A New AI Lexicon: Algolinguicism

Translating Language Justice to Digital Platforms

Illustration by Somnath Bhatt

A guest post by Mashinka Firunts Hakopian. Mashinka is a visiting Mellon Professor of the Practice at Occidental College and an Associate Director of Research for the Future of Democracy program at the Berggruen Institute. Her book, Algorithmic Bias Training, or, Lectures for Intelligent Machines is forthcoming in 2021 from X Artists’ Books.

This essay is part of our ongoing “AI Lexicon” project, a call for contributions to generate alternate narratives, positionalities, and understandings to the better known and widely circulated ways of talking about AI.

Which languages and language-users are prioritized by digital platforms?

Speakers of non-dominant languages are disproportionately subject to algorithmic harms.¹ They confront content moderation algorithms that “only work in certain languages”² on platforms that structurally omit non-Western nations from governance considerations. I call this tendency algolinguicism — a matrix of automated processes that minoritize language-users outside the Global North and obstruct their access to political participation. This essay addresses digital platforms as sites of algolinguicism. Which languages are accorded weight in the development of a platform’s algorithms? Which speakers are afforded the right to participate on a given platform, and how do linguistic hierarchies materially impact their lived experience? Whose languages are digital platforms taught to speak?

Defining Language Justice

In organizing spaces, language justice is understood as the right to engage in political participation in one’s own language(s). As the collaborative Antena Aire describes it, the concept provides a political foundation grounded in “the right everyone has to communicate, to understand, and to be understood in our language(s).”³ Practicing language justice can take the form of simultaneous interpretation services; materials distributed in each language spoken by impacted community members; multilingual media making; and beyond. In my own experience with Armenian interpretation for housing rights coalitions in Los Angeles, language justice initiatives can enable the formation of tenants’ associations by people who don’t speak a common language, or make it possible for community members whose preferred language isn’t English to participate in tenants’ union organizing.

Language justice unsettles the dynamics of linguicism, a term coined by Tove Skutnabb-Kangas referring to ideological structures that advantage dominant language-users by inequitably distributing power and resources between discrete linguistic communities (Skutnabb-Kangas 1988).⁴ The educational system is one arena where the effects of linguicism are keenly felt — specifically in instances where speakers of non-dominant languages are penalized by evaluation metrics calibrated for native speakers of official language(s). Linguicist formations are inevitably also raciolinguistic. They co-construct “language and race in ways that position racialized populations as inferior to the normative white subject” and, by extension, to the Anglo-European language user (Flores 2021).⁵ In one crucial example, continued attempts to impose English-only voting legislation aim to disenfranchise non-English speaking and predominantly non-white voters, denying them access to translated election materials and language services at polling sites.⁶

Linguicism pervades digital platforms whose algorithms are coded for engagement in dominant languages, with little consideration given to circumventing rights abuses that impact speakers of minoritized languages. At the same time, an algorithmically driven “networked public sphere”⁷ now serves as the infrastructure for much political organizing (Tufekci 2017). Against that backdrop, the effects of algolinguicism are keenly felt within authoritarian regimes where digital platforms become “the place to organize” for the opposition, as Azerbaijani journalist Arzu Geybullayeva puts it.⁸ In this arrangement, one of the few possible spaces for oppositional discourse is also a space that poses unique dangers for participants whose primary language is excluded from algorithmic harm mitigation. With that in mind, how do we confront linguicism in these spaces and translate the principles of language justice to digital platforms?

To which algorithm do we appeal for the right to communicate and to be understood in our languages?

“Countries Like Ours”: Algolinguicism and Azerbaijan

To draw out the stakes of these questions, we can look at digital platforms’ failure to intervene in Azerbaijan’s recent state-executed digital offensives. Facebook’s response, in particular, presents an instructive case study in the mechanisms of algolinguicism and the urgency of digital language justice.

In a 2018 report, the Institute for the Future analyzed the rise of state-sponsored campaigns of digital repression and disinformation across the globe, concluding that these practices represent “a new form of human rights abuse.”⁹ The report identifies Azerbaijan, a post-Soviet republic in West Asia, as a key offender. Notably, its autocratic regime ranks 167th out of 180 nations in the World Press Freedom Index. Its president, Ilham Aliyev, was the inaugural “Organized Crime and Corruption Person of the Year,” as determined by the Organized Crime and Corruption Reporting Project.¹⁰

Azerbaijan’s expansive “cyber militias” and coordinated attacks can be traced directly to its state apparatus. In one example, a state-funded initiative sought to train youth to “take an active part in the information war,” deploying tactics that game algorithmic outcomes toward “quantitative success” for the state.¹¹

