Early humans and the balance of power: Homo habilis as prey

Abstract

It has been argued that Homo habilis was responsible for the earliest episodes of stone‐tool making, animal butchery, meat eating, and the reversal of the predator–prey relationship with carnivores. Assessing the empirical foundation of these premises is of utmost relevance to understanding the role that H. habilis played in our evolution. A powerful position for H. habilis, regarding carnivore–hominin interactions, requires that this hominin could cope with predation hazards. This should be reflected in bones of H. habilis impacted by scavengers instead of flesh‐eating predators. Determining carnivore taxon‐specific agency on the modification of hominin bones is crucial for solving this dichotomy. Artificial intelligence (AI) tools, through computer vision (CV) methods, have proven successful at differentiating carnivore taxa using images of bone surface modifications (BSMs). The application of CV methods to the remains of the holotype and other specimens of H. habilis documents with unprecedented reliability that Olduvai Hominin (OH) 7 and OH 65 were consumed by leopards. This has consequences for our understanding of the role played by H. habilis on the emergence of the Oldowan archeological record, and of the evolution of behaviors that led to a fully terrestrial adaptation and a shift in the balance of power between carnivorans and hominins.

The traditional view regarding Homo habilis as the primary agent in stone‐tool making and animal butchery has long shaped our understanding of human evolution. Recent advances in artificial intelligence (AI) methods have provided unprecedented insights into carnivore–hominin interactions through the analysis of bone surface modifications (BSMs). The application of these methods to some H. habilis fossils shows that these individuals were preyed on by felids, questioning their trophic role.

INTRODUCTION

Predation has been one of the major drivers in the evolution of primates. ¹ By evolving in the dry seasonal forests and woodlands of the savanna biome, ² , ³ Miocene apes must have faced a higher predation risk than modern tropical rainforest primates, where the carnivore guild is smaller. This must have impacted early hominins, which are documented for the first time with the general spreading of C4‐dominated ecosystems. ⁴ , ⁵ Not only is there evidence of C4‐plant consumption in the most forest‐adapted hominin discovered (Ardipithecus ramidus), but the bones of the fauna coexisting with this species—as well as several hominin bones (T. White, personal communication, 2025)—show abundant conspicuous evidence of large mammal carnivore damage, ⁶ as would be expected if those hominins were living in a seasonal forested habitat of a savanna biome and not in a tropical forest one. ⁵ Carnivore–hominin interaction must, therefore, have shaped our evolution, given that African savannas contain the largest mammal predatory guild and largest carnivoran biomass of all biomes.

The large accumulations of hominin fossils in the South African caves have been variously explained as the result of carnivoran predation on our ancestors, ⁷ although the most likely explanation is probably multicausal and palimpsestic. ⁸ If true, this would imply that, at least, the first half of hominin evolution took place with a subordinate position of our ancestors with respect to other large mammal carnivorans. This could be partly justified by the extended plant‐based dietary behavior of these hominins and their niche adaptations. It has been argued that at some point this changed and that there was a shift in the balance of power, sometime during the second half of our evolution, with the earliest forms of encephalized hominins. ⁹ , ¹⁰ Traditionally, this moment has been linked with the intrusion of early humans within the predatory guild. ⁹ , ¹¹

A transitional phase has been argued by some to have existed during the initial stage of hominins as carnivorans, coinciding with the emergence of stone tool use and taphonomic evidence of animal carcass butchery, which can be framed as the scavenging phase of human evolution, mostly focused on within bone marrow and brain exploitation. ¹² , ¹³ However, it should be emphasized that no uncontroversial evidence of stone tool use or carcass butchery and/or exploitation exists prior to 2.6‐million‐year‐old (Ma). ¹⁴ , ¹⁵ , ¹⁶ Current taphonomic evidence from anthropogenic sites also underscore that such a transitional phase might as well have not existed, since by 2 Ma hominins had regular/frequent primary access to small and medium‐sized animals. ¹¹ , ¹⁷

Homo habilis was originally posited as the most likely candidate to be (a) the author of the earliest stone tools (the Oldowan), (b) the earliest carnivoran hominin, and (c) the key player in overturning the carnivore–hominin balance of power. Although early Homo fossils, some of them resembling H. habilis, go back to 2.8 Ma, ¹⁸ no evidence of their trophic behavior can be reconstructed from the poorly preserved archeological record >2 Ma, ¹⁹ assuming that they were the authors thereof. Furthermore, in the past few years, several discoveries have questioned the prominent role of this hominin in each of the aforementioned propositions. First, fossils of African Homo erectus have been discovered in 2 Ma deposits. ²⁰ , ²¹ At Olduvai, the locality with one of the highest densities of H. habilis remains and where the species was originally defined, ²² postcranial remains similar to H. erectus and penecontemporaneous with H. habilis have also been discovered. ²³ There is now no certainty about the authorship of the Oldowan archeological record. Second, a partial postcranium of H. habilis was discovered and it showed an extremely primitive anatomy, ²⁴ probably not compatible with some of the behaviors inferred from the taphonomic analysis of the anthropogenic Oldowan sites. ¹¹ , ¹⁷ What is, therefore, the role played by H. habilis in the evolution of extant humans?

Although some postcranial fossils from Olduvai have been attributed to H. habilis, there is uncertainty about this attribution. ²⁵ , ²⁶ The only exception is Olduvai Hominin^a (OH) 62, which appeared associated with dental remains. ²⁴ Although the deletion of the epiphyseal ends of its long bones in this specimen is suggestive of carnivore ravaging, the lack of definitively identifiable traces of carnivore damage on their cortical surfaces prevent further analysis. In contrast, such modifications can be observed on the juvenile OH 7 specimen and they are preserved well enough to allow its study. The same applies to the OH 65 adult specimen.

In this study, we test these hypotheses: (1) if there is a shift in the balance of power along the Plio–Pleistocene boundary between hominins and carnivores, one would expect that the carnivore signal on H. habilis remains should more likely be caused by durophagous scavengers; (2) if H. habilis was not the pivotal hominin in shifting the balance of power, its remains should also display taphonomic signatures of its predators, and not (only of) its scavengers.

