Key Takeaways
- Anthropomorphic language—describing AI as if it can “lie,” “scheme,” or “want”—creates confusion about agency and accountability in policy.
- Research shows people readily assign human traits to machines (the Eliza effect, CASA paradigm), which can lead to over‑reliance, emotional attachment, and even mental‑health harms.
- When statutes use verbs like “decide,” “see,” or “recognize,” they blur the line between the model’s output and the human choices that shape data, objectives, thresholds, and deployment.
- Precise, operational vocabulary (detect, classify, score, rank, optimize) clarifies who is responsible for building, testing, monitoring, and liable for AI systems.
- Policymakers should attribute agency explicitly to developers, deployers, and operators, mandate concrete transparency disclosures, and scrutinize whether human‑like design cues are necessary for a given use case.
Introduction: Anthropomorphism in AI Discourse
Science‑fiction imagery has long shaped how the public talks about artificial intelligence, leading headlines and marketing copy to portray AI systems as if they can “lie,” “scheme,” or “want” things. As the article notes, “Headlines frequently borrow science fiction tropes to describe AI systems as if they can ‘lie,’ ‘scheme,’ or ‘want’ things.” Companies reinforce this by calling chatbots “assistants” or “agents,” and policymakers have adopted similar language, sometimes embedding it directly in legislative text where tools are said to “make” consequential decisions or perform tasks “normally requiring human intelligence.” While such shorthand is vivid, it risks implying a level of agency that does not exist, obscuring the fact that AI outputs are the product of human‑driven design, data selection, and deployment choices.
Why Anthropomorphism Is a Governance Problem
Humans instinctively attribute human‑like qualities to machines—a tendency demonstrated by the Eliza effect, where users empathize with simple pattern‑matching dialogue, and the “computers are social actors” (CASA) paradigm, which shows people respond to digital systems using the same social cues they use with other people, even when they know they are interacting with software. This framing can lead users to treat AI as independent decision‑makers rather than tools built, refined, and deployed by people and institutions. Large language models, for instance, generate outputs by predicting the next most likely token based on training‑data patterns; they possess no intent, consciousness, or understanding. Labeling false outputs as “hallucinations” lumps together disparate issues such as missing source material, unclear prompts, data manipulation, or plausible‑sounding but unverified text. Design choices—realistic voices, named personas, personality cues, emotionally expressive chat interfaces—further increase perceived human‑likeness and trust, encouraging users to project stereotypes, overshare, rely excessively, or form emotional attachments. Researchers have documented cases where prolonged interaction with companion‑style chatbots intensified delusional thinking, emotional reliance, and severe mental‑health deterioration, while large‑scale data link heavy use of personalized chatbots to heightened loneliness and dependency over time.
Policy Consequences of Anthropomorphic Framing
When policy language adopts anthropomorphic terms, several governance risks emerge. First, accountability becomes diluted: describing an output as the system’s “judgement” suggests the AI itself chose a course of action, whereas in reality humans select training data, objectives, thresholds, interfaces, and deployment contexts. As the text explains, “When a policy reads, ‘AI makes a decision,’ the wording can blur the distinction between the output generation and the decision to use the model’s output.” Second, implementation suffers because verbs like “see” or “recognize” do not specify measurable performance criteria, error tolerances, or failure modes, making it hard for regulators to verify compliance or for procurers to enforce specifications. Third, anthropomorphic phrasing reinforces industry narratives that market ordinary data‑processing as autonomous intelligence, allowing firms to shift scrutiny away from business practices, dataset sourcing, and profit motives. If a system “decided,” responsibility appears ambiguous; if a company configured, tested, and deployed a model with known limitations, liability is clearer. Finally, such language conflicts with established governance principles like the OECD AI Principles (human‑centered values, transparency, accountability) and the NIST AI Risk Management Framework’s Generative AI Profile, which call for accurate characterization of system capabilities and limits.
Toward Technically Precise Policy Vocabulary
To mitigate these problems, the article advocates replacing folk‑psychological verbs with operational terms that describe what systems actually do. Statutes should prefer words such as “detect,” “classify,” “cluster,” “score,” “rank,” and “optimize,” reserving “decide” for the human or institutional act that produces legal or material effects. Agency must be attributed explicitly to developers, deployers, and operators: legal text can specify which entity determines a system’s purpose, sets decision thresholds, and decides whether and how outputs influence consequential decisions, while establishing avenues for civil‑society oversight. Many state laws already name who must test, disclose, and monitor systems; this practice should be expanded. Transparency requirements should be concrete—mandating documentation of data provenance, model architecture, known limitations, training‑data categories, evaluation benchmarks, and performance disparities—especially when full technical explanation is infeasible due to opacity or trade secrets. Policymakers should also name actors across the AI supply chain rather than abstracting them into vague references to “AI,” clarifying oversight responsibilities. For example, instead of stating that “AI verifies identity,” a statute could specify that “a particular contractor deploys a face‑matching algorithm trained on specified datasets by a particular developer under defined thresholds.” Finally, regulators should scrutinize whether human‑like design cues are necessary for a given use case; emerging laws in California and New York that require disclosures when users interact with simulated personas illustrate how to curb over‑trust and emotional dependence in higher‑risk domains such as mental health, employment, education, or public benefits.
Implications and Conclusion
Precision in legislative language will not erase all governance challenges—information asymmetries and power imbalances persist—but it can improve institutional outcomes at the margin. By swapping anthropomorphic shorthand for operational, accountable terminology, policymakers clarify who builds and deploys AI systems, how they function, where risks arise, and which actors bear documentation, testing, monitoring, and liability obligations. This strengthens procurement, enforcement, and public understanding, reminding stakeholders that AI systems are multipurpose technologies embedded in organizational decision chains, not autonomous agents. As the article concludes, “Policy should describe the technology accordingly.” Implementing these recommendations will help ensure that AI governance remains grounded in the real‑world sociotechnical processes that actually shape AI’s impact on society.