New study maps where AI species identification fails

General purpose AI use in species identification. Can it be useful for cosnervation?

A small marsupial photographed by a camera trap in Colombia was identified by an AI model as a mammal whose known range does not include the country, even though the model had been told the animal was seen in Colombia. A new study in the Biodiversity Data Journal uses this kind of error as a starting point to ask whether general-purpose AI is ready for conservation work. 

The marsupial, Metachirus myosuros, was recorded in the Morrocoy protected areas in Montes de María, where the conservation organisation Sembrandopaz has been monitoring wildlife. The AI proposed Bassariscus sumichrasti, a species whose known range does not include Colombia. And the expert Dr Juan Fernando Díaz Nieto of EAFIT, confirmed the correct identification to be Metachirus myosuros.

Metachirus myosuros
Camera-trap image recorded during biodiversity monitoring in the Morrocoy reserves in May 2026. A general-purpose multimodal AI system proposed Bassariscus sumichrasti despite receiving the Colombian locality and a description of the animal’s locomotion, posture, facial structure, tail proportions, and possible colour pattern. Expert review identified the animal as Metachirus myosuros. Dr Juan Fernando Díaz Nieto, Universidad EAFIT supported the identification. Photo credit to Fundación Sembrandopaz.

Two other cases from the same fieldwork showed different problems. The armadillo Cabassous centralis has been known to science for decades, yet few verified photographs of it are publicly available.

Armadillo Cabassous centralis
Armadillo Cabassous centralis. Credit to Fundación Sembrandopaz.

The porcupine Coendou vossi, recognised as a distinct species only in 2025, was given an older, superseded name that still circulates in public resources.

Porcupine Coendou vossi
Porcupine Coendou vossi. Credit to Fundación Sembrandopaz.

“If a rare animal is confused with a more familiar one, its presence might go unnoticed,” said study author Gustavo Nicolas Paez Salamanca.

If AI assigns an incorrect species name or suggests an animal outside its natural range, that error could enter biodiversity records and affect how researchers interpret species distributions.

Testing 10,000 photographs

To test identification accuracy in a systemic way, the study evaluated OpenAI’s GPT-5.2 on 10,000 research-grade iNaturalist photographs of vertebrate species recorded in Colombia: amphibians, birds, fishes, mammals and reptiles, with 2,000 images per group. The model was told only that the animal was seen in Colombia and was required to name a family, genus and species for every image, with no option to decline.

Under a benchmark requiring species-level predictions, GPT-5.2 correctly identified 9.5% of these images, or roughly 1 in 10.

Performance varied widely by group, with exact species-level match rates of about 17-19% for mammals and reptiles and under 5% for amphibians, birds and fishes.

– said Paez Salamanca

It turns out, the errors were not random. Species with more publicly available photographs were generally identified more accurately, while recently described species were harder to recognise. Tests with other AI systems showed similar general patterns.

“This suggests that improving wildlife identification is not simply a matter of choosing a different AI application,” Paez Salamanca said. “The challenges are also connected to how much information exists about each species and how that information changes as scientific knowledge advances.”

Implementing AI could present a problem that goes well beyond species identification. Musters and de Snoo (2026) make a similar point in a recent paper on predicting ecological change with deep learning: conservation AI may need an entirely different kind of data, not just more of it. See our earlier blog here.

A role for AI, with safeguards

infographic
Visual poster presenting the key takeaways from the study. Infographic generated with ChatGPT-5.5 and provided by Gustavo Nicolas Paez Salamanca. The photos in the poster are from the trap cameras by Fundación Sembrandopaz.

The study does not argue for abandoning AI. It proposes using AI as a first step, with candidate identifications checked against current classifications and known ranges. Uncertain or high-stakes records should go to specialists, and AI tools should be allowed to say when an image can’t be reliably identified.

AI can help researchers process large numbers of wildlife photograph. But its greatest value for conservation comes when automated predictions are combined with reliable biodiversity information and scientific expertise.

“Although the research focused on Colombian animals, the challenges it identifies are relevant to biodiversity monitoring in other regions as well,” he concluded.

