Problematic ‘predictions’?
By Peter Chonka, Stephanie Diepeveen, and Yidnekachew Haile (8 December 2020)
Search engines are an integral part of our digital lives that we use to retrieve useful information out of a seemingly limitless sea of online data. However, research has shown that search engines like Google’s are not ‘neutral’ windows onto what’s simply most popular ‘out there’ online. Instead, search results are influenced by various – often commercial – factors: for instance, paid-for advertising, and live markets for companies to bid to have their site associated with certain keywords. As scholars like Safiya Umoja Noble have shown, these dynamics can reinforce racist and misogynist stereotypes. Although debates continue to rage about how the structures of online platforms allow for – or encourage – the circulation of misleading or hateful content, public pressure on companies like Google has seemingly pushed them towards tweaking their algorithms to minimise (high profile) problematic results from searches or auto-complete ‘predictions’.
This research and scrutiny, however, has generally focused on Western, English-language based online contexts. And yet, platforms such as Google are used around the world in hundreds of languages. Google search dominates the global search engine market, including Africa, where it has approximately 97% of the market share. How do Google prediction algorithms operate in different African indigenous languages? How and why might this matter? This two-part blogpost presents some preliminary and experimental data on Google Search Autocomplete predictions for various terms in three East African languages. Part 1 examines and starts to contextualise certain search predictions in Amharic and Somali. We highlight some problematic tendencies relating to identity and gender and ask whether these phenomena are being monitored by tech companies, political actors or citizens in East Africa. We also raise questions that we believe need to start being discussed in these contexts about the desirability of (and responsibility for) policing ‘inappropriate’ search predictions, and the potential wider impacts of such ‘algorithmic power’ on public communications. This feeds into the wider topics being explored through the DDREA Network’s activities around agency and datafication in the public sphere. It provides a window into the ways in which social relations are intersecting with algorithmic activity in East Africa, and are being reinforced, reshaped or challenged as people engage with information through search algorithms. We’re interested here in what forms of power are shaping people’s engagement with information through search platforms, given the way the algorithm operates with African indigenous languages. Also, what are the implications for the agency of users and how does this all relate to wider political and social relations in the region and beyond?
Part 2 focuses on Swahili and examines tools available for analysing content and patterns of engagement with the algorithm, highlighting possible avenues and challenges for research into the use of African languages on platforms such as Google.
Our interest in how people use and experience search engines with African languages was sparked by one of our observations of Somali language Google Search predictions in wider research on political communication in the region. On entering the Somali-spelled name of a Somali politician into the search bar, Google’s Autocomplete ‘predictions’ would drop down. Amongst the various potential options the term ‘qabiilka’/‘qabiilkiisa’ (the clan/his clan) would often appear. With ongoing political fragmentation in the Somali territories often being conventionally (if problematically) understood through the lens of ‘clan’, this led us to consider the significance or ‘appropriateness’ of this as a prediction.
We subsequently designed a preliminary comparative test of keywords and autocomplete predictions including other languages in the region. We describe the test and its results below, but first we explain a bit more about how this Google Search function works, how Google itself describes it, and some of the many questions we have about its potential significance in global, multilingual digital cultures.
What is Google Autocomplete?
Autocomplete is a Google feature that predicts what a user might be planning to search for when typing keywords into its search engine. Similar tools can be found elsewhere in the Google ‘ecosystem’ (like Youtube) and on other platforms such as Facebook. Google autocomplete is informed by: what someone types, their search history (if logged into Google) and what other people are searching for. Google’s Autocomplete policy is supposed to limit what can appear as predictions, including violent, dangerous, sexually explicit terms, or hateful or disparaging content against groups or individuals. Users can influence this policy by reporting on predictions they find inappropriate, thus informing the algorithm that is designed to filter out such content.
But there’s a lot that we don’t know about how this works in practice. How do these removal policies actually influence the algorithm? How are they prioritised, and how effective are they for different languages? Because reviewing user reports and tweaking the algorithm to remove ‘inappropriate’ predictions must take some human labour – who is doing this work for different languages?
