💬 NLP

How Anthropomorphic Language Impacts Public Perceptions of AI

This study finds that, contrary to concerns about misleading expectations, using anthropomorphic language in public-facing AI discourse does not substantially alter participants' immediate perceptions of AI technologies compared to non-anthropomorphic descriptions, though long-term or naturalistic effects remain possible.

Original authors: Betty Li Hou, Sophie Hao, Sunoo Park, Tal Linzen

Published 2026-06-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Betty Li Hou, Sophie Hao, Sunoo Park, Tal Linzen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to explain a very complex machine to a friend. You have two ways to describe it:

The "Robot Friend" Approach: You say, "The machine thinks about your problems, wants to help you, and decides what you should buy." (This is anthropomorphic language—giving human traits to a non-human thing).
The "Toolbox" Approach: You say, "The machine processes data to generate suggestions based on patterns it found." (This is non-anthropomorphic language—describing it as a tool).

This paper by Betty Li Hou and her team at NYU and Boston University asks a simple question: Does the way we talk about AI actually change how people feel about it?

The Experiment: A Taste Test for AI

The researchers set up a "taste test" for public opinion. They gathered 815 people and gave them a "briefing packet"—a collection of articles about AI that a regular person might read online.

They split the participants into groups and gave them different versions of the same story:

Group A read stories where the AI was described like a human (it "thinks," "wants," "decides").
Group B read the exact same stories, but the words were swapped to sound more like a calculator (it "processes," "outputs," "is used").
Group C (The Control Group) read stories specifically designed to scare people, warning that AI might soon be as dangerous as nuclear weapons.

They also tested two different types of AI:

Chatbots (LLMs): Like the ones that write essays or talk to you.
Recommendation Systems: Like the algorithms on Netflix or Amazon that suggest what to watch or buy.

Before and after reading, the participants answered questions about:

Who is to blame if AI makes a mistake?
Will AI take our jobs?
Can we test AI to make sure it's safe?
Is AI good or bad for society?

The Big Surprise: The "Robot Friend" Didn't Work

The researchers expected that describing AI as a "thinking, deciding" being would make people more afraid, or perhaps more trusting, or change their minds about who is responsible for errors.

But that didn't happen.

Whether people read the "Robot Friend" version or the "Toolbox" version, their opinions stayed exactly the same.

They didn't suddenly think AI was more dangerous.
They didn't suddenly think AI was more trustworthy.
They didn't change their minds about who should be blamed for mistakes.

It's as if you told someone a story about a car either as "a car that wants to drive fast" or "a machine that uses fuel to move fast." The listener's opinion on car safety didn't change based on the wording.

The Real Driver: The "Doomsday" Packet

However, when the researchers gave the Doomsday Packet (the one explicitly warning of catastrophic risks), the results were totally different.

People who read the scary warnings did change their minds. They became more worried about AI taking over, less confident that we can test it safely, and more negative about its impact on society.

The Analogy:
Think of the "Robot Friend" language as a flavoring spice. You can sprinkle it on the soup (the text), but it doesn't change the main ingredients. The "Doomsday" packet, however, was like adding a whole new, spicy ingredient to the soup. That actually changed the taste.

What This Means (According to the Paper)

The paper concludes that substance matters more than style.

If you want to change how people feel about AI, you have to change what you tell them (e.g., the actual risks or benefits), not just how you describe the AI (e.g., calling it "smart" vs. "complex").
The immediate effect of using "human-like" words to describe AI is very small, almost non-existent.

A Small Caveat

The authors are careful to say this only measures immediate reactions. They note that if people hear "AI thinks" every single day for years, or if they interact with AI constantly, those small effects might add up over time. But for a single reading session? The "human-like" words didn't move the needle.

In short: Calling a calculator a "thinking genius" doesn't make people suddenly believe it's alive or dangerous. But telling them "this calculator might blow up the world" definitely does.

Technical Summary: How Anthropomorphic Language Impacts Public Perceptions of AI

Problem Statement

Public discourse regarding Artificial Intelligence (AI) frequently employs anthropomorphic language—attributing human capabilities, characteristics, and agency to non-human systems. Critics argue that this framing sets misleading expectations, inflates claims, and fuels hype, potentially distorting public understanding of AI capabilities and risks. Specific concerns include the projection of misplaced assumptions onto systems (e.g., interpreting "hallucinations" as perceptual experiences), overestimation of abilities leading to job displacement fears, and the creation of high-risk scenarios through over-reliance, particularly among vulnerable populations. While prior work has identified the prevalence of such language in media and research, there is a lack of controlled empirical evidence demonstrating whether anthropomorphic descriptions causally shift public perceptions of AI regarding responsibility, safety, job displacement, and societal impact.

Methodology

The authors conducted a survey experiment with $N=815$ participants recruited from Prolific (US-based, English-speaking). The study utilized a pre-survey/post-survey design to measure changes in attitudes after exposure to briefing packets.

