Chapter Three: Feminist AI Projects
Introduction
The Xenofeminist Manifesto urges feminists to wake up to the complexities of the digital age:
The excess of modesty in feminist agendas of recent decades is not proportionate to the monstrous complexity of our reality, a reality crosshatched with fibre-optic cables, radio and microwaves, oil and gas pipelines, aerial and shipping routes, and the unrelenting, simultaneous execution of millions of communication protocols with every passing millisecond.
In this spirit, a counter-hegemonic feminist movement is emerging that seeks to disrupt the status quo of AI. It may still be marginal, but its intent is radical – to develop ethical and equitable algorithms, while following in the footsteps of Haraway, who recognised the transgressive potential of intelligent machine technology. However, where Haraway has been criticised for ‘being much stronger at providing evocative figurations of a new feminist subjectivity than she is at providing guidelines for practical emancipatory politics’, these new technofeminists are seeking real-world tools for liberation and change, by creating AI systems free from algorithmic bias. As such, they are addressing Sandra Harding’s key question: can technology be used for feminist means?
There are a growing number of AI projects operating in this space – such as Data Feminist Principles, Feminist Principles of the Internet, Feminist AI and Mimi Onuoha’s Library of Missing Data Sets. The projects are all interdisciplinary, meeting at the intersection of technology, art, design and politics, allowing for more open-minded and critical questioning when it comes to envisioning an ethical, equitable and feminist future for AI. This chapter focuses on two ongoing projects - Feminist Data Set and Queer AI. Both focus on re-working the machine-learning process to build a feminist/queer chatbot, recognising that the data input must be intersectional and diverse to reflect on situated and community knowledge. Rather than seeing AI systems as impenetrable black boxes, they want to improve them, making them more sensitive to the lived experience of those who do conform to the norms of the white patriarchal system.
Feminist Data Set
Caroline Sinders is the founder of Feminist Data Set and a key innovator in the field of AI research, so I reached out to interview her in March 2021. A Gen Z American based in Berlin, she argues that ‘AI on its own is nothing scary, in the way that malware on its own is just a bunch of lines of code. The issue is the raw data that feed the neural networks.’She started Feminist Data Set after working at major tech company IBM Watson, where she was surprised by the lack of critical engagement from her co-workers about the data sets that they were building together. She explains: ‘I had a lot of questions, critique and feedback, and I remember one day my boss said, “You know, it’s not your responsibility to question everything we’re doing; it’s your job to make sure the project doesn’t fail.”’ For Sinders, this reaction highlighted the success/fail binary inherent in the field of technology and the lack of value placed on constant adaptation and evolution to minimize bias, which – in classic Marxist terms – alienates workers from the product of their labor and removes them from any form of accountability.
Responding to this issue, she set up Feminist Data Set to explore an alternative methodology where ‘every step of the machine-learning process is thoroughly re-examined through a feminist lens.’ She does this by running collective and iterative workshops, relying on communal decision-making:
I don't want to be the person saying, “This is feminist data”; that doesn't feel very feminist to me. I think an equitable form of data collection has to come from different groups of people working collaboratively...it becomes bigger than any individual decision.
There are seven steps in the workflow: 1) data collection, 2) data structuring and training, 3) creating the data model, 4) designing a specific algorithm to interpret the data, 5) questioning whether a new algorithm needs to be created to be ‘feminist’ in its interpretation or understanding of the data, 6) prototyping the interface, and 7) refining. The long-term aim of the project is to produce a feminist chatbot, but for now she is more concerned with auditing data and, if necessary, locating alternatives.
When I ask her to define ‘intersectional feminist data’, she responds:The big thing is trying to explain to people that intersectional feminism can manifest in writing without the word ‘intersectional’ being used. It is not just the topic of a text, but the approach it takes that is critical. For example, an intersectional article on wage inequality would break down how black women, indigenous women, Latinx women, Asian women, white women, trans women and trans men, and non-gender binary people are all paid differently. But a non-intersectional article would say women are paid less than men.
As with Joy Buolamwini’s case study from Chapter Two, intersectionality is key in understanding the discrimination of black women through algorithmic bias, which is why data need to reflect on these differences, and to see intersectionality not as ‘the morcellation of collectives into a static fuzz of cross-referenced identities, but a political orientation that slices through every particular, refusing the crass pigeonholing of bodies.’
Because Sinders is working with intersectional feminist data, she says that ‘what I find really interesting – and this is a big point of the project – is that a lot of older feminist literature can't really be in the data set, because it's not intersectional.’ This creates a point of generational tension between her and Haraway, as she sees ‘A Cyborg Manifesto’ as an older feminist text with outdated language - ‘it is not everything I want it to be right now’. She goes on to explain that ‘language and how we describe equity are constantly evolving, which is necessary, important and urgent. Sometimes, things written just four or five years ago can already feel slightly outdated, as if they are not acknowledging all the different kinds of harms that are in the world.’ Although this challenges the perception that ‘A Cyborg Manifesto’ is a proto-intersectional text that articulates a worldview ahead of its time, Sinders does acknowledge in the Feminist Data Set manifesto that she is ‘inspired by the cyborg manifesto, our intention is to add ambivalence and disrupt the unities of truth/information, mediated by algorithmic computation’.