Attesting to the flourishing of initiatives like these, 2020 disclosures from former Facebook data scientist Sophie Zhang revealed that a network of coordinated inauthentic behavior linked to Azerbaijan’s ruling party on the platform comprised 1,000 accounts and roughly 8,000 pages.¹² They were dedicated to producing pro-governmental content and engagement, orchestrating attacks against opposition, and amplifying the algorithmic reach of disinformation. Facebook waited 14 months to investigate. In the intervening period, Azerbaijan launched an ethnic cleansing campaign targeting the Armenian-populated Republic of Artsakh. I have written elsewhere about the impacts of this campaign — which included thousands of deaths, the seizure and occupation of ancestral lands, and the displacement of 100,000 Indigenous Armenians.¹³

The platform’s failure to intervene in state-affiliated campaigns linked to ethnic cleansing is consistent with the U.N.’s earlier findings that it played a “determining” role in the incitement of genocidal violence against the Rohingya minority in Myanmar.¹⁴ At the time, Facebook had “few Burmese-speaking moderators” and had not trained a hate speech detection algorithm in Burmese.¹⁵

Explaining why the platform dallied before taking action against Azerbaijan’s campaign, Facebook’s VP of Integrity, Guy Rosen, noted, “We prioritize stopping the most urgent and harmful threats globally. Fake likes is not one of them.”¹⁶ Notably, Rosen’s remarks omitted that the inauthentic engagement was issued from a state apparatus pursuing ethnic cleansing in a neighboring West Asian region. It seems reasonable to assume that the misunderstanding may be rooted in the total absence of Azerbaijani speakers from the platform’s global workforce at the time.¹⁷

Later, the platform’s own April 2021 Coordinated Inauthentic Behavior Report belatedly revealed that the inauthentic behavior was in fact linked to the state’s Defense Ministry and also targeted Armenia during the period of Azerbaijan’s military offensive in the Armenian-populated Artsakh.¹⁸ Regional experts characterized the scenario as a chilling militarization of digital platforms, asking, “what comes first — the tweet or the drone strike?”¹⁹ By contrast, Facebook’s representatives initially glossed the activity as so much trivial computational propaganda.

The rationale for inaction? Harm mitigation should focus on “top countries” and “priority regions like the United States and Western Europe.”²⁰ This rationale aims to legitimate sociotechnical systems that exclude protections for users of minoritized languages. They do not, after all, communicate in the lingua franca of “top countries.” Algolinguicism at once enables state violence and repression to flourish, and hinders political participation for oppositional actors. In that respect, it enacts what Mimi Ọnụọha has called algorithmic violence, a structural harm that an “automated decision-making system inflicts by preventing people from meeting their basic needs.”²¹

In one example of state-sponsored repression, the editors of oppositional media outlet Mikroskop (who were forced to flee Azerbaijan) describe having to moderate “700 hostile comments” for a single post.²² Other dissident commentators in Azerbaijan have had to appeal to global NGOs when state-sponsored digital abuses and deplatforming attempts occur. Modes of redress within the platform itself are complicated by the fact that “Facebook has not translated all of its tools and instructions into Azeri.”²³ This amounts to a failure to account for a non-dominant language, or the abuses that might impact its users.

Consider users whose preferred language is Azerbaijani; who lack the resources needed to routinely moderate 700 communiqués from bots and bad actors; who have no connections to non- or intergovernmental agencies; or who have limited access to robust translation services. For those users, a digital platform’s linguicism can be prohibitive to political participation. By most estimates, the preferred language of over 80% of Azerbaijan’s population is Azerbaijani, and over half of Azerbaijan’s population is monolingual.²⁴ Algolinguicist approaches to platform governance hinder political participation for millions of monolingual or non-English-speaking people across Azerbaijan, Armenia, and Artsakh.

In April 2021, still another investigation revealed that state-sponsored Azerbaijani disinformation networks continue to propagate on the platform.²⁵ This comes roughly three years after the Institute for the Future’s report documented digital repression in Azerbaijan, and six months after Sophie Zhang’s public disclosures. Responding to the unrelenting deluge of revelations about the Azerbaijani state’s unchecked platform abuses, Arzu Geybullayeva observed, “Facebook isn’t interested in countries like ours.”

Here we confront a disjuncture in global platform governance: while platform companies are predominantly sited in the US, the reach of what Kate Klonick calls the “new governors” is undeniably planetary.²⁶ At the same time, platform companies have ultimately abdicated responsibility for building capacity to circumvent rights abuses on a global scale. In the current example, the platform’s content moderation algorithms “only work in certain languages.”²⁷ Whether or not digital platforms espouse interest in “countries like ours” (that is, “small and non-western countries”), they operate as de facto political actors there. Nevertheless, as of August 2020, Facebook “did not have any full-time or contract operations employees who were known to speak Azeri, leaving staff to use Google Translate to try to understand the nature of the abuse” [emphasis added].²⁸ Despite its status as a de facto political actor and a venue for state-sponsored communication, the platform was ill equipped to understand the primary language used by the overwhelming majority of the nation’s speakers.

