5. Discussion

Demonstrating just what it means when an image classifier 'breaks' and the implications of the ability to break these apps very easily highlights the need for improvement, but improvement in a way that considers the ethical ramifications at stake. One possibility worth pursuing is a subject of ongoing research by Huffer especially, the application of qualitative and quantitative methods from forensic taphonomy to assess, often from photographs alone, whether human remains offered for sale are in fact 'old', or instead artificially aged? For law enforcement, an application determining if an item offered for sale is or is not likely to have been at one time buried, exposed to water, sun, or plant/animal activity could be very useful. These are all aspects of taphonomy and are used by archaeologists and forensic anthropologists alike to assess how human or faunal remains have been modified between death and recovery (e.g. Schotsmans et al. 2017). In a human remains trafficking context, such evidence is crucial to assess whether the dealer in question is selling authentic specimens or misrepresenting an older item from archaeological or ethnographic contexts as something 'harmless' and generally legal (e.g. a prepared anatomical specimen 'for study purposes only'). Overall, work of this nature will further the application of one aspect of the 'forensic archaeometry' approach to investigating antiquities looting, as described in Temiño et al. (2018).

To date, it has only been possible to record taphonomy of trafficked material in the course of an investigation itself, when law enforcement turns to university or museum-based experts for help with in-person evaluation of seized material (e.g. Pokines et al. 2017; Pokines 2015; Gill et al. 2009; Halling and Seidemann 2016; Seidemann et al. 2009). However, we raise the question of whether taphonomic assessment through machine learning on a corpus of human remains for sale or display on social media, when compared to photographs of specific, verifiable modifications seen on museum specimens, is possible? Preliminary results of such research will be presented in a future article, but we argue that using machine learning for these purposes is not only ethical but less prone to misclassification, while still useful to anti-trafficking efforts.

One aspect of the osteoarchaeological research of one of us (DH) targets how to better separate authentic from inauthentic ethnographic modified human remains. This includes determining if/when a resin or plastic replica has been modified to look enough like real bone so as to fool some collectors, although anecdotal evidence suggests that the human remains collecting community is generally not tolerant of individuals selling replicas as real bone. Huffer's osteoarchaeological research is also investigating means to distinguish how and when real skulls have been modified through such means as artificial patinas, machine carving, or the application of decorative accoutrements thought to be culturally appropriate or to enhance the 'exoticism' of the item being sold (currently focused on Dayak and Asmat examples, see Huffer and Charlton forthcoming). In other words, this is the modification of 'real' human remains to serve modern market demands. If a dealer or collector had the means or wherewithal to stumble upon our classifier and use it, might our work allow middlemen and dealers to produce better fakes? Our current understanding of the dynamics of the online market suggests that many collectors are not well versed enough in human osteology to ascertain even age and sex estimates reliably for the skulls they collect, let alone how to find and use machine-learning tools. Evidence exists that most collectors and dealers determine the authenticity of what they buy from others on social media via word of mouth (see Figure 6).

Figure 6
Figure 6: Communication between collectors regarding how authenticity is determined

However, if one were to condense the ability to accurately classify a modern, possibly grave-robbed, skull actively being modified to resemble a c.1800s ethnographic specimen (through the application of carving, staining, attaching decorations, etc.) in an easy-to-use app (and, e.g., TensorFlow For Poets is arguably straightforward), then what would stop a tech-savvy dealer from running the classifier until the desired result was obtained and thereby making better fakes, more able to fool customs officers and experts?


Internet Archaeology is an open access journal based in the Department of Archaeology, University of York. Except where otherwise noted, content from this work may be used under the terms of the Creative Commons Attribution 3.0 (CC BY) Unported licence, which permits unrestricted use, distribution, and reproduction in any medium, provided that attribution to the author(s), the title of the work, the Internet Archaeology journal and the relevant URL/DOI are given.

Terms and Conditions | Legal Statements | Privacy Policy | Cookies Policy | Citing Internet Archaeology

Internet Archaeology content is preserved for the long term with the Archaeology Data Service. Help sustain and support open access publication by donating to our Open Access Archaeology Fund.