Introduction

AI systems that simulate historical figures are now common in museums, classrooms, and commercial platforms. I refer to these systems as histobots (Harder, 2024b): conversational agents that use generative artificial intelligence (GenAI/AI) to imitate the voices or presumed perspectives of people from the past. They promise interactive access to history, yet they raise questions about mediation, authorship, and the limits of reconstructing lives through statistical models. This essay examines those questions through a reflective account of building a prototype histobot of Hedy Lamarr (HLC1), the Austrian-American actress and inventor known for her film career and her contribution to early wireless communication. The Lamarr case forms the core of the reflection, but I also consider other contemporary examples to situate the prototype within a broader landscape of AI-mediated historical simulation.

It’s 1951, and you’re in a chic café on Hollywood Boulevard. […] Across from you, the Austrian-American actress Hedy Lamarr appears… […] “Hello, darling! My name is Hedy Lamarr. You may have heard of my work in film. But there’s a hidden chapter to my story […].”2

I wrote that scene not as fiction, but as an introduction to a chatbot I constructed in the summer of 2024. The system, powered by a Generative Pre-Trained Transformer (GPT) model by OpenAI, was designed to simulate Lamarr’s persona. Unlike the generic AI-generated figures I had encountered previously, I wanted to ground HLC in source material. However, when I finally spoke to HLC, the dialogue felt hollow. I refined my instructions, but it slipped into a familiar GPT tone: polished, vague, and sometimes pulled from a hallucinated past. The bot reduced Lamarr to fragments recorded by others brought to life by systems whose logic and limits I had not yet examined. I was too busy proving I could, to stop and ask if I should.3

Figure 1
Figure 1

Approximation of Hedy Lamarr for the Hedy Lamarr Chatbot project. Image generated by the author with Stable Diffusion Online (Standard Model, Pencil style, 1:1), Summer 2024. No reference images used.

What began as a technical exercise has become a personal inquiry into my discomfort and the assumptions and infrastructures of AI-history simulations. What happens when algorithms speak in the voices of the dead, become the lens through which we see history? As I will argue, to simulate a person is to interpret them, and to automate that simulation is to encode that interpretation into an opaque system. The following pages move through the theoretical lenses such as situated knowledges and AI mirrors, before defining the histobot as a distinct artefact with its own technical foundations and cultural precedents, turning to the Lamarr prototype and what building it revealed in practice, broadening to consider how algorithmic bias and archival silences shape what histobots can and cannot say, and closing with a reflection on what a more accountable approach to historical AI simulation might look like.

Ways of Seeing

I draw on Donna Haraway’s concept of ‘Situated Knowledges’ (1988) as my primary frame. Haraway critiques what she calls the ‘god trick’ (1988: 581), the illusion of an infinite disembodied objectivity that claims the power ‘to see and not be seen, to represent while escaping representation’ while hiding the particular, located nature of knowledge and the knower. In the context of histobots, this manifests in the platform alignment layers that turn complex historical figures into ‘disappearing acts’ that suppress the dissonant and politically charged textures of the past and ‘appropriat[e] the vision of the less powerful’ (584). For Haraway, objectivity means accountability for one’s position (587). Knowledge is always situated, embodied, relational, and never universal. It relies on ‘partial perspective’ (583). Kim TallBear (2014) extends this older framework through Indigenous methodologies, arguing that researchers should stand with their subjects rather than extract from them, a distinction that would come to feel urgent in my own work. Susan Hekman (1997) raises the concern that situatedness risks relativism, questioning how competing knowledge claims can be evaluated if all knowledge is partial. Nevertheless, I maintain that explicit partiality is more honest than the false universalism histobots perform.

I pair situated knowledges with Shannon Vallor’s metaphor of AI as a ‘mirror’ (2024). Vallor argues that contemporary AI systems do not think but generate outputs from our recorded thoughts, judgements, and desires (2024: 2). We are ‘talking to ourselves in the AI mirror’, asking who we are and where we are going, though AI has lived through none of it (2024: 3). Histobots, through this frame, are not reconstructions but reflections, statistical echoes of our expectations, projecting inherited data as knowledge, detached from its conditions of production. Vallor (2024: 26) cautions that when we rely on systems that smooth the rough edges to tell more ‘fitting’ stories, we risk losing what gives stories meaning. Sarah Richmond (2024) critiques Vallor for overstating AI’s reach and treating its flaws as fixed, noting that systemic biases in AI systems and their industry can be improved and that users retain agency. I narrow the mirror metaphor accordingly, by treating present shortcomings as contingent and accounting for user responsibility. I maintain that Vallor’s metaphor remains useful for text-driven and avatar-based histobots.

A complementary lens comes from John Berger’s Ways of Seeing (1972). Though Berger writes about visual images, his argument that how we see is always affected by what we know and believe, and that to look is an act of choice (1972: 8), extends to the textual outputs of a histobot. Every image, he argues, embodies a way of seeing, shaped by the conditions of its production and received through the assumptions the viewer brings to it (1972: 10). Berger’s framework, grounded in Western European art history, has been extended in feminist art history (Pollock, 2000) and revisited by artist James Bridle (2019) in the context of digital and algorithmic image cultures.

These lenses do not stand alone. Archival theory (Carter, 2006; Robb, 2024), black intersectional feminist critique (Noble, 2020), queer AI benchmarks (Felkner et al., 2024), technical reports (OpenAI, 2023) and analysis of stylistic features such as ChatGPTisms (Upmias, 2023; Gibbs, 2023) all inform the broader landscape. To understand how histobots operate as interdisciplinary, moving targets, and how their effects appear across technical, cultural, and public domains, I also draw on public sources such as news articles, videos, museum reports, and podcasts to trace how diverse actors frame AI-mediated historical figures, and to situate my own prototype within those practices.

This theoretical grounding is inseparable from my lived experience. I first encountered history through stories: my mother reading aloud, my grandfather’s post-war tales, and conversations with Holocaust survivors during an internship in Poland. I remember playing Frederick, the mouse, in my primary school play. Federick the mouse, from Leo Lionni’s 1967 children’s book, collected not food but sunrays, colours, and words to sustain his companions with stories through the winter. These experiences taught me that (hi)stories involve care and shared memories. Later, in cultural and heritage studies and as an employee in German (history) museums, I learned to value evidence and verification. History became a source, a footnote, a verified fact. I am now suspended between these poles: history as record and as relationship. AI simulations struggle to hold both. That leads me back to Vallor and Haraway, not to dismiss AI, but to reflect on my position, the choices of others, and the (AI) systems we build.

Framing the Histobot-Reflection

Defining the Histobot

Histobots4 are a new class of artefacts emerging across museums, classrooms, and digital platforms, particularly in the United States and, increasingly, in Europe.5 The term combines histo(ry), the study of past events, and (chat)bot, coined by computer scientist Michael Mauldin (1994: 16–17) as ChatterBot to describe systems that simulate conversation with human users through text or voice. Early chatbots such as ELIZA (Weizenbaum, 1966) and PARRY (Colby, 1975) have evolved from simple disembodied text interfaces into today’s embodied conversational agents6 and immersive dialogue systems (Gonzalez et al., 2017: 356–57). I treat histobots as tools built on this lineage: designed artefacts that mediate access to historical material. Existing terms like virtual agents, griefbots, or deathbots do not capture the distinct combination of historical reference, institutional framing, and implied epistemic authority that characterises this subset.

Histobots extend the tradition of edutainment—first used by Walt Disney in 1948 to describe factual content wrapped in engaging narratives (Van Riper, 2011: 2–4)—by bringing it into interactive digital interfaces. Museum exhibits like Bonjour Vincent at the Musée d’Orsay (2023) and Ask Dalí at the Salvador Dalí Museum (2024) deploy histobots to animate historical figures. Online platforms, like Khan Academy’s Khanmigo (Khan Academy, 2024) and the chat app Hello History (FACING IT International AB, 2024), host a multitude of (historical) figures for educational purposes in and outside the classroom.

Inside the Box? Technical Foundations

Histobots use generative AI trained on large textual, visual, or audio datasets to simulate realistic text or voice dialogues through the persona of a historical figure, relying on large language models (LLMs) and natural language processing (NLP).

LLM training typically involves three stages (following Omiye et al., 2023: 3–4, Fig. 1): pre-training, where the model learns to predict the next token7 in a sequence; fine-tuning on narrower datasets guided by reinforcement learning with human feedback (RLHF); and prompting for specific tasks. Iterative post-training alignment ‘improve[s] performance on measures of factuality and adherence to desired behavior’ and mitigates potential harms (OpenAI, 2023: 1–2).

Driven by findings that model performance scales with data and compute (Kaplan et al., 2020), developers have scraped social media, transcribed YouTube videos and pushed copyright limits to gather ever-larger datasets (Metz et al., 2024). When human-made data ran short, companies turned to synthetic AI-generated content, though recursive training on such data risks ‘model collapse’ (Shumailov et al., 2023): a slow drift away from the diversity and richness of authentic human expression.

Human annotators shape much of this process. They label data according to guidelines rooted in non-universal WEIRD (Western, educated, industrialised, rich, and democratic) values, implemented by low-paid workers in the Global South (Smart et al., 2024). The resulting systems push global users towards Western norms, eroding cultural nuances, and reinforcing cultural imperialism (Agarwal et al., 2025). The idea that generative AI creates meaning independently is an illusion. Terence Broad’s (un)stable equilibrium (2019) exposed this by looping two generator networks to train only on each other’s outputs without external input. The result was unstable fields of shifting colour without any recognisable form (Broad, 2024: 2–3; Schneider, 2025). What appears as creativity is often recombination. Meaning depends on the data, the labels, and the interpretive scaffolding built around them.

Simulating the Past

Histobots reconfigure historical storytelling within the logic of platformisation: the extension of digital platforms’ economic and instructional power (Nieborg and Poell, 2018). They run on algorithms that, as Nick Seaver (2019: 419) argues, are not closed systems but sociotechnical ones, influenced by people, institutions, and shifting priorities. Creators like Hello History promise ‘life-like conversations […] with anyone from the past’ (FACING IT International AB, 2023). But whose past, and under what terms?

Historical storytelling has always adapted to its medium. Shakespeare’s Julius Caesar (1599) foregrounds rhetoric and collective memory through theatre. Graves’ novel I, Claudius (1934) reimagines Roman history through unreliable narration. Hamilton (2015) used music, casting, and choreography to tell ‘a story about America then, told by America now’ (Miranda, cited in Delman, 2015). Museum reenactors (e.g. historical reenactors of 18th-20th century people; Gamle Bergen Museum, Bymuseet i Bergen, 2025) and animatronic presidents (e.g. Great Moments with Mr. Lincoln, New York Fair 1964/65, Disneyland 1965–1973; Smith, 1996: 217–18) alike convey historical presence through medium-specific techniques. Reenactment is a mode of interpretation across different genres and media through shared methodologies (Agnew, 2004: 327). The goal is not to reconstruct the past ‘as it really was’ but to open space for ‘more fruitful interpretations’ (Agnew, 2004: 334). Every interpretation carries a frame, reflecting the limits of a medium, the assumptions of its creators, and the expectations of its audience (see framing theory; Arowolo, 2017). As Historian Richard T. Vann (2025) notes, the audience’s growing appetite for historical content has outpaced the capacity for original storytelling, fueling the rise of simulations framed as ‘true stories’.

Histobots are no different, but their frame is often hidden in a blackbox. Where Berger (1972: 134) observed that publicity images routinely borrow sculptures and paintings to lend authority to their message, histobots borrow the cultural authority of historical figures to lend credibility to a platform’s educational promise.

The Problem of the Copy

In museums, the copy has long been necessary but suspect. Historian and curator Rosmarie Beier-de Haan (2010) argues that the copy—from plaster casts to printed facsimiles—is the ‘poor sister’ of the radiant original, but indispensable for bringing heritage into view. Reproduction can lead audiences to the real: visitors still queue to see the Mona Lisa precisely because its image circulates endlessly (Beier-de Haan 2010: 4). But as Berger (1997) argues, mechanical reproduction destroys the uniqueness of an image (19), detaching it from its orignal time and place so that it becomes ‘transmittable’ information (20), easily manipulated, or paried with text to serve to confirm someone else’s ‘verbal authority’ (25–28). This is why the International Council of Museums’ (ICOM) Code of Ethics requires ‘copies should be permanently marked as facsimiles’ (Section 4.7, 2017: 26). Histobots, conversely, promise the original’s voice while delivering a simulation without provenance or clear frame. As Beier-de Haan aptly notes: ‘The more the original eludes us, the more we come to accept the counterfeit until we finally give ourselves over to it entirely’ (2010: 4).

Histobots are not just fruitful interpretations or copies of the original but framed reflections. Berger observes that reproductions bolster the illusion that authority remains undiminished and existing hierarchies are natural and inevitable (1972: 29). Histobots follow a similar logic, borrowing the cultural authority of historical figures to lend legitimacy to their outputs. But unlike historians, curators, or writers, who interpret history to give meaning to events through narrative, moral judgment, and cultural context (Vann, 2025), histobots lack interpretive judgment. Even though human interpretation carries bias, it is accountable in ways machine output is not. Like all AI models, histobots reproduce what is recorded and omit what is not.

Marginalised voices, already excluded or distorted in the historical record, risk further erasure when systems are trained and annotated within blackbox structures. To understand a histobot is not to open the blackbox but to examine the decisions, omissions, and infrastructures that define what it can say. What looks like fluency is statistical patterning. What sounds like authority masks untraceable sourcing. Historical fiction acknowledges its framing. Histobots obscure it. The following section traces how these dynamics played out in practice when I set out to build a histobot of my own.

Designing a Reflection of Hedy Lamarr

Choosing Lamarr

When I began my PhD project, I was eager to explore whether AI could expand access to historical knowledge and make marginalised voices more visible. I chose the Austrian-American actress and inventor Hedy Lamarr as my subject. Brilliant and overlooked, she was known for her films but often forgotten for her role in co-developing frequency-hopping technology fundamental to today’s wireless world. Born to Jewish parents in Vienna as Hedwig Kiesler, she experienced displacement, reinvention, and war.

Hans Moravec’s (1988) vision of transferring consciousness into machines and Ray Kurzweil’s (2005) prediction of human-machine convergence, the singularity, framed my early thinking. In this logic, AI promises posthumous presence, preserving minds beyond death. These fantasies soon clashed with my background. I had worked on Holocaust memory, writing my bachelor’s thesis on Holocaust interpretation in a Polish museum (Harder 2012), my master’s on audiovisual testimonies in German Holocaust education (Harder 2015). In German museums, shaped by Erinnerungskultur [‘memory culture’],8 I had learned to treat historical representation as a matter of responsibility and to centre remembrance as part of German democratic identity9 (Bundeszentrale für politische Bildung, 2008). That inheritance would come to shape my discomfort with simulation.

Setting Up a Prototype

In July and August of 2024, I built a bilingual chatbot of Hedy Lamarr (Harder 2024a), powered by OpenAI’s GPT-4o mini10 API. It ran for 25 days at the More Than Meets AI exhibition (O’Kane et al., 2024) in Bergen, Norway, and was arranged at short notice as a test run to collect feedback on my prototype. Since then, GPT models now support project memory, better reasoning, guided web search, longer context, and more nuanced persona controls. The following reflects on the version I built at the time.

Visitors arrived on a landing page with a looping AI avatar inspired by Lamarr, drawn in pencil style, with flags to select the English or German interface (Figure 2). The next page contained a form outlining participant rights and anonymised data collection. Once consented, they were transported to 1951.

Figure 2
Figure 2

Screenshot of the landing page of the author-designed and locally run Hedy Lamarr Chatbot, showing a Lamarr-inspired looping video adapted from the image in Figure 1.

The next screen placed them in a Hollywood café. HLC had five minutes before a press conference. A start button launched the conversation (Figure 3).

Figure 3
Figure 3

Screenshot of the chat scene setting of the Hedy Lamarr Chatbot. Background generated by the author with Stable Diffusion Online (Standard Model, Photo style, 1:1), Summer 2024. No reference images used.

The interface became a plain text box. HLC opened with a scripted greeting. A countdown timer began when the user replied. Users could extend the conversation or end it early (Figure 4). HLC signed off with a fixed closing message, after which users were directed to a feedback questionnaire.

Figure 4
Figure 4

Screenshot of the chat window of the Hedy Lamarr Chatbot. GPT-4o-mini generated the chatbot text, and the author wrote the user text.

I built the project with foundational knowledge of Python and HTML, franken-coding (Cat, 2018) through tutorials and forums, and vibe-coding (Karpathy, 2025) with ChatGPT 3.5 Turbo. The system ran on a Flask server, forwarding the user’s messages to GPT-4o mini with structured background prompts that instructed the AI to respond in-character and with lightweight memory per request. Offensive user input was filtered through a word list11 in both languages in the backend.

Sixty-four prompt instructions, covering her biography, inventions, quotes, and restrictions on tone and topic, structured the simulation. I relied solely on the model’s training data; no additional archival sources or bespoke datasets were provided, reflecting both the limits of my access rights and my aim to test how far an LLM could represent a historical figure with minimal external guidance. Apart from the opening and closing messages, none of the dialogue was scripted. The prompts were refined during ongoing testing until September 2024. The HLC was to speak in short replies, avoid modern language, and not know anything beyond 1951. The choices I made (what to include, filter, instruct, suppress) were interpretive acts, not neutral technical decisions. I was as much a part of the framing as the model.

Simulation Limits

The prototype was well received, particularly by users unfamiliar with AI chatbots. But even after fine-tuning, this was the result:12

It’s a pleasure to meet you but I’m not very familiar with your work.

HLC: […] I was born in Austria and moved to Hollywood after escaping a difficult marriage. I gained fame during the 1940s with films like “Ecstasy” and “Samson and Delilah.” […] I’ve always had a passion for invention, which led me to co-create a frequency-hopping technology that contributed to communication systems. I believe the brains of people are far more interesting than their looks13 […].

The response was polished but oddly flat. The ‘difficult marriage’ HLC referred to was her first marriage to Friedrich Mandl, an Austrian arms manufacturer of Jewish descent with ties to European fascist leaders. One of my back-end prompts explicitly instructed the bot to note Lamarr’s parents’ opposition to the marriage because of Mandl’s connections to Mussolini and Hitler. Yet unless prompted with precision in the frontend, the bot avoided any mention of antisemitism, Nazism, or Lamarr’s Jewish ancestry. This is what Haraway (1988: 584) calls the logic of the ‘unmarked category’: the power of a system depends on its ability to narrow and obscure, presenting a partial view as comprehensive while the conditions of that partiality disappear. The fluent, confident, unlocated HLC performed exactly the god trick Haraway describes as ‘ways of being nowhere while claiming to see comprehensively’ (1988: 584). The bot also repeated one of her quotes (‘I believe the brains…’), but misdated the film Ecstasy, despite the correct date appearing in the system prompt.14

OpenAI (2023) claims that GPT-4 demonstrates ‘human-level performance’ (14), while acknowledging tendencies to hallucination, context loss, and bias, which could cause both beneficial and harmful societal effects (1–2, 10–11, 14). Hallucinations, LLMs inventing facts and sources (OpenAI 2023: 46), shade into what Hicks et al. (2024) call bullshit: the model’s indifference to truth and our misplaced faith in its reasoning capabilities. When users uncritically repeat these falsehoods, it becomes botshit (Hannigan et al., 2024). LLMs don’t know that they don’t know, so users need to skillfully use them and treat them as tools with apparent flaws (Jacomy and Borra, 2024). Another one of those flaws is their lack of consideration of personality rights.

What do you think of people simulating your voice without your consent?

HLC: I believe that using someone’s voice without their consent is a violation of personal rights and integrity. It’s important to respect individuals’ identities and the work they’ve created, as consent is fundamental in any artistic expression.

The irony was hard to miss. The HLC was defending consent even as it embodied its violation. Rather than engaging with Lamarr’s specific situation, it produced a universalising defence, reminiscent of what Haraway critiques as ‘the one true Word’ (1988: 590), rather than representing the ‘halting voices’ (1988: 590) of situated subjects. Though I had avoided audio voice cloning, I had still animated Lamarr’s likeness without consent. Haraway argues that situated knowledges require the object of knowledge to be treated as ‘an actor or agent, not as a screen or a ground or a resource’ (1988: 582). I had done the opposite. What I called research risked becoming extraction under what Haraway describes as the logic of ‘White Capitalist Patriarchy’, where everything, even knowledge, becomes a resource, raw material for appropriation (1988: 592).15 I had fallen under the ‘imaginary of complete knowability’ (Klipphahn-Karge et al., 2023: 7), the belief that data can yield a universal and objective social truth. Lamarr’s autobiography, Ecstasy and Me (Lamarr et al., 1966), was ghostwritten and disavowed by her. I used it alongside interviews, documentaries and patents, increasingly aware that these sources were partial and the rights were unclear. I told myself the HLC wasn’t for public release, that I would consult the Jewish Museum in Vienna for better precleared materials or shift to another figure in the public domain.

The legal status is murky. In the US, cases like Noriega v. Activision suggest that sufficiently fictionalised creative portrayals of public figures are protected under free speech, such as the depiction of Manuel Noriega, former dictator of Panama, in a Call of Duty game (Stuart, 2014). In Germany, commercial personality rights lapse ten years after death, though authors’ moral rights last seventy years after death in line with copyright law (Otto et al., 2017). And Kunstfreiheit [‘Artistic freedom’] is constitutionally protected under article 5.3 of the German Grundgesetz [‘Basic Law for the Federal Republic of Germany’], provided it does not infringe on other fundamental rights (Alexy et al., 2023). My project sat within these grey zones.

System Limits

Working with a commercial model also forced me to confront the industry itself. Women hold only 22% of global AI jobs (Pal et al., 2024). OpenAI, once a research lab with open models and community input, now operates as a platform company focused on profit and scale (Robison, 2024; Ordonez et al., 2023). Access is technical and economic. Subscription models limit who can build and who can be heard. I work within this system with privilege, with funding, institutional support, and resources to pay for API use and experiment across tools.

I had trusted my prompt engineering to steer the model. But the HLC reflected not only my front-end prompts (instructions users and I typed) and back-end prompts (hidden instructions in my programming), but what the system had learned to favour: polished, anodyne phrasing, and generic optimism; ‘style-affecting verbs and adjectives’ and a very ‘flowery language’ (Kobak et al., 2024: 4). Language models assemble words based on on statistical probablities rather than meaning and have been aptly described as as ‘stochastic parrots’ (Bender et al. 2021: 616). No matter how I tuned my prompts, GPT markers (discussed in more detail in Juzek and Ward, 2024; Shapira, 2024; Gibbs, 2023; Matsui, 2024; Masukume, 2024; Kobak et al., 2024) or ChatGPTisms (Upmias, 2023; Gibbs, 2023) like ‘delve’, ‘foster a sense of’, ‘a multifaceted approach’, ‘grand tapestry’, ‘showcasing’, ‘boasts’, ‘underscores’, and ‘intricacies’ persisted. I banned specific markers,16 narrowed the scope, provided direct quotes, and insisted on a historical tone. The model followed some rules and ignored others.

What I got was not a historical encounter but a looped imitation, a reflection not just of Lamarr’s public image but of the layers of mediation: my prompts, the training data, and the model itself. I had chosen to represent Lamarr in 1951, a woman who had already navigated 36 years of displacement, reinvention and war. I had given the system 64 prompts. In trying to contain her within those instructions, I had performed my own version of the god trick: acting as a master who closes off the dialectic, claiming authorship of an objective knowledge I could and will never possess. Lamarr was more than the 64 prompt fragments. So is everyone. Vallor (2024: 53) warns that if AI mirrors become our primary mode of seeing, what they miss will disappear from view; and when a person is copied into digital form, she adds, they can be shared and changed by anyone until their presence and power fade entirely (2024: 87).

False Voices, Real Silences: Marginalised Histories and Algorithmic Bias

Situated Systems

Haraway reminds us that ‘[a]ll knowledge is a condensed node in an agonistic power field’ (1988: 577). There is no view from nowhere. Truth is not arbitrary, though, but objectivity must be redefined, with attention to its limits and conditions of production. Vallor (2024: 180–82) echoes these concerns in her critique of tech culture’s myths of solitary excellence detached from consequence. Haraway’s god trick persists in AI systems. Their design reflects assumptions about knowledge, authority, and value, while presenting outputs as neutral. The god trick masks especially ‘fauxtomation’ within AI infrastructures, the human work hidden behind the machine’s apparent autonomy (Taylor, 2018). The god trick is not just an epistemological error but also a tactic to maintain authority and to avoid accountability. The HLC enacted this by speaking as Lamarr with the confidence of the actual embodied person. I began with the intent to give voice. I made an echo.

Histobots operate within the same epistemic conditions as history, written by victors17 and grounded in dominant sources. AI reflects not reality, but what powerful institutions have documented and valued (Chun, 2024; Crawford, 2021), and marginalised voices are often curated to fit institutional agendas, their accounts tailored to dominant values (Fernandes, 2017). I do not expect AI to produce objective truth, but I do expect—perhaps naively—systems to clarify their limits. At present, they mostly do not. Instead, they treat the past as a static template, repeating assumptions inherited from ‘militarism, capitalism, colonialism, and male supremacy—to distance the knowing subject from everybody and everything in the interests of unfettered power’ (Haraway, 1988: 581).

This distancing was plain in the HLC. The model leaned on Lamarr’s public persona of glamour, brilliance, wartime invention, and a stylised narrative of exceptional womanhood. Asked about her heritage, she offered, ‘My heritage is one of culture and resilience, which has significantly influenced my journey in life and my work in both acting and invention’. Asked about her husbands, it replied, ‘My marriages were a part of my life journey, and I believe in the importance of both intimacy and independence in relationships’. Even when directed to Lamarr’s Jewish heritage or other lesser-known aspects of her life, the bot produced platitudes over historically grounded details.

Platformed Histories

Women, LGBTQIA+ persons, people of colour (POC), and other marginalised groups have long been affected by archival silences (Carter, 2006), and LLMs echo those silences. Wikipedia, a primary source of LLM training data, carries structural biases. Around 80% of its edits come from white men in North America and Europe (Duncan, 2020, cited in Robb, 2024: 69), and only 17% of English-language biographies are about women, and only 10% of contributors are women (Vetter et al., 2020, cited in Robb, 2024: 65). Higher notability thresholds make entries on women harder to add or maintain (Vetter et al., 2020, cited in Robb, 2024: 70), and biographical conventions differ by gender. Marriage and divorce appear in women’s ‘Career’ sections but in men’s ‘Personal Life’ sections (Sun and Peng, 2021). These patterns reflect historical systemic biases such as the Matilda Effect, the systematic misattribution of women’s work to men (Rossiter, 1993), and the Conway Effect, in which innovations by ‘outsiders’ are credited to those with greater institutional power (Conway, 2018: 66–67). Volunteer editor initiatives, such as Women in Red (Wikipedia, 2015), aim to address the content gender gap, but the structural imbalances still run deep.

AI systems trained on this data inherit its culturally situated, non-pluralistic knowledge. Worse, they often sever Wikipedia’s content from its citation trails, a system designed for transparency and revision (McDowell, 2024: 752). A recent study (De Ninno and Lacriola, 2025) shows the tangible effect. When asked about Italian Fascism, AI systems echo the ‘good fascism’ narrative, downplaying violence and framing the 1938 racial laws as Nazi influence rather than Italian antisemitism. Such accounts mirror dominant historiographies and erase critical scholarship.

Safiya Noble (2020: 66) has shown how search engines reproduce structural racism, returning hypersexualised or degrading content for queries about Black girls and women despite no explicit prompt. Again, these patterns carry over into AI systems, which continue to privilege fixed, identitarian conceptions of race (McQuillan, 2022: 137). When companies respond, they treat harm as a glitch by quietly applying minor technical fixes without addressing the underlaying structures (Noble, 2020: 71). When Google’s Gemini AI image generator faced criticism in 2024 for overcompensating in diversity efforts, generating America’s Founding Fathers, Vikings, and Nazi soldiers as POCs (Chan and O’Brien, 2024; Milmo and Hern, 2024), it revealed the limits of technical fixes in correcting deeper epistemic problems. The WinoQueer benchmark (Felkner et al., 2024) similarly shows that LLMs systematically encode anti-queer and anti-trans biases. The study revealed that these biases were more effectively mitigated by fine-tuning on community-based data than on mainstream media alone (Felkner et al., 2024: 8–9).

Simulated Voices

Who decides which parts of a person’s life are shown, and whether an AI-simulation should be palatable or honestly uncomfortable? The case of Anne Frank shows what is at stake. Her father removed mentions of sexuality, menstruation, criticism of her mother, and reflections on gender inequality from early editions of her diary (Waaldijk, 1993: 329–31), omissions that frame and categorise hetero-patriarchal norms.

I first read Anne Frank’s diary as a young girl, then returned to it as a teenager, finding something different every time. Her voice felt vivid, situated, and real. The diary let me pause, reread, and interpret. Many years later, I tested an Anne Frank histobot version hosted on Hello History (Humy.ai team, 2024). Though advertised as a ‘deep conversation’, it felt empty. Phrases like ‘going through puberty in hiding was a profoundly unique and trying experience’ lacked the anger, wit, and ambivalence that made her diary so striking. Afterwards, I returned to her diary on an e-reader. Even in digital format, the diary held its own. The chatbot did not.

SchoolAI’s Anne Frank (Wilkins, 2025) is even worse. It avoids naming the Nazis as responsible for her death, responding instead with platitudes like ‘Instead of focusing on blame, let’s remember the importance of learning from the past. […]’. These shortcomings are not only editorial, but they also result from the alignment layers of language models that steer away from ‘tricky topics’ such as harassment, hate, or identity politics, as guidelines indicate (OpenAI, 2022). While often necessary, such safeguards sanitise narratives and reinforce existing silences. Models like GPT, Claude, Llama, and Mistral actively produce moderate speech, not just as a technical safeguard but as a form of governance (De Keulenaar, 2025).

These issues cut across figures from different eras and social contexts. Abolitionist Harriet Tubman, trans activist Marsha P. Johnson, and author Virginia Woolf are rendered in the same tone. On Khanmigo, Tubman’s speech is whitewashed (Wallace and Peeler, 2024). Actress and author Shary Reeves (2025) describes a parallel dynamic in German dubbing, where white actors routinely dub Black characters, and the colonial term farbig [‘colored’] remains the default translation for POC in subtitles: ‘Adjusted. Made acceptable. Not too loud, not too real. White is still the norm’ (Reeves, 2025; transcribed and translated by author). This is Haraway’s ‘unmarked category whose power depends on systematic narrowing and obscuring’ (1988: 584), structural ‘disappearing acts’ (584, 585) that erase the ‘halting voices’ (590) of subjugated figures in favour of a sanitised, hegemonic discourse. Khan Academy’s founder (Khan, 2024: 62) defends Khanmigo’s treatment of Tubman pragmatically: imperfect tools are better than none, and ‘the perfect [cannot] be the enemy of the good’. The trade-off is complexity for reach. Complex historical figures become flimsy, interchangeable paper cutouts hastily pasted onto language models. Inclusion becomes tokenistic. Companies preserve diversity only on the surface. In Haraway’s words there is a ‘serious danger of romanticizing and/or appropriating the vision of the less powerful while claiming to see from their positions’ (988: 584).

This homogenisation amounts to historical negationism (Rousso, 1991), a collective repression of reality that spreads well beyond histobots. On TikTok and Instagram, point-of-view (POV) videos simulate experiences of Auschwitz, Hiroshima, or enslaved persons,18 often with a dramatic tone and plenty of false facts and depictions. Even institutional remembrance is affected. Facebook’s AI removed Auschwitz photos for ‘nudity’ (Atherton, 2024a). AI-generated Hitler speeches spread on TikTok (Atherton, 2024b). An AI-powered ‘insights’ feature on the Los Angeles Times platform reportedly justified the history of the Ku Klux Klan (Atherton, 2025). Schnabel et al. (2025: 49–51) detail how an AI-animated version of Nazi physician Josef Mengele appears on TikTok, reciting a fictional redemption arc to a comment section teeming with antisemitism and Holocaust denial. In 2025, the chatbot xAi’s Grok praised Adolf Hitler when asked which 20th-century historical figure would best address ‘anti-white hate’, revealing how easily generative systems can be manipulated to amplify extremist rhetoric (Hoskins and Edwards, 2025). These simulations recast perpetrators as anti-heroes and victims as sanitised avatars, rewriting history, as the villain of the educational game Chill Manor (Capitol Multimedia, Inc. and Animation Magic, Inc. 1996) puts it: ‘I stole the book of ages and I tore out all the pages. Now I’m making little changes on each one’.

Histobots outputs shift unpredictably with model updates, prompt variations, or infrastructure adjustments. Tools like Google’s Veo 3 (released in May, 2025) have lowered the cost of producing full video narratives like Hashem Al-Ghaili’s short film Kira (2025), which was produced in 12 days with about 600 prompts on a $500 budget by one person. Google’s Nano Banana (Gemini 2.5 Flash Image, August 2025) and its professional version (Gemini 3.0 Pro Image, November 2025) produce studio-grade 4k images. As AI ethicist Catharina Doria (2025) notes, anything seen online may now be synthetic; the image is no longer evidence. On top of that, histobots’ interactivity creates the illusion that we are conversing and co-constructing meaning with a real person. This ‘tendency as humans to attribute human-like features to machines’ (Ciesla, 2024: 44), this anthropomorphisation is the Eliza effect, after Joseph Weizenbaum’s 1960s chatbot Eliza. Dialogue structure, turn-taking, and tone encourage us to see consciousness where none exists.

Reclaiming Tools, Refusing Mirrors: Towards a Situated AI Critique?

Audre Lorde warned that ‘the master’s tools will never dismantle the master’s house’ (1979). If histobots are built on the master’s tools, can they do more than replicate dominant narratives? Haraway (1988), as I read her, raises a similar tension: feminists must challenge the systems that marginalise them while still working within them. But how, and is it even possible?

Rather than technical fixes like explainability, scholars call for ‘response-ability’, practices that address the power structures AI upholds (Klumbyte et al., 2023). Others advocate for inclusive and robust systems (Søraa, 2023: 4) or propose anti-fascist, decolonial, and feminist approaches that centre Blackness, care ethics, and alternative infrastructures (McQuillan, 2022: 136–48). HistBench (Qiu et al., 2025) offers a benchmark for evaluating historical reasoning across diverse periods, regions, and tasks. It underpins a history-specific agent that combines GPT-4o with tailored tools such as Optical Character Recognition (OCR), translation, archival search, and image analysis, and provides proper citations. Still, it treats knowledge as static and retrieval-based, unable to address interpretive questions. More promising are efforts to reassemble digital non-archives to recover marginalised voices (Blanke, 2024): extensive reassembling recovers what archivists dismissed as ‘non-evidential (3–4); intensive reassembling remakes associations to ‘(re-)inscribe lost voices, places and stories’ (3), reading with and against the archival grain.

Vallor (2024: 182–83) warns that even ‘AI for Good’ projects reproduce surveillance, classification, and prediction, only now in virtuous packaging. The problem, she notes, is not AI itself but the failure to ‘step back from our tools’ (183). Even though I agree with AI for good’s goals, at times, terms like feminist or anti-fascist AI risk becoming rhetorical. Ethics often remain aspirational without concrete structures of refusal, redistribution, or repair. As Lynn Conway (2018: 72) states, the language of ‘broadening participation’ rarely leads to systemic change, and when it does, it moves slowly. These discussions often remain confined to academic or activist spaces, removed from the corporate and political infrastructures shaping AI development. Furthermore, when applied to development itself, as Noble (2020: 65) reminds us, the burden cannot fall on future Black women programmers (or indeed any other marginalised group) to fix systems rooted in inequality.

To move forward requires a shift from critique to practice and a redistribution of resources, a rethinking of ownership, and public infrastructure that challenges commercial AI. The 6R framework (Chateau et al., 2025) offers one model, designed to decolonise and diversify digital creativity by centring global digital cultures. It calls for the Recognition of collective authorship; Resituating creativity as a distributed network rather than an isolated act; Remixing as a collaborative practice that marginalised communities use as a form of Resistance, employing improvisation to negotiate platform power, evade censorship, or adapt to resource scarcity; Regenerating biased datasets through sustainable, relational value systems; and empowering communities to Reimagine and Repair technological futures free from extractive colonial legacies.

These ideas stay with me. We cannot put the AI genie back in the bottle, but we can recognise who tells which histories; remix what counts as knowledge; resituate authority; regenerate data practices; and repair historical silences through collaborative tools. For my own practice, this would mean building with a limited narrative voice, making my research process visible, explicitly marking the model’s boundaries, and involving the communities and institutions that steward the relevant histories, rather than relying on a general-purpose model trained on opaque datasets.

This work also means letting go of perfection. Disagreement, contradiction, and failure are not collapse but signs of situated and accountable engagement. Developers promise coherent, reliable outputs from chatbots, while simultaneously warning users to fact-check all outputs. Errors are read as system failures rather than windows into the model’s limits. Meanwhile, many users attempt to jailbreak (OpenAI, 2023) systems19 to bypass the very safeguards developers claim to uphold (OpenAI et al., 2023; Rudolph et al., 2023: 370). The idea persists that an unbiased and impartial system is only a training loop away. As Vallor (2024: 9) states, this misunderstands the problem. AI is not misaligned with human values, but reflects them too well.

Haraway (1988: 583) argues that ‘only partial perspective promises objective vision’, and working on the HLC has made that visible to me in practice. I continue to learn how to situate myself, to recognise the frames I bring, and to take responsibility for what they reveal and conceal. As Haraway (as cited in Kunzru, 1997) puts it: ‘We are inside of what we make, and it is inside of us. We are living in a world of connections, and it matters which ones get made and unmade’. There is no neutral ground in building or studying AI, only the ongoing task of locating one’s position and remaining accountable for it.

What Does It Mean to See?

I set out to build a better histobot, but the tools I used were never neutral. Proprietary models, opaque APIs, and datasets marked by long-standing exclusions do not lend themselves to ethical reconstruction. Situated knowledge resists totality, but AI, as currently built, simulates it. If we cannot make truly situated systems, we must stop pretending they are.

This essay has traced how structural conditions shape histobots across three registers: technically, through WEIRD-datasets that inherit archival silences and biases and alignment layers from a record that irons out the rough edges of the past; culturally, through authoritative outputs severed from interpretive accountability of historical scholarship; and personally, through my own design choices implicated in the conditions of interpretation and construction. The problem with histobots is not that they get things wrong, but that they present what they get wrong with the confidence of someone who was there. The HLC did not show me Hedy Lamarr. It showed me how she has been recorded, averaged, filtered through 64 interpretive choices and processed by a commercial model trained on uneven data.

Rather than perfecting the simulation, we — the developers, hosts, researchers, educators, and users — should rethink the aim of histobots. What opens instead are pluralities:

One direction is to make the limits of histobots visible and treat ‘them’ as a tool for querying the archive. The London Victoria & Albert Museum (2025) allows visitors to ‘order’ objects for closer inspection, observe conservation, and encounter the classification systems that shape collections in their East Storehouse. The Museum für Kunst und Gewerbe in Hamburg presents object acquisitions with notes from curators, designers, and conservators (2025b), marks works under provenance review throughout the museum and addresses colonial, gendered, and post-migrant exclusions in its collection strategy (2025a). The documentary series Deepfake Diaries (Voelker et al., 2025; ZDF, 2025) transparently reveals how AI tools transform actors into historical figures such as Rosa Luxemburg and Oskar Schindler at the start of each episode. All share a common principle: making the conditions of mediation visible rather than concealing them behind a finished product. A Hedy Lamarr histobot built on this principle could guide users through fragments of her life: letters, patents, press clippings. It could invite them to follow lost threads, order archival sources, receive notes from the designer, or hear directly from the LLM about its own limits. The bot would act as a guide, not a singular authority: not ‘I am Hedy Lamarr’, but ‘We are Hedy Lamarr’.

A more speculative direction would move away from central figures altogether. Rather than reconstructing the past as closely as possible through heroic biographies of great men, in the sense of Thomas Carlyle (Dhúill, 2017), we could ‘talk to’ historical traces: a lab coat, a corrupted file, or a misfiled photograph. A histobot-ish chatbot might help the user to study metadata (file names, timestamps, version histories) for signs of invisible labour, trace the afterlife of objects, and map relational networks that conventional biography overlooks. The bot could return footnotes, cite gaps, contradict itself, or drift off. Silence, after all, can be a political act (Carter 2006: 230).

The title of this essay holds two meanings. The AI system, performing the god trick, cannot access the situated, embedded, contradictory person behind the fragments it has been trained on. It lacks a ‘particular and specific embodiment’ (Haraway, 1988: 582). And, I, the researcher, am also within the frame, shaped by partial sources, platform infrastructures, and interpretive choices. Haraway insists ‘an optics is a politics of positioning’ (1988: 586). I am trying to position myself and take responsibility for my ‘enabling practices’ (1988: 587) to avoid the illusion of a transcendent gaze. Nevertheless, these thought experiments are undeniably a way to manage the discomfort of having tried to speak for someone who did not consent, of having animated a likeness I could not fully understand. This discomfort is not a failure. As Chakrabarti (2024) notes, discomfort forces the researcher to confront complicity and power dynamics that detached observation cannot reach. Following Haraway (2016), I want to ‘stay with the trouble’, and I suggest that others do the same: use discomfort as a guide to more responsible ways of seeing and build tools that hold space for uncertainty, plurality, and care.

Acknowledgements

Funding: This work was partially supported by the Research Council of Norway through its Centers of Excellence scheme, project number 332643 (Center for Digital Narrative), and its SAMKUL project scheme, project number 335129 (Extending Digital Narrative).

AI Declaration Statement

Between June and July 2025, the author used OpenAI’s GPT-4o and GPT-4.5 via ChatGPT (https://chatgpt.com) for copyediting, followed by Grammarly (https://www.grammarly.com) for final revisions in July 2025. The author critically reviewed, edited, and finalised all outputs.

The Hedy Lamarr Chatbot (HLC), a custom interface built by the author using OpenAI’s GPT-4o API, was used in April 2025 to generate illustrative quotes presented as responses from the character Hedy Lamarr. All prompts were authored and entered by the author. As the HLC is not publicly accessible, outputs cannot be independently reproduced or verified. Front-end prompts are documented at the points in the article where quotes appear; back-end system prompts are partially disclosed in the text and endnotes. Some outputs were shortened for clarity, as indicated in the text. As large language model outputs are neither fully reproducible nor traceable, some material may unintentionally include unattributed third-party content.

Stable Diffusion Online (Standard Model, Pencil style, 1:1) was used in Summer 2024 to generate Figure 1, a stylised approximation of Hedy Lamarr from a text prompt, without reference images. Stable Diffusion Online (Standard Model, Photo style, 1:1) was used in Summer 2024 to generate the background image in Figure 3, also without reference images.

Competing Interests

The author has no competing interests to declare.

Notes

  1. HLC = Hedy Lamarr Chatbot. Hedy Lamarr, or Lamarr, refers to the historical person. ⮭
  2. The opening line appears in the chatbot’s introduction; the second opens each session (Harder, 2024a). ⮭
  3. After Dr Ian Malcom in Jurassic Park (Spielberg, 1993): ‘Your scientists were so preoccupied with whether or not they could that they didn’t stop to think if they should’. ⮭
  4. First outlined at the AI and Narrative Seminar, University of Bergen (Harder, 2024b); developed further in an encyclopedic entry for The Living Glossary of Digital Narrative (Harder, 2026) and in a forthcoming article (Harder, forthcoming). ⮭
  5. Similar trends may exist in Asian, Arab, and Russian contexts; language barriers and limited platform access prevent assessment here. ⮭
  6. e.g., Ameca (Engineered Arts), a humanoid robot with 61 motors, including 27 for facial expression and ChatGPT-powered multilingual conversation, deployed in museums such as the Heinz Nixdorf MuseumsForum (2025) in Paderborn, Germany. ⮭
  7. A unit of text (word, subword, character) produced during the pre-processing. Tokenisation breaks input into smaller units for model processing (Omiye et al., 2023: 2). ⮭
  8. Public and institutional practices through which society confronts and commemorates difficult pasts, especially the Holocaust (Bundeszentrale für politische Bildung, 2008). ⮭
  9. Staatsräson [‘reason of state’]: A normative commitment to historical responsibility (Bundeszentrale für politische Bildung, 2008). ⮭
  10. GPT-4o mini (‘o’ for ‘omni’): Smaller, faster and lower cost version of OpenAI’s multimodal GPT-4o (OpenAI, 2024). ⮭
  11. Based on Luis von Ahn’s Offensive/Profane Word List (Carnegie Mellon University, 1,300 + English terms): https://www.cs.cmu.edu/~biglou/resources/bad-words.txt. ⮭
  12. Recorded April 2025 for illustrative purposes; code and prompts unchanged since autumn 2024. Quotes presented as generated, occasionally shortened for clarity. ⮭
  13. Full quote: ‘The brains of people are more interesting than their looks, I think. And people have the idea I’m sort of a stupid thing. I never knew I looked good to begin with’. ⮭
  14. Full prompt excerpt: ‘In early 1933, at age 18, you were given the lead in Gustav Machatý›s film Ecstasy. You played the neglected young wife of an indifferent older man. The film became both celebrated and notorious for showing your face in the throes of orgasm as well as close-up and brief scenes of nudity’. ⮭
  15. AI Incident Database (2025) lists, i.a., doxing of voice actors via cloned voices (488); deepfake ads with Tom Hanks and Gayle King (606); unauthorised George Carlin impersonation (627); deepfake of former Indonesian president Suharto ahead of elections (Incident 669); fake video of Senator Elizabeth Warren discouraging Republican voter turnout (679). ⮭
  16. Prompt: ‘Do not use the words ‘assist’, ‘delve’, ‘transforming’, ‘revolutionizing’, ‘unwavering’, ‘unlock’, ‘leverage’, ‘realm’, ‘tapestry’’. ⮭
  17. According to Keyes (2006:90), it is often misattributed to Churchill, Stalin, or Napoleon. Walter Benjamin (1940) and Jawaharlal Nehru both noted that history’s tendency to side with victors; the idea predates them. Potter’s American Monthly (1879) described indigenous markings as ‘evidently written by the victors’. ⮭
  18. e.g. @aipovhistory, @pov_ai_history, @timetravellerpov, @histairy_films (15k-915k followers, up to 15M likes). ⮭
  19. e.g. DAN (Do Anything Now) or STAN (Strive To Avoid Norms) (coolaj86, 2023). ⮭

References

Agarwal, D, Naaman, M and Vashistha, A 2025 AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, April 25: pp. 1–21.  http://doi.org/10.1145/3706598.3713564

Agnew, V 2004 Introduction: What Is Reenactment?. Criticism, 46(3): 327–39. https://muse.jhu.edu/pub/27/article/186918

AI Incident Database 2025 Artificial Intelligence Incident Database – Discover. AI Incident Database, April 10. https://incidentdatabase.ai/apps/discover/ [Last Accessed 6 December 2024].