Original source:

Paez Salamanca GN (2026) Is conservation AI ready? Quantifying the challenges of automated species identification in Colombia. Biodiversity Data Journal 14: e208137. https://doi.org/10.3897/BDJ.14.e208137

Pensoft at iBOL 2026: A Week in Bangkok 

Learn about Pensoft’s highlights at the 10th International Barcode of Life Conference (iBOL 2026).

From 2 to 6 November 2026, Pensoft Publishers took part in the 10th International Barcode of Life Conference (iBOL 2026) in Bangkok, Thailand. The conference brought together around 350 participants from the international barcoding community.

Our team was joined on site by Scott Edmunds, who works with us on the development of the Research Ideas and Outcomes (RIO) journal.

  • Pensoft's team at iBOL 2026
  • Opening ceremony at iBOL 2026.

Pensoft on display

Our stand drew visitors with its beautifully designed and illustrated materials and stickers. Here, delegates could learn more about our journal portfolio and wider range of services, including ORPHO, our newly launched collaborative scientific writing tool. We showcased it with a Taxonomic Treatment and an ERGA Genome Report, and visitors found both examples relevant to their own work. 

  • Visitors at Pensoft's stand at iBOL 2026
  • Visitors at Pensoft's stand at iBOL 2026

Alongside the demos, visitors could explore the journals most relevant to barcoding and biodiversity genomics. Research Ideas and Outcomes (RIO) publishes outputs from across the research cycle, from project proposals to data and software, while the Biodiversity Data Journal (BDJ) covers data papers and taxonomic treatments. On the molecular side, Metabarcoding and Metagenomics (MBMG) focuses on DNA-based methods such as eDNA and species identification, and Natural History Collections and Museomics (NHCM) addresses the care, digitisation and study of natural history collections. 

  • Pensoft's stand at iBOL 2026.
  • MBMG displayed on the big screen.
  • Journal materials.
  • Journal materials.
  • Journal materials.
  • ORPHO banner

The stand also showcased our mycology-focused journals, MycoKeys and IMA Fungus, as well as ZooKeys, PhytoKeys, NeoBiota, on biological invasions, and Acta Ichthyologica et Piscatoria, on ichthyology and fisheries. 

Beyond journals, delegates also learned about Advanced Books, our platform for publishing open-access monographs, edited volumes and other book-length scholarly works. 

  • BGE+ and BEAGLE project materials.
  • TETTRIX materials.
  • Biodiversity Meets Data brochures.
  • BAGLE project materials.

In addition, visitors found out about Pensoft’s involvement in several Horizon Europe-funded projects, including BGE+, Biodiversity Meets Data (BMD), BEAGLE and TETTRIX.

Metabarcoding and Metagenomics editorial meeting

  • Metabarcoding and Metagenomics journal editors.
  • Metabarcoding and Metagenomics journal editors.

With so many Metabarcoding and Metagenomics (MBMG) editors attending iBOL, we used the occasion to hold an in-person editorial meeting, followed by a dinner. The journal’s editorial work normally happens online, so it was a real pleasure to sit around the same table and connect in person.

The meeting gave us the chance to catch up on the journal’s progress and to talk about its direction, while the dinner kept the conversation going in a more relaxed setting. 

Celebrating scientific excellence

Autumn Sylvestri, the winner of Pensoft's award for Best Young Researcher at iBOL 2026.
Autumn Sylvestri, the winner of Pensoft’s award for Best Young Researcher.

Once again, Pensoft sponsored the best young researcher talk award, one way for us to support early-career scientists who are shaping the future of the field. This year’s award went to Autumn Sylvestri of Duke University for her presentation “Using Environmental DNA to Assess the Biodiversity Recovery of Restored Wetland Ecosystems.” 

As part of the award, she received a free publication in Metabarcoding and Metagenomics. 

Familiar faces at the conference

Beyond the stand, the programme gave us the chance to hear from several editors and authors we work with. Here are just some of our encounters:

We were also glad to welcome our dear friend Hugo de Boer at the Pensoft stand. He is a subject editor for PhytoKeys and Metabarcoding and Metagenomics, and the lead author of the open access Advanced Books title Molecular identification of plants: from sequence to species, a practical guide to identifying plants with DNA.

