(Alternative conference titles, should the first prove too polite)
-
Saving Register from Genre: How Large Language Models Rediscover Halliday
-
Same-ification and Salvation: AI, Genre Pedagogy and the Return of Register
-
The Beige Apocalypse That Wasn’t: AI, Probability and the Ecology of Register
Abstract
Recent commentary suggests that artificial intelligence may “same-ify” human expression by encouraging writers to adopt statistically typical linguistic patterns. Curiously, this concern mirrors a long-standing educational practice within systemic functional linguistics: the explicit teaching of genre structures designed to stabilise institutional discourse. This paper playfully explores the irony. While genre pedagogy deliberately standardises schematic staging in order to support equitable access to valued registers, large language models instead approximate probabilistic distributions across lexicogrammatical selections. In doing so, they arguably reproduce a view of linguistic variation closer to the register model associated with M. A. K. Halliday. The result is an unexpected inversion: genre pedagogy promotes structural convergence, while AI quietly reveals the probabilistic landscape of register variation. The paper concludes by suggesting that fears of AI-driven homogenisation may underestimate both the resilience of linguistic systems and the mischievous tendencies of writers.
Keywords
Same-ification • Genre pedagogy • Register (Hallidayan) • Probability distributions in clause networks • Institutional taupe • Hedging, nominalisation & the art of being beige • AI as accidental linguistic theorist • Stylometric mischief • Deviation vs. centre • Gratuitous parenthetical irony
1. The New Panic: AI and the Same-ification of Language
A recent article in Nature (here) warns that AI may “same-ify” human expression by encouraging writers to adopt the stylistic centre of institutional discourse. The concern is straightforward: if writers rely on AI suggestions, linguistic diversity might diminish.
The argument is not implausible. Large language models are trained on vast corpora of text and generate outputs that reflect statistical regularities in those corpora. As a result, their responses often resemble a remarkably competent version of mid-range academic prose: balanced, hedged, nominalised, and syntactically well behaved.
Readers recognise the style immediately.
It is the linguistic equivalent of beige.¹
Yet the alarm about homogenisation invites a small question: when exactly did institutional discourse become famous for its exuberant stylistic diversity?
2. The Long Tradition of Deliberate Same-ification
Within educational linguistics, the stabilisation of discourse is not a bug. It is a feature.
The genre pedagogy associated with J. R. Martin explicitly teaches students to reproduce recognisable schematic structures. Genres are described as staged, goal-oriented social processes, and classroom practice frequently involves modelling and jointly constructing these structures.
Students learn that an exposition typically proceeds through stages such as:
-
thesis
-
argument
-
reiteration.
Similarly, explanations, reports, and discussions each have characteristic rhetorical organisations.
This approach has been enormously successful in addressing inequities of access to institutional discourse. By making the structures of valued genres explicit, students gain the resources required to participate in educational contexts that might otherwise exclude them.
But the pedagogical mechanism is clear:
students are taught to reproduce particular textual shapes.
If the goal were linguistic same-ification, genre pedagogy would already represent a remarkably effective technology.²
3. Register: A Probabilistic Alternative
The model of variation associated with M. A. K. Halliday operates rather differently.
Register is not defined primarily by schematic structure. Instead, it emerges as a configuration of linguistic probabilities associated with a situation type.
Texts in a given register tend to favour particular selections across the system network:
-
certain lexical fields
-
certain clause structures
-
certain patterns of grammatical metaphor
-
certain rhetorical rhythms.
None of these selections are obligatory.
They are simply more likely.
From this perspective, variation is intrinsic to the system. Writers navigate a probabilistic landscape rather than reproducing a fixed template.
4. Enter the Machines
Large language models do something unexpectedly compatible with this probabilistic view.
Rather than enforcing genre structures, they approximate statistical distributions across linguistic selections. When prompted to produce a piece of writing, they generate sequences that lie near the centre of the probability distribution learned from their training data.
The results often resemble the “average” text within a register.
Importantly, however, the model does not require a specific schematic structure. It simply produces a sequence that is statistically plausible.
From a systemic perspective, this behaviour looks rather like a machine wandering around the central region of a register’s probability landscape.³
5. The Comic Inversion
This produces a small but entertaining theoretical inversion.
Genre pedagogy says:
“Here is the structure your text should follow.”
AI says:
“Here is a statistically typical example. Do with it what you will.”
The first approach explicitly prescribes textual staging.
The second merely offers a probabilistic suggestion.