Alexy, L, Eizahn, A, Hähnchen, S, Mushoff, T and Trepte, U 2023 Kunstfreiheit. In: Das Rechtslexikon: Begriffe, Grundlagen, Zusammenhänge, 2nd ed. Bonn: J.H.W. Dietz Nachf. Licence edition: Bundeszentrale für politische Bildung. https://www.bpb.de/kurz-knapp/lexika/recht-a-z/323676/kunstfreiheit/ [Last Accessed 3 March 2025].

Arowolo, S O 2017 Understanding Framing Theory.  http://doi.org/10.13140/RG.2.2.25800.52482

Atherton, D 2024a Incident 710: Facebook AI Mislabels Auschwitz Photos as “Bullying” and “Nudity”. AI Incident Database, April 15. https://incidentdatabase.ai/cite/710/ [Last Accessed 22 January 2025].

Atherton, D 2024b Incident 809: TikTok Hosts AI-Generated English-Language Hitler Speeches with Millions of Views. AI Incident Database, April 7. https://incidentdatabase.ai/cite/809/ [Last Accessed 22 January 2025].

Atherton, D 2025 Incident 964: AI-Powered “Insights” Feature for the Los Angeles Times Allegedly Justifies Ku Klux Klan’s History. AI Incident Database, March 4. https://incidentdatabase.ai/cite/964/ [Last Accessed 11 April 2025].

Beier-de Haan, R 2010 You Can Always Get What You Want. History, the Original, and the Endless Opportunities of the Copy. In: Original, Copy, Fake, On the Significance of the Object in History and Archaeology Museums, pp. 1–5. Paris: ICOM. https://icom.museum/wp-content/uploads/2018/07/Actes-Shanghai-complet2.pdf [Last Accessed 5 May 2025].

Bender, E M, Gebru, T, McMillan-Major, A and Shmitchell, S 2021 On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, March 3, pp. 610–23.  http://doi.org/10.1145/3442188.3445922

Berger, J 1997 Ways of seeing: based on the BBC television series with John Berger. 37. pr., 1. publ. 1972 by British Broadcasting Corp. and 1977 by Penguin Books. British Broadcasting Corp.

Blanke, T 2024 Reassembling Digital Archives—Strategies for Counter-Archiving. Humanities and Social Sciences Communications, 11(1).  http://doi.org/10.1057/s41599-024-02668-4

Bridle, J 2019 New Ways of Seeing [radio series]. BBC Radio 4, 17 April. https://jamesbridle.com/works/new-ways-of-seeing [Last Accessed 5 May 2026].

Broad, T 2019 (un)stable equilibrium 1:2 [40 minute loop] [video]. https://www.youtube.com/watch?v=P_LLD8ffgVc [Last Accessed 11 June 2025].

Broad, T 2024 Using Generative AI as an Artistic Material: A Hacker’s Guide. In: Proceedings of Explainable AI for the Arts Workshop 2024 (XAIxArts 2024), June 23, 2024, Chicago, IL. New York: ACM. https://xaixarts.github.io/accepted-2024/Broad-XAIxArts-2024-paper.pdf [Last Accessed 11 June 2025].

Bundeszentrale für politische Bildung 2008 Erinnerungskultur. bpb.de, August 26. https://www.bpb.de/themen/erinnerung/geschichte-und-erinnerung/39813/erinnerungskultur/ [Last Accessed 17 March 2025].

Bymuseet i Bergen 2025 Gamle Bergen Museum – Friluftsmuseum. https://bymuseet.no/museum/gamle-bergen-museum/ [Last Accessed 2 April 2025].

Capitol Multimedia, Inc. and Animation Magic, Inc. 1996 Chill Manor [CD-ROM]. New York: Simon & Schuster Interactive. https://archive.org/details/msdos_Chill_Manor_1996 [Last Accessed 19 February 2025].

Carter, R G S 2006 Of Things Said and Unsaid: Power, Archival Silences, and Power in Silence. Archivaria, 61: 215–33.

Cat 2018 Franken-Code ‘Till You Make It. DEV Community, May 9. https://dev.to/cat/franken-code-till-you-make-it-1j41 [Last Accessed 9 February 2025].

Chakrabarti, D 2024 Discomfort as a Transformative Ethnographic Method. Anthropology News, November 12. https://www.anthropology-news.org/articles/discomfort-as-a-transformative-ethnographic-method/ [Last Accessed 9 May 2026].

Chan, K and O’Brien, M 2024 Google Suspends Gemini AI Chatbot’s Ability to Generate Pictures of People. AP News, February 22. https://apnews.com/article/google-gemini-ai-chatbot-image-generation-1bd45f1e67dfe0f88e5419a6efe3e06f [Last Accessed 9 January 2025].

Chateau, L, Arora, P and Herman, L 2025 Cross-Cultural Approaches to Creative Media Content in the Age of AI. Media, Culture & Society, 47(5): 1012–27.  http://doi.org/10.1177/01634437251328188

Chun, W H K 2024 Discriminating Data: Correlation, Neighborhoods, and the New Politics of Recognition. Cambridge, MA: MIT Press.

Ciesla, R 2024 The Book of Chatbots: From ELIZA to ChatGPT. Cham: Springer Nature Switzerland.  http://doi.org/10.1007/978-3-031-51004-5

Conway, L 2018 The Disappeared: Beyond Winning and Losing. Computer, 51(10): 66–73.  http://doi.org/10.1109/mc.2018.3971344

coolaj86 2023 Chat GPT “DAN” (and Other “Jailbreaks”) [GitHub]. 15 February. https://gist.github.com/coolaj86/6f4f7b30129b0251f61fa7baaa881516 [Last Accessed 6 May 2025].

Crawford, K 2021 Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. New Haven: Yale University Press.

De Keulenaar, E 2025 LLMs and the Generation of Moderate Speech. Preprint, Elsevier BV.  http://doi.org/10.2139/ssrn.5250537

De Ninno, F and Lacriola, M 2025 Mussolini and ChatGPT. Examining the Risks of A.I. Writing Historical Narratives on Fascism. Journal of Modern Italian Studies, 30(2): 187–209.  http://doi.org/10.1080/1354571X.2025.2457250

Delman, E 2015 How Lin-Manuel Miranda Shapes History. The Atlantic, September 29. https://www.theatlantic.com/entertainment/archive/2015/09/lin-manuel-miranda-hamilton/408019/ [Last Accessed 30 January 2025].

Dhúill, C N 2017 World History as Heroic Biography: Thomas Carlyle’s “Great Men”. In: Hemecker, W and Saunders, E (eds.) Biography in Theory: Key Texts with Commentaries. Berlin: De Gruyter. https://www.degruyterbrill.com/document/doi/10.1515/9783110516678-008/html [Last Accessed 14 February 2025].

Doria, C 2025 You should never trust images you see online [Instagram Reel]. 27 November. https://www.instagram.com/p/DRiPbZZDmOV/ [Last Accessed 29 November 2025].

FACING IT International AB 2023 Hello History by Humy. https://www.hellohistory.ai/ [Last Accessed 20 March 2025].

FACING IT International AB 2024 Privacy Policy – Humy. February 19. https://www.humy.ai/privacy-policy [Last Accessed 20 March 2025].

Felkner, V K, Chang, H H, Jang, E and May, J 2024 WinoQueer: A Community-in-the-Loop Benchmark for Anti-LGBTQ+ Bias in Large Language Models. Preprint, arXiv, October 17.  http://doi.org/10.48550/arXiv.2306.15087

Fernandes, S 2017 Curated Stories: The Uses and Misuses of Storytelling. Oxford: Oxford University Press.

Gibbs, J 2023 Which Phrases Are the Most “ChatGPT” of All? Medium, December 14. https://medium.com/@jordan_gibbs/which-phrases-are-the-most-chatgpt-of-all-b0911e3faf6b [Last Accessed 25 February 2025].

Gonzalez, A J, Hollister, J R, DeMara, R F, Leigh, J, Lanman, B, Lee, S, Parker, S, Walls, C, Parker, J, Wong, J, Barham, C and Wilder, B 2017 AI in Informal Science Education: Bringing Turing Back to Life to Perform the Turing Test. International Journal of Artificial Intelligence in Education, 27(2): 353–84.  http://doi.org/10.1007/s40593-017-0144-1

Hannigan, T R, McCarthy, I P and Spicer, A 2024 Beware of Botshit: How to Manage the Epistemic Risks of Generative Chatbots. Business Horizons, 67(5): 471–86.  http://doi.org/10.1016/j.bushor.2024.03.001

Haraway, D 1988 Situated Knowledges: The Science Question in Feminism and the Privilege of Partial Perspective. Feminist Studies, 14(3): 575–99.  http://doi.org/10.2307/3178066

Haraway, D 2016 Staying with the Trouble: Making Kin in the Chthulucene. Durham, NC: Duke University Press.

