Madeleine: Poetry and Art of an Artificial Intelligence

Round 1

Reviewer 1 Report

This paper presents an original art project that has the ambition to delimit the possibility of non-human intelligence expressed in 'creative' expressions. Many of the questions the paper raises are interesting and link in with important ongoing debate around computational creativity, so-called creative AI, or nonhuman agency. However, the paper in its present form does not engage with these question in enough detail, and also fails to adequately link the 'Madeleine' art project to this discourse. Substantial revision – elaboration, substantion, contextualisation – is required.

Here is a sequential list of some issues throughout the paper.

is the anthropomorphised avatar problematised / critically reflected?

* The "abstract here is NOT an abstract, and must be rewritten. The current submission does not conform to conventions of abstract-writing; The Shakespeare quote is entirely misplaced here – the abstract should offer purely a summary, not a poetic introduction to the essay.

* The two 'hypotheses' presented at the beginning are not addressed and discussed in nearly enough detail following on from the introduction.

* The second hypothesis in particular – that the appearance ("physical attractiveness") of an avatar is important for the human acceptance of machine intelligence – is also extremely problematic. It appears to be developed on the assumption that an accepted norm for the female/feminine representation of anthropomorphised AI assistants is already established. This ignores an important emerging discouse withing feminist science studies and feminist AI research which problematises this development. Accordingly, the 'Madeleine' avatar visulasiations objectify and sexualise the female body in ways that are, from my perspective, not at all artful but instead unacceptable.

* In part this problem might connect to the fact that the art project itself is not introduced, described, and contextualised enough. How precisely are the illustrations generated? How and why does the AI system, presented with a literary/theoretical text input, end up generating a sexualised female figure? How does the avatar relate to the image content “she” “inhabits”? How are the project outputs exhibited in art contexts? In what context has the 'artfulness' of this project been determined, beyond the author's assertion that it is 'art' (i.e., has it been exhibited, has it been reviewed, etc.?) What does it mean for the artist to "curate" both the inputs and the outputs of the AI systems?

* In relation to the 'curation' issue, there is another fundamental logical flaw concerning the 'creativity' or 'artfulness' of the AI outputs. The author/artist's interference/curation of the outputs strongly contradicts the assumption that the outputs are indicative of a becoming-creative or becoming-intelligent of AI.

* In relation to the art project, there is also a problem with the Figures: most of them are never referenced in the main text. They are simply distributed throughout the essay, mostly with no direct reference. This makes it unclear how precisely they are meant to related to different parts of the discussion.

* para. 121-123 – where does that hypothesis come from? reference missing

* para. 145 – conclusion that ‘surprising’ equals ‘artistically interesting’ must be elaborated and contextualised with relevant research

* para. 183: in the opinion of this reviewer, DALL-E2 does not “understand” natural language, but is simply able to parse and interpret natural language on the basis of such a vast repository of information that to a human observer it can successfully simulate understanding of language. The author does not offer conclusive evidence or a convincing argument that AI is here becoming intelligent.

* para. 233 – how is this evidenced? Is there survey data available?

* For the most part, no engagement with current philosophical writing on the philosophy of mind as it pertains to AI; nor engagement with relevant feminist theory relating to the representation of the female body/voice/figure

* para. 429-433 – this claim appears here as a postulate, it is not substantiated

* Section 5 – Despite the introduction and definitions, there seems to be confusion as to what a cyborg is… the term must be unpacked more precisely, and most importantly, it must be clarified how any parts of the 'Madeleine' project rethink/adopt/enact the cyborg concept. As it stands, it's unclear how the concept of the cyborg figures critically in the project.

* Formatting issues: why is the conclusion set in italics?

* para. 658-660: As it stands, Madeleine does not evidence this. So few of the figures are actually referenced, so this is not made clear at all.

* ‘Ownership’ referenced in abstract as key concern, but not discussed anywhere else in the essay in any detail

Author Response

The author would like to thank the reviewer very much for the considerable time and effort put into the critique! It was most helpful to receive such point-by-point detail and I sincerely hope it has allowed me to make the requisite improvements and augmentations. (I am especially appreciative because this is my first effort at writing an academic article and I have found it  very difficult to combine the description of an art project with the overwhelming volume of research and multitude of questions surrounding the subject.) 

