Get By With a Little Help frpom AI

Not long ago I found myself asking an artificial intelligence system for a little help, in particular whether my articles sounded as though they had been written by artificial intelligence. I didn’t think there was anything particularly strange about this request at first because I was interested in the answer, and when you’re interested in an answer it’s surprisingly easy to overlook the absurdity of the question. It was only later, when I thought about what I’d actually done, that it began to niggle at me. Without much thought, I’d asked a machine to examine articles I knew I’d written and decide whether I still sounded sufficiently human. And this is apparently a perfectly normal use of modern technology. (Consequently nobody has burst through the door to check whether I’m all right mentally, and there hasn’t been any sort of intervention…yet)

This whole conversation started because I wanted to look at how my writing had changed over time. I’ve been writing blogs for around 8 years, and although I occasionally return to an older article because something has reminded me of it, I don’t normally sit down and read them in order. Once I’ve finished an article, checked it, published it and spent several minutes wondering why a sentence I revised more times than necessary still contains a wrong word, my attention usually moves elsewhere. Actually, to be fair, I then spend quite a lot of time looking at how many impressions and views it has received. And I’m usually disappointed.  Still, there’s always another subject that interests me, another claim that doesn’t seem quite right or another confident statement resting on more than shaky foundations. So I push my disappointment to one side and start on the next.

there’s always another subject that interests me, another claim that doesn’t seem quite right or another confident statement resting on more than shaky foundations

Looking back through the articles in date order, though, was more revealing than I expected. Some of the earlier pieces, including Ask a Silly Question and Call It What It Is, feel much more like thoughts being worked through on the page. They don’t begin by announcing a carefully constructed argument and then marching the reader obediently through it. They begin with something that doesn’t feel right. In Ask a Silly Question, the apparent simplicity of an unemployment figure begins to fall apart once questions are asked about what’s being measured, who’s included and what the number is really being used to represent. Call It What It Is starts somewhere quite different, with the language we use to make harm sound more acceptable, but it also develops by moving through examples and asking what sits underneath them. At this point I wasn’t trying to establish an overarching philosophy when I wrote either piece. I was just following an idea because it had caught my attention and wouldn’t leave me alone.

That’s still how quite a lot of my writing begins. Something interests me, surprises me or, quite often, irritates me. To be honest, irritation has probably made a greater contribution to my blog than any formal writing technique. I’ll hear somebody repeat a statistic that sounds far too tidy, watch a complicated problem reduced to a brightly coloured dashboard or read a strategy that declares success before anyone has agreed what success means (particularly true of AI). I then spend some time (probably not enough) wondering whether I’m being unfair. Occasionally I am. More often, the thing becomes less convincing the longer I look at it, and sooner or later I start writing because it’s more therapeutic than shouting at a PowerPoint presentation, which is what most of my real job entails.

Something interests me, surprises me or, quite often, irritates me. To be honest, irritation has probably made a greater contribution to my blog than any formal writing technique.

As the production of articles continued, some of the questions became more familiar even when the subjects appeared to change. When I wrote about truth, evidence, measurement, statistics, organisational behaviour or, later, artificial intelligence, I thought I was moving between different topics. I can see now that I wasn’t moving as far as I imagined. The same concerns kept appearing, although I didn’t always recognise them at the time. What are we actually measuring? What assumptions have slipped into the argument without being noticed? When does a provisional estimate become an established fact?

In The Truth Is Out There, the concern was how easily reality can be neatened, filtered and reshaped until it becomes more convenient. How Guesses Become Gospel looked at what happens when uncertain claims are repeated often enough to acquire an authority they never deserved. Breaking the Cycle moved into the consequences for organisations, particularly when decisions are made first, evidence is gathered afterwards and evaluation arrives at roughly the same time as everyone involved has moved to another job. It Doesn’t Add Up brought several earlier articles together and considered the difference between producing an answer and understanding one. And in Looking at the Right Thing, it was the danger of carrying out an analysis correctly while answering the wrong question. Of course there have been more articles written than this but they mostly follow the same themes and conclusions.

