PaperlayerView exact revision on arXiv

Source-linked AI summary

The Many Morals of AI: The cautionary bias of North American news narratives in global perspective

Andrew Piper, Edie Pearman

arXiv:2610.18573v1cs.CLcs.AI

TL;DR

Evidence on cross-regional AI news narratives remains limited, despite news media’s role in shaping how people understand technological change. This study analyzes AI story morals across four regions and finds that North American cautionary framing is closely associated with broader regional negativity, though it is not fully reducible to sentiment.

  • Problem

    Cross-regional evidence on AI news narratives is limited, and common computational approaches can miss underlying narrative structures relevant to how people understand AI’s social role.

    Source: §1 · pp. 1–2

  • Method

    The authors analyze 12,924 news stories from four regions, using large language models to infer each article’s implicit story moral.

    Source: §1 · p. 2

  • Results

    North American AI coverage’s cautionary distinctiveness is substantially reduced after accounting for broader regional negativity, which also extends beyond AI-focused articles.

    Source: §4 · p. 4

  • Takeaways & Limitations

    The findings underscore the need to distinguish narratives prominent in local media environments from a global assessment of AI’s risks and benefits.

    Source: §4 · p. 4

    LLM-based moral inference depends on models’ interpretive priors, and regional categories compress internal diversity within media systems.

    Source: Limitations · p. 5

Abstract

from arXiv · show

The stories told about artificial intelligence (AI) play a central role in shaping public understanding, policy debates, and expectations about technological change. News media are especially influential in this process. Using a cross-regional corpus of 12,924 English-language news articles about AI published since 2022, we examine how storytelling about AI varies across media systems in four regions: North America, China, India, and the Middle East. Leveraging large language models, we extract the implicit narrative lessons conveyed by each article---commonly known as story morals---capturing value-laden interpretations that extend beyond thematic content or sentiment. We find that North American AI coverage is uniquely characterized by narratives of caution, while under-emphasizing narratives of innovation and collaboration that are more prevalent elsewhere. However, this pattern is largely explained by a broader regional negativity bias in North American news sentiment, rather than by heightened concern about AI itself. Our findings bring into view the culturally conditioned nature of AI news narratives.

1 Introduction

News narratives can shape how people understand AI’s social role and the balance of its risks and rewards. This study uses cross-regional story-moral analysis to address limits in prior evidence and finds that North American AI coverage’s distinctive cautionary framing is strongly associated with broader regional negativity.

1 Introduction

  • News stories frame technological change through moral and emotional registers of risk and reward, shaping public attention and understanding.The paper notes that these interpretations can shape how publics understand AI’s role in society.Source: §1 · p. 1
  • Prior AI-news research has limited regional diversity, and common computational methods can miss articles’ underlying narrative structures.Some studies also rely on qualitative analysis of small samples.Source: §1 · pp. 1–2
  • The study analyzes 12,924 news stories from North America, the Middle East, China, and India using story-moral inference to recover their implicit evaluative lessons.Story-moral inference asks why a story was told rather than identifying its topics, themes, or issue-specific emphases.Source: §1 · p. 2
  • North American AI news emphasizes caution more than innovation or collaboration, a pattern strongly associated with broader regional negativity.The authors caution against treating this framing as a global consensus about AI’s risks or social meaning.Source: §1 · p. 2

2 Data and Methods

The authors assemble a four-region English-language news corpus, filter articles for substantive AI focus, and use two language models to infer story morals and sentiment. Human review finds most extracted moral keywords at least moderately relevant.

2.1 Data Collection

  • The collection spans 12 major English-language news organizations in North America, China, India, and the Middle East, with queries running from January 1, 2022, to September 9, 2024.The authors initially retrieved 48,201 articles; their best automated filter had weighted F1 = 0.85 and retained 8,924 AI-focused articles, alongside 4,000 sampled rejected articles for comparison.Source: §2.1 · p. 2

2.2 Inferring Story Morals using LLMs

  • GPT-4o and Gemini-2.0-Flash generated moral statements, moral keywords, sentiment valence, and candidate industries using randomly assigned prompt variations.Both models’ outputs were aggregated into the final dataset.Source: §2.2 · pp. 2–3
  • GPT-4o and Gemini keyword rankings showed low-to-moderate agreement, with extrapolated RBO of 0.32 for the top 10 and 0.40 for the top 100 keywords.The reported p value was 0.9.Source: §2.2 · p. 2
  • 96% of GPT-4o keywords and 82% of Gemini keywords received at least moderate relevance ratings from human annotators.Two independent annotators rated keywords from a random sample of 100 AI articles.Source: §2.2 · pp. 2–3

3 Results

North American AI coverage is more cautionary than coverage elsewhere, but this difference shrinks substantially when broader regional sentiment is accounted for. North American news is also more negative across AI and non-AI articles, with negativity varying by industry.

