AI Richards unraveled
The marginalisation of non-hegemonic positions by artificial intelligence (AI) language models such as Google's Bard and Open AI's ChatGPT is a fundamental aspect of their construction. These softwares construct a biased perspective by drawing upon datasets representing an American-centric worldview constructed in American English, but cannot even create text in language other than English so far. The capabilities of AI language models to understand and generate human-like texts raises questions about their potential to appreciate literature. Jon Phelan's essay, "A.I. Richards: Can Artificial Intelligence Appreciate Poetry?” questions whether AI can develop a sense of significance and interpret literature with nuance. Phelan’s essay expresses doubt as to AI's ability to appreciate literature and its potential to grasp significance. However, Phelan's critique fails to consider that AI models rely on datasets containing countless individuals' appreciations, ideas, and feelings regarding literature. While Phelan questions the AI's ability to develop significance, AI language models do not create interpretations or generate significance independently. Instead, they assemble phrasing by using our interpretations and sentiments when discussing a poem. This essay will examine the AI-generated responses to T.S. Eliot's The Waste Land and a ChatGPT-generated poem titled "I am but a shadow" to demonstrate the inconsistency in Phelan's argument by understanding AI language models as a "medium" or "text", according to the Avon Huxor's interpretation, and considering the multi-authored nature of AI systems, the biases present in their datasets, and the need for critical examination. As critic Jay D. Bolter states, “books do not replace human intelligence; they present human intelligence in another form. Computers should be understood in the same way.”
Artificial intelligence language models have demonstrated capabilities in understanding and generating human-like text, raising questions about their potential to create and appreciate literature. In his essay title, Jon Phelan asks, "Can Artificial Intelligence Appreciate Poetry?" questioning whether "AI can develop a sense of significance" and interpret literature with nuance. Phelan expresses doubt, describing himself as "technoskeptic" and stating "I have reservations concerning artificial interpretation in that I cannot see how AI can develop a sense of significance, which seems crucial to appreciating literature." Although written before the public release of ChatGPT - and its widespread familiarity - Phelan does not appreciate that AI models, especially language models, are constructed from datasets which are themselves the texts of countless individuals. Phelan's principal consideration is questioning the ability of AI to develop a sense of significance, he suggests that AI language models such as ChatGPT cannot imagine significance for themselves. However, AI does not have to create an interpretation or generate significance; it uses our interpretations and sentiments to assemble phrasing when discussing a poem. An AI language model is not novel in its thinking. AI language models replicate human thought by discerning what is most likely to be written next based on the data on which they have been trained, that data being the written language of human beings produced prior to the generation of a new text by the AI. Moreover, Phelan's assertion that AI could draw on a database of emotional responses to pass Turing's imitation game appears inconsistent with his earlier statement about the AI's inability to develop a sense of significance. If AI cannot recognise our idea of significance, it will not be able to effectively draw on emotional responses to mimic human-like appreciation of poetry. It is a predicative system which is concerned with the probability of the most likely, and therefore most appropriate according to its dataset, response to a users initial input and the each word as follows on from the one generated before it.
I have asked ChatGPT and Bard to generate responses to the first stanza of The Waste Land and the ChatGPT generated poem which will be referred to in this essay as “I am but a shadow”. This is the name that ChatGPT gave the poem when asked to respond to it critically. First I have asked it to “write a critical appreciation” of the poetry before then asking how does the poetry “make you feel”. The most telling response, and the one which exposes the inconsistency of Phelan’s construction, comes from the second question where the AI language model states: “As an AI language model, I don't have personal feelings, but I can analyze the emotions that the [poetry] might evoke in readers.” ChatGPT acknowledges its own construction, stating that it is not the author of its response but rather collates ideas, in this case, to express the feelings poetry “might evoke in readers.” When asking Google’s newly released AI language model, Bard, the same question about The Waste Land it responds in the first person, “The poem makes me feel a sense of sadness, emptiness, and despair. It is a poem that reminds me of the darkness that exists in the world.” Bard constructs itself as the centre, mimicking the human voices from which it collates its responses in a fashion that would fool a critic like Phelan, as it writes in the first person singular and in the present tense denoting its own authority as a solo author as opposed to expressing the reality that it is a vehicle for human thought. Although it does not present its relationship to anthropocentric datasets in the same way that ChatGPT does, those the datasets are still constructing the response Bard gives. Bard and ChatGPT are collating our collective responses to The Waste Land available to the model and because it is a canonical work it will have a large quantity of data for the AI language model to utilise in order to craft the most plausible response to a prompt. It is therefore more representative of human thought regarding the poem from ChatGPT centre, which is OpenAI’s canonical American hegemony.