The recent advances in the specificity of taphonomic agent determination, through the use of artificial intelligence (AI) tools, has enabled taphonomists for the first time to approach taxon‐specific determination of carnivore damage on bone with high confidence. ²⁷ , ²⁸ , ²⁹ , ³⁰ , ³¹ , ³² , ³³ , ³⁴ , ³⁵ , ³⁶ , ³⁷ , ³⁸ , ³⁹ We will use these methods here to determine which carnivore(s) modified the OH 7 and OH 65 cranial remains. We are not including postcranial remains found in Bed I, not even those also found on FLK NN3, where OH 7 was discovered, because their attribution to H. habilis is not definitive. ²⁵ , ²⁶ OH 8 has been described as belonging to an adult individual, ²⁵ whereas the OH 7 hand and skull bones are from a juvenile individual.

MATERIALS AND METHODS

Materials

The bones that we analyzed here are those that correspond to OH 7 and OH 65. OH 7 was discovered in 1960 by Jonathan Leakey at FLK NN ⁴⁰ and is currently stored at the Museum and House of Culture of Dar es Salaam, National Museum of Tanzania (NMT). OH 65 was discovered in 1995 by the Olduvai Landscape Paleoanthropology Project (OLAPP) ⁴¹ and is currently stored at the Natural History Museum of Arusha, NMT.

OH 7

Following the work by L. S. B. Leakey et al., ²² where OH 7 was presented as the holotype of H. habilis, this specimen is composed of a mandible, two parietal bones, several hand bones, and an upper molar associated with the mandible. Part of this specimen was previously described by L. S. B. Leakey ⁴² , ⁴³ and the OH 7 mandible and skull have been extensively studied by Tobias. ⁴⁴ Despite the fact that the OH 7 bones have been mentioned or studied in several works, ²⁶ , ⁴⁵ , ⁴⁶ , ⁴⁷ , ⁴⁸ , ⁴⁹ , ⁵⁰ , ⁵¹ , ⁵² , ⁵³ , ⁵⁴ , ⁵⁵ , ⁵⁶ , ⁵⁷ we enumerate here the list of OH 7 bones initially presented by M. D. Leakey ⁴⁰ in Olduvai Gorge, Volume 3 as part of a juvenile individual: a mandible, which is not complete; the left (nearly complete) and parts of the right parietal; a piece of the occipital bone; fragments of the petrous bones (right and left); different skull fragments of a small size; and 21 bones of the hand, with some phalanges, the scaphoid, and the trapezium included. These bones are the ones that were found in situ, and “one upper molar, a terminal phalanx and a broken capitate bone, found on the surface, probably also belong to this individual.” ⁴⁰ Subsequently, Day ⁵⁸ updated and reduced the list of the OH 7 hand bones as follows: of the 21 hand bones enumerated by M. D. Leakey, ⁴⁰ six of them were discarded (they are nonhominin and one is not a hand bone) and of the rest (15), two were adult, seven were juvenile, and six were classified as of uncertain age; of the 13 bones that correspond to OH 7, one is left‐sided and four are right‐sided and “at least two hands are involved.” ⁵⁸

At the National Museum of Dar es Salaam, the bones that were stored as part of OH 7 and that we had access to were the following: the mandible with its dentition (A); the left and right parietal bones (B and C); two fragments of the petrous bone (D and E); three terminal phalanges, one being the thumb (F); four middle phalanges (H); two proximal phalanges (J and K); three carpal bones (G, L, and M), comprising a scaphoid (G), a trapezium (L), and a capitate (M); and the base of second metacarpal (I) (Figure 1). The last one does not appear in the figure presented by Napier ⁴⁸ with the hand bones of OH 7. The hand bones classified as juvenile by this author are three terminal phalanges (F) and four middle phalanges (H).

FIGURE 1.

OH 7 bones stored at the National Museum of Tanzania in Dar es Salaam. (A) Mandible. (B) Left and (C) right parietal bone. (D, E) Two skull fragments. (F) Three terminal phalanges. (G) A carpal bone (scaphoid). (H) Four middle phalanges. (I) The base of the second metacarpal. (J, K) Two proximal phalanges. (L) A carpal bone (trapezium). (M) A carpal bone (capitate). Note that the letters given to the OH 7 bones here do not match the ones given to the bones at the National Museum.

FLK NN, where the OH 7 specimens were found, is situated in Bed I, and was discovered in 1960. Bed I was dated to between 1.99 and 1.75 Ma; ⁵⁹ , ⁶⁰ , ⁶¹ , ⁶² specifically, the FLK NN site is located between tuffs IC and IB, dated to 1.84 and 1.85 Ma, respectively, ⁶⁰ , ⁶³ being one of the oldest sites in Olduvai Gorge. ⁴⁰ , ⁶⁴ At FLK NN, four different levels were documented initially. FLK NN levels 1–3 have been taphonomically studied ⁶⁵ and more recently reinterpreted by Domínguez‐Rodrigo et al: ¹¹ FKL NN1 is a palimpsest, where hominins were not the main agent responsible for the bone accumulation, and where the moderate modification of bones is attributed to carnivores, namely, felids; ⁶⁶ FKL NN2 has been interpreted as a carnivores accumulation ⁶⁵ where felids played an important role; ⁶⁷ and FLK NN3, where OH 7 was discovered, “is also a natural accumulation, probably formed by carnivores.” ⁶⁴ With the exception of three bones, the OH 7 remains were discovered widely scattered “over an occupation floor on a palaeosol which was on the surface of a clay immediately overlying Tuff IB.” ⁴⁰ , ⁴⁴ An exhaustive description of the distribution of OH 7 bones can be found in Tobias. ⁴⁴

Archeofaunal bones at FLK NN3 exhibit a very good preservation of the cortical surfaces ⁶⁴ and this allowed a detailed analysis of bone surfaces with confidence. OH 7 cortical surfaces also display good cortical preservation, with the exception of three fragments (see below).

