Constants and Variables: How Does the Visual Representation of the Holocaust by AI Change Over Time

Today, we can access an unprecedented volume of information about the Holocaust, which is freely available online. It includes historical collections digitised by heritage institutions such as the Shoah Foundation or the Arolsen Archives (Shandler 2022; Stone 2017), reference pages regarding Holocaust events or personalities in online encyclopaedias e.g. Wikipedia (Makhortykh 2017; Wolniewicz-Slomka 2016), and reflections of individuals visiting Holocaust memorials and sharing their experiences via social media, for example Instagram (Hinckley and Zühlke 2022; Zalewska 2017). Unfortunately, this “post-scarcity” (Hoskins 2011) ecosystem of Holocaust memory also includes a multitude of online content offering inaccurate information about specific details of the Holocaust, or promoting a rather one-sided view of this complex historical phenomenon (Grabowski and Klein 2023). In some cases, such content not only gets details incorrectly, but actually propagates antisemitism or denies the Holocaust (Allington 2017; Guhl and Davey 2020).

To cope with the abundance of Holocaust-related information and also to be able to filter out content spreading hate and distorting historical facts, we need new technological solutions. Even if we would like to follow the analogue practices of careful selection and moderation of content used to represent the Holocaust (e.g. Holtschneider 2014; Hansen-Glucklich 2014), these practices are not feasible for digital platforms that must process billions of existing web pages and integrate information about new ones. Under these circumstances, we cannot rely on human curators anymore, and instead, have to adopt automated solutions that are capable of processing, filtering, and ranking information about the Holocaust in a few milliseconds. These solutions are increasingly powered by artificial intelligence (AI) systems,^[1] which take into consideration many factors, such as content features or user engagement with specific content sources in order to prioritise information regarding specific subjects, including the Holocaust.

Search engines, for instance Google or Yandex, are one of such AI-driven solutions which have a major impact on how individuals and societies learn about the Holocaust. By crawling the Internet and processing crawled data, search engines decide what information sources go first when individuals enter search queries dealing with different aspects of the Holocaust (Makhortykh, Urman, and Ulloa 2021; Pfanzelter 2015). While doing so, search engines have to make choices: for instance, if a user is searching for general information about the Holocaust (e.g. “What is the Holocaust” or just “Holocaust”), then what should the first source be? Should it be a Wikipedia article or a museum exhibition webpage? A related question is whether news websites offering information about the recent developments regarding Holocaust memory shall be prioritised over websites of historical institutions? And if the user is searching from a specific location (e.g. Germany), should websites of local or foreign institutions appear in the top search results?

Answering these questions is not a trivial task either for the AI behind the system, which is expected to give the answer to its users, or the system designers developing the AI. The situation is further complicated by changes in what information sources and interpretations the search engines prioritise according to the searched topics. These changes reflect the evolving nature of what the AI powering the search engines sees as particularly relevant in relation to a specific subject. Without accounting for the changes in relevance,^[2] the selection of information provided by the search engines can easily become outdated, in particular regarding rapidly evolving topics (Ulloa et al. 2023). However, the changing perceptions of what is or is not relevant make AI systems even less transparent, and also stress the importance of understanding how the relevance of specific pieces of information about the Holocaust for AI changes over time. In other words, what are the constants and variables in the AI vision of the Holocaust?

To answer this question, we conducted a series of AI audits in 2020 and 2021, aiming to investigate how the visual representation of the Holocaust by search engines evolves over time. The recent addition to the field of algorithm auditing (Mökander 2023), AI audits investigate the performance of AI-driven decision-making systems, such as search engines. In the course of the audits, we were particularly interested in how the perception of relevance changes between search queries in Latin script (i.e. “Holocaust”; same spelling in English and German languages) and in Cyrillic script (i.e. “Холокост”; same spelling in Russian and Bulgarian languages). Our interest in comparing the two was due to the profound differences in Holocaust memory practices in Western Europe (including Germany and the UK) and Russia;^[3] hence, we wanted to know whether these differences translate into different perceptions of the Holocaust by AI.

For the practical implementation of the audits, we used a virtual agent-based auditing approach^[4] (for a detailed discussion of the method, see Ulloa, Makhortykh, and Urman 2022). For consistency, we deployed our agents using a set of IP addresses located in Germany. We collected data from three search engines – Bing, Google and Yandex – which are among the most frequently used search engines in Western and Eastern Europe. For each search engine, we programmed our agents to enter in Latin and Cyrillic scripts a selection of search queries, which dealt with the different aspects of the Holocaust. For each of the queries, the agents retrieved the top 50 image search results (for consistency, we used .com versions of each search agent). These images were then attributed with the help of authoritative sources (e.g. the United States Holocaust Memorial Museum collections) to determine whether they are related to the Holocaust, whether these are historical evidence or recent images (e.g. photos of contemporary memorials), what aspect of the Holocaust are shown (e.g. liberation of the camps), and from which Holocaust site these images are from.