I didn’t plan that journey, or for my articles to link together in that way. I wish I could claim that I sat down years ago with a carefully drawn map showing how one article would lead into the next, but that would be nonsense (although might make me sound much more clever than I am). At the time, they looked like separate articles about separate matters. The links only became obvious when I took a step back and looked at the writing as a whole. To be fair, that was one of the more interesting observations to come out of my conversation with AI. It noticed that the later articles were increasingly connected, that they referred back to earlier pieces and that they drew together ideas I’d previously treated separately. It also noticed that the writing had become more structured and that I sounded more confident about certain subjects.

I wish I could claim that I sat down years ago with a carefully drawn map showing how one article would lead into the next, but that would be nonsense

So far, that all seemed reasonable – the AI response seemed to make sense. It’s obvious that if somebody writes regularly over a number of years, it would be rather disappointing if nothing changed. We would usually assume that continued practice leads to some improvement, although anybody who has attended enough organisational meetings will know this isn’t guaranteed (one thing I don’t miss being self-employed). I’d expect my writing to become clearer as I gained experience, to become better at recognising when I was repeating myself, even if I didn’t always manage to stop. I’d also expect the connections between ideas to become more obvious because I’d spent longer thinking about them.

But then the AI assessment started to go sideways. Some of those developments were described as features that might make the later articles look as though they’d been written, or at least heavily shaped, by AI. The clearer organisation, the greater confidence, the connections between articles and  the reduction in repetition were all treated as possible signs of AI involvement. Now there’s a problem with this conclusion because some of the articles considered had received no input from AI at all. They’d been written in the ordinary way, which in my case means having an idea, thinking about it for longer than is probably healthy, writing far too much, removing some of it, putting part of it back and eventually deciding that further interference is more likely to make it worse than better

And to be honest, I found that more annoying than I expected. There’s something deeply uncomfortable about spending years trying to improve as a writer, only to discover that improvement may be used as evidence that you didn’t write something yourself. We’ve spent decades telling people to organise their thoughts, make connections, avoid needless repetition and explain themselves clearly. I’m sure those were once considered signs of competent writing. But they’re apparently now grounds for suspicion. If an argument develops coherently, the writer may have had help. If the conclusion follows from what came before, something unusual has clearly occurred. (Perhaps the author understood the subject, although obviously we shouldn’t rush to outlandish explanations.)

We’ve spent decades telling people to organise their thoughts, make connections, avoid needless repetition and explain themselves clearly. I’m sure those were once considered signs of competent writing. But they’re apparently now grounds for suspicion.

This doesn’t mean that I think AI has no recognisable influence on writing. It often does. There’s plenty of writing online that appears to have been constructed from headings, neat lists and short declarative statements designed to land with tremendous importance while saying very little. Everything is a journey, every problem is a challenge, every challenge is an opportunity, everything appears quietly, and every opportunity comes with a kicker and three key takeaways. I don’t think the world has ever been so neatly organised. So I’m sure those articles are very efficient, but I’ve rarely reached the end of one while still caring what happens next.

The difficulty instead comes when we start treating any polished writing as artificial. A single article can certainly give that impression, particularly when it’s removed from everything surrounding it. If somebody reads Looking at the Right Thing without knowing anything about the earlier articles, they may see a deliberate connection between statistical ideas, measurement and AI and assume that the piece was built around a carefully supplied instruction. If they read It Doesn’t Add Up on its own, they may see several strands brought together into a single discussion and reach a similar conclusion. What they won’t see is that those strands had already been appearing for years. The article didn’t invent the pattern - it just recognised it.

This is one of the reasons I’m doubtful about attempts to identify authorship entirely from the style of one article. A piece of writing doesn’t always tell you where its ideas came from or how long they’ve been developing. It can’t show the reader all the earlier conversations, experiences, mistakes and unresolved irritations that sit behind it. It doesn’t explain that the author has returned to the same question repeatedly, sometimes without noticing. When one article is isolated from the rest, this sort of development can look suspiciously like something manufactured by AI.

