Cultural Heritage Datasets, Artificial Intelligence and the Ethics of Non-Intervention

It is a well-known fact that machine-learning algorithms exacerbate biases inherent in the datasets on which they were trained. In recent times, this fact has found ample evidence, e.g. in Cathy O’Neill’s book “Weapons of Math Destruction” (2016), in Kate Crawford’s and Trevor Paglen’s “Excavating AI” (2019), or in articles such as “Data and its (dis)contents” (2020) and “Large image datasets” (2020). Hölderlin already knew: “Wo aber Gefahr ist, wächst das Rettende auch” (“But where there is Danger, Salvation also grows”), and the debate around biases in datasets and the (social) devastations resulting out of the application of machine-learning models trained on such datasets has resulted in such remarkable endeavours as Gebru et al’s “Datasheets for Datasets” (2021), Pushkarna et al’s “Data Cards: Purposeful and Transparent Dataset Documentation for Responsible AI” (2022) as well as Mitchell et al’s “Model Cards for Model Reporting” (2019). All contributions aim at providing context for datasets as well as machine-learning models, and at enabling fairness, accountability and transparency.

Cultural Heritage Datasets differ from contemporary, industrial datasets in many ways. They are heterogeneous with respect to the time period covered, the place or regions incorporated in them, or the cultural contexts in which they have to be located. Most cultural heritage datasets grow over time, because lately digitised items are added incrementally, whereas industrial datasets are often single-shot provisions. The importance of rights and provenance, and the use of linked open data are distinctive features of cultural heritage datasets. Their intended use is often research, in opposition to commercial deployment of contemporary datasets. Specifically trained cultural heritage practitioners have produced metadata, whereas underpaid workers, masked behind the euphemistic term “Mechanical Turk”, often do the labelling of present-day data material.

The intention to “avoid [dataset] uses that could result in unfair treatment of individuals or groups” and to “mitigate unwanted societal biases in machine learning models” (Gebru et al. 2019) can also be understood as a call to provide more balanced datasets, which would statistically aim at a representative composition of the population covered by such a dataset, and ethically at the avoidance of an unfair social impact. This call has also reached the cultural heritage community. An exemplary chapter published in 2022 discusses “Ethical Implications of Implicit Bias in AI“, targets “guidance for the […] ethical, bias-free application” of AI tools, and advises “repeated interventions to give statistically balanced results”.

The problem is obvious: In historical data, there are biases that no longer reflect contemporary ethical values. Most of history (as well as its documentation in cultural heritage institutions) does not match contemporary sensitivity to racism and discrimination; by the way, this applies to most current historical events as well. Whereas it is possible – though cumbersome and therefore expensive – to provide context or even to intervene into metadata produced in libraries, archives and museums, it is questionable whether the intervention into a historical dataset is feasible or actually sensible.

If we take a step back, we can note that the humanities have always put a strong emphasis on faithfulness to the original or to the sources. Historico-critical editions as produced by the humanities are faithful to the text and embed it into a critical apparatus; historians are trained to transcribe sources absolutely faithful to the original and to apply source critique as a methodological tool. Such an approach points back to the establishment of the humanities (including law) as scientific disciplines during the 19th century, and to their descendance from scholasticism with its reverence for holy texts.

Datasets produced in cultural heritage institutions inevitably reproduce the social and ethical biases characteristic of our shared history: The vast majority of the creators of books digitised by libraries or historical records digitised by archives have been white, socially privileged, literate men. History has been written – more often than not – by the winners of wars and not by the defeated people, or by social or ethnic minorities present in former societies. Intervention into such datasets is nearly not feasible: The production of a balanced dataset where female and male creators are incorporated in equal parts would by definition not be representative. The same applies to the use of terms which are now considered as “toxic” by contemporary ethical values: The fact of the matter is that we no longer use terms like “Neger”, “Hottentotten”, or “Kaffern”, but that people have done so historically. It is true that such terms can still hurt people today and repeat the violence done to people of colour throughout history. But it is equally true that we cannot erase or change history. Rather, we have to learn to deal with it. It is therefore not sensible to erase such terms from historical documents or to avoid incorporating such documents into datasets.

Rather, we are well advised to follow the example of historians and philologists, to leave the data as is, and to provide context via data sheets and model cards – and this at the risk that the use of such a dataset for the development of AI applications is not harmless in either case. We are well advised to use data sheets to create the necessary context and to inscribe our own positionality into them. The social and ethical biases in such datasets are issues that might not be as pressing as in contemporary datasets, and their real-world impact is often low. The integration of the view of collection “subjects” (as demanded by Kirk et al. 2022) is often not feasible, except for datasets coming from colonial contexts, where descendants of the people represented therein might have a strong interest in an appropriate construction of the dataset. Such an ethics of non-intervention conforms with the aspiration to learn to deal with history in an appropriate way, and to fight against the omnipresent forgetfulness of history.