Thursday, March 6, 2025

Oh no AI

 

A friend of mine ( I use the word friend in the widest possible sense ) has some weird ideas. He was telling me how good AI was and was ignoring my eyebrow of disapproval. He wagged his finger at me and said 'I’ll show you how good AI is.'



So he took my first novel Absolution, (which was shortlisted for the New Blood Dagger in the year of its publication. It's 120,000 words long ) Within 5 minutes AI had read the book and constructed a fifteen minute podcast where two people "discuss the book". 

He said, I would be amazed how accurate it was.

I was amazed at how wrong they were.

But wrong in a very interesting kind of way. 

I’m sure we have all listened to a panel where you know within five minutes that the chairperson has only pretended to read the books. You can sense the vague questions that try to provoke a precise answer.

In the AI version, the answers come back as the question rephrased. 

A deaf friend once taught me sign language for what comes out the back end of a bovine male and you can smell that a mile away when someone is pretending to have read your book. I have to confess I only listened to five minutes of the podcast because at that point I had to go and look up what book they were talking about. Because maybe I hadn't really remembered the book I had actually written.

Firstly, they stated on more than on one occasion that the entire book was based on the murder of Linda Cruikshank. A name I didn't recognise and was rather shocked as it's the name of my friend. Then I remembered she gave me a bottle of Prosecco if I killed her in the book. She's just the first body found, she has very little significance in the story. She's a bit player in the backstory. 

But AI took that murder as the baton of 'this is what the book was about' and kept running with it. As I write this blog I realise that giving the same book to AI could generate 4 or 5 different books as it totally misses the point of the one that's actually written!

AI couldn’t cope at all with the 23 year timeslip between the prologue and the rest of the book. Normal crime readers would recognise it  as a set up and be on the look out for a 23 year old character that might be that baby.  The podcast thought that the baby belonged to one of those characters in present day. Fortunately it didn't make the mistake of the baby being her own mother....

Secondly, the AI couldn’t get a hold at all of the way Scottish people talk to each other. It kept talking about the internal conflicts that threatened to pull the investigative team apart. In reality it was banter. I think there is one argument, but that’s over the ordering of a friend egg roll and that didn't interfere much with the investigation.

Thirdly, they kept talking about one of the character's decline into mental illness. I wondered where that came from. He was a bit fond of a dram or two. 

And, whether this is AI or not, it called one of the characters Trud  ( as in trudge) when her name was Trude ( as in Judy)  and its spelling was a slight nod to parents who might be European though not British. 

There is a hostel for homeless men and the podcast kept talking about knives going missing, and I genuinely have no idea where that came from.

It totally missed the plot of the book and missed the emotional hook of a woman who had just given birth, lying in hospital after having had acid flung in her face. ( Her clothing in European sizing not UK sizing). And the young male cop who’s detailed to sit outside her door becomes a bit fixated on her, enough to give him a fondness for a nippy sweetie from Glenfiddoch. 

The AI totally missed all of that, and kept talking about things like the investigative drive, multi-layered tensions, the high level of threat but using those words in the vaguest of terms. And then repeating them back.

So back to the chap who did this. He is very bright and a colleague at work. It took him about 2 years to design a poster. He ran it past me and said 'Is that okay?'  I said no because images of such things are now often looked at on a phone or a tablet so the writing has to be very clear and distinct. He said it was staying the way it was, so nobody can read it. Even up on the practice wall it's still difficult to read.

My very narrow minded conclusion is that when it comes to creative fiction AI might be cleverer than the average physical therapist ( or maybe a Y chromosome??) but it's no near as crafty as the average crime writer ( or the average XX chromosome).

I think some of you were in the audience when Mr Shazam, Ajay Chowdray, was asked when AI will be able to write a good crime novel. He said within 3 or 4 years it would be able to have a good go at a Marple, a Poirot or a Bond as it has a wealth of reference data to draw from. But it would still struggle to do anything original. 

So I presume the secret is that we all stay on our toes and mix it up as much as possible and then the demon AI doesn’t stand a chance. 

This blog was written without the help of artificial intelligence or indeed any intelligence at all.

 

4 comments:

  1. Caro, I was recently doing some research on romantic comedies and it was almost comical how much AI got astonishingly wrong, including listing Titanic as a rom/com.

    ReplyDelete
  2. One often heard "caution" about AI is that it hallucinates (meaning it lies through its digital teeth). That would be somewhat creative, I suppose. Actually the thing is a cliche generator and about as infuriating as a friend of mine who said, when I asked him what he thought of the movie "Lawrence of Arabia" that we had just seen on a big screen: It's long. (and he wasn't joking!)

    ReplyDelete
  3. This is fascinating, Caro. I couldn't have imagined how an AI app would respond to a mystery novel, and yet once I read the response you describe, it makes sense. Missing flashbacks, focusing on unimportant characters, not understanding idioms and banter, misinterpreting characters' behavior. The human brain is still a hell of a lot more subtle than Chat GPT, thank goodness.

    When my husband was planning our trip to Japan, he fed info about what kinds of things we wanted to see, how long we had, how much we wanted to travel per day, and a few other parameters into Chat GPT and asked it to recommend an itinerary. He said it didn't do too badly. Needless to say, that didn't become our trip--he looked at guide books, asked friends, and checked websites. But for those kinds of questions, AI can be a help. It's a far cry from writing books, though.

    ReplyDelete
  4. Your post should be required reading in every writing program where some students (and teachers) bemoan the end of human creativity. For example, it's hard to imagine how AI --no matter how fast it can create works "in the style of..." will capture the essence of the demons, psychoses, single malts, and fantasies that drive so many to create. [Present company excluded, of course.]

    ReplyDelete