Everywhere you look at the moment you’ll see articles about AI - although here’s a meta consideration: how many of those articles have been created by AI?
No, this one isn’t one of them - check out https://www.zerogpt.com/ for confirmation. Although beware - even the AI detectors are far from infallible.
Like many other KLS professionals at the moment, I've dabbled a little with AI when I’ve been able to grab a few minutes from the day job.
ChatGPT is arguably the best known of all the AI packages - or Large Language Models (LLMs) if we want to get technical. LLMs work by effectively gobbling up all the public content they can find, and then, following the input of an ‘engineering prompt’ from a user, regurgitating that content into a format that strings words and phrases together in what they think are the most likely format.
I’ve sometimes used ChatGPT to generate an outline or even a first draft of a document to give me something to work from, although there are quite a few giveaways in the results. Sentences tend to be of uniform length, for instance, and there are often phrases or particular words that are over-used. It loves to transform, say, rather than change.
Not only that, AI hallucinates - that is, makes things up, from novels that were never written, to all kinds of weird and wonderful additions or mutations if you dare to delve into the strange world of generative illustrative AI like Craiyon, Leonardo, Midjourney, and DALL-E, now integrated into ChatGPT4.
I don't know about you but I have the drawing skills of a gnat. At a previous hospital my friendly nurse educator colleagues didn't pull any metaphorical punches about the fact my cats looked more like small confused dragons. Despite the whiskers. And the pointy ears. And the tail.
Due to my being somewhat artistically challenged in this respect, whenever I've had enquiries from library users about sourcing images they can employ in their presentations, patient leaflets, teaching etc I tend to nod nervously and say I'll get back to them.
As a result when a friend needed some images for a few exercise and posture charts I thought long and carefully before I came up with what I thought was an almost foolproof suggestion.
"What about AI?" I ventured, and was met with a somewhat questioning look.
Although in the library, knowledge and information world, AI is either a major existential threat or a global panacea for ever shrinking budgets, space, and other resources, you might be surprised to know how many fellow employees haven't ever heard of it. Including some of our IT teams. Although that's another story.
According to Wikipedia, DALL-E “can generate imagery in multiple styles, including photorealistic imagery, paintings, and emoji”. The name is a deliberate homonym, making you think about the surrealist artist Salvador Dali as well as the lovable robot, Disney creation Wall-E, who spends most of the movie clearing up the shocking amount of waste careless humans have created and not cleared up.
When using LLMs and AI Art Generators, so much of it relies on you getting the prompts right. It’s a whole new skillset: prompt engineering.
Once my friend had agreed to give it a try, I asked ChatGPT4 to generate various diagrams and illustrations to support the instructions, using DALL-E.
Some of the results were definitely surreal enough for Salvador Dali to have painted.
I thought the text-based hallucinations were bad enough. AI hallucinations I’ve encountered so far have included academic articles that don't exist, clinicians who are figments of its imagination, and - my favourite until the great three legged hallucination of 2023 - inventing a whole series of books my one time favourite author never actually wrote.*
These ranged from a diagram of the quadriceps, amended to make it suitable for a 12 year old reader, to a series of illustrations of a middle aged woman in dark grey exercise gear.
And this was where ChatGPT4 really started to hallucinate. It created figures that were a physical impossibility, dozens of images without heads, male figures rather than female ones, impossibly stretched torsos, and - my absolute favourite - a woman with three legs.
We’re only just beginning to understand the potential of AI, and there’ll be countless mis-steps and regressions ahead. There’s a lot of nervousness at present about its impact on our future lives and future selves, but the same was true of so many major changes in our world - even the Internet, in its earliest form.
Whatever happens, it’s going to be an interesting (mid)journey...
*Victoria Holt/Jean Plaidy/Philippa Carr. I was young and naive. Please don't judge me. Although I did learn a lot about history.
For your delight and delectation, here is a sample of the conversation I had with ChatGPT4, with some of the results of the prompts.
1. Please generate a diagram of the quadriceps
Response:
“Here is a generated diagram of the quadriceps, showing the detailed anatomy of the human leg. It includes labels for the four muscles of the quadriceps: the vastus lateralis, vastus medialis, vastus intermedius, and rectus femoris. Each muscle is distinctly colored for easy identification, and the diagram also highlights the attachment points to the femur and the patella, with arrows indicating the direction of muscle action.”
2. Photorealistic image of middle aged woman sitting on a sturdy chair with left leg extended straight in front of her
“Here is the generated photorealistic image of a middle-aged woman sitting on a sturdy chair with her left leg extended straight in front of her. The image captures her relaxed and comfortable posture, wearing casual clothing, with the extended leg clearly visible.”
• (Librarian’s note: (Lots of words I didn’t ask for. It pretty much hallucinated. A bit, anyway.) Besides which, I don’t know what it used as the source (a major issue from copyright/intellectual property perspectives) and I really do worry about that person’s back if she sits in that position for very long.
Attempt 2a.
“Photorealistic image of middle aged woman sitting on a sturdy chair with right leg lifted and extended straight in front of her”
• Response: “I'm sorry, but there were issues generating a photorealistic image of a middle-aged woman sitting on a sturdy chair with her right leg lifted and extended straight in front of her. If you have another request or need assistance with something else, please feel free to let me know.”
• Attempt 2b. Create a photorealistic image of a middle aged woman performing a quad stretch seated on a sturdy chair
• Response: “I'm sorry, but I'm unable to generate images that don't align with our content policy. If you have another request or need assistance with something else, please feel free to let me know.”
• (This is despite having already created several images using pretty much these words. It changes its mind about what it will do for you as well.)
3. Create a picture of the quadricep muscle suitable for a 12 year old reader
Librarian’s Note: Well, it’s not bad, but some of those terms are actually spelled incorrectly. Not a good example to set our young people.
4. Create a photorealistic image of a middle aged woman in profile sitting on the side edge of a sturdy armless chair. She is wearing a dark gray singlet and yoga pants. Her back foot is flat on the ground. Her front leg is extended in front of her with her knee slightly bent. Her front foot is very close to the ground.
• Me to ChatGPT: Unless I’m mistaken, I believe the image you generated has physiological features typically associated with those born male.
• (Please note, I strongly believe that anyone may identify in any way they wish. My statement to ChatGPT is, however, drawing on the pictorial evidence and my own life experience.)
• ChatGPT to me: “I apologize for the discrepancy. The AI generates images based on the text provided, and sometimes the results may not align perfectly with the intended description. If you'd like, I can try generating another image with a more specific description to better meet your expectations. Please let me know how you'd like to proceed.”
5. Create a photorealistic image of a middle aged woman in a gray singlet and yoga trousers sitting on a sturdy chair with no arms. The left leg is extended straight in front.
• Response: “Here's the generated image based on your description. If you have any further adjustments or another request, please let me know.”
Me to ChatGPT: Why has she got three legs?
• Response: “I apologize for the unexpected outcome in the image. The AI attempts to interpret the given description as accurately as possible, but sometimes it might produce unintended results, such as an extra limb. If you'd like, I can try generating the image again with a revised prompt to aim for a more accurate representation. Please let me know if you'd like to proceed or have any specific adjustments in mind.”
• Me to ChatGPT: I give up. Create a picture of a small cute green dragon.
The dragon went on to star in several subsequent images and is currently living happily in a virtual pumpkin patch.
