, globalisation is ebbing, free trade is contested in its very homelands, and a multipolar world is taking shape in which the languages worth reading will multiply. Universities are producing scholars who read only English for a world where English will not be enough. Editorial boards will inherit that loss first.

When a board laments it cannot read its authors’ own voice, perhaps part of the problem lies with the board itself.

That is a serious charge, and it comes with a serious test. Would Progress in Human Geography referee a submission written in French, Italian or Japanese, accompanied by a machine-translated English version? The journal’s own guidelines recognise the ‘translational’ services that make it possible. A yes ends the problem. A no would tell us what the policy actually protects.

Either way, the choice is the editors’ to make. If the voice matters, as they claim, the journal should take papers in the languages authors think in. If it does not, the declaration form should go, and the threats with it.

Multilingualism | Clément Godbarge

Dear Editors: If You Want My Voice, Give Me Back My Language