How can AI help palaeontologists study fossils?

Artificial-intelligence methods can help palaeontologists process large sets of scans, photographs and measurements. A model may mark likely bone in a CT volume, group images by visible features or draw attention to a possible fossil in an aerial image. These are ways to sort and inspect evidence. They do not turn an uncertain object into a confirmed fossil or replace expert examination.

Finding bone in CT data

A CT scan is a stack of measurements through an object. In some fossils, bone and surrounding rock differ in density enough for software to distinguish them. Machine-learning segmentation can suggest which voxels belong to bone, a cavity or matrix. This can save time when a specimen contains many slices or tightly packed structures.

The model may fail where bone and stone have similar density, where cracks resemble boundaries, or where a fossil is crushed and mineralised unevenly. Training data also matter: an algorithm developed on one type of rock, scanner or specimen may not work well on another. Researchers inspect the original slices, correct the segmentation and record which areas were uncertain. A digital skull is only as reliable as the scan and the segmentation behind it.

Three-dimensional workflows make these results easier to measure and share. See how 3D scans and models are used in palaeontology for the difference between a surface model, a CT volume and a printed copy.

Comparing fragments and proposing a fit

Machine learning can compare shapes, surface texture or measurements across a collection and rank pieces that might be related. It can help a researcher search a large catalogue more quickly, especially when fragments are incomplete or have been labelled inconsistently. The ranked result is a shortlist to inspect, not a demonstrated anatomical connection.

Two bones may look similar because they share a broad shape, come from related animals or were altered in similar ways after burial. A proposed fit still needs to agree with anatomy, scale, geological age, locality and the condition of the specimen. Taphonomic alteration such as breakage, abrasion and mineral change can affect the signal in an image, so the preservation history of a fossil matters when algorithms compare it.

Classification and the problem of uneven data

Image classifiers can sort teeth, tracks or bones into groups based on patterns they learned from labelled examples. This can help flag candidates for further study, but performance depends on the quality and coverage of those examples. A collection dominated by one region, time period or well-known group can cause a model to perform poorly on underrepresented fossils.

Labels may also contain old identifications that later change. If a model learns those errors, it can return a confident but misleading classification. Researchers need to know how the data were selected, how uncertain examples were treated and how the result performs on material not used to train it. Confidence scores do not repair a biased or incomplete dataset.

Searching landscape images

Computer-vision systems can scan aerial, satellite or drone photographs and rank areas whose colour, texture or surface pattern resembles known fossil-bearing exposures. This may help teams prioritise field checks across a large landscape. The image records surface appearance, not the fossil itself. Vegetation, modern disturbance, shadows and geology can all produce misleading matches.

A candidate location must still be visited, mapped and evaluated by geologists and palaeontologists. The outcrop, rock type, access and stratigraphic setting determine whether a predicted signal is relevant. Software can guide attention, while field observation establishes what is actually present.

Why generative images are not fossil evidence

A generative image system can produce a plausible-looking skeleton from a prompt or partial reference. Plausibility is not evidence. Such an image may blend features from different animals, invent missing parts or give a speculative reconstruction the appearance of a documented specimen. It should not be presented as a scan, photograph or direct result of machine learning applied to a fossil.

Scientific reconstruction should identify which parts are measured, which were restored by comparison and which are artistic choices. The same distinction applies when software fills a gap in a surface or completes a shape. A prediction can be useful for testing a question, but the output remains an inference until supported by independent material.

Make the workflow reproducible

Researchers should preserve the raw scan or image, software version, training data where available, model settings and manual edits. They should test a model on independent examples and report where it fails, not only the cases it classifies correctly. Keeping the original file lets other teams check whether a result came from the specimen or from later processing.

Validation should include difficult examples: incomplete fragments, unusual preservation, small specimens and material from localities absent from the training set. A model that succeeds on familiar, well-preserved bones may fail on crushed or mineral-stained material. Researchers can compare automated results with annotations made independently by specialists and report disagreement instead of hiding it behind a single accuracy score.

It also matters which question the model was built to answer. A tool trained to outline bone in a CT scan is not automatically able to identify a species, infer behaviour or restore a complete skeleton. Each new task needs its own suitable data and tests. The palaeontological claim should remain no broader than the evidence and validation allow.

AI is most useful when it handles repetitive searching or measurement while leaving interpretation open to review. The fossil, its label, its position in the rock and its preservation history remain the basis for the claim. Algorithms can make that evidence easier to examine, but they cannot supply information that was never recorded.