OH 65

Following the work by Blumenschine et al., ⁵⁹ OH 65 is composed of a nearly complete maxilla with all of its teeth and the lower face, and it was found on the western part of Olduvai Gorge in Bed I. The minimum age of the location where the specimen was discovered is 1.785 Ma. ⁵⁹ Regarding the locality where OH 65 was found, these authors indicate that there was a stream near a lake margin during low lake level. ⁵⁹ Several faunal remains and stone tools were also found at this site. In contrast with OH 7, OH 65 is an adult individual. ⁴¹

Methods

Taphonomic analysis

At the National Museums of Dar es Salaam and Arusha, a taphonomic analysis was conducted of all the OH 7 bones and the OH 65 maxilla to record the following information: (1) bone surface modifications (BSMs); (2) the preservation of the cortical bone surfaces; and (3) the breakage patterns of the bones. The conspicuous BSMs (marks 1–7 in the case of OH 7 and mark 1′ in the case of OH 65) detected were documented and photographed with a binocular Optika microscope and a 3 Mpx digital camera (OptiCamB3). Marks were photographed with a different magnification depending on the size of each one (from 15× to 30× for tooth pits and from 10× to 15× for score‐like grooves in the case of OH 7 and 45× [and 40×] for the tooth pit documented on OH 65) to fit the mark′s border to the image frame. Because of the characteristics of the microscope and the camera used to take the images, the focus‐stacking technique had to be employed to obtain focused images without blurry areas. For that, multiples images of the same mark were taken, focusing on different areas of the surface for the purpose of obtaining the focused image and overcoming problems with the depth of field. The images were then merged using the focus‐stacking tool of Adobe Photoshop. In the case of scores that are larger in size (marks 3–6 or A–D in the Supporting Information), this procedure was done in sections. Regarding OH 7, the bones that present conspicuous marks and that have been analyzed in detail are the mandible, the right parietal, and one of the middle phalanges. The marks of OH 7 and OH 65 that were classified as made by a carnivore agency were then analyzed in detail using AI methods; these were restricted to the two tooth pits found on the OH 7 mandible and to one tooth pit found on the OH 65 maxilla. The AI methods employed here used only different carnivore taxa for the analysis of tooth marks. Carnivore agency on the OH 7 fossil had been determined previously under adequate magnification. It has been experimentally shown that carnivore tooth marks can be easily identified after a few hours of training. ⁶⁸ Fortunately, they lack the typical signatures that other marks like cut/butchery marks or trampling/abrasion marks have, which can be difficult to differentiate. ⁶⁹ Carnivore tooth marks can be also easily distinguished from biochemical marks. ⁷⁰ Therefore, no sophisticated AI model is needed to identify a carnivore agency. It is only when seeking carnivore taxonomic identification that AI contributed to the more specific detection of agency.

The experimental sample used for the AI application

A total of 1496 images of tooth marks, including tooth pits and scores, were used for the analysis. The experimental sample initially used in Domínguez‐Rodrigo et al. ³⁵ has been increased by adding new marks to the crocodile collection and introducing a new group of carnivores. Since no canid was originally added to that reference sample and given that large canids were also present during the East African early Pleistocene, ⁷¹ we decided to incorporate a modern large canid (wolves) sample as a proxy. Although an extensive description of the experiments carried out with the four carnivores (crocodiles, hyenas, leopards, and lions) used in the initial experimental sample can be found in Domínguez‐Rodrigo et al., ³⁵ we compile here a summary of the experimental data set of the four groups of carnivores (crocodiles, hyenas, leopards, and lions) and introduce the data of the expanded crocodile sample and the canid sample. The tooth mark sample is finally made up of crocodiles (n = 124), hyenas (n = 364), leopards (n = 544), lions (n = 264), and wolves (n = 200).

The tooth mark sample of crocodiles comprises two distinct experimental collections. The first (n = 84), detailed in Baquedano et al. ⁷² and Domínguez‐Rodrigo et al., ³⁵ was conducted by two authors of this study (Enrique Baquedano and Manuel Domínguez‐Rodrigo) and involved crocodiles (Crocodylus niloticus) from Faunia Zoo (Madrid, Spain). Female crocodiles of various sizes were provided with articulated limbs of suids (pigs and boars) and bovids (sheep and cows), along with scapulae and pelves, over a 4‐month period. Carcasses were fed to the crocodiles weekly and retrieved after 15 h of exposure. Following the experiment, bone remains underwent cleaning as per previously cited protocols. The second sample of crocodile tooth marks (n = 40) was obtained from an experiment conducted by Edgard Camarós at Altamira Zoo. Two adult dwarf crocodiles (Osteolaemus tetraspis), both male, were utilized in this experiment. Carcasses, comprising partly defleshed limbs of adult pigs and a pelvis, were collected after 10 min of exposure.

The tooth mark sample of hyenas also consists of two distinct experimental collections, detailed in Prendergast and Domínguez Rodrigo ⁷³ and Domínguez‐Rodrigo et al. ³⁵ The first collection was obtained from an Eyasi spotted hyena den, predominantly featuring bones modified by hyena cubs. The second collection resulted from experimental work conducted at the private Madrid Safari Reserve (Spain), involving three striped hyenas (Hyaena hyaena).^b One male and two female hyenas, aged 6–18 years, were included. Bones were collected after 1 h of exposure to prevent complete breakage. Bones provided to the hyenas were disarticulated and defleshed bones of adult deer (Cervus elaphus), with cleaning conducted following established protocols.

The tooth mark sample of leopards encompasses two experimental collections detailed in Domínguez‐Rodrigo et al. ³⁵ The first collection, generated at Madrid Zoo (Spain), involved three male Persian leopards (Panthera pardus saxicolor), aged 4–9 years, fed with complete adult sheep limbs (Ovis aries). The second collection, carried out at Bahari Zoo (Dar es Salaam, Tanzania), involved feeding articulated, fleshed goat (Capra hircus) forelimbs and hindlimbs to a leopard (Panthera pardus).

The tooth mark sample of lions also comprises two distinct experimental collections. The first collection, generated at Madrid Zoo (Spain), involved three adult Asiatic lions (Panthera leo persica) fed with slightly defleshed bovine (Bos taurus) limbs, which were collected after 12 h of exposure. The second collection, generated at the private Cabárceno reserve (Cantabria, Spain), involved 11 adult lions fed with equid (Equus ferus caballus) limb bones, which were collected after 1–4 days of exposure. In both cases, bone cleaning followed established protocols.

Regarding the tooth mark sample of wolves, a total of 125 pits and 75 scores, obtained from an experiment described in Moclán et al. ⁷⁴ at the El Hosquillo reserve (Cuenca, Spain), were included. Five captive Iberian wolves (Canis lupus) were involved in the experiment, fed with disarticulated limb bones of cervids, ovicaprids, and suids. Carcasses were collected after 3 months of exposure, and bone cleaning followed established protocols.

