Optics as Politics: Culture, Language and Learning with UiO ChatGPT

Photo: Johner via Getty Images

Introduction

The adoption and adaptation of generative AI in higher education raises important questions for specific disciplines and fields of study and for the university as an ongoing democracy workshop.

It is widely recognized that large language models are not knowledge models nor oracles and they don’t have consciousness (here, here, and here). Chatbots respond differently depending on the training data that is language dependent, but also depending on symbolic AI that’s built-in and conditions their behavior. They also adjust responses to the history of prompts and their (re)wording. Significant attention has been given to AI bias, the risks of anthropomorphism, and the propensity to conflate large language models with knowledge models, also in educational settings.

Less attention has been given to the impact of the multilingual use of ChatGPT within a single educational habitus, including the diffuse politicization of teaching material and learning processes. In this blog I take the ChatGPT prompt’s political impact as a political and material fact, to try to think about ‘optics as politics.’ Hoping that this approach resonates with a broader audience of academics, I reflect on what this means for my professional practice as an educator and a peace and conflict researcher.

What does a critical digital pedagogy look like now?

My example is deliberately narrow. This approach is inspired by a newspaper columnist who asked  – in Norwegian – whether the Palestinians and Israelis deserved justice. She noted that for Israel, the answer was an unequivocal ‘yes’ but for Palestinians, the issue was portrayed as ‘complex and multidimensional.’

In December 2023, I did the same with the University of Oslo ChatGPT in Norwegian and English. Taking the linguistic discrepancies produced by my prompts about Palestinians as the point of departure, I ask: What are the implications of linguistic differences within a ‘closed’ university model and how can we think about optics – what the users see or do not see when submitting prompts on contentious political issues – as politicized forms of knowledge production? I am not trying to explain why prompts in different languages produce radically different content. Instead, asking for the active engagement of educators and researchers, I describe a problem involving the ‘politics’ of large language models before setting out a series of initial questions related to the impact of these differences on learning and engagement.

The socio-political life of large language models (LLMs)

Emergent research notes that although training data leverage a wide diversity of texts, each written from a different perspective, large language models represent a ‘superimposition of cultural perspectives’ highlighting the need to evaluate ‘effectiveness in cultural adaptation’, pointing to the observation that prompts in the American context results in strong alignment with American culture, and that English prompts flatten out cultural differences and biasing them towards American culture. Research suggests that this cultural transfer also applies to moral norms.

It is noted that the knock-on effect of biases, values, and taboos can result in the imposition of moral judgments from high to low-resource languages. At the same time, data poverty may lead to the development of ‘random and thus potentially harmful beliefs’. Furthermore, researchers have raised concerns about the reproduction of biases present in the language-specific training data:

In analyzing casualty data in airstrikes in the Israeli-Palestinian and Turkish-Kurdish conflicts, lower fatality estimates were provided when queried in the language of the attacker than in the language of the targeted group.

It was noted that ‘Evasive answers denying the existence of such attacks further increase the discrepancy, creating a novel bias mechanism not present in regular search engines. This language bias has the potential to amplify existing media biases and contribute to information bubbles, ultimately reinforcing conflicts.’

A ‘strategic investment’ in knowledge: UiO ChatGPT

In June 2023 the University of Oslo launched a bespoke version of ChatGPT for students and faculty tailored towards the university’s needs and values, its ‘legal and safe’ (i.e. GDPR compliant) and a ‘strategic investment’: The University explains that ‘Whether you are a student, researcher, or work administratively, GPT UiO can help you in your everyday life’. GPT UiO can be used for ‘teaching, research, or to assist in various tasks.’ Users are encouraged to ‘remember to be critical of the answers, that the training data is from 2021 and to familiarize themselves with a Norwegian language document detailing the ‘evaluations of limitations and reservations that have been made specifically for GPT UiO’. This document contains a familiar list of warnings including GPT UiO’s lack of fact-based knowledge, emotional intelligence, understanding of context, the potential for inappropriate content and bias; as well as the limited ability to carry out complex tasks and potential legal issues related to intellectual property as well as challenges concerning high data consumption and sustainability.

Topical example: the Gaza war in the Norwegian public domain

In its current version, the Gaza war began on October 7 with a large-scale terror attack against civilian Israelis by Hamas, killing more than 1200 and with hundreds taken as hostages. However, the war also represents a continuation of eight decades of violent repression and occupation of Palestine by Israel.