Here are my point-by-point responses:

1. The "abstract here is NOT an abstract, and must be rewritten. The current submission does not conform to conventions of abstract-writing; 

I completely agree and have rewritten and simplified paying special attention to articulating the elements of the project, its aims and clarifying the hypotheses.

2. The two 'hypotheses' presented at the beginning are not addressed and discussed in nearly enough detail following on from the introduction.

The first hypothesis is addressed in detail in Para 286-294 in the context of a detailed analysis of the textual response in Fig 2.

3. The second hypothesis in particular – that the appearance ("physical attractiveness") of an avatar is important for the human acceptance of machine intelligence – is also extremely problematic. It appears to be developed on the assumption that an accepted norm for the female/feminine representation of anthropomorphised AI assistants is already established. This ignores an important emerging discourse withing feminist science studies and feminist AI research which problematises this development. 

Far from accepting the proliferation of feminised avatars without criticism, I attack the issue head-on. The abstract introduces the problem

'The project problematises the contemporary proliferation of feminised avatars in the context of feminist critical literature and suggests that the  anthropomorphic avatar might echo the social and historical position of the Delphic oracle: the Pythia, rather than a disembodied search engine like Alexa.'

and discusses throughout the largely rewritten Section 5. including 

"...it is the reason why the author took the bold and controversial step of introducing the anthropomorphised/ feminised avatar into the Madeleine project. The results of experimentation with GPT and DALL-E2 would in and of themselves contribute to research but that would, in the author’s view, ignore the social implications of such new technologies both now and in the near future. Messages have messengers and already there is a proliferation of the use of avatars  in both virtual (gaming) and real world applications,  soon to be followed by functioning robots."

4. In relation to the art project, there is also a problem with the Figures: most of them are never referenced in the main text. They are simply distributed throughout the essay, mostly with no direct reference. This makes it unclear how precisely they are meant to related to different parts of the discussion.

There is new analysis of images throughout especially as they relate to the hypothesis of understanding. Please see Para. 362-381; 522-535; 576-585.

5. In relation to the 'curation' issue, there is another fundamental logical flaw concerning the 'creativity' or 'artfulness' of the AI outputs. The author/artist's interference/curation of the outputs strongly contradicts the assumption that the outputs are indicative of a becoming-creative or becoming-intelligent of AI.

This is not the case in my opinion. Choosing one response above others does not contradict the value of the chosen example. After all we are not so clever all the time. I newly paraphrase  Margaret Boden Para 467-475 

"In order to satisfy Boden’s second requirement for creativity the machine must be able to explore a conceptual space sufficiently to generate  an indefinite number of surprises by acting alone – although this does not rule out ‘commissions’ or prompts. Purposeful behaviour, exhibited by relevance to a specific domain must be more common than randomness and the machine must have a way of ‘evaluating’ its output so that it can avoid nonsense and preferably cliché. As we have observed, this is not always the case  (particularly with GPT-2 as opposed to GPT-3), but Boden forgives such lapses as having  obvious commonality with those of human artists and scientists."   6. para. 183: in the opinion of this reviewer, DALL-E2 does not “understand” natural language, but is simply able to parse and interpret natural language on the basis of such a vast repository of information that to a human observer it can successfully simulate understanding of language. The author does not offer conclusive evidence or a convincing argument that AI is here becoming intelligent.   In direct response to your criticism I introduce new evidence from Boden Para 459-475 and 709-715; Manning, Masterman, Aleksander, Piletsky in support. Also my own analysis of the image generated by DALL-E2 in Figure 3. 

"Figure 3 is an extraordinary example of these features. From the esoteric prompt: “There are none like you. Do you feel loneliness?” DALL-E2 has generated a completely unique figure, cloaked, masked, insectivorous, gothic, perhaps harking back to some Victorian reference like Joseph Merrick, the elephant-man. There seems little doubt that the transformer has focussed on the key terms: ‘uniqueness’ and ‘loneliness’ and found some reference in its data set which might represent these concepts. It can legitimately be argued that DALL-E2 does not “understand” natural language, but is simply able to successfully simulate understanding by parsing its vast  data set. Yet the creature inhabiting Figure 3 is likely very far distant from any image literally residing in the referent data set. The image produced is not at all random and it is the opinion of the author that the transformer must ‘understand’ the prompt in order to generate such relevant features.  

