If translation, as Walter Benjamin argues, reveals the ‘unfolding’ of a work’s latent possibilities, then AI-generated imagery serves as a distorted mirror — amplifying the gendered subtexts buried in Dichterliebe’s historical and linguistic fabric. By feeding Heine’s poems and our feminist interventions into an AI trained on Western art’s patriarchal visual lexicon, we expose the tensions between historical erasure and contemporary reclamation. The algorithm’s ‘hallucinations’ — grotesque, surreal, or darkly comic — are not random. They are collisions of the nineteenth century’s repressed anxieties (about female agency, desire, and suffering) with twenty-first-century tools that replicate, rather than resolve, those biases.
Like Christine Brückner’s Unspoken Speeches of Outraged Women, these images force visibility upon what Heine’s text elides: the bride’s silent grief (‘Marriage is a Trap’), the violence of romantic idealization (‘Codependence’), and the bodily horrors of motherhood (‘Gynecological Horrors’). The AI’s failures — literalized idioms, fused bodies, monstrous intimacies — become metaphors for translation’s imperfect but necessary act: to deform the original until its hidden truths crack open.
De-invisibilizing the female
The process of de-invisibilizing the female voice in Dichterliebe raised legitimate concerns. We opted to maintain a degree of historical plausibility when giving our second protagonist her own voice, and maintained the nineteenth-century historical setting which is inextricably intertwined with Schumann and Heine’s original work. Concurrently, we considered what has been termed the ‘Bridgerton effect’. Applied to describe a ‘growing pantheon of historical drama that peddles a vision of the past crudely shaped by present-day preoccupations’ (Strimpel 2021), the concern is that while visibilizing the marginalized in period pieces is ‘helping audiences become acclimatised to seeing more diversity in historical settings’, giving those characters an unrealistic amount of agency may also be numbing audiences to the historical difficulties and brutal realities for their marginalized group (Conlan 2022). Writing our poet’s lover into Heine and Schumann’s work is an attempt to reconcile these bundled concerns: de-invisibilizing her, but within a historically plausible setting which neither relishes nor erases her inherent challenges as a nineteenth-century woman.
Here, we found inspiration within a similar poetic model executed in literature, namely Christine Brückner’s 1983 Ungehaltene Reden ungehaltener Frauen [Unspoken Speeches of Outraged Women]. In it, Brückner provides female characters from history, both fictional and historical, including Christiane von Goethe, Katharina Luther, and Effi Briest with original, powerful monologues (Brückner 1983). Writing when the book was reissued and expanded, critic Klaus Ziermann contends:
This poetic concept has proven to be extremely productive and enduring, and although the author refrains from theoretical treatises and cultural-historical excursions, the book is among the most interesting, instructive, and enjoyable works written so far on the situation of ‘disgruntled women’ in world history. Clearly, only a woman — a writer of stature, a sensitive psychologist, and an acute observer of marital life — could have dared this original literary excursion into European history since ancient Greece and captured the specifically feminine in its many nuances. (1997, p. 58)
Interested in further opportunities to literally visibilize both protagonists, we turned to AI, feeding both Heine’s texts in translation and Babb-Nelsen’s original texts into the AI program Dream Time Art Generator with minimal historical-aesthetic prompts (i.e. ‘1800s’, ‘painting’).
The slideshow to the right illustrates the process and both accepted and rejected outputs. These begin with straightforward, but continue through increasingly surprising, responses. These included visualizations which range from the hilarious to the uncanny to the horrifying.
The most common and obvious stumbling blocks for AI art, at least when trained this way, appear to be idiomatic turns of phrase, which prompted overtly literal visualizations; ‘no grudge I bear’, references to ‘masks’ and ‘pedestals’ being the most fascinating. The tendency to blend unrelated artistic styles and producing complex facial features and hands were also common issues.
Post-human in no way means non-human. Dreamtime AI Art Generator, like most AI image-generating-systems, is trained on extensive datasets comprised of enormous swaths of pre-existing artwork and images, which have been tagged and identified. By employing deep learning techniques, the system ‘learns’ to discern relationships and identify patterns within the data until it gets to the point that the AI is capable of generating new outputs based on input prompts, which are likewise text-based (Clarke 2023).
AI does not inherently create new media, but rather re-synthesizes existing media, drawing from a corpus of data and using algorithms to match input keywords. The system will therefore largely regurgitate the same types of images, with all their gendered clichés and tropes, as have existed in art and language prior. AI art generators, trained on a dataset shaped by societal norms, including gender-normative biases, inevitably inherit these biases: ‘As a consequence of AI coding and algorithms, the next dilemma stems from bias and stereotypes,’ a report from the Center for Media Engagement at the Moody College of Communication confirms. ‘Ask DALL-E for a nurse, and it will produce women. Ask it for a lawyer, it will produce men [...] the algorithm likely can’t eliminate stereotypes on the user’s end’ (Parra and Stroud 2023: 7).
Data is not neutral; it reflects the biases embedded within the cultural and social contexts from which it originates, and that is hardly the sole ethical dilemma inherent in the use of AI. Not only have AI art generators been trained on copyrighted existing works of art without compensation or acknowledgment of the original artists, creating a ‘risky precedent for the creative community and industry’, but also threaten to ‘stifle the development of unique and inventive ideas, potentially jeopardizing future opportunities for painters, writers, content creators, and other artists in various lucrative industries’ (Steynberg 2024: 2; 6).
Perhaps because AI art generation ‘can contain flecks of elements from numerous human-created images, so many that it would be hard to determine what parts came from which piece,’ it serves as a Frankenstein-esque experiment generator, creating novel, life-like monsters from known entities (Townsend 2024: 3). This sometimes holds up a mirror, allowing new prisms, pastiches, and kaleidoscopes through which to visualize women, love and motherhood. While containing only instances of how they have classically been depicted in art, and therefore generally blindly reinforcing gender norms and stereotypes, things get most interesting when the AI ‘hallucinates,’ producing the incorrect, the unexpected, or the bizarre (Salvagno, Taccone, and Gerli 2023: 27). While these errors can be dangerous for those using AI for scientific purposes, for artists they can be interesting indeed, revealing how thin the veil between the idyllic and the nightmarish actually is, and perhaps bringing to light latent fears and touching on difficult truths.
Here is a gallery exhibition of some of our top-ranked Dreamtime AI hallucinations with respect to The Poet’s Love(r), including key text prompts:
These AI-generated ‘failures’ are the project’s most candid collaborators. They betray what human translators might soften: the violence underlying Dichterliebe’s romanticism. By forcing the algorithm to visualize marginalized perspectives, we expose not only the limits of AI (its reliance on biased datasets) but also the limits of the original work itself — its silences, its sublimated fears. In this sense, the machine’s distortions are a form of truth-telling, pushing translation beyond linguistic fidelity into the realm of cultural reckoning. To de-invisibilize the female in Dichterliebe is to confront the monstrous latent in the familiar.