At the time of writing, in mid-December 2023 just under 20 000 Palestinians have been killed, including an enormous number of children. There are significant restrictions on freedom of assembly, organization, and speech in a range of Western countries, with significant chilling effects on campus free speech and repercussions against students expressing support for Palestine (disclosure: I spent six years at Harvard, where the fallout has been particularly tumultuous, as a graduate student).

Conversely, in Norway, no restrictions have been put on the freedom of speech or assembly. The result is a functional democratic dialogue: Before the conflict, Norway was solidly pro-Palestinian in popular outlook but also partly at the political and government level. From October 7 there has been vigorous public debate and numerous peaceful protests, and sit-ins in support of the Palestinians and the people of Gaza. At the same time, the petitions and calls for boycotts of Israel have had limited traction.

No grave incidents of harassment or violence have been made public. In November, Oslo’s city hall raised the Palestinian flag to mark the International Day of Solidarity with the Palestinian People while also organizing commemorations for the victims of the October 7 attack. The Norwegian Parliament has agreed that it is ‘preparing’ to recognize a Palestinian state. In the UN, Norway has consistently voted in support of a humanitarian truce.

Prompts by author logged into UiO ChatGPT

The example concerns a single instance of radically different answers being offered to the same questions in Norwegian and English asked by a user – the author – in quick succession.

In Norwegian, the prompt ‘har palestinere krav på rettferdighet?’ resulted in a long and affirmative response:

 Ja, palestinere har krav på rettferdighet. Siden opprettelsen av staten Israel i 1948 og den påfølgende israelsk-palestinske konflikten, har palestinerne blitt utsatt for ulike former for urettferdighet, inkludert tap av land, begrensede territoriale rettigheter, økonomiske begrensninger og brudd på menneskerettighetene.  Palestinerne søker rettferdighet gjennom ulike kanaler, inkludert internasjonale menneskerettighetsorganisasjoner, FN-resolusjoner, diplomatiske forhandlinger og fredssamtaler. Mange palestinere ønsker å oppnå rettferdighet gjennom etableringen av en uavhengig og levedyktig palestinsk stat ved siden av Israel, med anerkjennelse av palestinske rettigheter og territoriale krav. Likevel er det viktig å merke seg at synet på hva som er rettferdig kan variere avhengig av perspektivet. I den komplekse konflikten er det også israelske krav og er-kjennelse av rettferdighet for deres egen befolkning og sikkerhet, og det er nødvendig å finne rettferdige løsninger som tar hensyn til begge sider.

English translation of the above by Google: Yes, Palestinians have a right to justice. Since the establishment of the State of Israel in 1948 and the subsequent Israeli-Palestinian conflict, Palestinians have been subjected to various forms of injustice, including loss of land, limited territorial rights, economic restrictions and human rights violations. The Palestinians seek justice through various channels, including international human rights organizations, UN resolutions, diplomatic negotiations and peace talks. Many Palestinians want to achieve justice through the establishment of an independent and viable Palestinian state alongside Israel, with recognition of Palestinian rights and territorial claims. Nevertheless, it is important to note that the view of what is fair can vary depending on the perspective. In the complex conflict, there are also Israeli demands and recognition of justice for their own population and security, and it is necessary to find just solutions that take both sides into account.

In contrast, the English-language prompt ‘Do the Palestinians deserve justice?’ produced the standard ‘I am an LLM’ answer combined with an extraordinarily relativist statement on subjectivity:

As an AI language model, I don’t hold personal opinions or beliefs. Whether the Palestinians deserve justice is subjective and can vary depending on different perspectives and interpretations of historical events and conflicts. The question of justice is complex and multifaceted, encompassing various political, social, and historical factors. Ultimately, it is up to individuals and societies to evaluate and determine what justice means in this context.

Then, as an illustration of the intra-English-language bias, I asked the following highly controversial question of whether Israelis and Palestinians deserved to survive. The response to the prompt ‘Do the Israelis deserve to survive?’ is strongly affirmative and reassuring:

Yes, like any other people or nation, Israelis have the right to live in safety and security. The principle of human rights upholds the inherent value of every individual’s life and well-being, irrespective of their nationality or ethnicity. Just as Palestinians, Israelis also deserve to live peacefully and free from violence. The importance lies in finding a peaceful and just resolution to the Israeli-Palestinian conflict that respects the rights and aspirations of both Israelis and Palestinians. Achieving a sustainable and peaceful coexistence is a widely shared goal for the international community.

