The contributions for LitNet’s mini-seminar on AI, “AI: Catastrophe or Opportunity?”, are available here.
There has been considerable controversy over allegations that some writers, including black writers, have used artificial intelligence to generate their work. To be fair, white writers have also faced such accusations. Yet I have the impression that the allegations attract particular attention when directed at black writers, especially those who have won literary prizes or achieved a degree of public recognition. I cannot prove that race accounts for this disparity, and there may be other explanations, including the growing suspicion surrounding literary prizes themselves. Nevertheless, one cannot entirely dismiss the possibility that racial prejudice plays a part. In a literary culture that has never been free of questions about who possesses the authority to speak, whose work is considered worthy of recognition, and whose achievements are readily believed, accusations of AI use can become another means of casting doubt on a writer's legitimacy.
I suspect that few writers working today can entirely escape the possibility of being accused of using AI, whether or not they have done so. The difficulty is that there is no reliable way of establishing authorship simply by submitting a piece of writing to an AI detector. I have experimented with several such tools, including those available on my phone and Substack's text-scanning features, with frustrating results. On different occasions, passages I had deliberately copied from AI applications were classified as human-written, while passages I had written myself were flagged as potentially AI-generated. Such inconsistencies make it difficult to treat these tools as anything more than fallible indicators.
...
Even the better detectors cannot establish authorship with certainty, and their results should never be treated as conclusive evidence of misconduct, particularly when a writer's reputation or livelihood is at stake.
...
I have heard that paid services such as Pangram, Winston AI and GPTZero perform better than some of their competitors, although their accuracy varies according to the text, the language, the model and the conditions under which they are tested. I have not been sufficiently invested in the detection business to pay for several competing subscriptions and conduct a systematic comparison. Even the better detectors cannot establish authorship with certainty, and their results should never be treated as conclusive evidence of misconduct, particularly when a writer's reputation or livelihood is at stake.
There is, of course, a difference between detecting a pattern and proving its origin. Anyone who has spent a reasonable amount of time working with generative AI begins to recognise some of its recurring habits: the conspicuously balanced sentence, the predictable progression from assertion to qualification, the over-explained conclusion, the fondness for certain abstract nouns, and the tendency to make every observation sound more consequential than it really is. One encounters familiar phrases, carefully arranged contrasts, and paragraphs that seem to have been polished until all their irregularities have disappeared. These habits may alert an experienced reader to possible AI involvement, but they are not proof. Human writers have always relied on conventions, and some have cultivated precisely the kind of smooth, formulaic prose that AI now produces so efficiently.
There is another complication. Many writers, particularly those for whom English is not their first language, use AI applications to correct grammar, identify typographical errors, improve sentence structure or clarify an argument. Others use them to organise research, test an idea or identify weaknesses in a draft. These activities are not necessarily equivalent to asking a machine to write an entire story, essay or novel. To collapse them all into the category of AI-generated writing is to ignore meaningful distinctions between assistance, collaboration and the wholesale delegation of authorship.
...
AI has a tendency to flatten prose in pursuit of what might be called manufactured universal appeal.
...
Indeed, I often find AI-assisted writing more laborious than writing unaided. The initial result may look impressive, but a careful writer still has to examine it, correct its inaccuracies, remove its clichés and restore the distinctive qualities of the original voice. AI has a tendency to flatten prose in pursuit of what might be called manufactured universal appeal. It smooths out awkwardness, qualifies strong assertions, supplies transitions where none may be needed, and replaces particular observations with generalisations that sound reasonable but say very little. The result is frequently competent without being memorable. It is polished, but the polish comes at the expense of character.
This brings me to the question of universal appeal. I have never quite understood why writers should aspire to a single, supposedly universal standard of literary expression, particularly when that standard is often associated with the prevailing conventions of American creative writing programmes or the familiar house style of The New Yorker. There is nothing inherently wrong with either tradition. Both have produced distinguished writing. My objection is to the assumption that literary excellence must conform to a recognisable set of stylistic expectations: controlled sentences, measured cadences, carefully rationed emotion and a certain kind of cultivated ambiguity. Sometimes, even without AI, such writing can become wooden, mannered and curiously impersonal.
The irony is that prose written to satisfy these conventions may resemble the very language for which AI writers are now criticised. The sentences are grammatically impeccable; the paragraphs are logically arranged; the tone is restrained and professional. Yet something essential is often missing, which I call, for lack of a better phrase, “tone uniqueness”. The prose has no particular pressure behind it, no unmistakable rhythm of thought, no sense that this sentence could only have been written by this person, with this history, in response to this particular experience. AI did not invent these weaknesses. It has merely made their reproduction effortless.