Queer AI
Based in Los Angeles, Emily Martinez is a queer Latinx new media artist, digital strategist and founder of Queer AI. Launched in 2018, a year after Feminist Data Set (which she acknowledges as a source of inspiration), Queer AI experiments with creating conversational chatbots trained on a corpus of queer and feminist theory and literature. The chatbots are supposed to transcend functionality to become complex, emotional and empathetic, embodying the messiness of what it is to be human. Unlike Tay, the bot in Chapter Two, they are designed for one-to-one conversations with humans, rather than a public forum like Twitter, where intimacy is hard to achieve.
The Queer AI manifesto advocates:The queering of communication technologies. The queering of machine learning algorithms.The careful and collaborative curation of training material by and for a diversity of sensibilities and actors.The cumming undone of logics and sense making.The slipperiness of language.The accidental poetry of nonsensical miscommunication er(r)o(r)s.Fluid autonomous playful uncompromising disobedient bots.Messy fleshy curious humans.User intent mismatches as erotic deviations.Existential pause.Deep learning for deep orgasms.
The manifesto implies that the queering of AI is more than just diversifying the data set but approaching it in a transgressive and playful manner that celebrates ‘the slipperiness of language’ and ‘the accidental poetry of nonsensical miscommunication errors.’ The potential for imperfect and erotic embodiment through machines offers the excitement of the unknown and resonates with the ‘non-innocence’ of Haraway’s cyborg.
This appreciation for the ‘slipperiness of language’ requires being offset against the necessity for ‘careful and collaborative curation of training material’. Martinez writes that Queer AI is ‘not interested in adding to the mega corpus of straight white dudes training neural nets with boring fembot fantasy tropes for the lonely machine love future full of rapey robot sex.’ She is referring to the young ‘tech bro’ white male demographic that dominates AI programming, critiquing their tendency to create predictable and soulless chatbots that reflect only on heteronormative desires and values.
By contrast, she is trying to decolonise the language of AI ‘For the very first bot I made, trained on Audre Lorde, I meticulously cleaned all of the input text (a book and a half), which took me two days to complete. The entire corpus weighed in at just under 200KB, which is considered “useless” by most standards, but it’s what I had to work with, so I went with it.’
The contrast between this Audre bot and the Tay bot is striking, as revealed in this small sample of conversation:
> Emily – sometimes I feel like I’m afraid of my own power.
> Audre – like I’m going to run the risk of being dominated by someone who is, well, more powerful than I am... like I’m going to lose it all, when in fact power is actually expanding exponentially every time I think about feminism and lesbianism.
It is significant that Martinez chose Lorde’s work as a model of appropriate queer literature, as Lorde was also an inspiration for Donna Haraway when writing ‘A Cyborg Manifesto’ (see Chapter One). The legacy of her language clearly lives on, without being ‘outdated’.
While the Audre bot was a small-scale experiment, the Queer AI chatbot – turned on in 2018 – was trained on 50,000 extracts about queerness, eroticism and sexuality taken from queer theatre, incorporating the work of authors such as Oscar Wilde, Caryl Churchill and Jean Genet. Two years later, the data set had diversified and grown to eight million pages. Martinez writes about the result of her experiment: ‘Spoiler! When you use a corpus of queer theatre (mostly from the 1980s, at the peak of the AIDs crisis) to train a language model, you will likely generate an algorithm that is biased towards expressing the legacy of trauma endured and experienced by the characters in those texts.’ It is interesting how she uses the word ‘biased’ in terms of expressing trauma, because this implies bias is inherent and inevitable regardless of the algorithm, whether it fuels discrimination. Arguably, wherever there is language, there is bias.
Martinez is in the process of making a toolkit that teaches beginners how to create their own text-generating chatbots. Like Sinders, she is less interested in big data, with its acceptance of hegemonic ‘universal’ language, than the sort of data that account ‘…for the nuances of small, marginalised, or intentional communities – their identities, lexicons, vernaculars, sexualities, and sub/cultures.’ She asks: ‘How do we prepare data sets and set up guidelines that protect the bodies of knowledge of our communities… rooted in shared, agreed-upon values?’ She believes that making a project community-facing has the potential to demystify AI and reduce the power of the major tech companies which dominate the field. It may be a small act of protest, but re-thinking how algorithms can be educated could prove to have a significant impact, if tech companies are pressured to adopt their tactics and make their models more ethical and equitable.
Like Feminist Data Set, Queer AI is a work-in-progress and will continue to evolve as language evolves. So far, it has a landing page with a growing community of users, and a functioning chatbot that has gone through its trial stages. Ultimately, however, the success of the project will be judged less by the quality of the automated conversations it generates than by the ongoing process of critique and debate. The aim is to provoke, interrogate and experiment, all in the name of protest.
Conclusion
This chapter started with a radical demand for feminists to take action against ‘the monstrous complexity of our (digital) reality’ and the criticism of Haraway for not providing ‘guidelines for practical emancipatory politics’. In this context, new projects are emerging that seek to fill the vacuum and challenge the dominant paradigm. While ‘algorithmic oppression’ may indeed be embedded within the ‘operating system of the web’, it is reasonable to expect that awareness of bias, discrimination and injustice will grow over time – not because of the goodwill of the big tech companies, but because of the persistence of activists like Caroline Sinders, Emily Martinez and others who hold them to account and offer up alternatives which push for a more ethical and equitable future.
TO READ HE FULL TLICK HERE