Leafhoppers vs AI - checking on progress

I'm currently writing a short article on Ribautiana species, so naturally rather than get on with it, I took the opportunity to be distracted by playing with AI. Recently Google released an update to its Gemini AI engine and this includes improvements to its (annoyingly named) image generator nano banana. Having spent a long time staring at photographs of Ribautiana and having zero artistic ability, I thought it would be useful to see if Gemini was capable of generating a labelled summary diagram of the differences between the UK species. You can see the result above. It didn't quite work but there is considerable promise.

First the easy part - the things it got wrong brackets (everybody loves to criticise AI):

• It didn't get the position of the dark dots on R. ulmi right, although it did get the idea that those dots are important for identification. 
• It made a mess of the apical dark marks on the wings of R. tenerrima.
• It made a complete mess of the aedeagi drawings. 

What it got right:

• It picked up on the fact that only two species of Ribautiana are visually distinguishable, at least in some cases, while the other three species generally are not. 
• It got the idea that the dark dots on R. ulmi are important and that the pale lines on the forewings of R. tenerrima are also significant.
• Most importantly, it correctly confirmed that dissection is important for definite identification of these species.

I didn't spend any time refining the prompts and getting Gemini to redraw the image - I suspect that if I had, the result would have been pretty good. I was really interested to see how AI is progressing and it certainly shows promise. And now, if you're suffering from lack of reality, here's the real thing:

 

Google Scholar Labs - literature search in a box?

The future is here, it's just not evenly distributed. You can either learn to live with AI or stick your head in the sand.

Google has rolled its Gemini AI into Google Scholar  - "Scholar Labs" - a literature search in a box. Scholar Labs is currently an experimental feature, but as Google wants to get us all hooked on Gemini so it can monetise it, it's not going away. The good thing is that all the references are linked so you can read them yourself and check the veracity, and we're talking about peer reviewed literature here (unless AI has already contaminated scientific publishing). The bad thing? Well obviously you don't know what papers/sources it's missing out - the unknown unknowns. I tried it out on a few of my favourites (try these for yourself, or substitute your own interests):

• Describe what is known about Laboulbeniales in Coccinellida. 
• Explain the ecology of Anthocoridae in the UK. 
• Describe recent developments in the taxonomy of Collembola in Europe and the UK resulting from DNA studies. 
• Explain the known roles of cephalic modification in Linyphiidae.

I was quite impressed by the results, although as always, an AI output is a starting point/prompt rather than a conclusive endpoint. It certainly helped me eliminate a few of my own unknown unknowns and gave me a framework for further thinking. As always with LLMs, it's all about prompt engineering so I seeded ChatGPT with an idea and asked it to write prompts for me, for example:

• Seed: Explain likely causes of invasion waves of Opiliones. 

GPT-5 mini suggested: 

Here’s a concise, effective prompt you can use (for an AI or researcher) to explain the causes of invasion waves by Opiliones (harvestmen):

"Explain the causes and mechanisms behind invasion waves of Opiliones (harvestmen). Cover ecological, behavioral, physiological, and anthropogenic factors that enable range expansions and rapid colonization, and describe how life-history traits (e.g., dispersal ability, reproductive rate, diapause), habitat preferences, climate change, land-use change, human-mediated transport, and interspecific interactions (competition, predation, mutualisms, enemy release) contribute. Discuss temporal and spatial patterns of invasions, typical stages of an invasion wave (arrival, establishment, expansion), and empirical methods used to detect and study invasion waves (genetic analyses, long-term monitoring, species distribution models). Finally, provide examples from the literature of notable Opiliones range expansions and suggest testable hypotheses and study designs to disentangle driving factors."

If you want a shorter version for a general audience or classroom task, use this:

"Describe why and how harvestmen (Opiliones) produce invasion waves. Explain key biological traits, environmental changes (e.g., climate and land use), and human activities that promote arrival, establishment, and rapid range expansion. Include examples, detection methods, and testable hypotheses."

Of course, any field of knowledge is subject to the inherent biases in that field:

• What is the likely impact of artificial intelligence on climate? (try this one for yourself and see if you think it gives unbiased results)

https://scholar.google.com/scholar_labs/search?hl=en

 

AI is coming of age - as this potentially useful tool demonstrates

You don't have to like AI but you do have to live with it - it's not going away. So while we wait for the singularity, let's make it useful.