Here, we implemented a protocol modified from the one outlined in Domínguez‐Rodrigo et al., ³⁵ utilizing color photographs instead of black and white images. Additionally, we combined tooth pits and scores for the analysis, diverging from our previous practice of separately analyzing the pits. This amalgamation enhances modeling power by augmenting the sample size beyond that of the tooth pit sample alone. Tooth scores were documented under 30×/32× magnification, while tooth pits were documented under various magnifications, ranging from 7.5× to 60× depending on the size of each mark (7.5×, 10×, 12.5×, 16×, 20×, 25×, 32×/30×, 40×, 50×, and 60×). The tooth mark image database was compiled using a Leica Emspira 3 digital microscope. However, a portion of the tooth mark sample included in the database was documented using different microscopes: crocodile tooth marks from Altamira Zoo were documented using a Hirox digital microscope (optics HR‐2016); leopard tooth marks from Bahari Zoo were documented using an Optika binocular microscope, employing the same procedure as that described for documenting the OH 7 tooth marks (see above); and lion tooth mark images from the Cabárceno sample were documented using an Optika binocular microscope, utilizing the original images included in previous work. All images and code are accessible in a public repository: ⁷⁵ https://doi.org/10.7910/DVN/YAEADV.

Deep learning (DL) analysis

DL methods have been used to derive classification models of the five reference modern carnivores (crocodiles, hyenas, leopards, lions, and wolves). The first four carnivores are the predominant large carnivores (as potential predators or scavengers of hominin‐sized primates) in modern African savannas. These carnivores were employed in the first generation of models used as both a reference BSM dataset and as classification architectures for archeological BSMs. ³⁵ The addition of wolves to this sample was intended as a proxy for large African Early Pleistocene canids.

We constructed model architectures using transfer learning (TL), based on sequential and residual DL models: ResNet‐50 (version 1.0), ⁷⁶ VGG‐19, ⁷⁷ and DenseNet‐201. ⁷⁸ We generated individual models, and then we used them together for the experimental sample classification, using ensemble learning (EL), intended only as a comparison with the performance of the individual models. The EL method used here consisted of using the three DL models as base learners, and then a stacking process was implemented by using a random forest and an extra‐gradient boosted tree as the meta‐learner. The number of estimators used in hyperparameter tuning was 100. We did not use EL for archeological BSM classification because we discovered that classification was more reliable when using images of BSMs preprocessed using the same method as each individual model. When using EL, BSM images to be classified had to be preprocessed using a standard normalization procedure, which was not the one used for each independent model.

Prior to the DL analysis, the original 1496 tooth mark images were divided into a training set (75%) and a testing set (25%). This split was randomized. The original training dataset was artificially expanded by using image augmentation, which is an efficient way of decreasing overfitting. This was done through the implementation of the following procedures: random shifting of width and height (20%), modification of shear and zoom range (20%), horizontal flipping, and a rotation range of 40°. Image standardization, prior to the use of images by the DL models, was carried out using each architecture′s preprocessing functions. All images were reshaped to 250 × 200 pixels.

As was done with the previous models using four carnivores, ³⁵ exploratory TL modeling was done using different combinations of optimizers (stochastic gradient descent [SGD] and Adagrad) and activation functions (ReLU and Swish). For the five‐carnivore data set, the best combinations involved the ReLU function and the Adagrad optimizer (with a learning rate of 0.001). The loss function selected was categorical cross‐entropy. ⁷⁹ Accuracy was the metric selected for the evaluation of the classification process. F1 score values were also obtained to assess balanced accuracy, given the imbalanced nature of the original data set. Training was done using mini‐batch kernels of size 32. Testing was done using mini‐batch kernels of size 20. Weight update was done using a backpropagation process of 100 epochs. Regularization methods targeting limiting overfitting were also added to the models’ architectures. Here, we used dropout methods (30%), consisting of the random dropping of 30% of neurons during training. Training and validation graphs were also used to monitor model building processes, their performance, and potential under‐ and overfitting.

The DL models were built using the Keras (2.4.3) application programming interface (API) with a Tensorflow (2.3.0) backend. Computation was carried out on a GPU HP Z6 Workstation using a CUDA computing (cuDNN) environment. All coding was done using Python 3.7.

Procedure for establishing a minimum confidence threshold for the archeological BSM classification

The resulting DL models were used to identify taxon‐specific agency in the carnivore modifications of OH 7 and OH 65. For this purpose, and in contrast with other classification methods, like geometric morphometric (GMM) analyses of tooth marks, we establish a minimum confidence threshold for mark classification so that our inferences are based on high‐power models. GMM analyses of tooth marks have exhibited enhanced discriminatory capabilities; however, their reliability remains questionable until statistically robust samples are utilized, encompassing the complete spectrum of tooth mark forms within each carnivoran taxon. Studies employing GMM methods may be influenced by biased mark selection, particularly toward larger marks, which could obscure agent‐specific allometric variation, and by allometric variation. This is further masked by the successful classification estimates achieved using these methods. Classical discriminant methods and machine learning (ML) algorithms may lead to high accuracy but low reliability due to extensive overlap of tooth marks from different agents. Even with sophisticated methodology, such overlap undermines the reliability of highly accurate models. Therefore, BSMs in the archeological record should be classified with a higher confidence, relying only on high‐power models.

In the DL methods used in the present study, high‐power modeling was achieved using a threshold with a minimum of 70% probability of correct classification for each mark, which constitutes a fairly reliable threshold if we consider that the reference framework involves five different carnivores. We also prioritize concurrence in agency identification through EL of each of the participating models.

RESULTS

General taphonomic analysis

The OH 7 mandible

The mandible displays good cortical preservation. It is not complete, given the absence of portions of the mandibular rami and corpus. The left mandibular corpus is complete but the right one ends at the level of the first molar. The right ramus is missing, and the left ramus is also missing except for a small part. The dental pieces that are present range from the right first molar to the left second molar. The left side of the mandible has a green fracture plane that starts at the ramus and extends to the anterior part of the mental foramen or left P₃. Next to this plane, and associated with it, there is a large triangular tooth pit (mark 1), which is located on the external/buccal face of the mandible (Figures 2 and 3A). The pit is under the left M₂, specifically in the anterior area of this tooth. The mandible also presents biochemical marks at the level of the left C₁ and P₃. The left canine bears surficial biochemical marks. The left mandibular corpus, in inferior view and on its lingual side, shows a crenulated edge (Figure 3B), with two small micronotches, one of them with a negative scar on the inside. The other micronotch, also present on the lingual side, is in the area close to the junction of both sides of the mandible. It presents green breakage, with moderate angulation, similar to that caused by pressure (static loading). On the right side, in inferior view and on its lingual side, there is also a micronotch between the right P₃ and P₄. The right mandibular corpus presents a green fracture plane at the level of the right M₁ on its lingual side and it is associated with another large triangular tooth pit (mark 2) (Figures 2 and 3C). The pit is located on the lingual face of the mandibular body, between the right P₃ and P₄. Both the pit on the right lateral (lingual surface) and the micronotch, which is in the lower part of this lateral, are located at the same level.

FIGURE 2.

OH 7 mandible with the two tooth pits that were documented and magnified separately. The tooth pits were photographed with a binocular Optika microscope and a 3 Mpx digital camera (OptiCamB3) with a magnification of 10×.

FIGURE 3.

Detail of the OH 7 mandible showing green breakage. (A) Green fracture plane and the tooth mark 1 on the left side. (B) Inferior view of the left mandibular corpus with the crenulated edge. (C) Green fracture plane and the mark 2 on the right side.

These two carnivore tooth marks (Figure 2) are the subject of analysis using DL methods (see below). Their triangular morphology is reminiscent of the same morphology documented among some types of felid tooth marks (Figure 4). ⁸⁰

FIGURE 4.

Examples of tooth pits recorded from the leopard experimental sample (A–F) and the two tooth pits documented on the OH 7 mandible (G, H). (A–D) Leopard tooth pits from the Bahari Zoo sample (Dar es Salaam, Tanzania; 30× magnification for A, B, and D and of 20× for C). (E, F) Leopard tooth pits from the Madrid Zoo collection (Spain; 50× and 20× magnification or zoom of 5.0 and 2.0, respectively). (G) OH 7 tooth mark 1 (30× magnification). (H) OH 7 tooth mark 2 (15× magnification). Photographs in A–D by Gabriel Cifuentes‐Alcobendas.

The OH 7 parietal bones

Both the left and right parietals also display good cortical preservation. The left one, which is almost complete, presents biochemical modifications. The ectocranial surface shows bioerosion and manganese in the form of small speckles, but also extended in patches. The endocranial surface, although it displays good preservation, is more damaged than the ectocranial surface, with substantial modification of its external cortical layer. Regarding the right parietal, both the endocranial and ectocranial surfaces display good cortical preservation. The ectocranial side also presents manganese staining and bioerosion. Indeed, the four conspicuous marks (marks 3–6) traditionally classified as tooth marks ⁴⁰ , ⁴⁴ , ⁴⁵ , ⁵¹ , ⁵² are biochemical marks (Figure 5; see extensive analysis in the Supporting Information).

FIGURE 5.

OH 7 right parietal with the four biochemical scores shown in detail (marks 3–6). The marks were photographed with a binocular Optika microscope and a 3 Mpx digital camera (OptiCamB3) with a magnification of 7×. See the Supporting Information for additional description and analysis.

The other OH 7 bones

Regarding the fragments of the petrous bones, both are well preserved, with one of them displaying green breakage. The three terminal phalanges also show good cortical preservation and present manganese in the form of small dots. Concerning the four middle phalanges all exhibit good cortical preservation and manganese. One of them has a pit on the dorsal face (Figure 6). This pit, with irregular edges and substantial exfoliation, is also the result of bioerosion, which would explain its complete preservation despite its small size. Two of the three metacarpal bones show good cortical preservation and manganese dioxide, while one of them has lost almost all the cortical surface. Regarding the two proximal phalanges and the base of second metacarpal, one of the proximal phalanges displays good cortical preservation while the other two bones display moderate cortical preservation, and the three of them present fractures and manganese.

FIGURE 6.

OH 7 middle phalange with a biochemical pit shown in detail (mark 7). The pit was documented with a binocular Optika microscope and a 3 Mpx digital camera (OptiCamB3) with a magnification of 20×.

OH 65

The nearly complete maxilla displays moderate cortical preservation, since it preserves about 50% of its cortical surface. Even if the complete dentition were present, the maxilla is not complete, given the absence of some portions. This fossil bone bears fractures around the edges. On its left side and associated with a fracture, there is a tooth pit (mark 1′) which is located on the external/buccal face of the maxilla (Figure 7). The maxilla presents biochemical alteration and marks, highlighting the presence of manganese in different parts of the bone. These biochemical modifications can be also observed on its dental pieces. We have decided to include OH 65 in the present work for two reasons: (1) OH 65 exhibits a tooth pit on its maxilla, which in this case is similar to that documented in the OH 7 specimen; and (2) there is no doubt that OH 65 belongs to the same species as OH 7, which is not uncontroversially the case in other specimens attributed to H. habilis. ⁴¹

FIGURE 7.

OH 65 maxilla with the tooth pit documented and magnified. The tooth pit was photographed with a binocular Optika microscope and a 3 Mpx digital camera (OptiCamB3) with a magnification of 30×.

Classification of the BSMs through DL models

The use of the three models (ResNet‐50, DenseNet‐201, and VGG‐19) on the five‐carnivore experimental data set yielded an accuracy of 88% (loss = 0.37), 81% (loss = 0.44), and 76% (loss = 0.68), respectively. ResNet‐50 was the most successful model, resulting from the combination of Adagrad as the optimizer and ReLU as the activation function. This combination was also the one that yielded the best results for the other two models.

The training graphs show a close match of the training and validation process (accuracy and loss), with a limited amount of overfitting during training for the three models. Despite this, the three of them yielded high accuracy in the classification of unseen testing images (Figure 8). The high accuracy of these models, especially of Resnet‐50, indicates that taxon‐specific agency can be determined with confidence through the use of these methods. The ensemble analysis applied to the testing experimental set yielded an accuracy of 0.85 of correct validation using an extra‐gradient boosted tree and an accuracy of 0.87 using a random forest.

FIGURE 8.

Model accuracy (left) and loss (right) for ResNet‐50 (upper), DenseNet‐201 (middle), and VGG‐19 (lower).