To illustrate how the perception of relevance in relation to the Holocaust by AI systems changes over time, we share some preliminary observations regarding the outputs for the “Holocaust” query in the Latin script. We find this case particularly interesting because, on occasion of such a general query, AI systems powering search engines have a particularly broad choice. Specifically, we were interested in what facets of the representation of the Holocaust remain stable constants across time, and which variables are subject to change from the viewpoint of the AI.

In concrete terms, we looked at what aspects of the Holocaust were shown by top image search results, and which Holocaust sites these images came from. In the case of Holocaust aspects, we observed substantial changes in relevance for Google and Bing. In 2020, the two search engines focused on images of liberated camps, thus reproducing the common pattern in the representation of the Holocaust in the post-WWII period. However, in 2021, these images became substantially less common in top search outputs. Instead, for both search engines, we observed more modern photos showing contemporary Holocaust sites; in the case of Bing, there was also an increase in images showing the evidence of mass murder (contrasted to the focus on the images of deportations in 2020). In contrast to the changes in relevance of Holocaust-related content on Google and Bing, in the case of Yandex, we observed a rather stable prioritisation of images of contemporary Holocaust sites between 2020 and 2021, with rather few historical photos.

In contrast to aspects of the Holocaust, which were treated by two out of three search engines as variables which are subject to change in relevance, we found that the selection of Holocaust sites from which the images were coming was rather constant. With the exception of Bing in 2020, where around 45 % of images came from Auschwitz-Birkenau, between 60 % and 80 % of outputs, independently of the engine and the year, were related to Auschwitz. While there was some variation in the visibility of other sites among the individual search engines – for instance, Bing giving more visibility to images from Bergen-Belsen, and Google with Yandex prioritising content from Buchenwald – the prevalence of Auschwitz-Birkenau was a constant.

While we are still finishing the analysis of the outputs to the “Holocaust” query in Cyrillic for 2021, we expect that it will likely follow its own set of constants and variables. Based on 2020 data, we observed substantive variation compared with the outputs for the query in the Latin script, in particular regarding the visibility of specific Holocaust aspects. Notably, we observed a higher number of images in top search results showing contemporary Holocaust memory sites for Google and Yandex (a trend which was less pronounced on Google in response to the Latin query) and substantially fewer images showing deportations of Jews on Bing. Similarly, we observed some variation in the visibility of individual Holocaust sites, with some sites absent in the search results for the Latin query appearing in response to the Cyrillic query; examples included Ravensbrück for Google and Mauthausen for Bing. At the same time, for all three engines, the prevalence of images from Auschwitz-Birkenau also remained constant also for the Cyrillic query.

These empirical insights contribute to our understanding of how the representation of the Holocaust by AI can change over time, but also raise a number of questions regarding our expectations about constants and variables in what the AI sees relevant for the Holocaust. The core question is whether we expect AI to reiterate the constants in how the Holocaust is remembered and represented, including the large gap in awareness about Holocaust sites which feature more or less prominently in the popular culture (e.g. Claims Conference 2020), or can we strive for it to advance the current state of popular knowledge about the Holocaust? The answer to this question will influence how we approach other related questions: for instance, shall the representation of the Holocaust by the AI change over time and if yes, then what should the rate of such change be? Shall the AI systems be able to decide on such changes themselves, or should human stakeholders shape their decisions? Furthermore, what can be the impact of the changes in the AI’s perception of relevance regarding Holocaust-related changes in Holocaust memory and education practices?

Similarly challenging are the questions regarding how diverse or fair the perception of information relevance by AI should be, and to what degree these concepts are applicable to the case of the Holocaust. There is extensive debate about the importance of embedding the values of fairness and diversity into the AI system design (e.g. Helberger, Karppinen, and D’acunto 2018; Robert et al. 2020; Chi, Lurie, and Mulligan 2021; Madaio et al. 2022). Still, so far, it rarely relates to the AI systems used in the context of genocide memory. To enable such a relation, it would first be needed to define what is meant by diverse or non-diverse (or fair and unfair) Holocaust memory: is, for instance, the higher visibility of images coming from Auschwitz-Birkenau in the outputs of search engines an indicator of the lack of diversity in what AI sees as information relevant for the Holocaust in general? Can the unequal perception of relevance regarding specific Holocaust sites or aspects by AI systems be viewed as a form of unfairness? How can we define the meaning of fairness and diversity in the context of Holocaust memory, and what groups of stakeholders shall be responsible for these definitions? Neither these, nor the earlier questions have easy and available answers. However, we argue that trying to answer them is essential for the future of Holocaust remembrance, which is increasingly shaped by AI, already in the present.