Learning the Biased Languages of Benevolence, Equity and the Machine

Art museums and the organisations that support them frequently use language which promotes their benevolence and their commitment to equity. I will discuss why I have recently begun to experiment with machine learning to better understand how organisational texts like annual reports and policies adopt these biased languages, and how this might inform long-standing and systemic concerns about bias and discrimination in baked into societal AI. 

This article is based on a paper given at the Transforming Collections All Partners Workshop at Tate Modern on 15 November 2023. The Transforming Collections project forms part of the larger Towards a National Collection programme funded by the Arts and Humanities Research Council. My research on this project, entitled ‘Patterns of Patronage’ is based at the Decolonising Arts Institute at UAL, the University of the Arts London.

We have entered an age of hyper-sensitivity to language and words, used within manufactured contexts, and manipulated out of context in ways where common understanding can often be entirely absent. We see it daily, it’s affecting how we think about ourselves and our beliefs, and others and their beliefs.

It’s the reason why museums have begun scrutinizing their gallery labels and catalogue descriptions and why this remains a core function of the Transforming Collections project. My research so far has been to reveal the power structures that enable the collections we study, principally through looking the mechanisms and relationships of acquisition.

There is extensive literature on the biases already baked into the AI we are made to use every day without our consent. I think of Wendy Chun talking about “algorithmic authenticity” and its fallacies: it is critical to remain alert when researching bias in a social context where race, gender and other branded identities have become commodified, largely thanks to the design of extractive, discriminatory algorithms intended for accumulating power.


The binaryism that AI emerges from – this or that – yes or no – is particularly pernicious. It rides roughshod over nuance, awkwardness, contradiction and the kinds of complexities that humans have evolved around. Binaryism is contrary to creativity.

Stephanie Dinkins’ studio experiment, Binary Calculations are Inadequate, demonstrates an alternative to binary machine learning where you only have the choice to indicate presence or absence. It is based around human-readable responses, descriptions and media collected from as wide variety of informed and uninformed sources as possible.

In this example I uploaded an image of Chila Burman’s Band of Gold (2013) from the British Council Collection. The machine asked me to: “Upload an image that illustrates close attention, concern or responsibility.”

Unchecked and unbridled AI development is already, and will remain, a constant challenge to museums attempting to navigate our own “algorithmic authenticity.” If we accept that art museums can play a fundamental, tangible and textured role in mediating our inward and outward creativity then let’s take a look at the scaffolding of the texts and languages that authenticate and prop us up.

Language of benevolence

The language of benevolence, particularly around gift-giving, heavily influences the politics and processes of representation in public art collections. Over the last decade in particular, the language of equity has been entwined with that of benevolence as an attempt to signal an intention to move art collections from an exclusive to an inclusive arena.

The aim of my experiment is partly experiential and partly epistemological. I have no interest in creating any kind of tool or solving a problem. I want to see how organisational texts and languages relate to interpretive and documentary ones and I want to experience this thing we call the machine.

Architectures of absence

In particular I want to address two project questions:

  • 2. How can the architectures, algorithms and relations of oppression that structure and define collections’ narratives be surfaced and transformed?
  • 3b. How might ‘art’, ‘nation’ and’ heritage’ be complexified and re-imagined?

I am creating, relative to the kinds of large language models (LLMs) that feed commercial and political AI, two small, critical and focused datasets to train the machine on the languages of benevolence and equity encoded overwhelmingly in public documentation: collections development policies, annual reports, minutes of meetings, acquisition archives, press releases, and prospectuses – the kinds of formal texts we produce to communicate with the public and colleagues in a formal way.

Training the machine

The training dataset comprises the collections development policies of all Transforming Collections contributors, collaborators and partners but is predominantly based on documentation from the Government Art Collection, Arts Council England Cultural Gifts Scheme, Arts Council Collection, British Council Collection, Art Fund and Contemporary Art Society. Chronologically they span from around 1911 to 2023. Once I train this dataset, we will apply the algorithm on a test set. The test set comprises similar public documents but just from Tate. No material from Tate will have gone into the training set.

So far, the training set has yielded well over 10,000 paragraphs of text. There are a few thousand more to add. With a little regret, my training of the machine has had to adopt the binary approach so I am only able to indicate where the language of benevolence or equity is present or absent. In September we conducted a preparatory experiment, creating a rough and ready dataset that combined some of the material I am about to use, plus the generic artwork descriptions used in former demonstrations of the machine learning being developed by CCI colleagues.

Fruit and the language of equity

I was amused that after a little training, the machine predicted that any artwork described as a still life with fruit must be about equity… In a way I felt I was not training the machine but getting to know it. I have been journaling these little happenings as I am more interested in the process rather than the result. I want to use a relational and dialogic process.

Let me end by walking you through some examples of benevolence and equity I have identified manually, just through reading. I’m just going to use excerpts from reports by the Contemporary Art Society for this.

Good for artists, good for the public

We cannot escape the pervading nature of the language of benevolence when we talk about the work we do. We believe it is the right thing to do, and we want others to believe what we do is good for artists and the public. Similarly, equity in reward and representation, while dominating discourses about acquisition today, have never been far away even if the actions have taken a long time to match the commitment:

From the 1919-24 Annual Report, p. 3

“These are difficult times for young painters and sculptors who have to live on the proceeds of their work, and the need for some form of public patronage was never greater than now.”

From the 1919-24 Annual Report, p. 4

“The Committee are anxious to start a separate fund for the purchase of work by contemporary artists of foreign nationality. It is felt that the public galleries in England are very deficient in representative works of foreign schools…”

From the 1984 Annual Report, Chairman’s Report, p. ?.

“If we could introduce a new public to the pleasures of buying contemporary art we would also be helping the many talented artists…”

What has changed?

Let’s zoom in just on the latest CAS report from 2022-23 which demonstrates how active both these languages remain a century on. Underlined in red show the frequency of examples:


“I would also like to acknowledge the generous in-kind support of the commission…”

“We offer our heartfelt thanks to the sponsors of these new awards.”

“Thanks to the outstanding generosity of…”

“with the indispensable support of the Friends of the Museum there.”

p. 7

“we would like to make a special mention here of the Friends of Reading Museum, whose generous contribution made it possible to acquire a wonderful new carving by Halima Cassell.”

“we are grateful to… for their friendship and wise counsel.”

“People come to us to gift works of art because they know we have the networks of relationships required…”

“In 2022, we were honoured to…”

Say thank you

“We also hope that you will enjoy exploring the online catalogue of all the gifts we have ever given.” (p. 8)

There is always the familiar last page courtesy – courtesy and thanks being at the other end of the giving relationship: Supporters and Patrons are listed on pp. 110-111.

Equity expressed in acquisitions

The language of equity is a key mode of communication in the recent annual reports of nearly all art institutions. It is prominent in the descriptions of virtually all recent acquisitions and gifts made by CAS, for example, this quote by curator Rosy Gray, Norwich Castle Museum & Art Gallery, on the acquisition of Ibrahim Mahama’s photographic prints via the Collections Fund at Frieze. She references “interconnected narratives” and “global structures.” (pp. 15-17).

The acquisition of Eva Kotáková’s collage Blankets, Monsters, Anna and the World, 2022. Presented by the Contemporary Art Society, 2022-23 to the Hatton Gallery, Newcastle University (p. 48-49).

“Blankets, Monsters, Anna and the World is created by a contemporary woman artist and adds diversity to the collection by addressing gender balance for future exhibitions.”

The reason material from art patrons and funders like CAS are so illuminating is that they absorb the language used by artists, curators and museums themselves. However, some organisational reports are written without referenced voices and with increasing emphasis on PR-style communication.

The machine does not know the difference

The machine will not know the difference between the two and “algorithmic authenticity” will remain unconfirmed.

What I have shown you are a few randomly picked sentences, but read in context. When I come to the machine learning, I will be able to identify the pervasiveness of these languages at scale. However, the context will be gone, and I will have whole sentences, paragraphs and phrases to contend with, without a sense of the whole – or rather the whole will feel like a decentred collective rather than a more comfortable well-structured, authored, referenced text.

I feel a sense of responsibility, ethically, but equally I want to let go of some of the constraints of the way we are expected to research art and public collections, and to be open to the emergent process, in the belief that the machine will also be teaching me something.


Wendy Hui Kyong Chun, Discriminating Data. Correlation, Neighbourhoods, and the New Politics of Recognition (Cambridge, MA: MIT Press, 2021), p. 147. I inserted the notion of commodified identities also being “branded identities.”

Contemporary Art Society, Annual Report 1919-1924. https://contemporaryartsociety.org/about/story. Accessed: 6 November 2023.

Contemporary Art Society, Annual Report 1984. https://contemporaryartsociety.org/about/story. Accessed: 6 November 2023.

Contemporary Art Society, Annual Report 2022-2023. https://contemporaryartsociety.org/about/story. Accessed: 6 November 2023.

Stephanie Dinkins Studio (nd), Binary Calculations are Inadequate. https://binarycalculationsareinadequate.org. Accessed 21 April 2023.

Tom Seymour, ‘Exclusive: UK shadow culture secretary to map out first national infrastructure plan for the arts’, The Art Newspaper, 9 October 2023. https://www.theartnewspaper.com/2023/10/09/exclusive-uk-shadow-culture-secretary-to-map-out-first-national-infrastructure-plan-for-the-arts. Accessed 9 October 2023.

1 Comment