The two tooth pits found on the OH 7 mandible were analyzed by applying DL models (Table 1). The ResNet‐50 model classified mark 1 as made by leopards with a probability of 0.97; DenseNet‐201 also classified mark 1 as made by leopards with a probability of 1/1.0; and VGG‐19 classified mark 1 as leopard‐inflicted with a probability of 1/1.0. Regarding mark 2, both ResNet‐50 and DenseNet‐201 classified the tooth pit as leopard‐made, with probabilities >0.95. In contrast, VGG‐19 classified this tooth pit as hyena‐made with a probability of 0.53, while it classified it as being a leopard tooth mark with a probability of 0.46. VGG‐19 is the worst performing model, and for this particular mark it does not seem certain in the identification since agency was determined by this model with probabilities lower than 70%, which should be considered the minimum threshold of confidence.

TABLE 1.

Classification of bone surface modifications (BSMs) through the ResNet‐50, DenseNet‐201, and VGG‐19 models.

Mark	ID mark	Model	Carnivores
			Crocodile	Hyena	Leopard	Lion	Wolf
1	OH7+30+1	ResNet‐50	0.006	0.007	0.97	0.003	0.02
		DenseNet‐201	0	0	1	0	0
		VGG‐19	0	0.0007	1	0	0
2	OH7+15+2B	ResNet‐50	0.02	0.002	0.95	0.02	0.005
		DenseNet‐201	0.0002	0.0006	0.99	0.005	0
		VGG‐19	0.001	0.53	0.46	0.0009	0
1′	OH65+45+1	ResNet‐50	0.04	0.08	0.13	0.0006	0.75
		DenseNet‐201	0.002	0.02	0.95	0.02	0.002
		VGG‐19	0.06	0.44	0.46	0.03	0.007
1′	OH65+40+1	ResNet‐50	0.0004	0.03	0.005	0.0006	0.96
		DenseNet‐201	0.007	0.04	0.91	0.04	0.0002
		VGG‐19	0.0001	0.0006	0.98	0.02	0

Note: Tooth marks 1 and 2 are both in the mandible of OH 7; mark 1′ is in the OH 65 maxilla. ID mark, identification mark. The name assigned to the images in the ID consists of the hominin′s name, followed by the magnification and the mark′s identification number. Values in bold are the carnivore selected by each of the models.

Regarding OH 65 (Table 1), the ResNet‐50 model classified mark 1′ with magnification of 45× as made by wolves with a probability of 0.75; DenseNet‐201 classified it as made by leopards with a probability of 0.95; and VGG‐19 classified it as made by leopard with a probability of 0.46. Once again, VGG‐19 does not seem certain in the identification of the mark. Regarding the OH 65 mark 1′ with magnification of 40×, ResNet‐50 also classified it as made by wolves with a probability of 0.96; DenseNet‐201 coincides in the classification of this mark with a probability of 0.91; and VGG‐19 also classified it as made by leopards with a probability of 0.98.

In summary, regarding OH 7, the high confidence of the best performing models in classifying the two tooth pits clearly indicates that the modifications resulted from leopard agency in both cases. Regarding OH 65, we used EL (majority voting) to assess agency, since one model (ResNet‐50) classified mark 1' as made by wolves, but the other two models classified it as made by leopards.

DISCUSSION

Taphonomy is all about biases in the fossil record. When dealing with potential agent‐specific taphonomic signatures, one could argue whether some agent′s conspicuous impact on fossil bones would be better preserved than others. In this regard, one could ask whether the actions of felid (i.e., flesh‐eating) and hyenid and canid (i.e., durophagous flesh‐eating and bone‐crunching) carnivores can be equally detected on fossil bones and paleolandscapes. It has been extensively shown that bone‐crunching carnivores leave a more abundant number of traces on their prey skeletons than strict flesh‐eating carnivores. ⁸¹ , ⁸² , ⁸³ , ⁸⁴ , ⁸⁵ , ⁸⁶ , ⁸⁷ , ⁸⁸ From a quantitative point of view, the bias should be against felids. A detailed taphonomic analysis of all the Bed I sites from Olduvai Gorge showed how most of the damage documented in the anthropogenic sites (i.e., those created by dynamic accumulation and carcass processing by hominins) was made by hyenas. ¹¹ At the same time, it was shown that felids had a more prominent role in accumulating and consuming carcasses at palimpsests (i.e., sites with multiple agency, where hominin input in carcass transport, accumulation, and consumption was marginal or nonexistent). The three archeological levels at FLK NN are examples of such palimpsests. In all of them, felids were identified as the primary agent of carcass accumulation and damage. ¹¹ In some archeological sites/levels from Olduvai Bed I, felid–hyenid interaction has left a record where the damage inflicted by both types of carnivores on the same carcasses was readily identifiable. ¹¹ At a paleolandscape scale, Bed I was characterized by a more humid environment hosting a more wooded and even forested ecosystem. ⁸⁹ There, carnivore signatures from medium‐sized (i.e., leopards) to large‐sized (i.e., lions) felids are more abundant than those of durophagous carnivores, as would correspond with such an environment. ⁹⁰ , ⁹¹ During Bed II times, a drying process led to a more xeric and arid environment, with a predominance of open ecosystems, with grazing herbivores better represented than browsing ones, and a predominant signature of hyenids over felids, as would correspond with such environments. ⁹² An additional objective assessment of this lies in the ability of computer vision (CV) models to discriminate felid and hyenid BSMs evenly. ²⁷ , ²⁸ , ³⁵ This is even the case when comparing interpretations derived from traditional taphonomic studies of bone assemblages and the identification of agency on the same assemblages using CV methods. ³⁹ The predominance of hyenid damage on bones postdepositionally discarded by hominins at DS (Bed I), as interpreted by skeletal profiling and anatomical distribution of carnivore damage, has recently been reassessed by CV methods, identifying a predominant hyenid role in the generation of the identified tooth marks. ³¹ In only one case at the same site, a wildebeest carcass showed traces of an initial modification made by a felid, inferred by the type of damage documented on an ulnar olecranon, and the subsequent CV and GMM analysis of the documented tooth marks confirmed a lion agency. ³⁴ All this together indicates that at Olduvai the potential of detecting felid and hyenid taphonomic signatures on bones is not biased and is taphonomically equally feasible.

