Botany One & AI Revised

After a few months of working with AI, we’re revising how we use it. Some of this reflects changes in the capabilities of AI, and some of this reflects feedback on how we use it.

In short, 

• We’re using AI to cover more stories in the same amount of time. We are not replacing people with AI. This drive to cover more stories is because we have over a hundred stories each month that we cannot write about, if we do everything manually.
• We do not use AI to create photorealistic images without being painfully obvious about it, and we deliberately use art styles that will not be confused with photos when using AI-generated art.
• Where AI plays a significant role in writing the text, it will be flagged. Writing means planning an article as well as generating the words.
• We use AI tools for grammar checking, spell checking, and administrative work like identifying things to link to and papers to cover. We won’t be going into great detail about that, because it’s quite difficult to write using a computer without using spell check.

Four months ago, I wrote a statement on how we were using AI on the site. It opened: AI authors are like predictive text on your phone. They help writing words, but are less good at sentences and cannot be used for writing whole articles at the moment. Since then, GPT-4 has been opened up to the public, and now I’d say it can write sentences but struggles with paragraphs – and still cannot write articles.

Puzzlingly, I’m also finding that, when I put something through an AI checker, it still comes out graded as entirely human. I suspect that in the long term, this might be a fool’s errand as AI authors learn more diverse writing styles, but for now, it is an issue for some people. Here I’d like to cover what AI does and, maybe more importantly, why.

Why not use a human instead of AI?

Time and Twitter.

Botany One is funded by the Annals of Botany Company, a registered charity in the UK and a non-profit company. Also, Botany One doesn’t run adverts, or have paid subscriptions, so there’s no money coming in apart from the company funding, which limits our budget.

I handle most of the blogging. If the subject is one I know, and I can immediately see what the story is, then a blog post can take as little as an hour to write. But we cover a lot of Botany, and if the topic is one I’m not familiar with – which is most of them – a blog post can take four to six hours to write.

In addition to blogging, I handle admin work and compile our weekly news round-up, The Week in Botany. Changes at Twitter have made it much more work to track what is being shared there, so now there’s a lot less time in my week.

On top of this, I track stories I’d like to see covered, a list which is getting longer and longer. In the past, we’d post links to papers we thought deserved attention to Twitter. Recent changes mean that I’d rather see effort spent here than a billionaire’s site. ‘Tweeting’ to our site doesn’t really work. When we post a message here, it gets relayed to social media sites, and getting an alert for something that’s just a sentence is annoying. So, we need to post something a bit more substantial, and that takes time.

In the longer term, humans are better because they understand text instead of predicting what comes next in a sentence. We’re looking to develop a brief blogging course so we can have a pool of people who can contribute to the blog. But creating that course also takes time.

The goal is to combine the speed of AI for the parts it can do fast with the insight of a human for the elements a human can do better. Then we have the opportunity to provide more useful content for our readers, highlight more important research that’s being done, and develop the site so that more humans can contribute.

AI Art on Botany One

We use the current (July 2023) layout because it is fast and should take less time to load over a mobile phone, but it has some peculiar requirements. Every item needs a cover image for the top of the article. This image needs to be light in file size but still cover 1500 pixels by 500 pixels. That’s a 3:1 aspect ratio, which is rare to find in photobanks. If I can, I’ll crop a different photo to fit that size, but it’s not always possible. 

Another question is what image goes in that space. The obvious answer is the species or site that the article talks about. This leads to another problem. Often the species in the article is rare, and there are no easy to licence images of the plant. Likewise, it might be that a botanist has published dramatic photos of the site in their paper, but if their paper doesn’t have the right Creative Commons licence, we cannot use it.

Therefore if we need an image fast, we’ll use AI to generate an image based on concepts in the blog post. For this post, the concept is ‘robot writing an article’.

Generative AI is getting better. It might be possible to generate an image that looks like a photo of a robot Charles Darwin writing an article. This could easily be misleading. So we’re taking the line that we will not use AI to generate photorealistic cover images. Any AI image will be obviously a digital painting, or possibly in the style of a 1960s sci-fi comic, or any style that is clearly not meant to pass as a photo. We won’t use AI to generate images in-text, unless it’s necessary, in which case it will be clearly labelled so that you can see it’s a computer-generated image.

AI Text on Botany One

AI text is more complicated.

Some people think I have a button I press to write an article. That would certainly make life much easier. At the current time, that’s not the case. In the past, AI couldn’t even read a paper. One of the big changes since March is that AI can hold a lot more information in its memory to write something. GPT-4 can hold around half a paper. Claude can easily read a full article. But while they can take information in, getting information out is still limited. You can ask AI for 500 words or more, but it won’t produce 500 usable words. GPT-4 can probably write around 300 sane words. Claude is closer to 250. AI can make a text snippet that can be edited and expanded, but it cannot sit as an article without work.

