AI

203. Will AI really kill us all? 

~ Neil MacDonald ~ 6 Comments

They have an acronym for it: P (E)—the Probability of Extinction (sometimes called P Doom). It hit the news again this week when Jacob Coxon, a researcher from the AI firm Anthropic, put the P (E) of the technology wiping out the human race at 10% https://www.theguardian.com/technology/2026/sep/09/ai-superintelligence-risks-warnings-scientists-politicians

What should we make of this? The first thing to say is that this is not a probability like the risk of dying in a car crash (around 1 in 100 over a lifetime in the UK) or from a meteor strike (1 in 700,000). We can estimate these probabilities because there’s a lot of data from past such events, and because we understand the processes involved.

There is no data behind the estimates of AI killing us all. There are no similar past events, and our understanding of the processes involved is sketchy. These are not probability estimates but scenarios, stories if you like.

But if these warnings are coming from people in the industry, shouldn’t we assume they know what they’re talking about? Maybe, but maybe not. It depends whether they’re speaking from within a corpus of expertise. When the team building the first atomic bomb in Los Alamos during the Second World War recognised, there was a chance the explosion could ignite the atmosphere, this was a potential risk indicated by the physics. The risk from AI comes from no such body of science. If anything, it’s an older body of fear: the fear of the End Days. It also needs to be said that invoking awe of the power of the technology probably does no harm to the prospects for Anthropic’s imminent stock market floatation (expected to be in the region of one trillion dollars), though this may be a cynical suggestion.

The term Artificial Intelligence is a misnomer. Chat-GPT and Claude and other AIs are not intelligent—they’re pattern recognising machines. This is what makes them so good at solving problems, like how proteins fold or how to construct sentences from prompts. They’re not smart (albeit not as smart as us yet) because that’s not what they do. They can find patterns that humans don’t see, and they can do it very fast. But this is not thinking.

For AI to live up to the Terminator threat, we’d need a different kind of machine—an Artificial General Intelligence (AGI), one that can think and act like (or better than) a human.

Many teams are trying to build an AGI, but nobody’s quite sure what the term means https://www.newscientist.com/article/2432005-what-is-artificial-general-intelligence-and-is-it-a-useful-concept/, or even whether it means anything at all. There’s no scientific consensus on what intelligence is, even in living things. Nobody is clear how soon, if ever, an AGI will be built. The 10% P (E) scare looks like hokum. This is not to say there is no threat, but we simply can’t say how big or when.

To become a Terminator, AI would need to be able to think, form plans antithetical to ours, carry them out, and evade counter-measures. None of this is on the immediate horizon. Sure, as a writer, I could invent lots of scenarios. But it doesn’t make them real.

Does this mean we shouldn’t worry? No, there’s lots to worry about, but it’s not the stuff making the headlines:

  • UNEMPLOYMENT. AI is already threatening millions of jobs, from call-centres to radiographers, from actors to lawyers.
  • SERVICES. Power-hungry data centres may threaten electricity supplies to ordinary people.
  • MISINFORMATION. In addition to deep-fakes, AIs are prone to “hallucinate” (make stuff up). The spread of chatbots use may also spread this misinformation.
  • INEQUALITY. The technology is being developed by large corporations, concentrating more wealth in the hands of a tiny elite.

If an extinction event is not foreseeably on the cards in the immediate future, the same is not true of accidents, catastrophic and small. As more and more systems (including critical infrastructure) are run by AIs, glitches, outages and sabotage may take them down. As a trivial example, I have a smart hot-water tank that’s supposed to economise by learning my use pattern. I’ve had to take it offline because its learning was rubbish.

Arguably, all the End Days hype is distracting us from the real problem.

202. AI or human?

~ Neil MacDonald ~ Leave a comment

The witch hunters walk the Earth again, no longer rooting out the devil but hunting the machine. Let’s take an example: Jamir Nazir’s story, The Serpent in the Grove. The story won the 2026 Commonwealth Short Story Prize, and was then accused of having been written with the help of AI. The author denies this.

Certainly. the prose groans under the weight of metaphor it carries.


Sun on galvanise is a cruel instrument. It beats until the roof talks back in a dry moan. The day the grove began to remember, the roof over Vishnu Mohammed’s shack groaned like a drumskin too tight for the heat. Inside, air clung thick as porridge skin: damp earth, woodsmoke, and the sour tang of fermenting cocoa. A soot-blackened lamp hung from a nail. No fan, no bulb, no hum – only the thin light slipping between warped boards and the breath of hills holding their heat like a secret.”


But is this the mark of the machine or the author’s voice? There’s no way to be sure, but there are known “tells” in AI prose.


• Uniform quality and style, together with repetitive phrases and sentences that are similar in length and structure. Human authors tend to write in “bursts” with varying structure.
• Continuity problems. AI may have problems with consistency, Characters, setting, and aplot points may change abruptly.
• Lack of emotional depth. AI may struggle with the nuances and quirks of human emotion and expression.