Experimental Design

Participants were randomly assigned to one of five conditions:

LLM-A: Briefing packet on Large Language Models (LLMs) using anthropomorphic language.
LLM-NA: Briefing packet on LLMs using non-anthropomorphic language.
Rec-A: Briefing packet on Recommendation Systems using anthropomorphic language.
Rec-NA: Briefing packet on Recommendation Systems using non-anthropomorphic language.
Doomsday: Briefing packet emphasizing catastrophic societal risks of AI (serving as a positive control for attitude shifts).

Stimuli Construction:
The briefing packets consisted of excerpts from news articles, press releases, and white papers (7–8 pages total). For the anthropomorphic conditions, the authors systematically edited the text based on a taxonomy of anthropomorphism involving two dimensions:

Verb Selection: Replacing human-associated verbs (e.g., "thinks," "decides," "wants") with less human-associated alternatives (e.g., "processes," "outputs," "is used to").
Agency Structure: Shifting the AI from the grammatical agent of an action to an instrument/tool (e.g., changing "The model performs a task" to "The model is used to perform a task").
The non-anthropomorphic versions maintained the same substantive content and factual claims but removed these linguistic markers.

Measures

Participants completed surveys before and after reading the packets covering five domains prominent in public discourse:

Responsibility for Harms: Ranking responsibility among the user, developer, and the AI system itself.
Loss of Control: Agreement with statements regarding humans losing the ability to control or turn off AI.
Job Replacement: Beliefs about AI replacing human doctors, teachers, and experts within 10–20 years.
Safety Testing: Perceptions of the capacity of testing to mitigate risks.
Societal Impact: Overall positive or negative impact of AI on society.

Additionally, participants completed a writing task describing a human-chatbot interaction before and after the treatment. These samples were annotated using an LLM judge (GPT-4) to count instances of anthropomorphic language, testing for a feedback loop where exposure influences the user's own language.

Analysis

The study employed Bayes Factor analysis to evaluate evidence for the alternative hypothesis (that anthropomorphic language causes a shift) versus the null hypothesis (no meaningful shift). This approach distinguishes between evidence for an effect, evidence for the absence of an effect, and inconclusive data.

Key Results

1. Effect of Anthropomorphic Language

Across all measured outcomes (responsibility, control, jobs, safety, societal impact), the study found no substantial immediate effect of anthropomorphic language on public perceptions.

Pooled Analysis: When pooling LLM and Recommendation System conditions, all outcomes except one (Teacher Competence, which was inconclusive) favored the null hypothesis ( $BF_{01} > 3$ ), indicating moderate evidence that anthropomorphic language does not shift perceptions.
Technology-Specific Analysis:
- LLMs: 7 out of 13 outcomes favored the null; the rest were inconclusive.
- Recommendation Systems: 10 out of 13 outcomes favored the null; the rest were inconclusive.
Interaction Effects: There was no consistent evidence that the effect of anthropomorphic language differed between LLMs and Recommendation Systems.

2. Effect of Risk-Focused Content (Doomsday Condition)

In contrast to the anthropomorphic manipulation, the Doomsday packet (explicitly warning of catastrophic risks) produced measurable and substantial shifts in perceptions.

Strong evidence was found for shifts in Societal Impact ( $BF_{10} = 147,324$ ) and Safety Testing ( $BF_{10} = 36.87$ ), with participants viewing AI's impact more negatively and expressing less confidence in safety testing.
This confirms that perceptions of AI are sensitive to reading materials and can shift within a single session, but the shift is driven by substantive risk content rather than linguistic framing.

3. Writing Sample Analysis

Exposure to the Doomsday packet significantly increased participants' own use of anthropomorphic language in their writing ( $BF_{10} = 2,671$ ). However, exposure to anthropomorphic vs. non-anthropomorphic packets did not significantly alter participants' subsequent use of anthropomorphic language in their own descriptions, suggesting the framing did not create a strong immediate feedback loop in language production.

Significance and Claims

The paper claims that the immediate effects of anthropomorphic language on public opinions of AI are modest to negligible. While critics argue that such language fuels hype and misunderstanding, this study provides empirical evidence that, in a controlled setting, the style of description (anthropomorphic vs. non-anthropomorphic) does not substantially alter views on responsibility, safety, or job displacement compared to the substance of the content.

The authors conclude that:

Substance over Style: Public perceptions appear more sensitive to the substantive content of a text (e.g., explicit warnings of risk) than to the linguistic framing of anthropomorphism.
Limitations of Immediate Effects: The lack of immediate shifts does not prove anthropomorphic language has zero effect. The authors note that smaller effects may exist, or that effects may accumulate over longer time scales, repeated exposure, or in naturalistic settings where discourse is pervasive.
Future Directions: The findings warrant further investigation into long-term exposure and population-level consequences of small perceptual shifts, rather than assuming immediate, large-scale attitude changes driven solely by linguistic framing.

The study does not claim that anthropomorphic language is philosophically appropriate or inappropriate, but rather addresses the empirical question of its practical impact on public perception, finding that its immediate influence is limited.