A piece of writing doesn’t always tell you where its ideas came from or how long they’ve been developing. It can’t show the reader all the earlier conversations, experiences, mistakes and unresolved irritations that sit behind it. It doesn’t explain that the author has returned to the same question repeatedly, sometimes without noticing.

You see, AI can produce a plausible article about truth, evidence or measurement when asked to do so. But what it can’t readily reproduce is the unplanned development of a person’s concerns over years of writing. It doesn’t become annoyed by the same misuse of statistics in three different settings and then realise much later that all three examples are connected. It doesn’t re-read something it wrote years ago and discover that it was trying to express an idea it didn’t yet fully understand. It can obviously find relationships within material presented to it, but that isn’t the same as living through the experiences that caused those relationships to form. A longer body of work contains the history of somebody’s thinking, including the detours, changes of emphasis and occasional recognition that what appeared to be a new idea had been hanging around for years waiting to be noticed.

There was another part of the assessment that also bothered me. My later writing was described as containing less uncertainty, and this was also suggested as something that could make it appear more like it was AI generated. And I can sort of understand where that idea comes from. We know that AI systems are notorious for stating things confidently, including things that are wrong, hallucinated or so badly misunderstood that a sensible person would have stopped halfway through the sentence and gone to check. I have written enough times that confidence without knowledge is dangerous, and fluent confidence can be particularly deceptive because poor reasoning sounds much more impressive after somebody has made it sound convincing.

What I don’t accept though is that confidence in itself is evidence of artificiality. Sometimes people (especially professional people) just sound confident because they’ve spent years working in a subject and have seen the same problems repeatedly. Personally, I don’t feel much uncertainty about whether definitions matter in analysis because I’ve seen what happens when they’re ignored. I don’t need to remain undecided about whether measures affect behaviour because they clearly do. Now I’m not claiming that experience makes anyone infallible, and I’ve been wrong often enough to know better than that, but there comes a point when pretending to be uncertain becomes less honest than stating what you’ve learned.

What I don’t accept though is that confidence in itself is evidence of artificiality. Sometimes people (especially professional people) just sound confident because they’ve spent years working in a subject and have seen the same problems repeatedly.

We’re certainly at the stage where writers are now being advised to introduce more hesitation so that they sound human. We may need to weaken conclusions, add doubts we don’t possess and leave arguments slightly rough around the edges to reassure readers that a real person was involved. Having spent years encouraging people to write with greater precision, we can now offer them a new piece of guidance: try not to sound too sure about the subject you know well, because knowledge can look terribly artificial these days. I’m sure universities will adapt. I wonder whether universities will eventually advise students not to demonstrate too much knowledge of a subject in case it appears suspicious. Perhaps doctoral candidates will be encouraged to include a small but measurable amount of confusion to reassure examiners that a real person produced the work. There’ll probably be a postgraduate module in strategic uncertainty before long. And given some of the training courses I've attended over the years, I genuinely wouldn't be surprised if for the rest of us somebody proposed a framework for "authentic uncertainty" , and delivered it in a training workshop involving a great deal of discussion about authentic journeys. Horrifying, but I digress. Maybe a topic for another article.

This pretense is where the irony stops being mildly amusing and starts becoming irritating. People quite reasonably (and often quite vocally) say that they don’t want AI doing their thinking or controlling how they communicate or what they read. They don’t want “AI Slop” and want human judgement to remain central. Of course I agree with them. But if human writers begin changing their natural style of authorship because they’re afraid of sounding like AI, then AI is already severely influencing their communication. It doesn’t have to supply the ideas or write the article. Its presence is enough to make the human writer alter a sentence, soften a conclusion or insert uncertainty that wasn’t there before (and is often unnecessary or in the worst case, harmful). So we can continue insisting that AI isn’t in control, but it has surreptitiously persuaded us to rearrange the furniture.