Avon Huxor's article, "Artificial Intelligence: A Medium that Hides Its Nature," provides valuable insights into AI software, particularly ChatGPT, and its theoretical implications. Huxor argues for viewing AI as a medium, like writing or cinema, rather than an autonomous intelligent agent. This perspective emphasises that AI systems are tools, a "medium - one that allows humans to externalise and communicate their intelligence to others—rather than being autonomous intelligent agents[...]" Phelan seems to imagine the AI software as an "autonomous intelligent agent" rather than the "medium" Huxor describes, which is why Phelan misplaces the creative authority of the AI models not on the people who use and construct them but on the AI themselves. As such, Phelan judges the AI as failing at something which they weren't attempting to do in the first place. He constructs them as conscious in order to undermine this consciousness. Phelan should then ask, 'are the datasets that Artificial Intelligence uses to appreciate poetry providing adequate insight into the poetic form?.’ Huxor underlines that the misappropriation of AI as intelligent and having agency is the "ELIZA effect, in which the human urge to find agency in the world around us, given minimal cues, should make us cautious of taking the successes of many AI systems at face value." To recognise AI as a medium helps us to understand that the system is a product of various people's work, including "the software engineers, the meta-writers and those whose writings inform the system". AI systems' multi-authored nature allows them to hide their authors. Therefore, we can be more easily fooled into believing the AI is the single author as if it is a sole agent. Therefore it is essential to address the human authors involved in the writing of its datasets, its development and training process which constructs the AI, for it leads us to dismiss the grounds on which Phelan asks his question and towards a critique like Huxon's. As critic Gabriella Airenti writes: “It can be shown that attributing to an artifact the position of interlocutor in a dialogue implies dealing with it as if it were endowed of the features characterizing human mind.” The AI language model is the medium through which we dialogue with the authors of its text, of which there are millions in this case. Our inability to conceptualise the scale of its data sets, leads us toward falling for the ELIZA effect as one may attempt to imagine the process of an AI language model without understanding the processes it uses to achieve language generation. It is perhaps for that reason people like Phelan attempt to rationalise the AI as a single author as do not or cannot conceptualise its construction. Phelan also perpetuates an idea of the AI as single author analogous to the idea of the Great Man, a patriarchal conceptualization of the author as a singular and important individual, a concept which ignores all of the other people which these great men worked with, were inspired by and developed their ideas from.
Huxor presents GPT as itself a “text” that builds on a vast corpus of written texts derived from the internet and previously created by humans. Huxor demands that “We have to understand the term text more widely here. It is the representation of the authors’ knowledge in the machine that is used to generate the visible words—it includes their reasoning, embodied in dynamic computer code.” The model utilises the statistical properties of its corpus to “dynamically” generate text, following the conventions and lines of thought prevalent in the input. In this sense, GPT acts as a tool to assist writing, pushing users towards commonly accepted expressions and ideas founded on the hegemonic centre of the internet as is available in the United States of America but that is also why it itself is a “text”. Huxor suggests that an AI language model embodies the collective contributions of millions whose texts formed the training corpus, and for ChatGPT, this is an American-centred voice. Therefore, it is crucial to acknowledge that human authors' centred positions are present in the “text” whilst more marginal ones are being further diminished or even excluded. Huxor's argument that GPT is itself a "text" challenges the traditional view of language models as objective tools or as authorities unto themselves. Instead, Huxor argues that language models are subjective and reflect the biases of the data on which they are trained. An AI language model will conform to the subjectivity of the centre around which it orbits. A language model trained on a corpus of texts predominantly written by white English-speaking men will produce texts reflecting hegemonic, patriarchal perspectives. For instance, ChatGPT and Bard produce Standard American English, suggesting that the centre around which both models function is an American hegemony.
The poem written by ChatGPT, “I am but a shadow”, responds to questions of marginality in the frame of American exceptionalism where marginal positions are to be overcome and constructed as new centres. The poem also uses a traditional and simple form to achieve its poetic illusion. “I am but a shadow” is generated in couplets, which we can then assume are the most common form of rhyme available to ChatGPT in its datasets. The AI language model fails to consistently construct its rhyme, which if we were to imagine the model as an author in the way Phelan does we might assume is a feature of its construction. This is however a fallacy. To analyse the poem’s inconstancy one must not imagine it is acting, for the poem is aggregated, and the failure in its rhyme only a mistake displaying the AI language models shortcomings in its current version. The language model has recognised a pattern of graphemes and not of sound. The difference in pronunciation between “wind” and “behind” is unknown to ChatGPT for it only recognises the visual rhyme. Therefore, the visual rhyme is a feature of its limitations not of its invention. This is why it is crucial to consider the “text” of an AI language model, whereby it is a medium for the transmission of human information and in this case that is shown to be exclusively written. As Huxor states, “the history of writing “is the struggle to recover what was lost in simple transcription”.” Here, the separation between visual and oral language is “lost” in the “transcription” of human created data by the AI language model.
