Why I dislike AI in speculative biology and palaeoart

This post departs from my usual posts in that it is not directly about speculative biology but about a tool. Let me be clear: I dislike AI and will explain why. Because of the ongoing discussion about AI in palaeoart, I added that subject to the title of the post.

'No AI' icon; you can get it from Wikimedia 

AI as a tool to produce images
Some say that AI allows everyone to paint. I really enjoy painting and the pictorial arts, so wouldn't it be wonderful if everyone could paint and enjoy that as much as I do? But supplying prompts to an AI image generator is not painting; it is merely steering a complex string of associations in a specific direction. Those associations, the core of most LLM models, are derived from countless images, mostly found on the internet. It is no secret that many images used to train the AI models represent copyrighted work and were often used without permission of their creators. With that in mind, you can ask yourselves whether the use of these images is morally and ethically justified. I think not, personally. In fact, you can argue that using the work of others without their permission is a form of handling stolen property.

We have all seen discussions about piracy being a copyright infringement: companies producing films, music, games or software all protested loudly when piracy started to cut into their profits. Their protests gradually made governments fight piracy. But the companies making AI are the ones making money from  piracy, which dramatically affects their point of view. but sometimes I wonder whether governments shouldn't do more to protect copyright, as they did with previous piracy... 
 

Proponents of AI may defend its use by saying that AI does the same as an artist who is influenced by the work of other artists. It is indeed acceptable for an artist to be influenced by others, but there are restrictions: the results are accepted when inspiration leads to creativity. When van Gogh was inspired by Japanese woodblock prints, the resulting oil paintings represented new creativity. In contrast, profiting from the work of others by selling illegal copies of their work is not accepted in art. As AI images are wholly based on associating elements of existing works, they fall in the latter category; there is no inspiration, just a complex way of mixing and copying existing art.          

I will not go into the enormous energy costs of AI (if users had to pay the real costs of producing AI images and videos, hardly anyone would use it); I will also not go into how AI is pushed without any discussion of the moral, social, environmental and economic consequences; but I will say that these aspects make me like it even less. 

DinoCon 2026 AI statement

AI at DinoCon 2026
AI won't be at DinoCon 2026 (I will be, selling images and The Book!). The organisers have made a very clear statement forbidding AI; you can find it here. The image above is what you see when you register as a vendor. I borrowed it from Andy Frazer of 'Dragons of Wales' fame. At the previous DinoCon, Andy brought  'no AI' stickers to DinoCon, and Darren Naish wondered out loud about the ethical and moral aspects of AI use. These anti-AI sentiments have gathered strength over the last year. Good!

AI as a research tool in speculative biology
I have tried ChatGPT a few times as a research tool to examine how useful the results might be. I asked questions that I had already researched the conventional way, meaning I used PubMed and Google Scholar to find relevant peer-reviewed published scientific papers. Two questions I put to ChatGPT were first how deep chromatophores lie in the skin of various colour-changing animals (chameleons mostly), and which kinds of mechanisms might cause quick colour changes in animal skin. For me these are normal things to ask if you take your speculative biology seriously.  

In both cases, what I found led me to conclude that there wasn't much known about these subjects, or at least that I failed to find exact answers in any paper. The AI replies reported more and were much more positive. In fact, ChatGPT congratulated me for asking smart questions, which I now feel is an irritating trick. ChatGPT did produce exact skin depth measurements for chromatophores, and volunteered to put the results in a table or graph 'suitable for publication'. I then asked for specific literature sources, and ChatGPT duly provided five papers. However, not a single one actually existed! I recognised traces of real papers, but only becasue I had already done the hard work. In science, you always have to dig until you find the original measurements that a later conclusion is based on. Well, if you try that with ChatGPT, you get what are mistakenly called 'hallucinations': just nonsense indicating that the association process is going off the rails. If I wouldn't have learned how to find factual scientific information I might have copied ChatGPT's nonsense and would have made a fool of myself. There are other examples of such nonsense in the blog 'Sauropod vertebra picture of the week' (SV-POW; here and here)  

Such experiences suggest that using the word 'intelligence' for what these models do is at best a gross exaggeration. Some AI mistakes are so stupid that no person with any understanding of the subject matter and with a normal ability to reason would ever produce such errors. I now prefer to think that the abbreviation 'AI' stands for 'Associative Incompetence'.