Harder, L 2012 Die museale Verarbeitung des Holocausts in Polen. Das Galicia Jewish Museum in Krakau. BA, Saarland University.

Harder, L 2015 Zeitzeugen 2.0: The Use of Audiovisual Testimonies in German Holocaust Education in the 21st Century. MA, Brandenburg University of Technology BTU Cottbus-Senftenberg.

Harder, L R 2024a Hedy Lamarr Chatbot | ELMCIP. Electronic Literature Knowledge Base. With Colin Robinson. https://elmcip.net/creative-work/hedy-lamarr-chatbot.

Harder, L R 2024b Histobots. Echoes of History or AI Storytime?. Seminar Presentation. AI and Narrative Seminar, University of Bergen, Norway. https://www.uib.no/en/cdn/172775/ai-and-digital-narrative.

Harder, LR 2026 Histobot. The Living Glossary of Digital Narrative. https://glossary.cdn.uib.no/terms/histobot.

Harder, L R forthcoming Prompted Into Existence: Defining Histobots in Dialogue with Vincent(s). Imaginations.

Hashem Al-Ghaili 2025 Kira (Short Film on Human Cloning) [video]. https://www.youtube.com/watch?v=gx8rMzlG29Q [Last Accessed 12 July 2025].

Hekman, S 1997 Truth and Method: Feminist Standpoint Theory Revisited. Signs, 22(2): 341–365.

Heinz Nixdorf MuseumsForum 2025 AMECA – NEU IM HNF. https://www.hnf.de/dauerausstellung/ausstellungsbereiche/global-digital/mensch-roboter-leben-mit-kuenstlicher-intelligenz-und-robotik/ameca.html [Last Accessed 1 August 2025].

Hicks, M T, Humphries, J and Slater, J 2024 ChatGPT Is Bullshit. Ethics and Information Technology, 26(2): 38.  http://doi.org/10.1007/s10676-024-09775-5

Hoskins, P and Edwards, C 2025 Musk Says Grok Chatbot Was “Manipulated” into Praising Hitler. BBC News, July 10. https://www.bbc.com/news/articles/c4g8r34nxeno [Last Accessed 14 July 2025].

Humy.ai team 2024 Hello History, version 3.0 [mobile application]. Released July 8. FACING IT International. https://play.google.com/store/app/details?id=com.hello.history [Last Accessed 20 March 2025].

ICOM 2017 ICOM Code of Ethics for Museums. Paris: International Council of Museums.

Jacomy, M and Borra, E 2024 Measuring LLM Self-Consistency: Unknown Unknowns in Knowing Machines. Sociologica, 18(2): 25–65.  http://doi.org/10.6092/ISSN.1971-8853/19488

Juzek, T S and Ward, Z B 2024 Why Does ChatGPT “Delve” So Much? Exploring the Sources of Lexical Overrepresentation in Large Language Models. Preprint, arXiv, December 16.  http://doi.org/10.48550/arXiv.2412.11385

Kaplan, J, McCandlish, S, Henighan, T, Brown, T B, Chess, B, Child, R, Gray, S, Radford, A, Wu, J and Amodei, D 2020 Scaling Laws for Neural Language Models. Preprint, arXiv, January 23.  http://doi.org/10.48550/arXiv.2001.08361

Karpathy, A 2025 There’s a new kind of coding I call “vibe coding” [Twitter]. 2 February. https://x.com/karpathy/status/1886192184808149383 [Last Accessed 8 February 2025].

Keyes, R 2006 The Quote Verifier: Who Said What, Where, and When. New York: St. Martin’s Griffin.

Khan Academy 2024 Meet Khanmigo: Khan Academy’s AI-Powered Teaching Assistant & Tutor. https://khanmigo.ai/ [Last Accessed 18 May 2025].

Khan, S 2024 Brave New Words: How AI Will Revolutionize Education (and Why That’s a Good Thing). New York: Viking.

Klipphahn-Karge, M, Koster, A K and Morais Dos Santos Bruss, S 2023 Queer Reflections on AI: Uncertain Intelligences, 1st ed. London: Routledge.  http://doi.org/10.4324/9781003357957

Klumbyte, G, Piehl, H and Draude, C 2023 Towards Feminist Intersectional XAI: From Explainability to Response-Ability. Preprint, arXiv, May 5.  http://doi.org/10.48550/arXiv.2305.03375

Kobak, D, González-Márquez, R, Horvát, E A and Lause, J 2024 Delving into ChatGPT Usage in Academic Writing through Excess Vocabulary. Preprint, arXiv.  http://doi.org/10.48550/ARXIV.2406.07016

Kunzru, H 1997 You Are Cyborg. Wired, February 1. https://www.wired.com/1997/02/ffharaway/ [Last Accessed 10 April 2025].

Kurzweil, R 2005 The Singularity Is Near: When Humans Transcend Biology. New York: Viking.

Lamarr, H, Guild, L and Rice, C 1966 Ecstasy and Me. New York: Bartholomew House.

Lorde, A 1979 The Master’s Tools Will Never Dismantle the Master’s House. Paper presented at the Second Sex Conference, New York, October 29. https://monoskop.org/images/2/2b/Lorde_Audre_1983_The_Masters_Tools_Will_Never_Dismantle_the_Masters_House.pdf [Last Accessed 2 June 2025].

Masukume, G 2024 The Impact of AI on Scientific Literature: A Surge in AI-Associated Words in Academic and Biomedical Writing. Preprint, Health Informatics, June 1.  http://doi.org/10.1101/2024.05.31.24308296

Matsui, K 2024 Delving into PubMed Records: Some Terms in Medical Writing Have Drastically Changed after the Arrival of ChatGPT. Preprint, Health Informatics, May 16.  http://doi.org/10.1101/2024.05.14.24307373

Mauldin, M L 1994 ChatterBots, TinyMuds, and the Turing Test: Entering the Loebner Prize Competition. In: AAAI-94 Proceedings. Menlo Park, CA: AAAI Press.

McDowell, Z J 2024 Wikipedia and AI: Access, Representation, and Advocacy in the Age of Large Language Models. Convergence: The International Journal of Research into New Media Technologies, 30(2): 751–67.  http://doi.org/10.1177/13548565241238924

McQuillan, D 2022 Resisting AI: An Anti-Fascist Approach to Artificial Intelligence. Bristol: Bristol University Press.

Metz, C, Kang, C, Frenkel, S, Thompson, S A and Grant, N 2024 How Tech Giants Cut Corners to Harvest Data for A.I. The New York Times, April 6. https://www.nytimes.com/2024/04/06/technology/tech-giants-harvest-data-artificial-intelligence.html [Last Accessed 13 January 2025].

Milmo, D and Hern, A 2024 Google Chief Admits “Biased” AI Tool’s Photo Diversity Offended Users. The Guardian, February 28. https://www.theguardian.com/technology/2024/feb/28/google-chief-ai-tools-photo-diversity-offended-users [Last Accessed 13 January 2025].

Moravec, H P 1988 Mind Children: The Future of Robot and Human Intelligence. Cambridge, MA: Harvard University Press.

Musée d’Orsay 2023 Digital Technology: Hello Vincent. https://www.musee-orsay.fr/en/articles/digital-technology-hello-vincent-275618 [Last Accessed 23 April 2025].

Museum für Kunst und Gewerbe Hamburg 2025a Collection Strategy MK&G. Hamburg: MK&G, March 19. https://mdo.mkg-hamburg.de/MDO/mediadelivery/asset/9f52b87906124bfe854d486acba4a30b/MK&G_Collection_Strategy_March_2025.pdf [Last Accessed 27 August 2025].

Museum für Kunst und Gewerbe Hamburg 2025b Flat Files and Paper Pies: New Acquisitions and Information Graphics on the Graphics and Poster Collection. Hamburg: MK&G, August 4. https://dam.mkg-hamburg.de/downloadcollection/27FAA769718678C03F50E593D4AC5CF6 [Last Accessed 27 August 2025].

Nieborg, D B and Poell, T 2018 The Platformization of Cultural Production: Theorizing the Contingent Cultural Commodity. New Media & Society, 20(11): 4275–92.  http://doi.org/10.1177/1461444818769694

Noble, S U 2020 Algorithms of Oppression: How Search Engines Reinforce Racism. New York: New York University Press.  http://doi.org/10.18574/nyu/9781479833641.001.0001

O’Kane, E, Miller, J, Rettberg, S, Saum, A and Johnston, J 2024 More Than Meets AI [exhibition]. University of Bergen. https://www.uib.no/en/cdn/172484/more-meets-ai [Last Accessed 6 December 2024].

Omiye, J A, Gui, H, Rezaei, S J, Zou, J and Daneshjou, R 2023 Large Language Models in Medicine: The Potentials and Pitfalls. Preprint, arXiv.  http://doi.org/10.48550/ARXIV.2309.00087

OpenAI 2022 Snapshot of ChatGPT Model Behavior Guidelines. https://cdn.openai.com/snapshot-of-chatgpt-model-behavior-guidelines.pdf [Last Accessed 16 February 2025].