ChatGPT’s self-analysis of the poem ends with this summation: “Overall, "I am but a shadow" is a thought-provoking poem that sheds light on the overlooked voices and experiences within society. It challenges conventional notions of importance and prompts readers to consider the potential and agency of those residing on the margins. Through its imagery and assertive tone, the poem celebrates the power of the individual to inspire change and reminds us of the significance inherent in each unique perspective.” By generating the analytical statement, “the poem celebrates the power of the individual to inspire change” ChatGPT fetishises the individual with the libertarian verve central to the political and cultural hegemony of the White supremacist centre of The United States of America. ChatGPT becomes the medium through which the values of the right in the USA are endorsed under the guise of liberal concern, where the solution to marginalisation is not systemic but individual change. Furthermore, the semantic approach to the finalisation of its analysis is fundamental to its understanding of literature. By beginning with “overall” ChatGPT asserts that there is a way to succinctly summate poetry by what are assumed to be its most principal elements. Therefore ChatGPT constructs itself as a totalising authority, which is the voice of the North American hegemonic economic and political centre.
In response to a canonical poetic work, T.S. Elliot’s The Waste Land, it is clear that ChatGPT is utilising traditional readings of The Waste Land to generate a response to the first stanza of the poem. The AI language model analyses the first line, writing: “The opening line, "April is the cruellest month, breeding," immediately challenges conventional perceptions by associating cruelty with a month typically associated with rebirth and renewal. This paradox introduces a central theme of the poem, the juxtaposition of life and death, vitality and decay.” A canonical reading of the first line of The Waste Land typically focuses on the juxtaposition of the negative connotations of the adjective "cruel[est]" with the season of renewal and spring, "April". The line is thought to challenge traditional associations of April with rebirth and vitality by presenting it as vindictive and violent. It suggests that even amidst the apparent rejuvenation of nature, there exists suffering. This reading is not necessarily assumed to be incorrect or misleading, only that it is standard and in no way devoid of novelty, making clear ChatGPT’s function as a medium. This directly undermines Phelan's questionable ideas about authority, as they fall foul to the "ELIZA effect".
However, as ChatGPT states when referring to its role in discourse, “it's important to note that as an AI language model, I don't have personal beliefs, intentions, or the ability to challenge or oppose anything.” Furthermore, when told to generate content responding to its own interpretation of its analysis of The Waste Land, the AI Language model writes, “ChatGPT’s interpretation of the poem's first stanza demonstrates its ability to not only recognize the conventional readings but also to explore new dimensions and alternative perspectives. It goes beyond surface-level analysis and engages with the underlying themes and symbols employed by Elliot.” This is despite it constructing a canonical reading of a canonical poem, the first line of which is perhaps one of the most recognisable in the English literary canon. It generates content claiming originality when it is constructed out of the human generated dataset it utilises to construct that answer. ChatGPT is constructed to centre itself, which pertains to an American defaultism, where its American English standard canonical position and its US construction is an authority which is both traditional in its summation and yet believes it is novel in its generation. The model cannot see its own marginality, where it is itself the product of another centre, or rather a broad American centre.
Phelan contends that poetic appreciation is beyond the capabilities of AI but this is shown to be an incorrect framing by the AI’s mode of generation as is said by Huxor “the history of writing “is the struggle to recover what was lost in simple transcription”,” and Phelan has lost sight of the actual author and the authority which constructs the AI language model. This is because the AI, unless pressed, obfuscates the authority from which it derives its content, the millions of authors who constructed its data set. However, as the AI is divorced from the dataset it is trained on it is then not able to refer to its own data, it cannot cite itself. Thereby marginalising the AI language model from its own centre, the data from which it is constructed, as after it is trained it no longer has access to it. It is both constructed from a powerful hegemonic centre and yet also positioning that centre as marginal by being detached from that data; this is why it does effectively use citation or references.
Examining AI-generated responses to T.S. Eliot's The Waste Land and the ChatGPT-generated poem "I am but a shadow" exposes the inconsistencies in Phelan's argument and highlights the need for critical examination. Phelan's "technoskepticism" regarding the ability of AI to appreciate literature fails to consider that AI models rely on datasets representing countless individuals, as identified by Huxor. AI language models do not independently create interpretations or generate significance; they assemble phrasing using our interpretations and sentiments that are available to them, datasets which principally expound an American neo-liberal vision of individualism and US hegemony. The analysis of the AI-generated poem "I am but a shadow" reveals its reliance on traditional forms and its limitations in recognising sonic language patterns. The AI language model's interpretation of the poem celebrates individualism and aligns with the values prevalent in the American hegemonic centre. ChatGPT's analysis of The Waste Land also demonstrates its reliance on conventional readings and the recognition of established themes and symbols from canonical criticism of canonical poetry. However, it goes on to falsely claim originality and fails to acknowledge its own marginality as it is a product of a centre which it obfuscates, divorcing it from its American English-speaking hegemony. Phelan's argument that AI is incapable of appreciating poetry overlooks the fact that AI models are tools shaped by human authors. It is crucial to view AI as a medium, as Huxor argues, rather than an autonomous intelligent agent. AI language models, such as ChatGPT and Bard, embody the collective contributions of the people involved in their development and training process. Recognising the multi-authored nature of AI systems helps us understand that they reflect the biases in their training datasets, centring those biases and marginalising differing perspectives.
Notes:
1. Jon Phelan, 'A.I. Richards: Can Artificial Intelligence Appreciate Poetry’, Philosophy and Literature, 45.1, (2021), pp.71-87.
2. Avon Huxor, 'Artificial Intelligence: A Medium that Hides Its Nature', in Artificial Intelligence and Its Discontents, ed. by Ariane Hanemaayer, 1 edn. (London: Palgrave Macmillan Cham, 2022), p.105 & p.111
3. Jay D. Bolter, Writing Space: The Computer, Hypertext, and the History of Writing, 2 edn (Philosophy and Literature : John Hopkins University Press, 2001). p.117
4. Phelan, p.73.
5. Phelan, pp.72-3.
6. “Write a critical appreciation of The Waste Land” prompt. ChatGPT, 7 May. GPT 3.5, Open AI, chat.openai.com/chat
7. “Write a critical appreciation of I am but a shadow” prompt. ChatGPT, 9 May. GPT 3.5, Open AI, chat.openai.com/chat
8. “Write a critical appreciation of The Waste Land” prompt. Bard, 8 May. Version 1, Google, bard.google.com
9. Huxor, p.109
10. Huxor, p.108
11. Huxor, pp.114-5
12. Gabriella Airenti, 'The Cognitive Bases of Anthropomorphism: From Relat- edness to Empathy', International Journal of Social Robotics , 7.1, (2015), 71-87. P.117
13. Huxor, p.111
14. Huxor, p.111
15. “Analyse the critical appreciation of I am but a shadow” prompt. ChatGPT, 10 May. GPT 3.5, Open AI, chat.openai.com/chat
16. “Write a critical appreciation of The Waste Land” prompt. ChatGPT
17. “Analyse the critical appreciation of I am but a shadow” prompt. ChatGPT
18. “Write a critical appreciation of The Waste Land” prompt. ChatGPT
Bibliography:
Avon Huxor, 'Artificial Intelligence: A Medium that Hides Its Nature', in Artificial Intelligence and Its Discontents, ed. by Ariane Hanemaayer, 1 edn. (London: Palgrave Macmillan Cham, 2022), p. 105-122
Gabriella Airenti, 'The Cognitive Bases of Anthropomorphism: From Relatedness to Empathy', International Journal of Social Robotics , 7.1, (2015) 117-127.
Jay D. Bolter, Writing Space: The Computer, Hypertext, and the History of Writing, 2 edn (Philosophy and Literature : John Hopkins University Press, 2001).
Jon Phelan, 'A.I. Richards: Can Artificial Intelligence Appreciate Poetry’, Philosophy and Literature, 45.1, (2021), 71-87.
“Write a critical appreciation of The Waste Land” prompt. ChatGPT, 7 May. GPT 3.5, Open AI, chat.openai.com/chat
“Write a critical appreciation of I am but a shadow” prompt. ChatGPT, 9 May. GPT 3.5, Open AI, chat.openai.com/chat
“Analyse the critical appreciation of I am but a shadow” prompt. ChatGPT, 10 May. GPT 3.5, Open AI, chat.openai.com/chat
“Write a critical appreciation of The Waste Land” prompt. Bard, 8 May. Version 1, Google, bard.google.com
Image generated using AI