Without belaboring the point, a clear link can be drawn between how digital platforms deprioritize non-Western nations; how they exclude those nations’ languages from platform governance considerations; and the algorithmically enabled harms that follow. Plainly put, algolinguicism creates the conditions of possibility necessary for digital autocracies to take bloom.

Beyond Algolinguicism

How can we imagine models of language justice calibrated to digital platforms? Linguistics scholar Helen Kelly-Holmes describes the internet as “a sociolinguistic machine that learns language through the language work that is carried out by users in online contexts.”²⁹ As platforms ignore rights violations in non-dominant languages, the burden disproportionately falls to minoritized language-users to perform the “language work” necessary to carve out spaces of political participation.

While Azerbaijan’s ethnic cleansing and disinformation campaigns were concurrently underway, an Armenian diaspora collective called She Loves designed a counter-offensive. Building on their performance series entitled The Rifles Our Ancestors Didn’t Have (2020), the collective initiated an ad-hoc digital campaign to label state repression and disinformation. Its premise deliberately invoked a Web 1.0 aesthetic: when a member encountered a post made by a bot or disinformation actor, they responded by commenting with a work of ASCII art depicting a rifle. The collective circulated the ASCII rifle to community members, urging them to “replicate this digital rifle and paste it under every online aggression,” and encouraging a form of collective guerrilla content moderation.

Nelly Achkhen Sarkissian and Adrineh Baghdassarian, the collective’s co-founders, contextualize these tactics through the ethos of “daké-dakin,” roughly translating to “strike the iron while it’s hot (do it with no delay and do it now).”³⁰ Daké-dakin recognizes the compressed timescales associated with action in conditions of crisis, against the timescales of state and industry actors who proceed with interminable delay. Appropriating the tools of “ASCI imperialism,”³¹ the collective’s members refuse the erasure of minoritized language-users, converting platforms into temporary and provisional spaces of transnational feminist struggle. Interventions like these enable us to build toolkits of community-driven tactics for multilingual liberation, to cultivate imaginaries beyond algolinguicism.

References

[1] Most recently, former Facebook employee Frances Haugen emphasized that harms associated with Facebook’s algorithms are “are far worse in regions that don’t speak English because of Facebook’s uneven coverage of different languages.” See Karen Hao, “The Facebook whistleblower says its algorithms are dangerous. Here’s why,” MIT Technology Review, October 5, 2021, https://www.technologyreview.com/2021/10/05/1036519/facebook-whistleblower-frances-haugen-algorithms.

[2] Billy Perrigo, “Facebook Says It’s Removing More Hate Speech Than Ever Before. But There’s a Catch,” Time, November 26, 2019, https://time.com/5739688/facebook-hate-speech-languages/.

[3] Antena Aire, How to Build Language Justice (Los Angeles: Libros Antena Books, 2020).

[4] Tove Skutnabb-Kangas, “Multilingualism and the Education of Minority Children,” in Minority Education: From Shame to Struggle Multilingual Matters, eds. Tove Skutnabb-Kangas and Jim Cummins (Philadelphia: Clevedon, 1988), 13.

[5] Nelson Flores, “Raciolinguistic Genealogy as Method in the Sociology of Language,” International Journal of the Sociology of Language 2021, nos. 267–268 (2021): 114. https://www.degruyter.com/document/doi/10.1515/ijsl-2020-0102/html.

[6] One recent example includes the current contestation around English-only voting materials in Iowa, which disproportionately impact Indigenous and Latinx voters in the state. See Shelby Kluver, “LULAC prepares for lawsuit on Iowa’s ‘English-only’ voting law,” WQAD8 ABC, September 28, 2021, https://www.wqad.com/article/news/local/vote/lulac-lawsuit-english-only-law-iowa/526-c3e9a969-d9b5-433d-aac9-caed0a63c9c6.

[7] See Zeynep Tufekci, Twitter and Tear Gas: The Power and Fragility of Networked Protest (New Haven: Yale University Press, 2017).

[8] Arzu Geybullayeva quoted in Julia Carrie Wong and Luke Harding, “‘Facebook Isn’t Interested in Countries Like Ours’: Azerbaijan Troll Network Returns Months After Ban,” The Guardian, April 13, 2021, https://www.theguardian.com/technology/2021/apr/13/facebook-azerbaijan-ilham-aliyev.

