The Tower of B[A.I.]bel is a Flawed Building of our own Faulty Design
We trained everyone to write the same way, then built machines that do it perfectly. Now we're mad about it.
Standard English is fundamental to good writing, right? It’s clear. It’s what gets praised, and what signals education, intelligence and competence. It’s what’s expected. But standardised1 English isn’t even most people’s English. It’s a learned performance that some can execute more easily than others depending on their linguistic background.
Well… now there’s a tool that makes that performance almost entirely effortless.
Large language models can take anyone’s prompt and output a response in the standardised voice that gets rewarded by society. I would say that this is one of the main reasons why people use it so often for writing tasks. Because we’ve all spent decades in education systems being taught that HOW you express something matters as much as (and, let’s be honest, very often more than) the content of what you’re actually saying.
We’ve created a society that correlates surface-level linguistic features with intellectual competence. How often do you see people mocked online for simple grammatical mistakes variations? Everyone watches this happen and internalises the lesson: you’re language will be judged, and by extension so will you. It is entirely rational to avoid this shaming however you can. So how can we act shocked (SHOCKED!) when people reach for technology that helps them meet those linguistic expectations?
We didn’t just stumble into this problem. We built it through mass literacy, standard language ideology, and an education system that still confuses conformity with intelligence.
We Built this Tower
Mass literacy is a relatively recent development. Compulsory schooling over the last 150 years has fundamentally changed who could read and write by expanding literacy from a small elite to entire populations.
From a broad landscape of living Englishes the one chosen for instruction was a specific variety, one which was (and still is) the one used by people in power. As literacy expanded, more people learned to read and write in similar ways, and diverse linguistic practices were funnelled into a single, institutionally legitimised variety of English. Over time, this standardisation became so entrenched that fixed spelling and grammar “rules” are treated as natural, and linguistic variation is interpreted as error. But written language wasn’t always this rigid. Spelling variations used to be common and accepted, and even Shakspeare Shakspere Shackspeare Shakespeare’s name was spelled multiple ways.
More recently the internet appeared, creating vast, machine-readable archives of this education-enforced standardised language. We fed LLMs the data, and they feasted on trillions of examples of the kind of writing that overall has relatively little linguistic variation (compared to spoken World Englishes, which vary enormously). The resulting ChatGPT voice isn’t really as alien as people make out. It’s just our own standard language amplified, automated, and mirrored back at us.
We built this tower brick by compliant brick, weeding out the variants, each example of standardisation and conformity supporting the next, until we created something that speaks in one voice. A voice we programmed it to have.
The Myth of Neutral English
Attitudes towards language are shaped by assumptions about correctness and legitimacy. The way one variety comes to be treated as the standard is not through explicit agreement but through shared, largely unconscious beliefs that feel like common sense. Because these assumptions are so widely held, people rarely recognise them as ideological at all, assuming instead that virtually everyone agrees. This is standard language ideology.
In primary and secondary education, inspection frameworks demand “Standard English” and as a result enforce real, organisational consequences for daring to diverge from it. In universities, marking criteria often conflates correctness with intellectual quality: in England, the Office for Students and SPaG (spelling, punctuation, and grammar) requirements embed this ideology into institutional practice.
The consequences of structural enforcement of standard language are stark for those on the wrong side of it. These linguistically marginalised groups, with their identity-confirming and perfectly legitimate ways of languaging, are not valued in prestige spaces like education and business. This is something I struggled with in my own experience teaching academic English to international students who were very often from formerly colonised countries and who had English as their first language. These were insightful, intelligent students whether they could “perform” academic English or not. Despite already being English speakers, these students were required to invest significant additional time, money, and emotional labour simply to be recognised as legitimate participants in education and professional life. This asymmetry is not incidental; it reflects a colonial linguistic order in which certain Englishes are treated as gateways to prestige, and others as barriers.
The standard isn’t neutral. It never was. It is the language of the powerful made mandatory.
The Perfect Standardisation Engine
LLMs being adept at producing standardised English creates an impossible double bind for many writers:
Don’t use AI → risk being judged as unprofessional based on language “mistakes” which are mostly just variations that don’t really impede communication.
Use AI → risk being accused of cheating or inauthenticity based on “flat” language with no voice.
For many writers—particularly those from minoritised linguistic backgrounds, working‑class communities, and neurodivergent folk—standardised written English is not a neutral skill deficit to be remedied. It is a structural barrier that demands additional cognitive labour, time, and emotional energy on top of the intellectual work itself.
Using AI for language adjustment assistance should not be seen as a shortcut around thinking or effort. It is a compensatory strategy for navigating systems that continue to reward linguistic conformity over intellectual contribution. The choice to use AI assistance for surface-level language expression is both rational and ethical.
Homogenisation by Design
But isn’t this just flattening linguistic diversity? Aren’t we asking people to conform rather than bring their authentic selves? I don’t actually disagree that generative AI assistance with writing could be homogenising writing and risks eroding the expressive, cultural, and cognitive diversity that education ought to protect and encourage.
But this is not a new or uniquely digital problem. It’s what the system has always demanded. We’ve already been flattening linguistic diversity for over a century through standardised education. AI has just made the process more accessible to people who previously struggled to perform it.
