AI prompting is creating a new form of digital anxiety

New research reveals that everyday interactions with large language models (LLMs) are producing a new psychological condition tied to uncertainty, repetition and the constant adjustment of instructions.

The study, Prompt anxiety and the algorithmic politics of uncertainty, published in AI & Society, examines how users of systems such as ChatGPT, Claude and other generative AI tools experience prompt writing as a form of wager. The paper links this behavior to older theories of gambling, labor and political power, arguing that AI platforms are turning uncertainty itself into a source of economic value.

How prompt anxiety emerges from unpredictable AI systems

Prompt anxiety, as defined in the study, is a state users enter when they try to control a stochastic system through wording, tone, formatting, examples and repeated attempts, while knowing that even a carefully written prompt may not produce the same result twice.

LLMs do not have fixed rules to generate answers. They produce text by selecting tokens from probability distributions, which means a prompt doesn't simply trigger a stable response. It opens a field of possible outputs, where small changes in wording can shift the direction of an answer. This, according to the author, makes prompt engineering feel less like ordinary software use and more like a repeated bet.

The paper draws on Walter Benjamin's analysis of gambling to explain the user's position. The prompt writer, like the gambler, operates in a compressed present. Each submission carries the hope that the next attempt will produce the desired result. A failed answer leads to another adjustment, another run and another attempt to find the right combination of words and the user keeps moving between control and chance.

In consumer-facing AI tools, users often refine prompts by changing tone, adding roles, asking models to reason step by step, adding examples or imposing output formats. These practices can improve results, but they cannot remove uncertainty. The gap between the appearance of control and the reality of probabilistic output creates the anxiety that now surrounds prompt-based work.

Prompt sensitivity encourages what the paper describes as a paranoid style of interpretation. Users begin to assign meaning to small changes in AI behavior, treating formatting, politeness, word choice or structure as possible hidden levers. In online communities, this has produced large informal bodies of advice about how to make AI systems respond better, even when the actual effect of such techniques remains unclear.

The study treats these practices as evidence of how people adapt to opaque systems. Users are trying to build practical knowledge in an environment where model behavior is only partly visible. Their strategies reveal both the limits of AI transparency and the growing burden placed on users to manage uncertainty.

Platforms turn uncertainty into value

According to the author, prompt anxiety is not merely a side effect of generative AI. It is increasingly built into the economic structure of AI platforms. Subscription models, token limits, usage caps, premium access and paid tiers all shape how users experience AI uncertainty. When outputs are inconsistent, users may spend more time, more tokens or more money trying to reach a useful result. The platform does not have to eliminate unpredictability. In many cases, it can profit from the user's repeated attempts to manage it.

The study links this to vector capitalism, a term used to describe an economic system in which knowledge, language and social relations are captured, encoded and monetized through vector-based AI systems. LLMs rely on vast amounts of collectively produced text, culture and communication, but access to the resulting systems is controlled by private platforms, creating a contradiction.

The intelligence embedded in AI systems depends on social labor, including writing, coding, research, moderation, user feedback and online communication. However, the benefits are concentrated inside commercial platforms that charge users to access tools trained on collective knowledge. Prompting becomes a new form of labor performed inside a system whose rules remain largely hidden.

The paper also highlights prompt marketplaces, professional prompt engineering and online prompt libraries as signs that uncertainty itself is being commodified. Prompts are packaged, sold and promoted as if the right wording can unlock stable results. This creates a market around the promise of control, even though the systems remain probabilistic.

Prompt engineering is not experienced in the same way across all settings. For instance, professional users with access to fine-tuned systems, low-temperature settings, testing procedures or internal tools may face less uncertainty than casual users of public chatbots. The paper, however, points to a clear broader trend: AI platforms are creating work conditions in which people must adapt to opaque, shifting systems.

The author uses LLMbench, a research tool developed to compare large language model outputs. It examines token probability distributions, entropy and cross-model divergence and its analysis shows that uncertainty is not just something users imagine. It can be measured in the generation process itself.

When two models receive the same prompt, they may produce different answers and distribute uncertainty differently across their responses. At some points, a model may be highly confident. At others, it may choose among several possible continuations. These moments of uncertainty are usually hidden from users behind a smooth chat interface, but they shape the final answer. The result is a system that looks fluent and confident while remaining unstable at the level of generation. The instability helps explain why users may become obsessed with prompt wording, why they rerun tasks, and why they develop elaborate theories about how models behave.

Why it matters for AI governance and digital labor

If prompting is becoming part of everyday labor, then the uncertainty built into AI systems becomes a workplace issue, a transparency issue and a governance issue. The research suggests that users should not be left to absorb the costs of opaque AI behavior on their own. When platforms sell access to tools that produce variable results, they should provide clearer information about model limits, output uncertainty, data use, pricing structures and the conditions that affect performance. Without that transparency, users may mistake system-level uncertainty for personal failure or lack of skill.

The paper also raises questions about the labor hidden behind AI systems. Generative AI depends on data creation, content moderation, model training, evaluation, infrastructure and user experimentation. Yet much of that labor is either invisible or unpaid. As prompting becomes a routine part of work, organizations may need to recognize the time and cognitive effort required to use AI effectively rather than treating the tools as simple productivity boosters.

The analysis further points to the environmental and economic costs of AI. Token-based systems may appear abstract to users, but they depend on large computational infrastructures. If platforms encourage repeated attempts, longer interactions and premium use without clear accountability, the social and environmental costs of AI use remain hidden.

AI failures are also politically important. Hallucinations, inconsistent responses and unexpected outputs are often discussed as technical problems to be fixed. The author argues that they also reveal the instability of claims that AI systems are neutral, reliable or fully controllable. These failures expose the limits of automated authority.

Online communities that share prompting techniques, jailbreaks and workarounds are another key part of the argument. The paper describes these communities as producing informal knowledge about AI systems. Some of that knowledge may involve superstition or mistaken assumptions, but it also reflects a collective attempt to understand and resist opaque technologies.

To address this, the study introduces the idea of revolutionary prompting - a term that refers to the possibility of turning individual prompt anxiety into a collective critique of AI production. That includes asking who owns AI systems, who benefits from them, who performs the labor behind them, and who bears the risks when they fail.

AI regulation should go beyond safety testing and content rules, the study suggests, adding that it should also address platform power, pricing transparency, labor conditions, data ownership, environmental accountability and democratic oversight of AI development. The key issue is not only whether AI can produce accurate answers, but how uncertainty is organized, monetized and shifted onto users.

The future of generative AI, the paper argues, will depend not only on better models, but on whether society can govern the uncertainty those models create.