On the plenary stage, Iliana Bista of the Senckenberg Nature Research Institute and Natural History Museum presented “From Genes to Ecosystems: Biodiversity Genomics in Extreme Environments.” Iliana is also an editor at Metabarcoding and Metagenomics.

Iliana Bista during her talk at iBOL 2026.
Iliana Bista during her talk at iBOL 2026.

Michał Grabowski then spoke about DNA-based biodiversity monitoring beyond environmental DNA. His latest paper in MBMG, “Empowering national capacity for DNA-based approach to species identification and biodiversity monitoring,” covers this work in detail.

  • Kevin Beentjes author at Metabarcoding and Metagenomics from its very start in 2018.
  • Dr. Ayman Khamis Elsayed, subject editor at Biodiversity Data Journal.
  • Prof. Xin Zhou at Pensoft's stand. He's an editor at Metabarcoding and Metagenomics.
  • Andrea Galimberti, author at Biodiversity Data Journal.

Looking ahead

End of iBOL 2026.
End of iBOL 2026. The next conference will be in Montpellier, France.

It was a pleasure to take part in iBOL 2026 and to meet so many authors, editors and partners in person. Our team left the conference energised by everything we heard. We thank the organisers and everyone who stopped by the stand, and we look forward to continuing the conversations. 

Digital Twins for Plants: How Ordinary Cameras Are Helping Preserve Living Collections 

Researchers have shown that ordinary cameras can build detailed 3D “digital twins” of plants which could help botanical gardens, museums and conservation efforts.

Imagine a fern, tucked away in a botanical garden, that unfurls. By next week it’ll look different and by next season it may even be gone. Plants are living organisms that are constantly changing and can be fragile which makes them hard to document. 

For instance, a pressed specimen in a herbarium captures a plant in two dimensions and at a single moment. While a photograph shows one angle at a time, leaving a lot of this detail hidden.

Researchers at the University of Tartu in Estonia have recently published a study in Biodiversity Data Journal showing a practical alternative: using ordinary cameras to build detailed, explorable 3D models of plants, from a single orchid flower to an entire tropical greenhouse. 

Overall view of 14 individual 3D Gaussian Splats, visualised in Unity Game Engine. Credit to Toivo Ylinampa and Urmas Kõljalg, 2026.

Why 3D matters for plants

Branching patterns, the spiral arrangement of leaves, the curl of a petal and the structure of a flower are all traits biologists use to identify species and understand growth, yet they are hard to judge from a flat image. A 3D model preserves them.

You can rotate it, zoom in and view it from angles the photographer never chose. For gardens and museums, that means a plant can be documented, shared and studied by someone on the other side of the world. 

What the researchers did

4D workflow of the digital twins for plans
4D workflow. Credit to Toivo Ylinampa and Urmas Kõljalg, 2026.

The team combined two techniques. The first, photogrammetry, works out the 3D shape of an object from many overlapping photographs. It produces solid, measurable models, which makes it good for scientific measurement.

The second, 3D Gaussian Splatting, is a newer technique that represents a scene as millions of tiny, soft-edged blobs of colour. It renders in real time and looks strikingly photorealistic, especially for wispy, tangled vegetation that traditional methods struggle with. 

3D model and Gaussian splat
Position 1. Top-left: Textured 3D model in Metashape. Top-right: camera positions. Bottom-left: Meshconfidence level in Metashape. Bottom-right: 3D Gaussian Splat. Credit to Toivo Ylinampa and Urmas Kõljalg, 2026.

Using modest equipment, mostly a compact Olympus TG-6 camera supplemented by a macro-lens camera, a smartphone and a drone, they applied both techniques to a range of subjects.

The researchers photographed a fern outdoors in a botanical garden on twelve days across a fifteen-day period, with roughly 8,500 photos in total. They captured the garden’s tropical greenhouse in fourteen overlapping sections, using about 20,000 photos, and stitched these into an interactive virtual walkthrough.

