A student submits a practice piece to the writing tutor in UniFluent, and a good deal of the feedback that comes back is written in the student's first language. Nobody in the company argued about that. It went in because feedback written in a language somebody reads easily is feedback that gets read.
There is also a setting for the correction standard, and British academic English is one of the options. We should say straight away that we have not established whether that setting governs the standard being applied or only the language the feedback is written in. Not knowing that about our own product is the sort of thing this exercise keeps turning up, and we would rather write it down than have it found.
The first of those facts is thinner than it looks, and we want to be careful about what we claim it proves. No theory produced it. It is a build decision that turned out to agree with a position we did not know we held.
The build came before the belief
Here is how we found out we held it. Asked near the start of this process what is actually changing inside somebody when they get better at English, we gave an answer in two parts. The first was that they have a better grasp of what is going on in English compared to their own language. The second, offered as the plainer version of the same thing, was that they know more words and understand how phrases in their language map onto English. Both halves are relational. Neither describes somebody who has swapped one language for another. Both describe somebody holding two languages at once and understanding each of them better because the other is there. We did not notice at the time that we had said the same thing twice in one answer.
Sixty-six questions later, with no warning that the earlier answer had been recorded, we were asked to agree or disagree with the statement that a student's first language is a resource rather than interference. We agreed without qualification, and said the realisation had happened over the course of this research. The ordering is worth stating precisely, because it is not flattering to us. The build came first. The instinct we now describe as a belief came second, and the belief itself came third, during a piece of work undertaken to test whether we had one.
Breaking the default without claiming the theory
This is the point where Phillipson's 1992 argument reaches us and does not land. He set out five assumptions embedded in how English was taught and called each of them a fallacy, and the first is the monolingual one: that English is best taught without the first language present (Phillipson 1992, p. 185). He is worth taking seriously partly because of where he was standing. He had spent nine years working for the British Council before he wrote it. The monolingual assumption is also the factory setting of commercial language software, which typically operates in the target language throughout and treats any other language as a scaffold to be removed. We inherited that setting and then broke it, for reasons that had nothing to do with Phillipson and everything to do with wanting our feedback to be legible.
We should be honest about the limits of what we can say next. We assumed that feedback which is easy to read gets read. We have no data showing that first-language feedback produces more learning than the same feedback in English would.
There is a name in the literature for the broader practice, translanguaging, which is García and Li Wei's term (2014) for drawing on a whole multilingual repertoire instead of keeping languages in separate compartments. We are naming it here in order to say that we are not claiming it. What we do is much narrower than what that word covers. Feedback arrives in the student's first language. We do not build lessons that move across a student's whole repertoire and we have no pedagogy for doing so. Calling one delivery decision by the name of a research programme would be exactly the sort of borrowed vocabulary this whole exercise is meant to avoid.
Kubota (2014) warns against celebrating plurality uncritically, and the phrase in her title for where the multi and plural turn can end up is neoliberal multiculturalism. Our own version of the worry is that a multilingual credential is cheap to announce and expensive to actually hold. A company that discovered its own multilingual credential halfway through writing a philosophy document should take that warning personally rather than cite it approvingly. We are one sentence away from putting "works in your own language" on a slide, and we know what that sentence would be doing there.
The fallacies we commit
Two of Phillipson's other four describe us accurately, and we would rather name them than have them named for us.
The native speaker fallacy (p. 193) lands on our stated goal. Asked what a seminar room would look like if the product worked perfectly, we said students would participate "as if they were doing it in their native language". Holliday named that ideology native-speakerism in 2005, and it is difficult to think of a plainer example than a sentence which sets somebody else's first language as the point of arrival. The belief underneath the sentence is defensible and we would defend it, because what we were describing was participation and not accent or score. The sentence we reached for to express it argues against the rest of this essay.
The same problem reaches the AI voice, where we have no position at all. Asked what accent it should use, the most we could say was that we imagine it is English, and we have not gone and checked. Making it selectable is possible. We do not have a reasoned account of which varieties of English should be on that list or why.
The maximum exposure fallacy (p. 209) lands harder. Our retention model assumes that more time in the app means more learning, and the streak is the machine built to run on that assumption. We lifted it from other language apps, and one of its purposes is to give us a reason to send a notification.
The only defence we have is a threshold, and it is a real one: a streak requires completing a whole lesson rather than a login or a single word, and we committed to that standard in this research before checking what the build actually did. The threshold survives the check. It survives a check made against our own account of the code rather than against the code, which is a piece of homework we still owe. And it does not touch the assumption underneath it. Nobody has used UniFluent for a hundred hours, so there is no efficacy data of any kind, which means that when we say more practice is better we are stating a belief in the grammar of a finding. That is the thing we most need to stop doing, and a threshold on a streak does not fix it.
If the first language is not the variable, our market is wrong
One more thing follows from all this, and it points somewhere uncomfortable for our own commercial plan. No-one's first language is academic English. Lea and Street (1998) made the case that academic literacy is plural and discipline-specific rather than a single transferable skill a student either has or lacks, and Wingate (2015) argued that provision built on that account belongs to every student rather than to international students alone. We did not arrive at either position. We found both convincing on reading them. What we would add is that the distance to travel varies, and that it varies by prior educational exposure more than by nationality. A student from a household where nobody read much has further to go than a student whose parents are researchers, and neither of them learned academic English at home. That is the reality of it.
If that is right, then a student's first language is not the interesting variable at all, and a product sold only to international students is aimed at the wrong half of the problem. We have not acted on this. Our route to market is international students, because that is where the institutional budget sits, and our stated belief and our sales plan do not currently agree.
Comprehension is not the same as learning
What we cannot tell yet is whether first-language feedback is doing any of the work we have just credited it with. The mechanism for comprehension is obvious: a student reads the feedback and understands it. The mechanism for learning is not, and the two are easy to confuse when the feature is already shipped. It is entirely possible that delivering feedback in the first language helps somebody understand a correction today and does nothing at all for the contrastive understanding we described as the thing that actually changes. We should also say plainly that the contrastive part is not built. Nothing in the product teaches a student how a phrase in their language maps onto English, which is the mechanism we named when asked what improvement consists of. Delivering feedback in a language is not the same as teaching across two of them, and we have not yet worked out what the second thing would look like.
The question we cannot answer honestly at this stage is what we would do if the evidence went against us. First-language feedback is now the part of the product that answers Phillipson best, and we noticed that after we built it. If a pilot suggested it slowed acquisition down, we would like to think we would remove it. We are aware that it is also the answer we most want to keep.