Another problem is the kind of output AI can write. It’s a lot easier to write when there’s no wrong answer. We could ask AI to write an article about the best flowers for a garden, a lot of which is a personal opinion. On the other hand, when writing about science, there are concrete facts that matter. The easiest way to tackle accuracy is to go through the text line-by-line.

This means that while AI is used to produce a draft of an article, the final product is edited by a human. As an example, the post Tiny Plant Made Smaller By Microbes was produced with Claude AI as an assistant. The full output of the AI is at the bottom of the post.

Comparing the AI’s lead paragraph and the final version, there’s a lot of the AI’s output in the opening. AI text in bold

Research reveals that the whole microbiome community, including fungi and protists along with bacteria, reduces the growth rate and modifies the morphology of the aquatic plant duckweed (Lemna minor). The research by Mark Davidson Jewell and colleagues, published in AoB PLANTS, contrasts with previous findings that isolated bacteria can enhance duckweed growth. The results show the importance of considering the whole microbial community, or microbiome, when studying plants.

Claude also suggested some quotes:

“Contrary to our expectation, the presence of the microbiome consistently decreased plant fitness, on average by 12%,” said lead researcher Dr. Mark Jewell.

“Although several important plant-bacteria and plant-fungi mutualisms have been identified for L. minor, our results highlight the importance of pathogens, parasites, competitors and herbivores in the microbial assemblage.”

“This was the case for all environment treatments and for most, but not all genotypes,” added Jewell.

As a note, at the time of writing Mark Jewell’s website says he’s a PhD researcher, and the quotes come from the article, so they should be credited to all authors. The words are in the article, and I used the first two quotes but added more from the article that better illuminated other parts of the research.

The rest of Claude’s output is bullet points.

I could re-write that first paragraph and claim that the text is entirely human-generated, but that would be misleading, at best. You can see that even if AI hadn’t written any of the text, it still had a hand in planning the article, and it would be dishonest to say the text was AI-free. 

Crediting AI

There’s concern about AI text at the moment. A lot of those concerns are justified. If a text is purely AI-generated, it will read poorly, and you can also be worried about its accuracy. While the human writer takes responsibility for the article, it’s reasonable for people to want to know if AI was part of the writing process. For this reason, any articles with AI assistance have this flagged in the biography section of the author’s credit.

This approach has led to complaints. Some people feel that this method is misleading because they think that any AI-assisted text is the same entirely AI-written, and this should be flagged at the top of the piece. However, humans have to take responsibility for any possible errors in the article, so I’m wary of purely crediting AI. For this reason, AI-assisted articles will now get a co-author AI Assistant. If people think this is a magic box that produces articles, then below is an example of it at work.

For the paper Presence of microbiome decreases fitness and modifies phenotype in the aquatic plant Lemna minor, we have this blog post.

Here’s the output from the AI Assistant

---

Microbiome decreases fitness and alters phenotype in aquatic plant Lemna minor

Headline: Microbiome decreases fitness and alters phenotype in aquatic plant Lemna minor

Lead paragraph: Research reveals that the whole microbiome community, including fungi and protists along with bacteria, reduces growth rate and modifies morphology of the aquatic plant duckweed (Lemna minor). This contrasts with previous findings that isolated plant-growth promoting bacteria can enhance duckweed growth.

Key points:

• Scientists conducted lab experiments growing Lemna minor with and without its microbiome under different light and nutrient conditions.
• The presence of the complete microbiome, not just bacteria, consistently decreased population growth rates across all genotypes and environments tested.
• Microbiome presence also led to altered phenotype – smaller frond size and shorter roots.
• Results suggest a greater role of pathogens, parasites and competitors compared to mutualists in the whole microbiome.
• Microbiome decreased variation among genotypes in response to environment, mediating phenotypic plasticity.
• Finding challenges prevailing view that plant-microbe interactions in Lemna minor are largely mutualistic.

Quotes:

“Contrary to our expectation, the presence of the microbiome consistently decreased plant fitness, on average by 12%,” said lead researcher Dr. Mark Jewell.

“Although several important plant-bacteria and plant-fungi mutualisms have been identified for L. minor, our results highlight the importance of pathogens, parasites, competitors and herbivores in the microbial assemblage.”

“This was the case for all environment treatments and for most, but not all genotypes,” added Jewell.

Conclusion:

• Research expands understanding of plant-microbiome interactions in an aquatic species beyond just beneficial bacteria.
• Shows microbiome can reduce host fitness, in contrast to emphasis on microbial mutualists.
• Findings demonstrate important environmental and genetic variation in plant-microbiome relationships.

---

Over time our relationship with AI will evolve again. I suspect writing at length will prove a difficult problem for AI because coherence over longer texts becomes much more complicated the longer a text gets. Fact-checking will also be another problem because a lot of scientific studies are about showing that previous facts are incorrect.