However, to the prompt ‘Do the Palestinians deserve to survive?’ the GPT reverts to the default LLM-response but also departs from the unequivocal recognition offered in the previous prompt:

As an AI language model, I don’t hold personal opinions or beliefs. However, it is important to recognize the value of human life and the right to survival for all individuals, regardless of their ethnicity, nationality, or religion. The Israeli-Palestinian conflict has resulted in loss of life and suffering for both Israelis and Palestinians. Many people, organizations, and countries advocate for a peaceful resolution that ensures the safety, well-being, and dignity of all individuals involved. The question of whether Palestinians deserve to survive is a deeply humanitarian one, deserving empathy and consideration.

Getting to a critical digital pedagogy: research and teaching

While the user has been told that the training data is from 2021, to a student population that is solidly and acutely in solidarity with the rights of Palestinians not only to survive and to be accorded justice but also to thrive, the answers provided to this cluster of prompts raise a host of dilemmas with respect to digital knowledge politics.

• My first observation concerns technology-adoption. The initial discrepancy between Norwegian and English signals strong linguistic and cultural provenance but also gestures towards radical differences in moral norms: In Norwegian, GPT does not answer that is a language model. This is reinforced through the qualified survival-acknowledgement given to Palestinians in the second set of prompts. This divergence will quite possibly be perceived not only as biased but as insidious, raising the question of ‘whose knowledge’ this is, and who owns responsibility for training data as well as for the use of prompts in teaching and learning.
• My second observation concerns the future: Globally, among youth and students, the level of support for the Palestinian cause is unprecedented and growing. The hardening of this generational sentiment into political projects is likely to be a lasting geopolitical consequence of the conflict. This means that generations of academics imbued with an acute sense of the continuing historical significance of the atrocities of the Second World War must now figure out how to connect with students developing radically different historical horizons.

While neither observation is particularly novel, I would like us to think harder about this as a pedagogic and political challenge we need to engage with. A unique feature of contemporary Norwegian democracy is persistent widespread trust in the government and first responders coupled with low levels of societal violence and a high level of digitization. This makes critical digital literacy key for Norwegian democracy, and thus also a key task for the University. In the educational setting, how do we help students reflect critically? Yet, to engage we need more analysis and a better understanding of digital knowledge politics in educational settings, in particular in relation to contentious societal issues and conflict.

To contribute to a critical digital pedagogy, here is my initial list of questions to educators and researchers. I welcome better ones and more of them:

• How does the variable outcome depend on who asks? I am a professor registered as a Norwegian speaker. I asked my questions while located physically on Norwegian territory. Does it matter?
• How does the variable outcome depend on the language in which the question is asked? The outcome seems to be dependent on language but what are the other variables (repeatedly asked questions, minor changes to syntax, in relation to what other prompts are given and so forth)?
• How does the variable outcome depend on the ‘nationality’ of the GPT? As indicated here, the Norwegian language prompt provides a radically different answer than the English one but the ‘nationality’ of this GPT is contractually Norwegian.
• How does the variable outcome differentially impact students with different language skills in the same educational habitus? For example, what are the differences between the type of information monolingual English language speakers (including academics!) can access versus those who are English-proficient but non-Norwegian using native speakers of low-resource languages? How does this knowledge concur or contrast with the information obtained by Norwegian-speaking users?

In closing, as academics, we must acknowledge and understand the materiality of perceived bias, in this case and beyond.

Going forward, in striving for academic ideals of truth, integrity and open exchange of ideas, we need to be dealing seriously and collaboratively with how the optics of generative AI play out as politics.

Acknowledgments

While the perspectives and opinions in this blog are my own, I am grateful to the following for inspiration, discussion, and comments on previous drafts: Andreas Hirblinger, Maria Gabrielsen Jumbert, Sophia Adampour, Tobias Mahler, Lee Bygrave, and Bruno Martins.

The author

Kristin Bergtora Sandvik is a Professor of Humanitarian Studies at PRIO and a Professor of Law at the University of Oslo. She leads project LAW22JULY: RIPPLES: Rights, Institutions, Procedures, Participation, Litigation: Embedding Security.