  • An Orchid imaged with a smartphone.
  • A small flower imaged from a wall, using tripod camera with a macro-lens and studio lights.

They also modelled a potted plant, an orchid photographed with an entry-level smartphone, close-up macro shots of flowers, and a tree filmed from a drone. Enjoy the orchid model below:

Watching a plant grow in 4D

The most striking result is the fern series, which the authors call “4D” because it adds time to the three spatial dimensions. Instead of fixing cameras in place, as many growth-monitoring systems do, the researchers aligned all photos from all days in a single project, using stable background features as reference points.

4D Fern growth timelapse
4D Fern growth timelapse, 12 3DGS reconstructions. Using Metashape’s multi-day global alignment, all the resulting individual day datasets share common scale and orientation without the need for manual adjustments. Credit to Toivo Ylinampa and Urmas Kõljalg, 2026.

Each day’s model then shared the same scale and orientation, so growth could be compared directly. The result is a time-lapse you can walk around, allowing growth to be tracked without touching the plant or installing expensive equipment.

After running thousands of photogrammetric scans I realised that not all the images have to be directly related. While two pieces of fern won’t share many comparable feature points, the surrounding area has lots of them — for free!

– said the lead author, Toivo Ylinampa of University of Tartu.

Having a digital twin?

Living collections are irreplaceable, yet vulnerable to disease, climate stress, renovation and age. A digital twin cannot replace a living plant, but it can record what existed, where it grew and how it looked at a given moment.

Rare species can be documented far more richly than a single photo allows, and people who cannot travel can still explore a garden. Repeating the capture over time can also reveal how plants respond to their environment, and because the method relies on consumer cameras and widely available software, even small institutions can afford to try it. 

The limitations

The effect of wind in 3D reconstruction.

The effect of wind in 3D reconstruction. Windy conditions on the left resulted in inferior 3D reconstruction. Credit to Toivo Ylinampa and Urmas Kõljalg, 2026.

Despite the tremendous potential of this methodology, the researchers encountered several limitations. Wind was the biggest problem, with even gentle leaf movement noticeably degrading model quality, so calm days were essential.

Also, the greenhouse sections were stitched together by hand, so the walkthrough was suited to visual exploration rather than precise measurement, and processing was demanding, with the fern alignment taking about seven hours and 48 GB of memory. The two techniques play different roles: photogrammetry is better for measurement (a macro flower measured 13 mm on the model against 13.2 mm with calipers), while Gaussian Splatting better captures delicate structures such as hanging flowers.

As these tools become easier to use, every botanical garden could one day have a digital twin helping to document and preserve its species. 

Original source and more 3D models: 

Ylinampa T, Kõljalg U (2026) Photogrammetry and 3D Gaussian Splatting for 4D plant growth capture and greenhouse digital twins. Biodiversity Data Journal 14: e191083. https://doi.org/10.3897/BDJ.14.e191083 

First Individual-Based Model of Basking Sharks Replicates Complicated Social and Environmental Drivers of Behaviour

First individual-based model of basking sharks sheds light on what drives their aggregations.

Basking sharks (Cetorhinus maximus) are the second-largest fish in the world, and they’re largely solitary, following highly individual migration routes. Yet off the coast of Ireland, they sometimes gather at the surface in groups of up to 1,000. Researchers are not sure why as to why that is. Food appears to be an obvious explanation, but observations suggest it isn’t the whole story. These gatherings also put the sharks at risk of boat strikes, which makes understanding them a conservation priority.

Previous models have captured long-range migration trends but struggled to reproduce behaviour at finer scales. Chlorophyll a, the usual stand-in for the sharks’ zooplankton prey, has also matched sightings data inconsistently.

Basking shark
Basking shark (Cetorhinus maximus). Photo credit to Nigel Motyer. Photo via Irish Basking Shark Group.