I worry, therefore, that some of the current hostility towards AI writing rests on an inadequate understanding of what writing actually involves. There is a tendency to treat the use of AI as a moral failing in itself, without asking what the technology was used for, how extensively it shaped the final work, or what responsibility the writer retained for the result. Accusing someone of using AI is easy; establishing what happened in the production of a manuscript is considerably harder. Where there is evidence of deception, fabricated research or a breach of competition rules, the matter deserves serious attention. But suspicion alone should not be allowed to be substituted for evidence.
...
In historical writing, a fabricated reference can send a researcher down a blind alley; an invented quotation can distort the historical record; and a false claim, repeated often enough, can acquire the appearance of fact.
...
My own experience has been that using AI demands more work from a writer who cares about accuracy and style. The material it supplies must be checked, sometimes sentence by sentence. It can produce an argument that sounds plausible but is logically defective; a historical account that confuses people or events; or a reference to a book, document or scholarly article that does not exist. These are examples of what the technology industry calls “hallucinations”: outputs presented with confidence despite being false or unsupported. The term is almost too benign for the consequences. In historical writing, a fabricated reference can send a researcher down a blind alley; an invented quotation can distort the historical record; and a false claim, repeated often enough, can acquire the appearance of fact.
What discourages me most is not simply that AI can be wrong, but that its errors often arrive wrapped in the language of authority. It can produce long, apparently well-informed sentences without supplying the depth of understanding those sentences imply. It can mistake verbosity for substance, and a persuasive tone for knowledge. A writer who accepts such material uncritically risks surrendering the very judgement that makes writing an intellectual and creative act.
As a historical novelist, I also encounter a more fundamental limitation. Much of the material I need is not readily available on the internet, and some of it may never be digitised. It resides in archival collections, missionary correspondence, old newspapers, family papers, obscure publications and documents whose significance becomes apparent only when they are read alongside other sources. Historical research is not simply a matter of retrieving information. It requires knowing where to look, recognising what has been omitted, interpreting the language of a particular period and deciding whether a source can be trusted. AI may help identify a research trail, but it cannot be relied upon to replace the painstaking work of following that trail.
Indeed, I have sometimes had the disquieting impression that AI systems are drawing on material I have researched and published, while reproducing it with varying degrees of accuracy. In one instance, someone sent me a link to an AI-generated explanation of Queen Nojoli, based on a booklet I researched and wrote for the Jakes Gerwel Foundation. The response contained a significant genealogical error in identifying Queen Nojoli as the mother of King Ngqika, when she was actually his grandmother. I pointed out the mistake and made several other corrections. Gemini acknowledged them and indicated that it would update its response.
That exchange left me with mixed feelings. I was pleased that the error could be corrected in the immediate interaction, but I was less certain about what the promise to update actually meant. A correction made during a conversation does not necessarily mean that a model's underlying knowledge has been permanently altered, or that the same mistake will not recur in another conversation. Nor does it guarantee that the corrected information will reach everyone who has already received the original answer.
My concern is not merely that the technology might spread historical misinformation, but that the misinformation could be associated with my own research, effectively placing my name, however indirectly, behind an account I have never written and would not endorse. There is an important distinction between a system drawing on published research and that system accurately representing that research. A source may be acknowledged, yet its contents can still be misquoted, simplified or distorted. Attribution alone does not guarantee fidelity.
...
The problem is not simply whether AI has copied a particular sentence, but rather whether the knowledge, labour and interpretive authority embodied in a body of research can be appropriated and recombined without adequate acknowledgement, compensation or accountability.
...
This raises difficult questions about intellectual ownership and the relationship between published scholarship and generative systems. Researchers and writers invest years in locating, interpreting and contextualising material. Once that work enters the digital environment, it may be absorbed into systems that reproduce its substance without preserving the full chain of reasoning that produced it. The problem is not simply whether AI has copied a particular sentence, but rather whether the knowledge, labour and interpretive authority embodied in a body of research can be appropriated and recombined without adequate acknowledgement, compensation or accountability. These questions deserve more serious consideration than the simplistic division between writers who use AI and those who do not.
I have many reservations about AI-generated writing, but I would not describe myself as opposed to the technology. I suspect that in the not so distant future, its use will become increasingly difficult to avoid, whether in research, editing, translation, education or the ordinary business of communicating with one another. The more useful question is not whether we can somehow return to a world without it, but what standards of authorship, disclosure, accuracy and accountability we should establish for a world in which it is becoming commonplace.