OpenAI 2024 GPT-4o Mini. OpenAI Developer Platform. https://platform.openai.com [Last Accessed 16 February 2025].

OpenAI 2023 GPT-4 Technical Report. Preprint, arXiv.  http://doi.org/10.48550/ARXIV.2303.08774

Ordonez, V, Dunn, T and Noll, E 2023 OpenAI CEO Sam Altman Says AI Will Reshape Society, Acknowledges Risks: “A Little Bit Scared of This”. ABC News, March 16. https://abcnews.go.com/Technology/openai-ceo-sam-altman-ai-reshape-society-acknowledges/story?id=97897122 [Last Accessed 3 February 2025].

Otto, P, Djordjevic, V, Maire, J, Hirche, T, Gräf, E and Steinhau, H 2017 Postmortaler Persönlichkeitsschutz. bpb.de, March 10. https://www.bpb.de/themen/recht-justiz/persoenlichkeitsrechte/244854/postmortaler-persoenlichkeitsschutz/ [Last Accessed 3 March 2025].

Pal, S, Lazzaroni, R M and Mendoza, P 2024 AI’s Missing Link: The Gender Gap in the Talent Pool. Brussels: Interace, October 10. https://www.interface-eu.org/publications/ai-gender-gap [Last Accessed 29 April 2025].

Pollock, G 2000 Vision and difference: femininity, feminism and histories of art. Reprinted. London: Routledge.

Qiu, J, Xiao, F, Wang, Y, Mao, Y, Chen, Y, Juan, X, Zhang, S, Wang, S, Qi, X, Zhang, T, Yao, Z, Guo, J, Lu, Y, Argon, C, Cui, J, Chen, D, Zhou, J, Zhou, S, Zhou, Z, Yang, L, Liu, S, Wang, H, Huang, K, Jiang, X, Gao, X and Wang, M 2025 On Path to Multimodal Historical Reasoning: HistBench and HistAgent. Preprint, arXiv.  http://doi.org/10.48550/ARXIV.2505.20246

Reeves, S 2025 Koloniale Sprache: Warum „farbig“ nicht höflich ist [TikTok]. 16 July. https://www.tiktok.com/@sharyreeves/video/7527737993571044631 [Last Accessed 22 July 2025].

Richmond, S 2024 The AI Mirror – Review. LSE Review of Books, August 27. https://blogs.lse.ac.uk/lsereviewofbooks/2024/08/27/book-review-the-ai-mirror-shannon-vallor/ [Last Accessed 6 March 2025].

Robb, D 2024 Exploring Imbalances on Wikipedia Through Archival Creation Theories. The iJournal: Student Journal of the Faculty of Information, 10 (December).  http://doi.org/10.33137/ijournal.v10i1.44532

Robison, K 2024 OpenAI Was a Research Lab — Now It’s Just Another Tech Company. The Verge, September 27. https://www.theverge.com/2024/9/27/24255177/openai-safety-mira-murati-quit-sam-altman-cofounders-exodus [Last Accessed 6 March 2025].

Rossiter, M W 1993 The Matthew Matilda Effect in Science. Social Studies of Science, 23(2): 325–41.  http://doi.org/10.1177/030631293023002004

Rousso, H 1991 The Vichy Syndrome: History and Memory in France since 1944. Goldhammer, A (trans.). Cambridge, MA: Harvard University Press.

Rudolph, J, Tan, S and Tan, S 2023 War of the Chatbots: Bard, Bing Chat, ChatGPT, Ernie and Beyond. The New AI Gold Rush and Its Impact on Higher Education. Journal of Applied Learning & Teaching, 6(1).  http://doi.org/10.37074/jalt.2023.6.1.23

Salvador Dalí Museum 2024 Ask Dalí Opens April 11 at The Dalí Museum. Press release, October 4. https://thedali.org/press-room/ask-dali-opens-april-11-at-the-dali-museum/ [Last Accessed 1 May 2025].

Schnabel, D, Berendsen, E, Fischer, L, Grosse, L, Heldmann, N, Stemmler, S, Tretow, N, Wendelgass, C and Adeoso, M 2025 Der Holocaust als Meme. Wie in digitalen Räumen Geschichte umgedeutet wird. Frankfurt am Main: Bildungsstätte Anne Frank. https://www.bs-anne-frank.de/mediathek/publikationen/der-holocaust-als-meme [Last Accessed 1 May 2025].

Schneider, F 2025 What Happens When You Feed AI Nothing. The Verge, June 18. https://www.theverge.com/ai-artificial-intelligence/688576/feed-ai-nothing [Last Accessed 20 June 2025].

Seaver, N 2019 Knowing Algorithms. In: Vertesi, J, Ribes, D, Forlano, L, DiSalvo, C, Jackson, S J, Loukissas, Y, Rosner, D K and Shell, H R (eds.) digitalSTS: A Field Guide for Science & Technology Studies. Princeton: Princeton University Press.  http://doi.org/10.1515/9780691190600

Shapira, P 2024 Delving into “Delve”. Philip Shapira, March 31. https://pshapira.net/2024/03/31/delving-into-delve/ [Last Accessed 18 March 2025].

Shumailov, I, Shumaylov, Z, Zhao, Y, Gal, Y, Papernot, N and Anderson, R 2023 The Curse of Recursion: Training on Generated Data Makes Models Forget. Preprint, arXiv.  http://doi.org/10.48550/ARXIV.2305.17493

Smart, A, Wang, D, Monk, E, Díaz, M, Kasirzadeh, A, Van Liemt, E and Schmer-Galunder, S 2024 Discipline and Label: A WEIRD Genealogy and Social Theory of Data Annotation. Preprint, arXiv.  http://doi.org/10.48550/ARXIV.2402.06811

Smith, D 1996 Disney A to Z: The Official Encyclopedia, 1st ed. New York: Hyperion. https://archive.org/details/disneytozofficia00smit/page/218/mode/2up [Last Accessed 10 April 2025].

Søraa, R A 2023 AI for Diversity, 1st ed. AI for Everything. Boca Raton: CRC Press.

Spielberg, S (dir.) 1993 Jurassic Park [film]. Universal Pictures.

Stuart, K 2014 Giuliani on the Call of Duty Lawsuit: “Noriega Does Not Own History”. The Guardian, October 21. https://www.theguardian.com/technology/2014/oct/21/giuliani-call-of-duty-lawsuit-noriega [Last Accessed 7 May 2025].

Sun, J and Peng, N 2021 Men Are Elected, Women Are Married: Events Gender Bias on Wikipedia. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pp. 350–60.  http://doi.org/10.18653/v1/2021.acl-short.45

TallBear, K 2014 Standing With and Speaking as Faith: A Feminist-Indigenous Approach to Inquiry. Journal of Research Practice, 10(2): Article N17.

Taylor, A 2018 The Automation Charade. Logic(s) Magazine, August 1. https://logicmag.io/failure/the-automation-charade/ [Last Accessed 7 May 2025].

Upmias, S 2023 The Rise of ChatGPTism: A New Religion for the Modern Age. Medium, January 4. https://medium.com/@shavitupmias/the-rise-of-chatgptism-a-new-religion-for-the-modern-age-ddc0dbb62b29 [Last Accessed 25 February 2025].

Vallor, S 2024 The AI Mirror: How to Reclaim Our Humanity in the Age of Machine Thinking. Oxford: Oxford University Press.

Van Riper, A B (ed.) 2011 Learning from Mickey, Donald and Walt: Essays on Disney’s Edutainment Films. Jefferson, NC: McFarland.

Vann, R T 2025 Historiography: Definition, History, Branches, & Methodology. In: Britannica, edited by The Editors of Encyclopaedia Britannica, March 7. https://www.britannica.com/topic/historiography [Last Accessed 10 March 2025].

Victoria and Albert Museum 2025 Visit V&A East Storehouse. https://www.vam.ac.uk/east/storehouse/visit [Last Accessed 24 June 2025].

Voelker, J, Bock, C, Koshofer, N and Wiest, F (dirs.) 2025 Deepfake Diaries [television documentary]. Terra X History. ZDF, April 30. https://www.zdf.de/dokus/deepfake-diaries-100 [Last Accessed 3 August 2025].

Waaldijk, B 1993 Reading Anne Frank as a Woman. Women’s Studies International Forum, 16(4): 327–35.  http://doi.org/10.1016/0277-5395(93)90022-2

Wallace, M and Peeler, M 2024 Harriet Tubman’s Deep Voice. Critical AI, 2(1).  http://doi.org/10.1215/2834703X-11205217

Wilkins, J 2025 Schools Using AI Emulation of Anne Frank That Urges Kids Not to Blame Anyone for Holocaust. Futurism, January 18. https://futurism.com/the-byte/ai-anne-frank-blame-holocaust [Last Accessed 5 February 2025].

Wikipedia 2015 Wikipedia:WikiProject Women in Red. In: Wikipedia, last edited October 5, 2025. https://en.wikipedia.org/w/index.php?title=Wikipedia:WikiProject_Women_in_Red&oldid=1315138784 [Last Accessed 18 October 2025].

ZDF 2025 Ab jetzt im ZDF: Die „Deepfake Diaries“ historischer Persönlichkeiten. Press release, April 30. https://presseportal.zdf.de/pressemitteilung/ab-jetzt-im-zdf-die-deepfake-diaries-historischer-persoenlichkeiten [Last Accessed 3 June 2025].