The usual response at this point is: well, people should just put the effort in and become better writers. But this misses the point entirely because judgements about language correctness are inseparable from social power and prejudice. Becoming a “better writer” in this context almost always means becoming more fluent in a very specific, standardised written register—one that some people acquire with relative ease because it already aligns closely with their home language, schooling, and social background, and that others have to work much harder to approximate. Telling people to simply “work harder” at writing ignores these unequal starting points and quietly reframes structural inequality as individual failure.
Why wouldn’t someone who is linguistically marginalised use available tools to meet the discriminatory demands of linguistic expectation? People often approach me after I give talks on this topic, staff and students, non-native or stigmatised English speakers, neurodivergent folk, and they tell me that they use AI to help them express their worthy ideas in ways their institution will accept, but that they can’t talk about it openly because of the judgment or punishment they risk for using AI.
An Anarchist Linguistics Alternative
A twitter thread by Jon Hudson (kindly reproduced publicly here) in 2024 caught my attention because it used the term “anarchist linguistics”. In the thread Jon describes what “destandardization” entails. His central argument is that standardised English is the State wielding language as a weapon and that standardising language erases linguistic diversity from the communities where it developed and replaces those roots with the ideology of whoever holds power. It becomes a gatekeeping mechanism consistently used against working class speakers and people of colour who speak stigmatised varieties of English.
Destandardisation is necessary to push back against this. It doesn’t mean “anything goes” but is a shift from asking “Is this correct?” to “Did communication succeed?” Communication is a negotiation, but as Jon argues, standardised languages don’t allow negotiation on an even playing field.
In their article “After Whiteness” J.P.B. Gerald, Vijay A. Ramjattan, and Scott Stillar propose “counterprescriptivism”—moving beyond merely describing language variation to actively challenging correction itself. They argue that “meaning-making is ultimately a negotiation of power, and if [people] can convey the meaning they seek, they should be able to assert their intent in the face of possible correction.” This about dismantling the artificial hierarchies that treat standardised English as inherently superior.
Destandardisation and counterprescriptivism demand that readers and listeners have to do the work of understanding. Discomfort doesn’t equal incompetence. Have you ever tried to read a novel or article written in an unstandardised variety of English? It is difficult at first, but then you adapt if you stay with it, just as you do when speaking to someone with an accent you’re not familiar with. We can do this. We can understand diverse written form. We just choose not to in most institutional contexts.
In education what this change requires is smaller classes, more engagement with students, and knowing them as people. You can’t negotiate meaning when you’re marking 200 essays and don’t know the people who wrote them. You can’t ask “could you help me understand by explaining it a different way?” when there’s no space for that conversation to happen. Destandardisation demands resources and relationships we’ve systematically stripped from education. It also means allowing students to express themselves in linguistically diverse ways, encouraging translanguaging, and “not limiting students’ language resouces to those the instructor has prior competency in.” (Lee 2016, p. 190)
We need to abolish the grammar police, but also recognise that doing so well requires education conditions we’re not currently creating.
What do we actually want?
We can’t really have it both ways. We need to choose.
Option 1: Double down on standardisation. Keep the standard as the litmus test of intellectual competency. And then accept the consequences: continued AI use to conform, ongoing inequality, and writing that all sounds broadly the same.
Option 2: Move toward destandardisation and voice. Not only accept, but also value more diverse written repertoires. Judge by ideas and reasoning. Treat the standard as one language variety among many (as that is what it is).
If we want authentic voice, we cannot keep treating standardised English as the only legitimate voice. It’s that simple and that hard. Let’s be clear about the contradiction: you can’t complain that AI homogenises writing while treating stigmatised, unstandardised English as ‘bad writing.’ Either standard English is the goal (making AI assistance rational), or linguistic diversity is the goal (and standardisation is the problem).
Pick one.
We built this. We can choose what comes next.
We built this problem ourselves through mass literacy that equated education with a single style of writing, through standardisation that treats linguistic variation as error, through assessment systems that confuse linguistic conformity with intelligence.
Standard language ideology creates conditions where AI becomes the rational solution to an unjust demand.
So here’s the task for educators: re-examine your rubrics, experiment with destandardisation, talk (and empathise) with students about the linguistic tightrope they’re being asked to walk. For institutions: decide whether you’ll reward sameness (and accept the consequences) or reshape assessments to value genuine diversity of voice.
The Tower of B[A.I.]bel is a fallible monument to our societal choices about language. We built it one grammar rule at a time, and now it speaks back to us in our own flattened voice. We can keep building upward, layer after layer of the same, or we can recognise that true linguistic diversity is valuable and desirable.
Humanity is diverse. We know this. We could allow writing to fully reflect that.
Note on terminology: I use “standardised” and “unstandardised” rather than “standard” and “non-standard” following JPB Gerald’s practice in Anti-Social Language Teaching (2020, p. 8). This framing emphasises that standardisation is a deliberate process enacted by those in power, not an inherent property of any language variety. In the excellent book Other People’s English the authors use the term “undervalued English”. In one of my workshops about AI in the writing process I encourage people to view the distinction as prestige vs. stigmatised English.