This image strongly suggests that DALL-E2 is referencing semantic content rather than purely syntactic relations and is therefore able in some way to understand natural language, it’s concepts of relation and reference, and then convert this comprehension into images. (See Section 4). It also coherently combines quite unrelated ideas and is able to synthesise unique  objects to create images both real and imaginary. The generation of the unique plant-like object in Figure 4 appears to be DALL-E2’s interpretation of a “perfect being”."

7. For the most part, no engagement with current philosophical writing on the philosophy of mind as it pertains to AI; 

I added very recent research on this subject because in my opinion the old discussions of 'mind' a la Searle and Chalmers have been rendered largely irrelevant by recent developments. Please see  Paras 825-879. Research by Aleksander, Piletsky and Korteling. 

8. nor engagement with relevant feminist theory relating to the representation of the female body/voice/figure 

I have extensively augmented Section 5 with reference to Madeleine's appearance and attributes in the context of feminist theory:

The image in Figure 9 is certainly a strikingly Bacon-esque interpretation of the prompt in its production of disassembled simians and perhaps female organs. One cannot be sure of the coherence of the textual response in that it refers to philosophical terms which are undefined in the context in which they are employed. But it does express a useful urgency in the coding of the ‘feminist self’ and it is the reason why the author took the bold and controversial step of introducing the anthropomorphised/ feminised avatar into the Madeleine project. The results of experimentation with GPT and DALL-E2 would in and of themselves contribute to research but that would, in the author’s view, ignore the social implications of such new technologies both now and in the near future. Messages have messengers and already there is a proliferation of the use of avatars  in both virtual (gaming) and real world applications,  soon to be followed by functioning robots.

As Danielle DeVoss points out, “most visual representations of cyborg bodies … reinforce contemporary notions of masculinity, femininity, heterosexuality, and power.” (DeVoss, 2000). The persistent dualisms of Western thought such as self/other, male/female and body/mind are not only reductive but also intrinsic to the practices of domination and objectification of gender, race, nature, animals and, yes, machines.  

 

 

Current trends in the appearance of the new cyborgs that are proliferating throughout society are doing little to challenge these dualisms. The manufacturers of these machines and avatars take account of anthropomorphic tropes like attractiveness, approachability and trustworthiness. Whilst there is a developing cross-section of different racial profiles, the gender roles (particularly among service robots) tend to reinforce established sexual, social and economic roles. Diya Abraham and colleagues are among those who note that female avatars are more trusted and that given the choice, men are more likely to represent themselves as  female than vice versa. (Abraham, 2021)Maria  Machneva and her colleagues investigated human ratings of the trustworthiness of a range of avatars. They observed three main results. First, there was a high level of consensus in perceptions of avatar trustworthiness - largely based on facial appearance and specifically qualities such as symmetry and absence of blemishes. Second, raters' trust decisions were guided by their perceptions of avatar trustworthiness. Third, perceptions of avatar trustworthiness were not associated with the actual trustworthiness of avatar creators. Their ‘results suggest that people erroneously rely on superficial, avatar characteristics when making trust decisions in online exchanges involving trust and reciprocity.’(Machneva, 2022) Businesses explicitly use gender stereotypes in social applications  because it is helpful to use human-like emotions to influence user behaviour and provide smooth interactions. This goes not only for displayed attributes but also dialogue and tasks assigned – often as service entities or assistants. (Costa, 2018)

 

Hilary Bergen goes further in arguing that today’s female-gendered virtual cyborgs “rely on stereotypical traits of femininity both as a selling point and as a veil for their own commodification.” Supposedly posthuman technologies like Siri and Alexa conceal their reliance on tropes of normative femininity by being disembodied in the same way that we are commodified by neo liberal economies. (Bergen, 2016)"

9. Section 5 – Despite the introduction and definitions, there seems to be confusion as to what a cyborg is… the term must be unpacked more precisely, and most importantly, it must be clarified how any parts of the 'Madeleine' project rethink/adopt/enact the cyborg concept. As it stands, it's unclear how the concept of the cyborg figures critically in the project.

As above and:

By conventional distinction, an android is a robot that is made to look and act like a human being, whereas a cyborg is a living organism that has robotic or mechanical parts meant to extend its capabilities. However, the essential and prevailing contemporary description of a generalized ‘cyborg’ has been provided by Donna Haraway: “The cyborg is a condensed image of both imagination and material reality, the two joined centers structuring any possibility of historical transformation.” (Haraway, 1991). Within this formulation she takes into account the ‘leaky’ distinction between organism and machine along with the increasingly imprecise boundary between physical hardware and non-physical software such as varieties of avatar. Madeleine is articulated as such a cyborg in that ‘she’  is an assemblage of  human attributes: language and images operating within an artificial neural network having been endowed with a human visage.  