I've been playing with Google's NotebookLM, a tool which collates, summarizes and analyzes sources provided, helping users quickly grasp key points and easily access answers to questions. While it carries the warning "NotebookLM can be inaccurate; please double check its responses", one of the best aspects is that it links statements it makes to original sources so that they can be checked (AI coming of age). It can assist in brainstorming and generating ideas (through question prompts), making it a very useful tool in my limited experience so far.

As a first test, I took a loose collection of notes I have amassed on the Hemipteran family Anthocoridae and stuck them into a notebook. It can be difficult to pull information from these easily so decent summaries would be useful. As an LM (Language Model), visual input/output is non-existent (unless I've missed something), which is a shame as this would enhance the usefulness greatly. Unfortunately, it is not possible to share the whole notebook publicly (though it can be shared with selected users), but I can share some of the outputs it generates, such as this Anthocoridae Study Guide; the result of asking it to "Describe Anthocoris viscii"; and the podcast-style Audio Overview (audio playback takes a while to start, just wait): https://drive.google.com/file/d/1DxBFzWq1w5BbfLpZU4oRbHFYWvt5300X/view

Here's another attempt: Leicestershire & Rutland Entomological Society: https://drive.google.com/file/d/1YVKOlXTVUnqi6JhF2n0Eb4t5qUSc1L_Q/view

And finally, for Diptera fans: https://drive.google.com/file/d/1HodYEFtP_Kau7SWrQaRi7PzAAGCd6bre/view

Obviously it's only as good as the sources provide - Garbage In, Garbage Out - but I'd suggest the best thing you can do is try it for yourself. I'm sure there are more useful things I can do with it than my stumbling attempts so far (and I'd welcome your ideas), but I'm even more sure that it's going to evolve and improve. Welcome to the future. 

DNA Barcoding - where are we, and why should I care?

"Most insects encountered in the field are initially entomological dark matter in that they cannot be identified to species while alive." The DNA revolution in entomology has been progressing diligently in the background but not making much of a splash with civilians. A new paper provides a good overview of the state of the art and makes a prediction: the big leap happens when DNA and AI start working together. 

Illuminating entomological dark matter with DNA barcodes in an era of insect decline, deep learning, and genomics. (2024) Annual Review of Entomology, 70: 185-204 https://doi.org/10.1146/annurev-ento-040124-014001

Abstract: 

Most insects encountered in the field are initially entomological dark matter in that they cannot be identified to species while alive. This explains the enduring quest for efficient ways to identify collected specimens. Morphological tools came first but are now routinely replaced or complemented with DNA barcodes. Initially too expensive for widespread use, these barcodes have since evolved into powerful tools for specimen identification and sorting, given that the evolution of sequencing approaches has dramatically reduced the cost of barcodes, thus enabling decentralized deployment across the planet. In this article, we review how DNA barcodes have become a key tool for accelerating biodiversity discovery and analyzing insect communities through both megabarcoding and metabarcoding in an era of insect decline. We predict that DNA barcodes will be particularly important for assembling image training sets for deep learning algorithms, global biodiversity genomics, and functional analysis of insect communities.

This set me thinking .. (contd)

 

More on the unstoppable rise of AI

I'm aware that AI identification divides opinion, but it you think it's going to go away, you've lost the plot. 

The latest AI incarnation I've become aware of is the beta version of automated picture recognition on the Danish Fungal Atlas website - https://svampe.databasen.org/en/imagevision

Denmark's mushroom atlas, Danmarks svampeatlas, is home to the well-known MycoKey software. The image recognition system is a logical extension of MycoKey. Trained on a Danish dataset of 250,000 images, this is pretty useful for UK users. Allowing people to identify fungi via AI - what could possibly go wrong? The software includes the disclaimer: 

Note that the system should be used with great care, and not as tool to identify edible fungi without involving knowledgeable humans with experience in fungal recognition. So please explore with curiosity and sanity.

It also flags when the ID results contain poisonous species. Beyond that, if people choose to ignore the warnings, well that's an argument in favour of natural selection as far as I'm concerned (although I'm pretty sure there are lawyers out there who would see it profitably differently). I've been playing with the AI, and as far as I can see it's pretty good. Although it lacks a probability score for search results that Obsidentify helpfully incorporates, it does have an unlabelled "confidence bar" for each hit. Each species returned does link to a useful guide text which includes spore details, etc. Obviously, if you think you'll be able to identify Cortinarius or Russulas from a mobile phone photo, you're probably not safe to be let out on your own. In future, it would be good if such apps could incorporate key ID information into the results returned, eg "check spore size" or "look for hairs on hind tibia". Even without that, unless you're a conspiracy theorist who thinks that this is all part of the great AI plot to take over, things can only get better. Because they're certainly not going to go away...