[9] Carly Nyst and Nick Monaco, “State-Sponsored Trolling: How Governments are Deploying Disinformation as Part of Broader Digital Harassment Campaigns,” Institute for the Future, 2018, 1.

[10] See “2021 World Press Freedom Index,” Reporters Without Borders, https://rsf.org/en/ranking; and “Ilham Aliyev: 2021 Person of the Year in Organized Crime and Corruption,” Organized Crime and Corruption Reporting Project, 2021, https://www.occrp.org/en/poy/2012/.

[11] Nyst and Monaco, “State-Sponsored Trolling,” 25.

[12] Craig Silverman, Ryan Mac, and Pranav Dixit, “‘I Have Blood On My Hands’: A Whistleblower Says Facebook Ignored Global Political Manipulation,” BuzzFeed News, September 14, 2020, https://www.buzzfeednews.com/article/craigsilverman/facebook-ignore-political-manipulation-whistleblower-memo.

[13] For further background, see my essay, “On the Struggle for Indigenous Self-Determination in the Republic of Artsakh,” Los Angeles Review of Books, October 25, 2020, https://www.lareviewofbooks.org/short-takes/struggle-indigenous-self-determination-republic-artsakh.

[14] Tom Miles, “U.N. Investigators Cite Facebook Role in Myanmar Crisis,” Reuters, March 12, 2018, https://www.reuters.com/article/us-myanmar-rohingya-facebook/u-n-investigators-cite-facebook-role-in-myanmar-crisis-idUSKCN1GO2PN.

[15] Perrigo, “Facebook Says It’s Removing More Hate Speech Than Ever Before.”

[16] Guy Rosen, September 14, 2020, https://twitter.com/guyro/status/1305613201803943936?s=20.

[17] Julia Carrie Wong, “How Facebook Let Fake Engagement Distort Global Politics: A Whistleblower’s Account,” The Guardian, April 12, 2021, https://www.theguardian.com/technology/2021/apr/12/facebook-fake-engagement-whistleblower-sophie-zhang.

[18] “April 2021 Coordinated Inauthentic Behavior Report,” Facebook, April 2021, https://about.fb.com/wp-content/uploads/2021/05/April-2021-CIB-Report.pdf.

[19] John Beck, “What Comes First — The Tweet Or the Drone Strike?,” Rest of World, October 11, 2020, https://restofworld.org/2020/tweet-and-a-drone-strike.

[20] Silverman, Mac, and Dixit, “‘I Have Blood On My Hands.’”

[21] Mimi Ọnụọha, “Notes on Algorithmic Violence,” GitHub, February 7, 2018, https://github.com/MimiOnuoha/On-Algorithmic-Violence.

[22] Fatima Karimova quoted in Wong and Harding, “‘Facebook Isn’t Interested in Countries Like Ours.’”

[23] Wong and Harding, “‘Facebook Isn’t Interested in Countries Like Ours.’”

[24] See Concise Encyclopedia of Languages of the World, eds. Keith Brown and Sarah Ogilvie (Oxford: Elsevier, 2009); and Oishimaya Sen Nag, “What Languages Are Spoken In Azerbaijan?,” WorldAtlas, August 1, 2017, https://www.worldatlas.com/articles/what-languages-are-spoken-in-azerbaijan.html.

[25] Wong and Harding, “‘Facebook Isn’t Interested in Countries Like Ours.’”

[26] Kate Klonick, “The New Governors: The People, Rules, and Processes Governing Online Speech,” Harvard Law Review, April 10, 2018, https://harvardlawreview.org/2018/04/the-new-governors-the-people-rules-and-processes-governing-online-speech.

[27] Perrigo, “Facebook Says It’s Removing More Hate Speech Than Ever Before.”

[28] Julia Carrie Wong, “How Facebook Let Fake Engagement Distort Global Politics: A Whistleblower’s Account,” The Guardian, April 12, 2021, https://www.theguardian.com/technology/2021/apr/12/facebook-fake-engagement-whistleblower-sophie-zhang.

[29] Helen Kelly-Holmes, “Multilingualism and Technology: A Review of Developments in Digital Communication from Monolingualism to Idiolingualism,” Annual Review of Applied Linguistics 39 (2019): 34.

[30] Email to the author, August 14, 2021.

[31] Daniel Pargman and Jacob Palme describe ASCI imperialism in their eponymous essay, “ASCI Imperialism,” in Standards and Their Stories: How Quantifying, Classifying, and Formalizing Practices Shape Everyday Life, eds. Martha Lampland and Susan Leigh Star (Ithaca: Cornell University Press, 2009), 177–199.