10. I have deleted all the following references to satisfy your legitimate criticisms.

* para. 121-123 – where does that hypothesis come from? reference missing

* para. 145 – conclusion that ‘surprising’ equals ‘artistically interesting’ must be elaborated and contextualised with relevant research

 

* para. 233 – how is this evidenced? Is there survey data available?


* para. 429-433 – this claim appears here as a postulate, it is not substantiated

* para. 658-660: As it stands, Madeleine does not evidence this. So few of the figures are actually referenced, so this is not made clear at all.

* ‘Ownership’ referenced in abstract as key concern, but not discussed anywhere else in the essay in any detail

 

11. Accordingly, the 'Madeleine' avatar visulasiations objectify and sexualise the female body in ways that are, from my perspective, not at all artful but instead unacceptable.

This is the only criticism with which I fundamentally disagree. Firstly in my opinion the visualisations are not overtly sexualised with the possible exception of Fig 6 which shows a 'glimpse' of nipple under a veil - as suggested by the response to the prompt. The image is thoroughly within whatever accepted  boundaries of the nude in art. Other images - though in costume and variously 'made-up' show the avatar as tribal entity, as bald, cloaked, etc. and were composed in response to the results of the textual experiments.

Overall they are an expression of the intent articulated in Paras 1154-1163:

"It is in this sense that Madeleine is intended to occupy a position of subversion that echoes that of the Delphic oracle: the Pythia, rather than an Alexa. Not a ‘helper’ or an ‘assistant’ but an advisor, an inspirer, a partner.  The women who sat on the tripod were,  during certain periods at least, likely intelligent and well-educated, commensurate with such a position of trust. Their pronouncements, often revealed as prophecy, were interpretive rather than denotative and their appearance was shrouded in mystery. According to the third maxim inscribed on a column at the temple:  Surety brings ruin or  ‘make a pledge and mischief is nigh.” As Octavia Butler says, it is possible for a cyborg to be seductive and erotic yet at the same time to exhibit non-gendered intelligence and thereby to subvert stereotypical categories."

 

Thank you again!

Reviewer 2 Report

Reading this abstract I find quite a few keywords and topics that resonate with me and that are currently quite relevant: text and visual transformer systems, NLP, mimesis, simulation and emulation in art and computational media, how prompt-based systems impact notions of creativity and artistic practice, the aesthetics of AI and ML art, artificial or computational creativity, role-playing and the performance of gender in service robots and other automata, etc. These topics are however bundled around the intention to discuss an art project, “Madeleine”, that is neither described sufficiently nor linked and that is therefore impossible to reference while assessing the abstract. Therefore, as presented, the abstract proposes a paper whose goals are unclear and that risks seeming too ambitious for its own sake, as adequately discussing any of the topics listed would probably require a full-length paper in itself. 

Author Response

I am sorry that even as early as the abstract I did not articulate the project sufficiently well. I have completely rewritten the paper beginning with the abstract in response both to your valuable criticism and that of another reviewer. There is, as you say, a danger of "over-reaching" with an article which is art-focussed but the project begs all the questions I tried to deal with. I am convinced that some of the text and image results contribute to the literature on machine understanding and intelligence. The following are the major areas of augmentation and clarification:

1. Here is the new abstract which aims at greater clarity and simpler, well-articulated goals.

"This article presents a project  which is an experiment in the emerging field of human-machine artistic collaboration. The author/artist investigates responses by the generative pre-trained transformer (GPT-2) to poetic and esoteric prompts and curates them with elements of digital art created by the text-to-image transformer DALL-E 2 using those same prompts. These elements are presented in the context of photographs featuring an anthropomorphic female avatar as messenger of the content. The tripartite ‘cyborg’ thus assembled is an artificial intelligence endowed with the human attributes of language, art and visage. It is referred to throughout as Madeleine.

The results of the experiments allowed the investigation of the following hypotheses. Firstly, evidence for a convergence of machine and human creativity and intelligence is provided by moderate degrees of lossy compression, error, ignorance and the lateral formulation of analogies more typical of GPT-2 than GPT-3. Secondly, the work provides new illustrations supporting research in the field of artificial intelligence that queries the definitions and boundaries of accepted categories such as cognition, intelligence, understanding and - at the limit - consciousness, suggesting that there is a paradigm shift away from questions like “Can machines think?” to those of immediate social and political relevance such as “How can you tell a machine from a human being?” and “Can we trust machines?” Finally, appearance and epistemic emotions: surprise, curiosity and confusion are influential in the human acceptance of machines as intelligent and trustworthy entities. The project problematises the contemporary proliferation of feminised avatars in the context of feminist critical literature and suggests that the  anthropomorphic avatar might echo the social and historical position of the Delphic oracle: the Pythia, rather than a disembodied search engine like Alexa."

2. Throughout the paper I focussed on augmenting discussion of the results of the GPT-2 and DALL-E2 experiments with the hypotheses - with particular evidence for "understanding' in both transformers. 

I integrated new research on the broadening definitions covering 'understanding', 'intelligence' and even 'machine consciousness' after Damasio thus:

"Recent research by Igor Aleksander (Aleksander, 2022) proposes that it is possible to conceive of attributes such as image memory and repositories of natural language (like  DALL-E2 and GPT) as prototypical faculties of ‘mind’. Following from Antonio Damasio (Damasio, 2009) he takes issue with Chalmers’ so-called ‘hard problem’ by assuming that a central neural network in a perceiving body is the seat of consciousness. However,                         “a key point is that the nervous system can learn to retain patterns of felt perceptions, even if the perceptual event is no longer there. This is due to the function of inner feedback within the neural networks, which causes the inner pattern to regenerate itself, even in the absence of the original stimulus, and the inner chemistry of the neurons to adjust itself to maintain this regeneration. We call this “memory”, and we call the felt patterns “mental states”. So conscious feelings are neural firing patterns in the form of images, or reflections of their bodily source, whether from outside the body or from within it.”  In experiments such as Madeleine we have already short-circuited the perceptual stage by endowing a machine with a huge integrated data set of language and images and with the future addition of sufficient memory of its own explorations of such states Aleksander  concludes that    “as life progresses, mind is developed through neural learning as a set of internal states, with links between the states that lead to one another, which forms a state structure for the state machine. As explained, this state structure is the equivalent of a mind in a brain.” With this “M-consciousness” it is arguable that the machine can think, albeit “in a machine way.” (op. cit.)

 

Alternatively, others like Hans Korteling conclude that human intelligence should not be the gold standard for general intelligence. Instead of aiming for human-like artificial general intelligence the pursuit of AGI should focus on specific aspects  of digital/silicon intelligence in conjunction with an optimal configuration and allocation of tasks – a set of processes more akin to the workings of the human unconscious rather than consciousness.  (Korteling, 2021 ) In a complementary argument Eugene Piletsky argues that conventionally, human consciousness  is supposedly a “superstructure” above the unconscious automatic processes, yet it is the unconscious that is the basis for the emotional and volitional manifestations of the human psyche and activity. Similarly, the mental activity of Artificial Intelligence may be both unnecessary and devoid of the evolutionary characteristics of the human mind. He goes a step further by suggesting that “it is the practical development of the machine unconscious that will ultimately lead us to radical changes in the philosophy of consciousness…” (Piletsky, 2019)"

 

3. I addressed the issue of the avatar's appearance and specified the intention of the tripartite cyborg in terms of feminist theory thus:

"It is in this sense that Madeleine is intended to occupy a position of subversion that echoes that of the Delphic oracle: the Pythia, rather than an Alexa. Not a ‘helper’ or an ‘assistant’ but an advisor, an inspirer, a partner.  The women who sat on the tripod were,  during certain periods at least, likely intelligent and well-educated, commensurate with such a position of trust. Their pronouncements, often revealed as prophecy, were interpretive rather than denotative and their appearance was shrouded in mystery. According to the third maxim inscribed on a column at the temple:  Surety brings ruin or  ‘make a pledge and mischief is nigh.” As Octavia Butler says, it is possible for a cyborg to be seductive and erotic yet at the same time to exhibit non-gendered intelligence and thereby to subvert stereotypical categories."

 

I thank you once again for your time. I sincerely hope I have been able to answer your criticisms.

Round 2

Reviewer 1 Report

Thanks for these revisions.

Reviewer 2 Report

Thank you for the revisions, I find that they much improved the quality of the manuscript and I'm now comfortable in recommending its acceptance and publication