FunGiPT I: AI usage as a Fungal Face ID 🖥️🤖🍄

Ezequiel A. Cruz-Campuzano - Fungaria

Hey there MycoFam! It’s Ezequiel here, back this time to dive into the growing relationship between fungi and artificial intelligence. This tool is definitely becoming more widely used, and even though AI has the power to help us unlock fungal potential faster, there are still some important concerns we shouldn’t ignore, ranging from misidentifying a mushroom to the unfair exploitation of public data. On this occasion, we’ll take a look at how fungal taxonomy can be aided by AI, though it’s worth emphasizing that this tool should never replace fungal taxonomists per se.

Disclaimer: This article explores the use of artificial intelligence in the identification of fungal specimens, but it is not intended as guidance for determining edibility. I strongly advise against relying on AI tools to identify fungi for consumption.

How AI Is learning to learn fungal taxonomy, and why it will always fall short

Species recognition via AI is mostly accomplished with deep learning models such as Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), are being trained to recognize fungi from images, acting like a kind of fungal facial recognition. These models analyze a huge number of pre-identified images, and learn to detect patterns in shape, color, and texture that correspond to a particular species (Ozsari et al., 2024; Tsang et al., 2025). 

So how do these systems actually work? CNNs mimic how the human visual cortex processes information. They use filters to scan across an image and detect features like edges, textures, or shapes at different levels of complexity. This layered structure allows them to gradually build a “hierarchical understanding” of the object in the image, from simple curves to full mushroom silhouettes. ViTs on the other hand, take a different approach. They divide images into small patches and analyze the relationships between those patches using an attention mechanism. This allows them to capture both fine detail and broader context, which can be helpful when distinguishing between species with subtle differences (Maurício et al., 2023).

Simplified identification of Aspergillus strains using CNN (Tsang et al., 2025).

Simplified identification of wild mushrooms using ViT (Picek et al., 2022).

Despite their promise, there are, to my understanding, three key factors that limit the reliability of these tools: training time, within-species variability, and photo quality or timing. And these are all deeply interconnected. Add to that other taxonomic nightmares like cryptic species, and we start to see just how tricky fungal ID can get for AI.

The first factor seems like an obvious one, right? The more you train, the better you get. Hence, even when these tools are currently far from perfect and will never be as trustworthy as a trained fungal taxonomist (we’ll come back to that), over time, they will accumulate more feedback and get closer to matching species accurately (Bashir et al., 2024). Still, it sounds deceptively simple: just run more training cycles, simulate more epochs, and voilà! Problem solved. But the fungal answer to that is a big, fat NO.

Even if we leave out the environmental concerns of water use in massive data centers, just flooding an AI with photos of mushrooms, conidia, or hyphae won’t do the trick. Why? Because fungal morphology is incredibly variable, even within the same species. Fungi are known for their high morphological plasticity: this means that the same fungus can take on very different shapes. Take for example some members of the Thelephorales, where several species that can form fruit bodies ranging from simple crust-like patches to complex, branched, coral-like or mushroom-like structures (Ramírez-López et al., 2015; Zmitrovich et al., 2018). That level of variability isn’t something a machine can easily pin down, especially when even crust-like species are hard to identify using microscopy and molecular data combined (e. g. Svantesson et al., 2019).

Then there’s the issue of photo quality or timing, and this one really messes things up. Imagine two photos of the iconic Fly Agaric (Amanita muscaria), which is a fairly easy-to-distinguish species. In one, the mushroom has lost its white or yellowish warts due to rain, but you can still vaguely tell it's red-capped with a white stem. In the other, the photo is underexposed or desaturated, and the cap looks more pinkish or orange. Or consider another version: the same unwarted A. muscaria left out in the sun a bit too long, losing its vivid red pigment and turning more orange. In both cases, an AI, just like an inexperienced forager, might confidently suggest something like Caesar’s mushroom (Amanita sect. Caesareae).

(From our sponsors: Speaking of the Amanita genus, go load up on some decarboxylated Amanita muscaria mushrooms from our sponsors Minnesota Nice Botanicals below. As this newsletter is intended for a mature and intelligent audience, use discretion and educate yourself before diving in headfirst, but myself and many others have benefitted tremendously from forming a relationship with ‘The other magic mushroom’ 🍄)

I’ll never forget something that happened to me a few years ago that really hit home on this topic. We were holding a fresh mushroom exhibition in downtown San Cristóbal (Chiapas, Mexico), open to the public. At some point during the event, a girl came up to me, excited to share some interesting mushrooms she had photographed that very morning. Then she casually mentioned that she had even eaten some of them, because she had identified them with an app (Please, NEVER do that!). As soon as she said it, alarms rang in my head and I got genuinely worried. She pulled out her phone and showed me a picture of the mushroom she had eaten. Even though it was a bit blurry and underexposed, it was still clear enough for ID purposes. It was unmistakably an orange scaber stalk (part of the Leccinum insigne complex), which made me breathe a little easier.

