There is something slightly odd about the current backlash against generative AI.
In academia, concerns are increasingly raised about Large Language Models (LLMs) flattening linguistic diversity, standardising expression and eroding the distinctive voice of individual writers. In software development, meanwhile, "vibe coders" — people who use AI to build software without necessarily possessing traditional programming expertise — have attracted considerable criticism from experienced developers.
Both groups have legitimate concerns.
But I cannot help wondering whether there is another factor sitting rather uncomfortably beneath the surface: professional insecurity.
When a technology suddenly allows outsiders to perform tasks that once required years of specialist training, those already inside the profession are unlikely to regard it neutrally.
That does not mean the criticism is wrong. But it may help explain some of its intensity.
The democratisation problem
Academic writing and software engineering have historically had relatively high barriers to entry.
To participate successfully in academic publishing, one must understand not only a subject, but also its conventions: disciplinary language, referencing systems, methodological expectations, journal structures and the peculiar rhetorical style of scholarly prose.
Programming has its own equivalent barriers: syntax, algorithms, debugging, architecture, frameworks and years of accumulated technical knowledge.
Generative AI weakens some of those barriers.
Someone who cannot write polished academic English can now substantially improve their prose. Someone who cannot program traditionally can describe what they want and, increasingly, produce a functioning application.
That is enormously empowering.
It is also potentially destabilising to professions whose status has partly depended upon possessing scarce technical skills.
The reaction is therefore unlikely to be entirely about quality. It is also about what happens when previously exclusive capabilities become widely accessible.
The linguistic diversity argument
One of the more interesting criticisms of LLM-assisted writing concerns linguistic homogenisation.
Large models naturally gravitate toward statistically common forms of expression. Asked repeatedly to "improve" prose, they can push writing towards smooth, grammatically conventional and remarkably similar sentence structures.
Individual eccentricities disappear.
Regional expressions disappear.
Unusual turns of phrase disappear.
Eventually we risk arriving at something resembling a universal form of polished internet English.
That concern deserves to be taken seriously. AI assistance does have a tendency to strip away stylistic variation and individual voice.
But there is an irony here.
Academic writing is already highly homogenised
Academia is a peculiar place from which to launch an unqualified defence of linguistic individuality.
Pick up twenty empirical psychology papers and their structures will often be remarkably similar.
Introduction.
Method.
Results.
Discussion.
The language will tend to be formal. Personal expression is constrained. Discipline-specific terminology is expected. Journals impose formatting rules. Reviewers regularly recommend removing colloquialisms, rhetorical flourishes and anything judged insufficiently "academic".
Academic prose is governed by conventions that are themselves almost algorithmic.
Academia therefore cannot simultaneously insist that authors conform to highly standardised writing conventions and then act surprised when machines become exceptionally good at reproducing them.
Perhaps what academics are actually worried about is not the loss of linguistic diversity.
Perhaps they are worried about the loss of evidence of thought.
And that is a much stronger argument.
Writing as evidence that thinking occurred
Academic writing serves two purposes.
The obvious one is communication: presenting an argument, experiment or theoretical position clearly enough for another person to understand it.
But writing also performs a less obvious epistemological function.
Writing forces us to organise thought.
Anyone who has written a difficult paper knows the experience. You think you understand an argument perfectly well until you attempt to express it in 200 words. Suddenly contradictions appear. Definitions become slippery. Claims that seemed obvious require evidence.
The struggle to write is partly the struggle to think.
This provides a more compelling reason to be cautious about LLMs.
The problem is not necessarily that the final sentence has been polished by a machine. The problem occurs when the machine performs so much of the intellectual structuring that the writer no longer has to confront those difficulties themselves.
There is an important distinction between:
AI helping me express what I have thought
and
AI doing the thinking that I subsequently claim to have expressed.
That distinction is considerably more useful than simply asking whether AI was "used".
And then there are the vibe coders
Software development is experiencing a remarkably similar argument.
"Vibe coding" broadly describes developing software by telling an AI what you want, running the resulting code and iteratively asking the model to fix or change things — sometimes without the user fully understanding the underlying code.
To an experienced programmer, this can be horrifying.
And sometimes for good reason.
Software that appears to work is not necessarily well engineered. Security problems, inefficient architecture and maintenance difficulties may remain invisible until much later. Rapid AI-assisted development may transfer work downstream to the people eventually responsible for understanding and maintaining the system.
But once again, there is another side.
Vibe coding allows people who previously could not create software to turn an idea into something tangible.
Researchers can build experimental tools.
Teachers can create educational applications.
Small businesses can automate processes.
Individuals can prototype ideas without employing an entire development team.
Dismissing all of this because the people involved are not "real programmers" starts to sound rather like academics complaining that someone is not a "real writer" because they used an LLM to improve their English.
The appropriate question surely isn't:
Did you use AI?
It is:
Do you understand what you produced sufficiently for the context in which you intend to use it?
For a prototype running on someone's laptop, the answer required may be very different from software controlling a medical device or processing financial transactions.
Could Small Language Models offer a compromise?
This brings me to Small Language Models (SLMs).
Most public discussion of generative AI revolves around increasingly enormous cloud-based systems attempting to know something about almost everything.
But academia might ultimately benefit from moving in the opposite direction.
A smaller model can potentially be deployed locally and adapted towards a narrower domain, institutional context or body of literature. Instead of relying entirely upon a global linguistic average, models could be designed around particular disciplines or purposes. Localised models could retain domain terminology and operate within a more defined stylistic ecosystem.
Imagine, for example, a psychology-focused model whose purpose was not to write your paper for you, but to:
- check clarity;
- identify ambiguous methodological descriptions;
- flag inconsistent terminology;
- challenge unsupported claims;
- check whether an argument follows logically;
- or suggest where prose has become unnecessarily opaque.
That feels much closer to intellectual scaffolding than ghostwriting.
And importantly, a locally deployed model could potentially give researchers greater control over both their data and the behaviour expected from the system.
This is a move away from the model as an all-knowing "oracle" and towards AI as a customisable tool.
The real question is agency
I suspect the academic AI debate has become too preoccupied with detecting machines.
A more useful question is whether the human researcher remains intellectually responsible for the work.
Did they choose the question?
Did they understand the literature?
Did they design the methodology?
Did they interpret the findings?
Can they defend the argument?
Can they explain why each conclusion follows?
If the answer is yes, then obsessing over whether an LLM helped restructure a paragraph starts to look increasingly difficult to justify.
Similarly, a programmer using AI does not cease to be responsible for the software merely because a model generated part of the code.
AI does not remove responsibility.
If anything, it makes intellectual ownership and accountability more important.
Perhaps we are defending the wrong thing
There is a danger that universities will attempt to preserve the outward appearance of pre-AI scholarship rather than the intellectual values underneath it.
Academic prose has never been valuable because it is difficult to write.
Programming has never been valuable because syntax is difficult to learn.
Their value lies in what those skills allow people to accomplish.
If AI removes some of the mechanical difficulty while leaving human judgement, creativity, scepticism and responsibility intact, that may be progress rather than decline.
And yes, some resistance probably is driven by job insecurity and professional gatekeeping. History suggests that professions rarely welcome technologies that weaken their monopoly over specialist skills.
But dismissing every criticism in those terms would be equally simplistic.
There are genuine questions about intellectual dependency, software quality, linguistic homogenisation and the outsourcing of cognitive work.
The interesting question is therefore not whether we should choose between humans and AI.
It is what kind of AI use preserves the parts of human expertise that actually matter.
Perhaps the future does not belong to ever larger models writing everything for us.
Perhaps it belongs to smaller, specialised tools that help us think without pretending to think on our behalf.
And, somewhat ironically, that may ultimately preserve far more intellectual diversity than simply banning the technology ever could.