The identification of a predominant felid agency in the modification of carcasses at FLK NN reinforces the attribution of the damage documented on OH 7, which was found amid the felid‐modified remains (see below). ¹¹

M. D. Leakey pointed out that both OH 7 and OH 8 had tooth marks “indicating that they had been damaged by scavengers”; furthermore, she described the marks found on the OH 7 right parietal as follows: “there are four roughly parallel grooves 4–5 mm wide on the parietal.” ⁴⁰ M. D. Leakey also indicated that the agent responsible for the tooth marks would have been a large‐sized carnivore and questioned that the hyenas might have modified the OH 7 parietal, since the bone is very slender; however, it must be noted that she mentioned that the left parietal “exhibits a crushed area, 7 × 5.5 mm. in diameter, situated almost centrally, in which the bone has become pulverized and from which there are a number of radiating cracks.” ⁴⁰ M. D. Leakey attributed these cracks to natural causes (e.g., sedimentary pressure); nevertheless, this area of the left parietal presents biochemical alteration (see Results  and the Supporting Information). Concerning the OH 7 skull fragments, Reader ⁵² also mentions the probable tooth marks, linking them to hyena activity. Tobias also interprets the parietal marks as tooth marks, but he does not specify the modifying agent. He argues that the marks present a “symmetrical pattern, the two outer members being suggestive of the impact of two large lateral incisors or canines and the two inner members of two smaller, medial incisors.” ⁴⁴ Davidson and Solomon, ⁴⁵ in contrast, propose that crocodiles are the agent responsible for the three of the “tooth marks” documented previously on the parietal. Njau and Blumenschine ⁵⁰ suggest that the interpretation proposed by Davidson and Solomon ⁴⁵ would be in line with what they propose to explain the interactions between hominins and crocodiles at Olduvai Gorge during the Pleistocene. However, in 2012 these authors questioned that the parietal marks (scores) were generated by crocodiles due to their large width, taking into account their experimental assemblage with crocodiles in captivity. ⁵¹ Concerning both the parietals and the mandible, Njau and Blumenschine ⁵¹ mention that these bones had been damaged by carnivores, specifically, by one similar in size to a leopard. In relation to the OH 7 hand bones, they argue that carnivore damage is “minimal and nondiagnostic,” referring to the tooth marks on “two phalanges of the OH 7 hand.” ⁵¹ However, as we indicated before (see above), only one conspicuous pit has been documented in one of the phalanges and it has been classified as a biochemical alteration. Regardless of the carnivore involved, a bite to such a small phalanx would have broken it, instead of leaving it intact.

Regarding the OH 7 mandible, Njau and Blumenschine ⁵¹ argue that some parts of it have been destroyed, presenting crenulated break edges, and that it bears a tooth score, as the result of the action of a leopard‐like carnivore. However, we have not documented any tooth score on the OH 7 mandible. These authors conclude, referring to OH 7, OH 8, and OH 35, that “in neither case is it possible to discern the consumer sequence, nor which carnivore, if either, was the predator.” ⁵¹

Given the evidence presented here, the most parsimonious interpretation is that OH 7 was preyed on by a leopard. FLK NN 1 was initially interpreted by M. D. Leakey ⁴⁰ as a living floor where stone tools and fossil bones were found in association. FLK NN 2 was understood as a paleontological site where only fossil bones were documented. ⁴⁰ FLK NN 3, the level in which OH 7 was found, was also interpreted as a living floor, where stone tools and fossil bones were in association with the hominin bones. ⁴⁰ Regarding FLK NN1, Binford ⁹³ pointed out that it might have been a natural carnivore‐made assemblage. FLK NN2 was interpreted by this author as a natural death place of large mammals. ⁹³ Bunn ⁹⁴ and Potts ⁹⁵ both also interpreted FLK NN2 as an accumulation made by carnivores, specifically by hyenids. Regarding FLK NN3, Binford ⁹³ interpreted the site as a place where hominins were accumulating faunal remains previously consumed by other carnivores, and Potts ⁶⁵ pointed out that hominins had been participating in the accumulation of at least some of the remains.

In the taphonomic reinterpretation of the FLK NN site carried out in the 2000s, it was concluded that FLK NN1 is a palimpsest, where the moderate modification of bones is attributed to carnivores, namely, felids; ² , ⁶⁶ FLK NN2 is a carnivore accumulation where felids played an important role; ³ , ⁶⁷ and FLK NN3 is also an accumulation formed by felids. ⁶⁴ Concerning FLK NN3, Domínguez‐Rodrigo and Barba ⁶⁴ concluded that there was no demonstrable relationship between the deposition of lithic artifacts and fossil bones, in the form of green‐broken fragments through dynamic hammerstone loading, percussion marks of other butchery marks. The prominent role of felids, most likely medium‐sized, as suggested by the size of the fauna accumulated, supported the idea that a leopard (or a similarly sized saber tooth felid) could have played a role in the deposition of remains at the site, as was the case of the nearby site of FLK North. ⁹⁶ , ⁹⁷ , ⁹⁸ , ⁹⁹ The predominant felid taphonomic signature on the three FLK NN archeofaunal assemblages lends support to the interpretation that OH 7 was consumed by a leopard, as identified by the DL models with high probability (Table 1). The predation by leopards on the small game represented in the FLK NN paleosurfaces underscores the paucity of primate bones at Olduvai, which would also be expected to have been preyed upon by a medium‐sized felid, as reflected in the accumulations of hominins and baboons in South African caves. ⁷ This can also be explained not necessarily by a paucity of primates in the lacustrine Olduvai basin, since Theropithecus remains have been recovered in most Bed I sites, but to the fact that leopards in open air settings accumulate mostly small antelopes. ¹⁰⁰ Only in caves and shelters that are being actively used by baboons as sleeping sites can leopards succeed on preying and contributing to the natural accumulation of primate carcasses. ⁷ , ⁸ , ¹⁰¹

Regarding OH 65 and considering the results of one of the best performing models and the EL (majority voting), the interpretation is that OH 65 was preyed on also by a felid, and specifically by a leopard. If we take both individuals (OH 7 and OH 65) as random representatives of the larger H. habilis population that lived at Olduvai, their convergent signal of having been preyed on by leopards would indicate the inability of this taxon to cope with the predation risks of a medium‐sized carnivore like a leopard.