But then she showed me what the app had matched it with, and the concern came back. It had identified the mushroom as a king bolete (Boletus edulis complex), which is not only wrong, but made me want to throw up on the spot. Luckily, in this case, the mistake was between two edible species. But it so easily could have been the other way around—and the consequences could have been serious, even deadly. I asked her to avoid using these apps for final identification, and instead contact any of the mycologists present for proper IDs.

Did you know that several Amanita have white variants? Take this Amanita "multisquamosa-02" for example (Garett Taylor, 2020). Molecularly speaking, it is very similar to something like A. pantherina. However, it is all white. This has also been documented with the Fly Agaric for example, showing how variation within a species can also mess with AI aided identification.

That’s where firsthand experience comes in. Knowing how to ID a mushroom correctly often depends on recognizing these subtle clues and oddball exceptions, a kind of Taxonomic Gut Feeling (TGF) that only comes with years of fieldwork, lab work, and repetition. And that’s something no AI, no matter how much it trains, is likely to replicate (Valdecasas, 2024). 

This concept of TGF might sound like something I just made up while writing this, but I actually first heard it last year at the MSA (Mycological Society of America, which of course only considers the U.S. and Canada as “America” 🤦🏽) annual meeting. Alfredo Justo, the world expert on the genus Pluteus, mentioned it, and if there’s anyone qualified to talk about difficult fungi it’s him, given that most species of said genus look nearly identical (though, to be fair, that could be said of a thousand fungal genera without exaggeration).

Right from the moment he described it, the idea stuck with me. As he explained, after looking at microscope slides over and over, and examining fruit bodies again and again, you start to build this inner compass, a kind of sixth sense that kicks in without you even thinking and quietly tells you: “this species is this.” Though I’m still far from being as skilled as Justo is with Pluteus, I’ve definitely felt something similar with a few species complexes I’m very familiar with. That intuitive skill is priceless, and is one of the many reasons classical taxonomy needs to be preserved.

How AI Can Truly Enhance Fungal Taxonomy

You might be thinking by now that I’m completely against using AI in fungal taxonomy, but that’s not the case at all. In fact, I see real potential for AI to support and even accelerate taxonomic work, as long as it's guided and trained by experienced professionals. In particular, I believe there are three key areas where AI can make a meaningful impact: I) Molecular identification of fungi, speeding up sequence comparisons, clustering, and database matching; II) Identifying specimens above the species level, especially helpful when genus or family-level traits are distinctive enough for reliable classification; III) Providing ID suggestions in community-driven platforms, giving users a helpful starting point that can later be refined or confirmed through expert review. I’ll briefly discuss each.

Molecular identification of fungi—Think of it like scanning a barcode at the supermarket. Our genes are made up of four nitrogen bases: Adenine (A), Guanine (G), Thymine (T), and Cytosine (C). Certain genes vary more between species than others. In fungal taxonomy, we usually sequence one of these variable genes, which is the ITS region for most fungi, to get the exact arrangement of these letters, known as a DNA sequence. That sequence is then compared to public databases to find the closest match, helping us identify the organism (Schoch et al., 2012). It’s basically scanning the genetic barcode of a fungus to get its name. Currently there are tools that do this, such as BLAST and QIIME2.

At first glance, it might seem like AI doesn’t bring much new to the equation. But the real power of AI in this lies in speed and scale. Fungal molecular data is being produced faster than ever, especially through metabarcoding, where thousands of ITS sequences are generated from environmental samples like water, soil, etc. Traditional tools struggle with these massive datasets, as they’re slow and not always precise. Recent deep learning models like FungiLT (Zervas et al., 2024) skip alignment entirely and classify ITS sequences with over 98.77% accuracy, processing thousands of reads in just seconds. In comparison, conventional tools can take up to an hour for the same job, with slightly lower accuracy (<95%). That’s a huge leap for ecological studies.

Even more interesting, tools like MycoAI combine classification with clustering, which helps when reference databases are incomplete or ambiguous. Rather than just returning a low-confidence match, these AI tools can still suggest a likely taxonomic placement based on patterns in the data (Romeijn et al., 2024).