A new study in Individual-based Ecology takes a different approach. Dr. Chelsea Gray of George Mason University and the Irish Basking Shark Group, with colleagues, built the first individual-based model of basking sharks. It simulates zooplankton directly, rather than relying on a proxy, and follows sharks around the Inishowen Peninsula in Co. Donegal from 1982 to 2018. The team tested four behavioural versions against three decades of public sightings: sharks seeking food only, sharks seeking other sharks only, sharks seeking both, and random movement as a control.

So what did 36 simulated years reveal, and what could it mean for protecting these animals? We reached out to Dr. Chelsea Gray to ask about her motivation, her biggest surprises from the simulations, and the insights the model could offer for conservation.


Basking shark
Basking shark. Image via Canva.

Chelsea Gray (CG): I am a member of the Irish Basking Shark Group and my previous research was focused on public perceptions of basking sharks as well as interest in basking shark tourism in Ireland. While completing that research, basking sharks were downgraded by the IUCN from ‘vulnerable’ to ‘endangered’. I was also interested in individual-based modelling.

I recognized that basking sharks, with their distinctive migratory and social behavior, were an ideal subject for an individual-based model (IBM), and that this kind of research could directly support conservation.

Ireland and other nations are currently developing or revising their ocean conservation strategies, so this is the right moment to build models and simulations that can inform policy. I have also carried out qualitative research on how to communicate IBM results to policymakers, so that this work can be put to practical use.

Basking shark
Basking shark. Photo credit to Nigel Motyer. Photo via Irish Basking Shark Group.

CG: What surprised me most was how well the model replicated the boom-and-bust cycle of shark aggregations seen in the real world.

This is a complex model, and it took considerable trial and error to get it running and to find settings that produced realistic results. When I finally compared the time series, particularly those of the Food and Food/Social submodels under Test A, the match with real-world sightings data was striking.

The version in which sharks sought both food and each other matched the sightings data most closely, suggesting aggregations emerge from a combination of environmental and social drivers.

A video of the model interface itself. Note: the sharks are not to scale, but are enlarged specifically for the video. Credit to Gray et al., 2026

The random-movement sharks never behaved like real sharks, under any settings. Their aggregations never exceeded two or three individuals, and they formed noticeably fewer aggregations overall. Testing made it clear that the emergent behaviour in the Food and Food/Social submodels resulted from the sharks’ decision-making. This was reassuring, because it shows the model’s results could not have occurred by chance.

CG: The model is still at an early stage. As our ability to measure the in-situ preyscape and to remotely sense phytoplankton and zooplankton improve, future iterations of the model will allow for more precise understanding of where the sharks are aggregating. In its current form, it can provide some insight into when (based on zooplankton availability) an aggregation might occur, but not where. It does, however, lay the groundwork for further development. It also serves as a proof of concept, showing that individual-based modelling is a valuable approach to simulating and understanding shark behaviour, even though it demands substantial effort up front.

Head of a basking shark.
Head of a basking shark. Credit to Irish Basking Shark Group.

The model also allows us to develop and test future scenarios, which is essential when considering climate change and marine protected areas. Currently, we don’t know much about the impact of climate change on shark migration, however we do have some evidence that their main prey (calanoid copepods) are shifting northwards.

Climate change could therefore impact the timing and location of future shark aggregations, and we will need to account for this when developing voluntary or statutory guidelines to reduce human threats such as ship strikes. Because the model is adaptable, it is particularly well suited to testing multiple scenarios of this kind. 

Individual-Based Ecology (IBE) fills a very important niche when it comes to environmental research. When previously looking to publish this article, it was difficult to find a journal that would be interested in a proof-of-concept model, especially one that doesn’t quite have policy application. IBMs also remain underused in environmental research, which makes it hard to find editors equipped to assess both the model design and ecological merits of a study.


The model is freely available in the paper’s supplementary material, and Dr. Gray welcomes questions and ideas from researchers interested in developing or experimenting with individual-based models of basking sharks.

Read the full paper:

Gray C, Rothman DS, Peters-Burton E, Smith C, Parsons ECM, Garretson A (2026) An individual-based model of basking sharks in Ireland. Individual-based Ecology 2: e176474. https://doi.org/10.3897/ibe.2.176474