All of which brings me back to this article because it would be rather dishonest of me to discuss the assessment of writing without turning the same attention on what I’ve just written. Now that you’ve read it - how likely is it that this article was written by AI? That question is harder to answer than the cheerful percentages produced by detection systems might suggest. If it’s considered in isolation, there are reasons it might attract suspicion. It explores a single broad idea, returns to earlier themes and links several previous articles together. It’s also more confidently expressed than some of my early writing because I’ve spent a considerable amount of time thinking about AI, evidence and critical judgement.

Against that evidence, I have told you that the article grew out of a real conversation that surprised and irritated me. It developed because an assessment of my writing changed when individual articles were viewed on their own rather than as parts of a longer journey. It returns to matters I’ve written about for years, not because they were supplied in an instruction, but because I seem unable to leave them alone. It probably spends longer getting to the point than an efficient editor would recommend, and there are thoughts here that could be shortened without damaging the argument. I’m leaving them because this is a blog, not an executive summary, and sometimes the route taken tells the reader as much as the final conclusion.

there are thoughts here that could be shortened without damaging the argument. I’m leaving them because this is a blog, not an executive summary, and sometimes the route taken tells the reader as much as the final conclusion.

Given the above I don’t think any reader (or any AI) can give an honest percentage for how much this article looks as though it was written by AI because there’s no reliable arithmetic behind such a number. You  could say five per cent, fifteen per cent or seventy-three per cent and make the figure sound appropriately scientific, but that would merely demonstrate another theme that has followed me through years of writing: numbers don’t become meaningful simply because somebody (usuallly AI) has presented them with confidence. If you ask ChatGPT it gives the answer:

“I'd rate the article approximately:

• Human-written with AI assistance: 60–70%

• Entirely human-written: 20–30%

• AI-generated then substantially edited: 5–10%

• Predominantly AI-generated: <5%

I would not regard this as an AI-generated article. I would regard it as a human article that openly incorporates AI into its development — and, ironically, one that is polished enough that a simplistic AI detector could very plausibly accuse it of being AI-generated.”

Not really very helpful, although it does contain an em-dash which is reassuringly AI!

My own position is that this article, reflects my established concerns, my earlier work and the gradual development of my writing, while also having been shaped through a discussion with AI about how best to express those thoughts. Whether somebody describes that as writing, editing, organisation or assistance will probably depend upon what they’ve already decided before reading it. I will still mark it as having been AI assisted because some of the content I have included has come from discussion with AI. That doesn’t bother me if people understand the context. Interestingly if AI involvement were based on a watermark like Claude then this wouldn’t show it. Again, a topic for another article,

Whether somebody describes that as writing, editing, organisation or assistance will probably depend upon what they’ve already decided before reading it. I will still mark it as having been AI assisted because some of the content I have included has come from discussion with AI. That doesn’t bother me if people understand the context.

What remains ironic is that writers are turning to machines to ask whether they sound human. Perhaps we don’t trust our own judgement, or perhaps we’re worried that readers won’t trust it. Either way, the machine that supposedly threatens human authorship is being invited to certify it. We write, the machine examines us, and if our humanity score is unsatisfactory we revise ourselves until it improves. Then, having changed our natural writing to satisfy an artificial judgement, we publish the result as more authentically human. Rather than AI shaping itself, it is shaping us.

And I’m really not certain where that leaves us, although I know it frustrates me. We may eventually become less concerned with superficial patterns and return to asking whether an article says anything worthwhile, whether its reasoning can withstand scrutiny and whether its ideas fit within the author’s wider work. I’d like to think so, I really do. Until then, writers may have to continue proving their humanity to machines, which is surely a task no human writer ever knew they needed.

So as the title of this article suggests, we’ve arrived at a peculiar reversal of responsibility, and perhaps the most accurate conclusion is also the simplest: We used to test machines to see whether they could pass for human. Now, human writing will only get by with a little help from AI.