To these I’d add metaphors that sound metaphory but actually make no sense. An example from the story might be “breath of hills holding their heat like a secret.” Sounds good, but what does that actually mean? However, human authors may also create metaphors that misfire.


I ran the story through a couple of AI detectors, and was less than impressed by the result. GPTZero returned a 100% probability that the story was the work of AI. I listed among the most AI sentences these two: “Vishnu was twenty-five wearing the face of fifty” and “No fan, no bulb, no hum – only the thin light slipping between warped”. The first sentence I found to be one of the more powerful metaphors in the piece, The second is ungrammatical but very human to me,

Copyleaks also set the AI probability at 100%. By contrast, EyeSift put the probability that this was AI generated at 39% (no strong AI evidence).


My own feeling is the story is irritatingly elaborate, but also with a simple thread that’s powerful and haunting in a way that could only be written by a human.

188. App on review: How good is ProWritingAid’s Manuscript Analyser?

~ Neil MacDonald ~ 4 Comments

Regular readers of this blog will know I’m a big fan of ProWritingAid for spelling and grammar checking. Now it has branched out with a whole AI-powered manuscript critique service, and they offered me a free trial (normal charge £50 or $50).

Manuscript Analysis is broken down into five sections:

  • About My Story: Gives you key information about your story’s genre, narrative elements, and competitive landscape.
  • Narrative Themes: Highlights the key narrative threads throughout your story, and flags themes that are working well along with themes that could use some adjustments.
  • Plot & Structure: Highlights plot points that are working well, as well as those that might need improvement.
  • Characters: Examines important characterization moments and highlights areas where a particular character is working well or could use some closer examination.
  • Setting: Analyzes how your use of setting contributes to the overall narrative structure of your manuscript. This section flags if you need to improve any aspect of your setting, and what you might do to make it stronger.

So, how good is it? The short answer is “not great.” On the plus side, it showed a reasonable grasp of the story and quite accurately identified three comparable titles (two of which were among those I’d already chosen). The “unique narrative structure” was commended, though I’m hard put to it to understand how an alternation between two narrators might be a unique narrative structure,

Less impressive was the actionable feedback. There were 20 suggestions on plot and structure. Of these, only two were moderately helpful, and some were plain wrong.

Lest this seem to be a fit of pique on my part, I’ll give a couple of examples. The AI was troubled by:

“The timing of Ansna’s second pregnancy and miscarriage is unclear. It seems to occur shortly after the previous miscarriage, which feels rushed and lacks emotional weight.”

There is, in fact, only one miscarriage, something a human reader would have understood. Another chapter is said to have little consequence or outcome:

“The initiation ritual, while descriptive, lacks a clear impact on Ansna’s character development or the broader narrative. The lessons learned seem to have little consequence. Show how the ritual’s lessons influence Ansna’s later decisions or interactions.”

In fact, the character recalls such lessons in five subsequent chapters.

On Character, the AI notes three areas of concern, none of which I accepted. For example, a conflict between two characters is said to be “undeveloped”. This is because it’s a conversation, not a conflict. One of the three is illuminating:

“Ansna and Kautia’s relationship is inconsistently portrayed. Ansna expresses deep love for Kautia, but their interactions often lack warmth or genuine connection, and Kautia’s feelings remain ambiguous.”

I spent some time considering this comment before rejecting it. Ansna is conflicted in her feelings for Kautia. Again, I decided a human reader would not have read this as an inconsistent portrayal. Indeed, no human reader has made such a comment.

It’s worth noting that Kindlepreneur ran Alice in Wonderland through the Manuscript Analysis tool. Some of the issues were similar. Alice is said to lack emotional depth:

“Despite the bizarre events, Alice rarely expresses strong emotions. Her reactions are often muted, which makes it hard for the reader to engage with her experience.”

The driving force behind this criticism is, perhaps, the emphasis in modern Western novels on emotional exploration. But Alice is a Victorian upper-class girl, stiff-upper-lipped and confident of the social rules, even as they are buffeted by the absurdity of the Wonderland creatures. She behaves entirely consistently with her background and class. Emoting would be entirely wrong. This failure to grasp the nuances of the setting underlies another critical comment:

“The symbolic meaning of Alice’s experiences is not fully developed. The lack of thematic depth makes the story feel somewhat superficial.”

I would beg to differ. First, this is a book for children, so too much symbolic depth would be inappropriate. Second, and more important, the fundamental symbolism is abundantly clear in the repeated challenges to Alice’s sense of order by the other characters.

This leads me to my conclusion about the app. If the tool makes interpretations that no human would, this is no surprise. The AI does not “understand” a story. It merely follows algorithms that look for patterns of word associations its database says are probable. The net result, in the present stage of development of the technology, is to default to tropes. Though it is an impressive leap to be able to (more or less) follow a narrative arc over tens of thousands of words, my judgement is that the release of this tool is premature.  More work will be needed to give the tool a better grasp of context and psychology.