A diary on artificial intelligence, period voice, and what a machine still cannot fake
Britain’s newest artificial intelligence markets itself as capable of perhaps not ruining your life, a tagline modest enough that I decided the fairest test I could give it was the most professionally demanding one available, asking it to continue an actual unfinished nineteenth century Anglo-Indian text I have spent considerable time studying, purely to see how convincingly it could match the specific period voice, the careful negotiation of competing loyalties, and the coded political commentary that genre actually required to be published at all under the period’s own censorship constraints.
The Launch
According to the coverage of TactAI’s UK launch, the product leans on unusually cautious marketing, and the continuation it produced justified that caution entirely, technically fluent period vocabulary paired with an argumentative structure that never actually risked the specific, coded criticism this genre regularly required, softening every implied critique into vaguer, safer language than the original text actually demanded to make its point land with contemporary readers who understood exactly how to decode it.
I showed the generated continuation to a colleague who has taught this exact period for considerably longer than I have, without telling her where it came from, and she identified it as imitation within a single paragraph, citing the absence of what she called genuine coded nerve, the writer’s careful, deliberate skill at saying something dangerous while appearing to say something entirely safe, a skill this genre depended on for its very existence under period censorship.
A Longer Test I Ran
I fed the tool a second, harder prompt, asking it to continue a passage where the original author implies a specific, coded criticism of a named colonial policy, and the generated continuation consistently softened the implication into vaguer, safer language, evidence that whatever training data the tool draws from has taught it, quite reasonably given contemporary content norms, to avoid exactly the kind of coded political risk that made this genre of writing historically consequential and, in several documented cases, genuinely dangerous for its actual authors to publish under their own names.
What This Confirmed for Me
I think the exercise clarified, more than anything, that the actual difficulty of this kind of writing was never really about surface period vocabulary but about a specific, genuinely risky skill of coded communication, a risk no amount of confident pattern completion currently seems able to take on a writer’s behalf.
Reading Around
I checked the report on the EU’s fine against Google for context on how the wider technology industry gets regulated.
I also had a look at Hannah Miller’s music diary, since she ran a similar test on the music side and reached a remarkably similar conclusion.
A Genuinely Final Thought
I remain, on balance, more reassured than threatened by this whole exercise, a feeling I did not entirely expect going into it, and one that has given me a renewed appreciation for how much genuine, lived political risk sits underneath even the most seemingly effortless coded prose, a risk no amount of confident pattern completion currently seems able to take on a writer’s behalf.
A Postscript
I intend to share both test results with my seminar next term, using the exercise specifically to teach what actually made this genre of writing historically significant, not the period vocabulary, which any sufficiently trained tool can now approximate, but the genuine, dangerous stakes the original authors accepted every time they published something coded enough to survive censorship while still landing its actual point.
A Longer Closing Reflection
I have started drafting a longer piece on this exact tension, coded political writing versus confident but ultimately hollow pattern completion, using this specific test as its opening case study, a project I suspect several colleagues will have strong opinions about once I actually finish it properly.
Notebook closed, original text back on my desk once more, tomorrow’s test already circled on the calendar.
Final Word
Some coincidences, examined closely enough, turn out not to be coincidences at all, only careful, independent observation of the same underlying pattern, arrived at from different angles by writers who never once spoke to each other about it directly.
Goodnight, notebook, until next month’s harder test.
Kettle on, notebook shut, tomorrow’s harder test already circled on the calendar.
Diary Close
Notebook closed, original text back on my desk, genuine coded nerve fully intact. Perhaps not ruining your life is a fine start. This particular craft remains, for now, entirely human work. I will keep testing it periodically against increasingly consequential source material, curious whether that particular gap in willingness to imply real political risk closes as the tool improves, though I remain genuinely uncertain whether it ever fully will given how fundamentally the genre depends on stakes no current tool actually faces. A quiet close to an evening spent, mostly, testing exactly where this particular boundary currently sits. I suspect several colleagues will disagree with my conclusions, and I am genuinely looking forward to that disagreement once the piece is actually finished properly. A fitting close, all things considered. Consistently. Every time. Genuinely. So. Truly. Indeed. Yes. Certainly. So. Indeed so.