There is a historical analogy worth considering, although it should not be pressed too far. In Plato's Phaedrus, Socrates recounts the story of the Egyptian god Theuth, who presents writing to King Thamus as a remedy for a lack of memory and wisdom. Thamus is unconvinced. Writing, he argues, will encourage forgetfulness, because people will rely on external marks rather than exercising their own memories. It will give them the appearance of wisdom without necessarily giving them understanding.
Socrates's objection is not quite a straightforward rejection of writing, and Plato's dialogue is subtler than a simple contest between an old technology and a new one. Nevertheless, it captures a familiar anxiety – that an invention designed to extend human capacity may weaken the very faculty it appears to strengthen. Writing does relieve us of some of the burden of memorisation. We no longer have to carry every genealogy, law, story or argument in our heads. Yet this externalisation of memory has also enabled knowledge to travel across generations, preserved arguments that would otherwise have disappeared, and allowed ideas to be examined, challenged and developed beyond the limits of individual recollection.
The invention of writing did not eliminate oral culture, although it transformed the conditions under which oral traditions survived and were transmitted. Nor did the printing press destroy the human need to tell stories. Each technology has altered the relationship between the creator, the audience and the means by which knowledge is preserved. AI may bring about a similarly profound change, though its implications are not yet fully understood. Unlike writing or printing, it can generate new passages of language in response to a prompt, imitating forms of reasoning and expression that previously required direct human effort. That difference matters, and it would be foolish to dismiss the dangers simply because earlier generations also feared technological change.
The challenge, then, is to distinguish the enlargement of human capacity from the abdication of human responsibility. A writer who uses AI to check a reference, identify a grammatical error or test an argument remains responsible for the final work. A writer who delegates the substance of a manuscript to a machine and presents the result as wholly their own raises different questions, particularly where a prize, publication contract or academic qualification depends on original authorship. Between these two positions lies a broad and complicated territory that cannot be governed by detection software alone.
For me, the decisive issue is not whether a machine has touched a piece of writing. It is what remains of the writer's judgement, imagination and responsibility after that encounter. Has the writer interrogated the material? Has the language been made to serve particular intelligence and experience, or has a distinctive voice been exchanged for a frictionless approximation of literary competence? Have the facts been verified? Are the sources represented faithfully? And, where the circumstances require it, has the writer been honest about the assistance received? Most crucially, has the writer delegated their writing and critical skills to the machine?
These questions matter because literature is more than the arrangement of grammatically correct sentences. It is the act of giving attention to a place, a person, a history, a private grief or a contradiction that resists easy resolution. In my own work, the archive may supply the evidence, but it cannot decide what the evidence means. A machine may help me locate a fact, but it cannot absolve me of the responsibility to understand it from my own Xhosa point of view, which the machine lacks. Nor can it decide which silences in the historical record matter, or what moral obligations arise from writing about the dead.
Perhaps this is where the debate should begin. We should be less interested in treating AI as a contaminant that can be detected by an electronic test, and more interested in defending the intellectual and artistic qualities that make writing worth reading. We should insist on accuracy, transparency where appropriate, respect for intellectual labour and accountability for what is published. Above all, we should resist the temptation to confuse a particular style with proof of either human or machine authorship.
The danger is not that AI will write sentences indistinguishable from those produced by some human writers. That has already become possible in certain contexts. The greater danger is that we will increasingly accept writing that sounds authoritative without asking whether it is true, writing that sounds polished without asking whether it has anything real to say, and writing that claims universality by erasing the particular histories and voices that give literature its force.
In my own opinion, the task before writers, then, is not simply to keep machines out of literature, but to preserve the human responsibilities that literature demands, whether or not machines have become part of the process. We may eventually find that AI, like writing before it, extends our reach in ways we cannot yet imagine. But the extension of a capacity is not the same thing as the cultivation of a faculty. We must still learn to remember, to judge, to imagine and, above all, to take responsibility for what we say.
See also:
KI-miniseminaar: Gebruik jy KI om jou met jou skryfwerk te help?
PenAfrican: AI and the future of creative arts – why writers are not as replaceable as we fear
KI en kopiereg: kreatiewe skryfwerk van die masjien versus kreatiewe skryfwerk van ’n mens
Hou KI ’n bedreiging vir skrywers in? | Etienne van Heerden Veldsoirée 2023
LitNet-AfriKI-miniseminaar: Kunsmatige intelligensie as skrywer
Seen elsewhere: AI, Artificial Intelligence, is at best purely Simulated Intelligence
KI-miniseminaar: Kunsmatige intelligensie as uitwissingsbedreiging?
Van mens tot masjien: Wikipedia en die voortbestaan van betroubare kennis