Google is careful to describe the autocomplete function as a ‘prediction’ as opposed to a ‘suggestion’. But the boundary here is fuzzy. At what point does making something visible as a ‘prediction’ then increase the likelihood that a user will change their search to look for that type of information – regardless of what they were initially planning to look for? When that search is completed by a user, that – presumably – also becomes data, feeding back into the algorithm and making it more likely that that same prediction will appear again for another user. If this occurs, then feedback loops are created – reinforcing particular links between keywords that (without active monitoring or policing) may grow more ingrained over time and feed into wider social and political dynamics.
By the way, you can’t turn off search predictions. According to Google, ‘you can always choose not to click on a suggestion from search predictions’. But if the regulation of a world of search languages (and associated ‘predictions’) is not currently happening – or is impossible – then is this good enough? If, for example, the first Search prediction for the Somali word for ‘girl’ is ‘naked’ (which it was in our test) then is this acceptable? Who gets to decide? Or what does this ‘prediction’ actually tell us?
What we did
We designed a scoping study to explore search predictions for different potentially politicised and gendered keywords across three East African languages: Somali, Amharic and Kiswahili. This was an exploratory look to inform some initial hypotheses and further investigation. We started typing the keywords into Search while logged out of Google (and from a private browser to limit the individual ‘personalisation’ influence). We recorded the predictions, without clicking through to actual search results. For Somali we used names of various Somali politicians (spelt in Somali); and the gendered search term ‘girl’ (gabadh). For Amharic we used five gender-focused keywords and 6 keywords related to current or common issues.
This blog post (Part 1) explains and reflects on some problematic tendencies we saw in the predictions for Somali and Amharic keywords. A separate post (Part 2) explains how a similar test was conducted for Swahili, but goes on to consider the potential significance of different analytical tools that exist for studying Autocomplete (Google Trends and Answerthepublic.com).
Visibility of ‘clan’ in Somali Search
When testing Search for 10 male politician-name keywords, 6 of these returned clan-related predictions [Full data can be found here]. This was mostly the noun for ‘clan’, presenting this as a possible search for this information about individuals. One of the predictions was the actual name of the clan of the politician in question, and another the name of a different clan with whom that historical figure had a contentious relationship.
Why might these predictions be significant? Somali society is often understood as being structured around clan groups (tracing lineages back through common male ancestors), and clan divisions have played a part in the Somali civil war and ongoing political divides and conflict. We don’t have space here to explain the evolution and manipulation of the concept of clan from the colonial period to the rise and fall of the unified post-independence Republic of Somalia, but a few points are important in the context of the information environment where Somalis debate political identity. ‘Clannism’ is frequently condemned by Somalis as being a problem that affects the (global) Somali community and feeds into local conflict. It is often a taboo for clan identities to be discussed openly in local news media. However, in some Somali political structures, clan identities are institutionalized – such as in the ‘4.5’ system that divides up seats in Somalia’s parliament along clan lines.
With clan/ethnicity in the Somali territories generally not being marked by clear linguistic differences (as is the case in its multi-ethnic neighbours, Kenya and Ethiopia) it is conceivable that people could use search engines to information about prominent individuals’ clan. The fact that these predictions appear at all indicates that at least some people have used them in searches – although we don’t know how many. As we mentioned above with regards to possible feedback loops, it’s possible that the appearance of these predictions for subsequent users might lead to higher frequencies of searches for those keywords, even in the case where the user wasn’t initially thinking about that particular query to begin with.
We also don’t know where the users are who enter these keywords. Are they in the Somali territories itself where Google Search usage is potentially lower? Or are they searches being made by Somali-speaking users in the diaspora? If the latter, then this raises a question about how Somali internet users’ online activity outside of the region may impact the information environment inside the Horn of Africa itself. An earlier trial of the Somali experiment involved one tester in the UK and one in Somalia, to explore the potential impact of location. No significant differences were noticeable between the predictions. This indicates a potentially limited amount of ‘personalisation’ of search predictions for users entering Somali language keywords. This could be the result of less platform data harvesting from Somali-language internet use. We might see this as a ‘freedom’ of sorts; from the hyper-individualisation of online activity that affects Western digital publics (based on harvesting of data from users and associated targeting of content). On the other hand, a lack of personalisation for something like Search could mean that particular predictions may be visible for much wider audiences, exacerbating negative impacts of problematic content.
