As an author and image designer, I use AI to proofread texts and generate images. I also created my website with its help. Having a ChatGPT Plus account, I ask "my best friend" about all sorts of topics every day. Whatever people say, AI is a major breakthrough for finding actionable information. It's a fantastic tool for discovery and generalization. Say goodbye to using search engines to find useful information. The time saved is considerable.
Based on chatbots from major AI tools available on the web, there are several dozen. In the space of three years, and since the arrival of the first one for the general public, ChatGPT, at the end of November 2023, they have made enormous progress and have impacted all technical and cultural fields.
However, it is good to know some basic rules of usage, some examples of which are:
- Be direct in your requests (prompt) without adding polite formulas, for example. Indeed, AI is primarily a statistical tool. This adds nothing, and in fact, AI performance degrades according to some studies (article: Being mean to ChatGPT increases its accuracy — but you may end up regretting it, scientists warn).
- AIs have ethical rules implemented by their creator that prevent them from asking sensitive questions.
- AIs can be subject to hallucinations. LLM-type AIs "hallucinate" primarily in certain specific contexts, where their operating method (predicting the most probable word) leads them to to invent plausible, but false, things.
- Be patient. When you're up against a wall and the AI doesn't respond as expected, you have to know how to get around the obstacle and try other approaches.
Chatbots are not yet the holy grail we hoped for. However, even though they provide invaluable services across all fields, their behavior can be disconcerting when we push them to respond to expectations that seem obvious to us but are not to AI. On my end, I have experimentally identified two major instances of failure.
Let me illustrate my point with image creation. Creating an image with AI becomes complex when you delve into the details with a precise and evolving description. Indeed, to enhance each article I publish, I include an image illustrating the topic. I've published dozens of images and therefore generated hundreds of images using AI. I generally need more than ten images to publish one that meets my criteria.
I think of AI as the equivalent of a dedicated but obtuse draftsman. I give it precise instructions (the prompt) and it generates an image. Most of the time, the first image is a good approximation of what we have in mind. But there's always some small detail that's off or missing. And that's where the battle begins.
AI has a remarkable ability to produce stunning images in terms of quality and hyperrealism. Unfortunately, it also suffers from a terrible inability to follow certain instructions that seem obvious to us. Sadly, I was forced to struggle with the AI on most of the images to get it to meet my expectations. Often, I had to finish the job with a drawing tool that couldn't make it do what I wanted.
I wrote an article about the errors that AI makes when generating drawings. The image at the top of that article perfectly illustrates my point. I had to be very patient and find workarounds to get around the AI's shortcomings, ultimately having to make pixel-by-pixel corrections.
To get to the bottom of it, I challenged myself to outsmart our virtual illustrator friend in just a few image generation iterations. Here's the context:
«"Generates the image of a man sitting at a table reading the title 'Flower' from a closed book placed in front of him on the table."“
The first image was almost perfect, except that the title was legible facing the camera. The instruction that the man should be able to read the title had not been followed. After repeating my request, the book was turned the right way up in the second image. I asked that the title be replaced with "Vagabond." The third image was correct, but the spelling mistake in the new title was still missing. (Too bad.)
Things got complicated when I asked the man to put a suit on. The book ended up on the wrong side of the screen again, facing the camera. It turned out that if I don't explicitly request a new image, the AI doesn't do it automatically. This isn't consistent with previous requests, which didn't explicitly ask for it to generate a new image.
And then, after hundreds of generated images, I learned for the first time that the chatbot uses an internal OpenAI tool specializing in image generation (DALL-E, I assume), and that it needs to be given the entire context for each new image generation, otherwise it loses data along the way. Indeed, following my request, the image showed the man in the suit, but the book title was legible on the camera side. I asked it to turn the book over. This time, the title was the right way up. But the man's hands were under the table. It was as if I had to choose between the book being the right way up or the hands being on the table.
This is a typical example of an AI's inability to repeat instructions. As soon as a modification is requested, the next image generation can alter a parameter of the previous image. What I'm describing here is a basic situation. Frequently, I've had to make manual adjustments by reworking the generated image, using copy-paste and other techniques in a drawing tool.
The exercise I described to you is not complex from a human intelligence standpoint. And yet, repeatability is not AI's strong point. However, from my perspective, it's essential. Otherwise, to what extent can we consider an AI reliable?
To conclude this presentation, I will give two contradictory points of view.
1 – I encountered the same difficulties with text correction. It took me two months to create a prompt with a 90% accuracy rate (%), while still limiting its correction capabilities. And it took me another month to realize I wouldn't be able to do any better. AI is therefore a good tool, but it has its limitations at the moment. Consequently, I supplemented the AI with a reference text correction tool like Antidote.
2 – In the field of AI use for cars, according to American statistics, there are about ten times fewer accidents for a car driven by an AI like Tesla's Autopilot than for a car driven by a human.
I would be delighted to hear about your experience, whether or not you have encountered these AI inconsistencies. Please feel free to leave a comment.
For those who would like to understand where some AI hallucinations come from, I share my experience with a recalcitrant creation of images by an AI.