Regarding the two hypotheses to be tested in the present study, we need to reject the first one (there was no shift in the balance of power) if we assume that the carnivore–hominin interaction documented on OH 7 and OH 65 was extended to other members of the species. From previous taphonomic analysis, it is known that FLK NN3 is a palimpsest where hominins and carnivores made use of the same space. ⁶⁴ In this regard, knowing the potential agent that modified OH 7 and OH 65 bones is of relevance in the context of the shift in the balance of power to address relationships between felids, hyenids, and hominins 2 Ma ago. With respect to the other hypothesis, we confirm the predatory role of leopards on H. habilis, assuming a primary role of these felids in the consumption of this hominin taxon, and the lack of taphonomic evidence of durophagous carnivores implied by the modification of the individuals analyzed. Both hypotheses could potentially indicate that H. habilis was still in a similar position as some australopithecines regarding their relationship with large mammal carnivorans. This would explain their primitive body, which included substantial adaptations to tree climbing. An arboreal component in their adaptive behavior would probably have been needed to buffer predation hazards.

This leaves us with the issue of what role (if any) H. habilis played in the formation of the archeological record. The information obtained after the taphonomic analysis of the Olduvai Gorge Bed I anthropogenic sites suggests that some hominins were already inserted within the carnivoran guild, with a predatory component. ¹⁰² , ¹⁰³ , ¹⁰⁴ , ¹⁰⁵ , ¹⁰⁶ , ¹⁰⁷ , ¹⁰⁸ , ¹⁰⁹ This shows that around 2 Ma ago, some hominins were capable of coping with predation risks. At the time that this is documented taphonomically, there is evidence of more modern hominins with complete terrestrial adaptation (i.e., H. erectus) who are contemporaneous and very likely sympatrically adapted to the same environments as H. habilis, even at Olduvai Bed I. ²³ The most parsimonious interpretation is that this larger and more anatomically modern hominin is responsible for the anthropogenic record. Regarding H. habilis, the hypotheses that future archeological work would need to test are:

• Could two anatomies as different as those of H. habilis and H. erectus (showing two drastically different modes of adaptation) generate the same behaviors leading to the same archeological records?

• If not, how could we differentiate those divergent behaviors, assuming that both left an archeological record?

• Hominins at the few taphonomically anthropogenic sites seem to repeatedly have had access to flesh carcasses and targeted bulk flesh exploitation. Could this be the record of one type of hominin anatomy, whereas the other anatomy and its behavior could be reflected in those palimpsests in which hominins were engaged in other activities (i.e., plant exploitation)?

• Can we identify two different “sympatric” technological spheres that could be attributed to different hominin taxa?

• Was H. habilis a tool user and an archeological record maker after all?

Until archeologists devise a way to test these and other alternative hypotheses, the attribution of authorship in the archeological record will remain an exercise of speculation. Until now, this association has been made exclusively on indirect grounds; the presence of a particular hominin taxon associated with either the sites or the strata in which they were contained was considered enough to claim a direct relationship (e.g., H. habilis—Oldowan). This is no longer a heuristically valid argument. At Olduvai, there are more Oldowan sites containing paranthropine than habiline remains. Recently, a more modern form (H. aff. erectus) has been sympatrically documented in Bed I. ²³ As a matter of fact, the Oldowan has also been associated with Paranthropus boisei. ¹¹⁰ The Oldowan was also argued to have emerged as the adaptive behavior of Australopithecus garhi, because it was the only hominin documented in the strata where the Oldowan was first discovered. ¹¹¹ Using the same logic, paranthropine teeth and skeletal remains documented at FLK West and BK, both containing Acheulian assemblages, could be inarguably used to posit that they were responsible for this technology. ⁴⁰ , ¹¹² , ¹¹³ We do not intend here to disprove that H. habilis was not responsible for the Oldowan record documented at Olduvai Bed I (because that is heuristically as impossible as to prove as the opposite), but to provide an alternative scenario, given its primitive anatomy^c and the ecological premise that any predator positioned high in the trophic scale has ways to deter other predators at the same trophic level from preying on them. The present work documents that this did not seem to be the case for H. habilis, if assuming its authorship of the Oldowan archeological record.

CONCLUSION

The 1.85 Ma holotype of H. habilis, the juvenile specimen of OH 7, and the 1.8 Ma adult specimen OH 65 both preserve evidence of carnivore damage in the form of two tooth pits and crenulated green breakage on its mandibular body in the case of OH 7 and one tooth pit in the case of the OH 65 maxilla. Given the small size of the OH 7 individual, a durophagous carnivore, like a hyena, would most likely have turned the mandible into fragments. Instead, the deletion/breakage of only the mandibular rami and lower part of the body is suggestive of a carnivore with less destructive dentition. There is unanimity in the three models (for one tooth pit) and the two best performing DL models (for the other tooth pit) that the type of triangular marks are most likely those inflicted by a leopard. This is reflected in probabilities >0.90 in the joint consideration of five different carnivore taxa. This lends reliability to the taxonomic identification of the modifying agent. This is also supported by the overall predominant felid taphonomic signature of the archeofaunal assemblages that constitute the three levels of FLK NN. Regarding OH 65, if a durophagous carnivore like a hyena had modified the maxilla, it probably would have been much more fragmented. This is corroborated by two DL models, which classified the documented tooth pit as made by a leopard.

The implications of this are major, since it shows that H. habilis was still more of a prey than a predator. It also shows that the trophic position of some of the earliest representatives of the genus Homo was not different from those of other australopithecines. This suggests that H. habilis was unable to fend off top predators from their kills (if we assume that OH 7 and OH 65′s fate is not an anomaly, but a random representation of the species’ interaction with other predators) and that the access to fleshed carcasses documented at some early sites might have been done by a different hominin species. Finding this other species and documenting the timing of the shifting in the balance of power with other carnivores remains challenging, but it is crucial to understand the evolution of our own genus.

AUTHOR CONTRIBUTIONS

Marina Vegara‐Riquelme, Enrique Baquedano, and Manuel Domínguez‐Rodrigo designed research; Marina Vegara‐Riquelme and Manuel Domínguez‐Rodrigo performed research; Marina Vegara‐Riquelme and Manuel Domínguez‐Rodrigo analyzed data; Marina Vegara‐Riquelme, Enrique Baquedano, and Manuel Domínguez‐Rodrigo acquired funding; Marina Vegara‐Riquelme and Manuel Domínguez‐Rodrigo wrote the paper; and Marina Vegara‐Riquelme, Enrique Baquedano, and Manuel Domínguez‐Rodrigo revised the manuscript.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

Supporting information

Supporting Information

Vegara‐Riquelme, M. , Baquedano, E. , & Domínguez‐Rodrigo, M. (2025). Early humans and the balance of power: Homo habilis as prey. Ann NY Acad Sci., 1553, 140–157. 10.1111/nyas.15321