3.1 Distribution of moral keywords by region

  • North American AI coverage disproportionately emphasizes skepticism, risk, exploitation, greed, and bias, while giving less weight to innovation and collective-benefit values.The study groups caution narratives around risk, responsibility, ethics, regulation, and harm, and innovation narratives around progress, opportunity, efficiency, and growth.Source: §3.1 · p. 3, Figure 3 · p. 3

3.2 The association of sentiment and cautionary AI framing

  • Negative sentiment was predicted for 33.1% of North American AI-focused articles and 40.8% of comparison articles, versus lower ranges in other regions.The reported probabilities were 12.8–14.7% for AI-focused articles and 14.9–24.5% for comparison articles in other regions.Source: §3.2 · p. 3, Figure 4 · p. 4
  • 15.1 percentage points before sentiment adjustment fell to 5.0 points afterward for the North America–other regions cautionary-moral gap.The remaining adjusted gap was positive when newspapers were omitted one at a time and in analyses of either annotation model.Source: §3.2 · pp. 3–4

3.3 Negativity Around AI Is Unevenly Distributed Across Industries

  • In North American coverage, media and entertainment, along with law, clustered among negatively valenced industries, while healthcare, semiconductors, retail, and finance were generally portrayed more positively.Politics showed the highest level of concern among the industries examined.Source: §3.3 · p. 4

4 Conclusion

The findings connect North American AI cautionary framing with broader regional negativity, while showing that sentiment does not fully account for narrative values. The authors argue that cross-regional comparison can help distinguish local media patterns from global assessments of AI.

4 Conclusion

  • North American AI coverage’s cautionary distinctiveness is substantially reduced when broader regional sentiment differences are taken into account.Greater negativity also appears in North American articles not centrally focused on AI.Source: §4 · p. 4
  • Narrative values are strongly associated with sentiment but are not completely reducible to it.The paper connects research on AI-news negativity with work on regional variation in media cultures (Ryazanov et al., 2024).Source: §4 · p. 4
  • The findings invite journalists to consider whether AI story morals follow from story evidence or familiar sentiment conventions, and urge policymakers to distinguish local media narratives from global risk assessments.The authors suggest that recognizing regional patterns can bring a wider range of perspectives into public debate.Source: §4 · p. 4

Limitations

The analysis is limited by model-dependent moral inference, compressed regional categories, and the observational nature of its mediation analyses. Its cross-cultural sample and data cutoff also bound how its findings should be interpreted.

Limitations

  • LLM interpretive priors and semantic frameworks constrain moral inference, so the authors call for further validation despite prior evidence of alignment with human judgments.The analysis controls for model and prompt effects and focuses on robust aggregate patterns.Source: Limitations · p. 5
  • Regional categories compress diversity within media systems, so the findings are not exhaustive characterizations of national or cultural attitudes toward AI.The sample was designed primarily for cross-cultural comparison.Source: Limitations · p. 5
  • Observational mediation analyses do not establish causal pathways, but provide evidence consistent with sentiment linking region and moral framing.The authors recommend investigating this relationship at publication or article level.Source: Limitations · p. 5
  • Narrative distributions may change after the data cutoff, although regional sentiment distributions may anchor story morals and make them more consistent than expected.Source: Limitations · p. 5

AI Disclosure Statement

The disclosure describes Codex's contributions and assigns responsibility for the final manuscript, analyses, and conclusions to the authors.

AI Disclosure Statement

  • OpenAI Codex assisted with idea development, outlining, language editing, code development, and data analysis, while the authors reviewed its contributions and took full responsibility for the final work.Source: AI Disclosure Statement · p. 5

A1 Article Counts by Publication and Region

Table A1 is titled as a breakdown of article counts by news source, region, and category.

A1 Article Counts by Publication and Region

  • Table A1 presents article counts by news source, region, and category.Source: Table A1 · p. 6

A2 Coefficient Tables

The appendix reports regression tables for article sentiment and cautionary moral framing. The cautionary-framing models are shown both before and after sentiment adjustment.

A2 Coefficient Tables

  • The appendix identifies the models in Tables A2 and A3 as those reported in Section 3.2.Source: A2 Coefficient Tables · p. 6
  • Table A2 reports a multinomial logistic regression of article sentiment using 25,848 annotations of 12,924 articles, with negative sentiment as the outcome reference.Coefficients are log odds; standard errors are clustered by article (HC1), with North America, AI-focused articles, Gemini-2.0-Flash, and prompt 1 as reference categories.Source: Table A2 · p. 7
  • Table A3 reports logistic regressions of cautionary moral framing in AI-focused articles using 17,848 annotations of 8,924 articles, with and without sentiment adjustment.The outcome is whether any extracted moral keyword is coded as caution; coefficients are log odds, with standard errors clustered by article (HC1).Source: Table A3 · p. 7

A3 Caution / Innovation Framing Exercise

The exercise coded stories for caution and innovation moral framings, which could appear separately, together, or not at all. Table A4 maps moral keywords to those categories.

A4 Prompt Notebooks

The prompt notebooks document LLM prompts for filtering and coding AI-news articles. They also describe prompt variations intended to reduce sensitivity to wording.

A4 Prompt Notebooks

  • All models used temperature 1.0 and a 512-token maximum; the prompt templates also specified article-text placement and JSON response formats.Source: A4 Prompt Notebooks · p. 8, §4 · pp. 8–9
  • The topic prompt asked for three comma-separated keywords, from general to specific, to filter non-AI-related articles.Source: A4 Prompt Notebooks · p. 8
  • Eight prompt variations were used for moral statements, keywords, industry candidates, and sentiment valence, changing question phrasing, article-text position, and role-statement inclusion.Source: A4 Prompt Notebooks · p. 8
  • The valence prompt classified AI's representation in a story as positive, negative, or neutral.Source: §4 · p. 9
  • Industry prompts asked for the principal industry or industries as single-word entries in a JSON object.Source: §4 · p. 9
  • Moral-statement prompts asked for a one-sentence interpretation of the story's moral.Source: §4 · p. 9
  • When included, role statements could address news-article interpretation generally or moral-keyword interpretation specifically.Source: §4 · pp. 8–9
  • Moral-keyword prompts requested one to three distinct keywords, avoiding terms such as “technology” or “AI.”Source: §4 · p. 9
Loading 2610.18573v1…