Simulation and Hyperreality

In “The Precession of Simulacra,” Jean Baudrillard describes simulation as a hyperreal in which models of a real cannot be traced to an origin. To explain what exactly a simulation is to Baudrillard, one of the first examples he discusses is people believing they were sick and thus producing symptoms as if they really were sick. The ability to produce “real” symptoms makes illnesses simulatable, though since they are not true symptoms, medicine cannot treat them (455). In this scenario, there is no root cause to the simulated symptoms other than the fact that they were produced. The real thing–the medicine–has no power in putting an end to the illness, giving it no other option than to continue being reproduced. Why simulations can continue to be reproduced rather than left as a single act is explained by Baudrillard’s application of the law of supply and demand: “As is the fact that power is in essence no longer present except to conceal that there is no more power. A simulation that can last indefinitely, because… it is no longer subject to violence and death… it, like any other commodity, is dependent on mass production and consumption” (470). With the presence of power functioning as a possible endpoint to simulation, the absence of power means there cannot be an endpoint; and because simulations only refer to their own production, rather than a real source, there is no starting point either. This implies that the form in which simulation can last indefinitely is a cycle determined by consumer behavior. In the age of AI and algorithm-driven media, biases fuel feedback loops that reinforce our own beliefs and distort language models’ understanding of reality–leading to a hyperreality where simulations become more real than the real.

Beginning with consumers becoming increasingly more persuaded by beliefs that align with their own, the reproduction of biases driven by human and AI interactions leads to a feedback loop of confirmation bias. In “How Human-AI Feedback Loops Alter Human Perceptual, Emotional, and Social Judgements,” Moshe Glickman and Tali Sharot run experiments to determine how biases within AI models affect the biases of consumers. The baseline of one of their experiments was to simulate a sample of professionals “utilizing a popular real-world AI system–Stable Diffusion. Stable Diffusion tends to over-represent White men when prompted to generate images of high-power and high income professionals” (Glickman & Sharot). This inherent bias of an over-representation of white men puts forward a false representation of reality, but can be perceived as a true reflection of reality if the consumer who interacts with the content also has this bias and is unaware of Stable Diffusion’s artificiality. This indeed led to confirmation bias when participants were more likely to choose white men over other races and genders, to be financial managers (Glickman & Sharot). This example of confirmation bias being a type of feedback loop is relevant to Baudrillard’s explanation of the successive phases of images, which describes the process of representation becoming a pure simulacrum. Specifically, the second stage of simulation, “It masks and denatures a profound reality” (456), depicts the state in which a false representation of reality not only takes precedence over true reality but also modifies it to eventually replace it. In this way, confirmation bias becomes a system whereby the more reinforced existing beliefs become, the less likely one is to agree with opposing beliefs. Additionally, with Stable Diffusion being exposed to many online users on social media and news outlets (Glickman & Sharot), this experiment shows how vulnerable we are to being misinformed by simply interacting with these platforms. At a deeper level, misinformation’s ability to alter reality depends on interactions with people and how their consumer behavior advances it.

To continue with the role that misinformation plays in feeding confirmation bias, identifying with claims that mirror personal beliefs can diminish online users’ exposure to real facts. In “Generative AI in the Era of ‘Alternative Facts,’” Saadia Gabriel, et al., test online users’ abilities to differentiate true and false claims using a simulated social media and news feed. Before implementing any preventative measures, they observed that “the agreement between the headlines’ political leaning and the user’s affiliations has significant effects on all three types of user behaviors (sharing, liking, and flagging), which indicates confirmation bias” (Saadia Gabriel, et al.). With users being more likely to share and like articles that align with their beliefs (Saadia Gabriel, et al.), these user behaviors not only reinforce personal political views as fact but also suppress opposing views, even if those may be more accurate. Since social media platforms and news feeds tend to boost the relevance of content with higher engagement, Baudrillard’s insight that belief stimulates simulation becomes more applicable. The actions consumers take to spread or report content, based on their personal beliefs, create demand for similar content. This encourages the cycle of simulation to continue to the third stage of simulation: “It masks the absence of a profound reality” (462) in which incorrect judgements of what is true can come to seem more real than truth itself, as correct information that opposes one’s views becomes less accessible. AI-generated disinformation similarly leads to users having a harder time identifying whether claims are true or false, especially if “it specifically targets them, highlighting the risk of personalization being exploited by malicious actors” (Saadia Gabriel, et al.). Personalized algorithms and confirmation bias feedback loops not only increase our exposure to artificial content but also lodge us deeper in simulations, making it increasingly difficult to recognize reality.

Just as consumer behavior fuels the spread of misinformation, it also shapes that data that language models learn from–creating biases that reinforce each other in a continuous feedback loop. In “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” Emily M. Bender, et al., explain how online participation on various platforms shapes the inherent biases that language models develop, especially as demographics become more concentrated. Specifically, the training data of language models is highly receptive to homogeneity given the “limited set of subpopulations [that] can continue to easily add data… this systemic pattern in turn worsen[s] diversity and inclusion within Internet-based communication, creating a feedback loop that lessens the impact of data from underrepresented communities” (613). The selective behavior of online users silences opposing ideologies, which often extends to the people associated with those beliefs. Conversely, opposition to certain individuals can influence rejection of ideas they represent. This bidirectional movement of biases makes it nearly impossible to identify a clear origin or endpoint within this feedback loop. Furthermore, since language models cannot generate meaning outside of their training data, they must assume the exclusivity of online spaces is an accurate reflection of the entire user base–when in reality it is the final stage of simulation, hyperreality: “It has no relation to any reality whatsoever: it is its own pure simulacrum” (456). The language model continuously produces in response to users’ demand for similar discourse–discourse that is itself a false construction of reality, shaped by simulations of bias.

Jean Baudrillard’s theory of simulation helps explain how human interactions with AI-generated images, politically-driven news headlines, and language models function as feedback loops that advance biases through the stages of simulation, ultimately producing a hyperreality.

Works Cited

Baudrillard, Jean. “The Precession of Simulacra.” https://www.homeworkforyou.com/static_media/uploadedfiles/Baudrillard%20-%20The%20Precession%20of%20Simulacra.pdf.

Bender, Emily M., et al. “On the dangers of stochastic parrots: Can language models be too big?” Proceedings of the 2021 ACM conference on fairness, accountability, and transparency. 2021.

Gabriel, Saadia, Liang Lyu, James Siderius, Marzyeh Ghassemi, Jacob Andreas, and Asu Ozdaglar. 2024. “Generative AI in the Era of ‘Alternative Facts.’”

Glickman, Moshe, and Tali Sharot. “How Human–AI Feedback Loops Alter Human Perceptual, Emotional and Social Judgements.” Nature Human Behaviour, vol. 9, no. 2, 2025, pp. 345–59, https://doi.org/10.1038/s41562-024-02077-2.