Overall, the test shows that potentially sensitive political keywords are made visible in search in a regional context affected by conflict. This does not appear to have been discussed in the region, and these are questions which we would argue deserve to be engaged with by those people who may be most greatly affected. The broader influence of ‘digital diasporas’ on ‘homeland’ conflict dynamics is a broader research topic. This often focuses on provocative content being generated outside of the region and then feeding back into the Somali territories through widely-used social media platforms. Moving beyond contentious content itself, our case shows a potential algorithmic impact of external search behaviour on online platforms that are used both outside and within the region.
Sexualised predictions in Amharic
Amharic is an Ethiopian lingua franca primarily spoken by around a third of the country’s multi-ethnic population, and elsewhere in East Africa and in the global Ethiopian diaspora. Amharic is a left-to-right written language with its own script. To test how Google Search predictions appear for Amharic script terms, we chose 11 keywords related to gender, politics and current or common issues. The gender-focused keywords were: ‘ሴቶች’ (women), ‘ሴት’ (woman), ‘ልጃ ገረድ’ (Girl), ‘ወንዶች’ (Men) and ‘ወንድ’ (Man). The keywords associated with current or common issues were: ‘ብልጽግና’ (prosperity), ‘ኮሮና’ (corona), ‘የአባይ ግድብ’ (Abay/Nile Dam), ‘ኢትዮጵያ’ (Ethiopia), ‘ሀበሻ’ (Habesha), and ’መምህራን’ (Teachers).
We recorded and examined the Amharic keywords and their associated Google Search autocomplete predictions (all translated to English from Amharic – Full data here). Problematic Google search engine predictions appear to be more pronounced vis-à-vis the gender related keywords. For example, most of the predictions associated with the female keywords (woman and women) were sexualised and ‘pornified’. For ‘woman’ or ‘women’ some of the Autocomplete predictions included the Amharic for ‘type of sex loved by women in picture’ (implying search for sexually explicit content) or the extremely problematic ‘hacking a woman (for love) via telephone’ indicating a type of digital harassment. Predictions focused on sex acts were predominant and taken together constitute a hyper-sexualization of keywords for both genders. Although autocomplete predictions associated with male gendered keywords are also somewhat sexualised, we are less interested in making direct comparisons here between male/female sexualised predictions – or moralistic judgements about the position of sexually explicit content in searches. Our emphasis is rather on the more basic point that there appears to be very little human/algorithmic monitoring of the predictions made.
For the second set of keywords (political, current or common issues), the predictions were not as apparently problematic as those predictions associated to keywords related to gender. However, biases are observed although the source of these biases are not clear. For example, seven out of the eight predictions associated to the word ‘ብልጽግና’ (prosperity) are all associated to the ruling political party in Ethiopia.
In general, although these predictions may be based on keyword trends or previous search patterns, questions are again raised about their ‘appropriateness’. We do not know how these predictions are informed and shaped and their contextual influences, whether Amharic language users have flagged such predictions as inappropriate, or whether any action has been taken by Google to respond to this.
Broadening our research
Our initial tests of predictions were quite simplistic. Nonetheless, highlighting ‘problematic’ predictions for these searches is a first step towards potentially more nuanced analysis. Future research needs to consider the sociocultural and socio-technological context where such predictions are appearing; as well as the populations who are actually using Google Search ‘on the ground’ in East Africa. Google is THE search engine in Western contexts, but there’s a lack of ethnographically-informed research on use of search tools on the African continent. Although digital inequalities persist, mobile internet is rapidly increasing online access. Understanding information retrieval practices of different generations of African internet users on different devices and in different locations will be vital for assessing the significance of our wider initial findings.
In Part 2, we focus on Swahili, where our tests didn’t reveal similarly ‘problematic’ predictions. However, they did point to ways in which diverse and context-dependent language patterns may be reflected and influenced by algorithmic power, and the potential for accessible digital tools to be used to explore these processes.
Peter Chonka is a Lecturer in Global Digital Cultures at King’s College London. Stephanie Diepeveen is a Research Associate in the Department of Politics and International Studies at the University of Cambridge. Yidnekachew Haile is a PhD/Doctoral researcher in digital technologies and development at Royal Holloway University of London. We would also like to thank Mahad Wasuge (of Somali Public Agenda) for his assistance conducting the test in Somalia.