Example of AI workflow for molecular identification with ITS (Liu et al., 2025)

Identifying specimens above the species level—Let’s say you find a mushroom in the wild, but it’s too young, degraded, or missing key features to confidently pin down to species. In these cases, identifying it to genus or family level is still incredibly useful. Many fungal groups have distinctive traits at these higher ranks, like the waxy fruit bodies of Hygrophorus or the sponge-like pores in Boletaceae. AI models trained on visual data can learn to recognize these broader patterns and group fungi accordingly, and give more trustworthy ID suggestions.

For example, Rahman et al. (2023) used CNNs to classify fungal genera of microscopic fungi, based on traits such as hyphal network, conidia, etc. achieving high accuracy. This kind of tool can assist in quickly sorting samples, for latter species-level resolution carried by a taxonomist. Another study by Bartlett et al. (2022) used machine learning models trained on a curated dataset of Hebeloma species and found that while species-level ID was challenging, the algorithms were especially good at assigning samples to the correct section of the genus, which is highly valuable for diverse taxa where morphological differences are subtle.

Providing ID suggestions in community-driven platforms—Ever uploaded a mushroom photo to iNaturalist and received an ID suggestion? That’s AI already at work. These platforms use Vision Transformers trained on hundreds of thousands of labeled images to propose identifications. While species-level predictions are still dubious, genus and family-level suggestions are surprisingly accurate, offering valuable starting points.

In Picek et al. (2022), the FungiVision app—backed by a ViTs and human-in-the-loop validation—-achieved nearly 93% accuracy on macrofungi and reduced classification errors by 46.8% compared to previous systems. Crucially, the platform integrates expert feedback: users or specialists can correct IDs, and those corrections are used to retrain the model, refining its species-level suggestions over time. This creates a virtuous loop: AI offers a helpful guess, experts verify or refine it, and the system learns from those confirmations

Screenshots from the iPhone app of the Atlas of Danish Fungi: (i) the application menu with the Name suggestions feature, (ii) image selection, and (iii) the fungi species suggestions with mushroom edibility info.

Take Home Message

In the end, AI isn’t the enemy of fungal taxonomy, but it’s no magic solution either. It’s a powerful tool that, when used wisely and in the right contexts, can genuinely help us explore fungal diversity more efficiently. But no algorithm can replace the deep, experience-based intuition that comes from years of fieldwork, microscope time, and taxonomic training. So let’s embrace the tech, but never forget the mycologists behind it.

Mycopreneur Incubator Today

The Mycopreneur Incubator is our weekly online professional development and networking forum for mycopreneurs, fungi researchers, the mushroom curious and the occasional fungally-informed AI agent (I’m looking at you Otter.AI notetaker)

We’ve had hundreds of mushroom entrepreneurs from 30 + countries attend one of our incubators in the last year, and as a regular attendee told me last week, “A lot of good things have come from the Incubator’. Come say hi and hang out for a bit, you never know who you’re going to meet or what new opportunity the community will unlock for you and vice versa.

Time: Today, Jun 26, 2025 3:00 PM East Coast U.S. / 12 pm West Coast U.S.

Link to join:

https://us06web.zoom.us/j/86062840898?pwd=6s5yzaGhPRiL6zEFeRLiPgjWCU5XLx.1

Meeting ID: 860 6284 0898

Passcode: 184253

If you plan on attending the Mycopreneur Incubator please consider making a financial contribution in the amount of your choice here:

This week’s Mycopreneur Incubator is sponsored by The Fungivore, hosts of the renowned MyColores and MycoChiapas mushroom expeditions to Oaxaca and Chiapas, Mexico.

The upcoming Fungivore trips to Oaxaca and Chiapas, Mexico in August and September will take you deep into some of the world’s best mushroom foraging terrains and connect you with regional experts on the mushrooms. World-renowned resident mycologists like Dr. Gordon Walker (Fascinated By Fungi), Chef Chad Hyatt, and the author of today’s newsletter Ezequiel Cruz of Fungaria will join you on the ground to provide real time insights and unparalleled regional mushroom knowledge in addition to many other surprises. I’ll personally be present for at least a portion of the MycoChiapas trip September 24 - October 2 as well. There’s no one else doing a tour like this and it’s right smack in the middle of peak wild mushroom season in Southern Mexico. Only 5 spots left (3 for MycoChiapas, 2 for the MyColores Oaxaca trip August 12-20). Check them out below and snag one of the remaining spots before they sell out.

Thanks for reading today’s Mycopreneur Newsletter, if you enjoyed it then please share it with a friend or post it on your social media and help us grow by subscribing if you haven’t done so already.

See you at the Mycopreneur Incubator today,

DW