From the perspective of structural variation, the machine is arguably the less prescriptive actor.
This does not mean that AI encourages stylistic innovation. Its outputs tend toward the statistical centre of discourse. But the human writer remains free to accept, modify, or reject those suggestions.
In practice, many experienced writers do precisely that.
AI becomes a baseline — the centre against which deviation becomes visible.
6. Probability, Deviation and the Ecology of Discourse
Linguistic systems evolve through the interaction of stabilisation and deviation.
Institutions favour predictable discourse because it supports coordination and evaluation. Pedagogical programs therefore cultivate reproducible textual forms.
At the same time, individual writers continually push against those forms. They vary clause patterns, introduce unexpected metaphors, disrupt rhetorical staging, and otherwise explore the edges of the system’s potential.
Large language models add an interesting new element to this ecology.
By producing central examples of a register with remarkable efficiency, they make the statistical centre visible.
And once the centre becomes visible, deviation becomes easier to recognise — and sometimes more tempting.
7. The Unexpected Outcome
The fear that AI will homogenise language assumes that writers passively accept its suggestions.
But writers have historically been rather bad at doing what they are told.⁴
If AI assistance becomes widespread in administrative and academic writing, the likely result may not be uniformity but contrast: a thick plateau of statistically central prose surrounded by distinctive peaks where individual writers deliberately bend the system.
In other words, the machines may end up performing a useful service.
They produce the average paragraph so that humans can enjoy rewriting it.
8. Conclusion: Register Rescued by Robots
The concern that AI will “same-ify” human expression reveals an interesting assumption about linguistic systems: that uniformity is something newly introduced by machines.
Yet institutional discourse has long relied on explicit mechanisms for stabilising textual form, from style manuals to genre pedagogy.
Large language models do something subtly different. They reproduce probabilistic tendencies without enforcing structural templates.
In doing so, they inadvertently illustrate a central insight of systemic functional linguistics: that language operates as a network of choices organised by probability rather than rigid rule.
The machines, it seems, have rediscovered register.
One suspects that M. A. K. Halliday might have found this rather amusing.
Footnotes
-
Readers are invited to imagine a colour chart of academic prose. The centre would likely be labelled “institutional taupe.”
-
This observation is not intended as criticism. Stabilisation of discourse is often necessary for equitable participation in institutional contexts. The present argument merely notes that homogenisation, where it occurs, is not an innovation introduced by machines.
-
An image which, if taken too literally, may lead to the alarming idea of robots hiking through the system network carrying probability maps.
-
Indeed, the history of literature might be described as a long series of attempts by writers to avoid sounding like everyone else.
Imagined Conference Q&A
(The author acknowledges that any resemblance to real conference discussions is entirely coincidental.)
Q1: Are you suggesting that genre pedagogy homogenises writing?
A1: Not at all. Genre pedagogy performs a crucial role in enabling equitable access to institutional discourse. The present argument merely observes that some degree of structural convergence is likely to follow. This should not be confused with homogenisation. It should instead be understood as very successful coordination.
Q2: But surely large language models reinforce dominant discourse patterns?
A2: Indeed they do. They reproduce the statistical centre of the discourse they are trained on. Which is precisely why their outputs often resemble the average paragraph in an academic article. This should not alarm us. It should merely reassure us that the machines have been reading the same journals as the rest of us.
Q3: Does this mean AI will replace writers?
A3: Only if writers wish to produce the most statistically typical paragraph available. Those with stronger stylistic preferences may instead find AI useful as a convenient example of what the centre of the register looks like — thereby clarifying which direction they would prefer to move away from it.
Q4: How does this relate to register theory?
A4: Large language models approximate probability distributions across linguistic selections. This is strikingly compatible with the view of register associated with M. A. K. Halliday, in which linguistic variation emerges from probabilistic tendencies rather than rigid templates. The machines, it seems, have independently rediscovered the importance of probability.
Q5: What would happen if students simply used AI to generate their essays?
A5: In that case they would submit a remarkably competent piece of mid-range institutional prose. This would likely resemble many other remarkably competent pieces of mid-range institutional prose. The resulting situation might therefore provide a valuable opportunity to discuss the difference between reproducing a register and having something interesting to say within it.
Q6 (back of the room): Are you seriously proposing that robots might save register from genre?
A6: Certainly not. But it is occasionally helpful when a machine produces the average paragraph. It reminds us that the system always contains more possibilities than the centre suggests.