Dreaming of Electric Sheep

Earlier this week at our local wildlife group Russell Parry gave an excellent talk on urban botany in Leicester. As part of this he discussed the frightening loss of plant species from the countryside and the concomitant(?) rise in urban species. As with all such changes, the causes are multifactorial, but one of the big drivers are management practices. In particular, one thing that Russell mentioned is the impossibility of replicating the Lammas system without appropriate winter grazing. No mowing regime can replicate the effect of light grazing over winter as this removes over-vigorous thugs, allowing delicate species to survive. Mechanical mowing is not a substitute as it cannot be carried out regularly through winter due to increasing rainfall, meaning the land is unavailable through flooding, or machinery cannot go on it without causing compaction (and loss of delicate species). 

This has given me much cause for thought since I listened to Russell's talk. We struggle with our small urban perennial meadow, where the thugs take over in the blink of an eye and winter flooding (and compaction) is an ever increasing problem. My first thought was around Rentasheep - for a monthly subscription would-be meadowers take delivery of a sheep (and portable fencing, presumably an electronic system with a collar on the sheep) for a few hours a month. Leaving aside issues of animal welfare, it didn't take long to figure out that this would be prohibitively expensive, and to be honest I can't see a tribe of itinerant door to door shepherds arising anytime pre-apocalypse. 

But there is a solution. Robot lawn mowers are all the rage and it wouldn't be that difficult for someone much cleverer than me to incorporate a slot for a mobile phone, so that with an AI image recognition appropriate app, a lawnmower with a suitable blade becomes an electric sheep (lightweight, although you could of course dial in the desired degree of poaching for optimum biodiversity in the same way that you set the cut height on a lawnmower). All the hardware and the software for this already exists, it's just a question of someone who cares enough stitching it all together. Surely it can't be long before I can pop into my local B&Q for a electric ovine (or mini-bovine)? 

Electric sheep are not limited to my urban pocket handkerchief meadow - it's also a practical and affordable management solution for rural landowners, not to mention hard pressed Wildlife Trusts. Of course, with rural crime what it is, there will need to be a second generation or robot guard dogs to prevent the electric sheep being nicked ... or do I mean rustled? Electric lamb chops anyone? 

Wild and Wacky (We are the Robots)

When I talked about spider recording in VC55 a couple of months ago, the discussion quickly became "imaginative", and by the end of the meeting, we had "decided" ;-) to equip a drone with an eVac and program it to fly a 10m grid covering the whole of VC55 taking samples. These would need to be sorted by hand to sort out the [insert taxon of choice, e.g. spiders] from the debris, and then sent off for DNA barcoding (High-throughput sequencing for community analysis: the promise of DNA barcoding to uncover diversity, relatedness, abundances and interactions in spider communities (2020) Development Genes and Evolution volume 230, 185–201). Not good news if you're an ecological consultant, great news for gene jockeys.

Mass sampling projects are "easy" to set up. "Hey, I know, let's sequence the whole of Wytham Woods." There was another one published yesterday (Changes in phenology and abundance of suction‐trapped Diptera from a farmland site in the UK over four decades. (2020) Ecol Entomol). What's not easy is processing the data they produce. It took the Danes 15 years to record the results of their malaise traps (The Swedish Malaise Trap Project: A 15 Year Retrospective on a Countrywide Insect Inventory (2020) Biodiversity Data Journal 8: e47255). Right now DNA barcoding is the only feasible way of tackling the mass of data such projects produce. But maybe not for much longer. A.I. (Artificial Intelligence) is coming: A light trap and computer vision system to detect and classify live moths (Lepidoptera) using tracking and deep learning. (2020) bioRxiv https://doi.org/10.1101/2020.03.18.996447, and: Successful launch of automatic insect counting in North Holland. I've been concerned about the impact of A.I. for some time. Well, not about A.I. as such, more about A.S. - the poorly functioning first generation systems - Artificial Stupidity. It's not there yet, but inevitably it will come.

Recently I've noticed a disturbing trend in entomology blog posts to include dodgy music videos. Until the A.I. finally takes over, it seems that We are the Robots: