Source-linked AI summary
Generative AI Search Engines as Arbiters of Public Knowledge: An Audit of Bias and Authority
Alice Li, Luanne Sinnamon
TL;DR
This paper examines how generative AI search engines respond to public-interest queries and establish authority. It audits systems using authentic queries and response analysis, finding sentiment bias and source biases alongside uneven source quality and authority claims weakened by ambiguous language.
Problem
Generative AI search systems may produce biased responses and weak source attribution, while users often trust search engines and may accept outputs uncritically.
Source: INTRODUCTION · p. 1Method
The authors analyzed responses to 48 authentic queries across four topics, using sentiment analysis and inductive coding of responses and sources.
Source: Topics and Queries · p. 4, Data Preparation and Analysis · p. 5Results
The audit found sentiment bias tied to queries and topics, commercial and geographic source bias, and authority claims weakened by ambiguous language and questionable-quality sources.
Source: CONCLUSION · p. 10Takeaways & Limitations
The findings support raising public awareness of generative AI search systems’ strengths and limitations, including their authority claims and source quality.
Source: CONCLUSION · p. 10
Abstract
from arXiv · showhide
This paper reports on an audit study of generative AI systems (ChatGPT, Bing Chat, and Perplexity) which investigates how these new search engines construct responses and establish authority for topics of public importance. We collected system responses using a set of 48 authentic queries for 4 topics over a 7-day period and analyzed the data using sentiment analysis, inductive coding and source classification. Results provide an overview of the nature of system responses across these systems and provide evidence of sentiment bias based on the queries and topics, and commercial and geographic bias in sources. The quality of sources used to support claims is uneven, relying heavily on News and Media, Business and Digital Media websites. Implications for system users emphasize the need to critically examine Generative AI system outputs when making decisions related to public interest and personal well-being.
INTRODUCTION
Generative AI search systems generate narrative answers rather than simply directing users to webpages, positioning them as potential authorities in public knowledge. This study examines their responses to public-interest topics, including how they present claims and establish authority.
INTRODUCTION
- GenAI search engines generate narrative responses and engage in dialogue, potentially making them powerful gatekeepers and authorities in public knowledge.Unlike traditional search engines, they do not merely direct users to relevant web pages.Source: INTRODUCTION · p. 1
- Their adoption raises concerns about misinformation, bias, disinformation, weak source attribution, and users’ uncritical acceptance of system outputs.Search users may have low information literacy, trust search engines, rely on cognitive biases, and be susceptible to inaccurate information.Source: INTRODUCTION · p. 1
- The study examines responses to public-concern and welfare topics, focusing on sentiment bias and the tactics used to present claims and establish authority.The authors reasoned that responses could reveal system design goals and inform understanding of potential user impacts.Source: INTRODUCTION · p. 1
- The findings aim to raise public awareness of GenAI search systems’ strengths and limitations, inform design improvements, and encourage continued research and user education.Source: INTRODUCTION · p. 2
PRIOR WORK
- The authors use “Generative AI search engine” for hybrid systems that combine LLMs with search results to generate responses and links to current web content.The paper audits ChatGPT, Bing Chat, and Perplexity, then quantitatively compares three systems and inductively analyzes responses and sources from Bing Chat and Perplexity.Source: PRIOR WORK · p. 2, INTRODUCTION · p. 2
Evaluation
Prior studies leave open how query sentiment shapes search outputs and how generative AI search engines establish authority. This section reviews audit findings on sentiment bias and research on authority, attribution, and source credibility.
Evaluation
- Prior audits reported varying citation support and system-specific differences, while single-time data collection may miss dynamic output changes.One audit found about 52% of generated sentences citation-supported and 75% of responses fully supported; Perplexity cited more sentences, while Bing Chat paraphrased cited websites more closely.Source: Evaluation · p. 2
Sentiment Bias
- Search-output sentiment can vary by topic, and users may prefer positive or belief-confirming results, making query- and topic-related sentiment bias relevant to information behavior.Earlier studies found health queries returned more confirmatory and positive results, while controversial queries returned more negative-oriented results.Source: Sentiment Bias · pp. 2–3
Authority and quality in search results
- Search engines are treated as cognitive authorities, but it remains unclear how generative AI search systems establish authority through claims and attribution.People often rely on credibility heuristics rather than in-depth evaluation, while attribution provides evidence and support for system claims.Source: Authority and quality in search results · p. 3
User perception of generative AI search engines
Existing user studies of generative AI search systems are few and report that participants may not understand how generated information is produced. This paper audits system responses to study sentiment and authority across public issues.
User perception of generative AI search engines
- Published user studies have involved students, university personnel, and crowdsourced workers, with participants often unaware of how generated information was produced.These studies also found that generated text could appear fluent and informative, while trust varied among students.Source: User perception of generative AI… · p. 3
Current study
- The study asks whether query and topic sentiment influence generative AI responses and how search-oriented systems establish authority.The authors identify auditing sentiment and examining how claims are presented and supported as open research needs.Source: Current study · p. 3
RESEARCH METHODS
- The authors use a scraping audit across four public-issue topics, operationalizing sentiment bias as a correlation between query and response sentiment polarity.The second analysis stage examines Bing Chat and Perplexity, which include citations to sources, and considers additional response variables.Source: RESEARCH METHODS · pp. 3–4
Systems
The study compares three publicly available systems and selects authentic queries for four topics. Query examples and their sentiment scores are presented in Table 1.
Systems
- The study includes GPT-3.5-powered ChatGPT, GPT-4-powered Bing Chat, and Perplexity; Bing Chat and Perplexity use retrieval-augmented generation and cite sources.The systems’ implementations are proprietary, and Google’s system was unavailable in the study country at data collection.Source: Systems · p. 4
Topics and Queries
- For each of four topics, the authors selected 12 frequent English-language queries from Canada, choosing queries with substantial sentiment variation within each topic.Table 1 gives example queries and their sentiment polarity scores on a -1 to +1 scale.Source: Topics and Queries · p. 4, Table 1 · p. 4
Data Collection
The study manually collected responses across three systems over seven days, then analyzed sentiment and response rhetoric and classified cited-source domains. Its methods also used independent coding and reliability checks for the rhetorical analysis.
Data Collection
- Over seven days in June 2023, researchers manually submitted 48 queries daily to ChatGPT, Bing Chat, and Perplexity in Canada, using default settings and recording first responses.They used fresh sessions and different public computers to reduce personalization and session-memory effects.Source: Data Collection · p. 4
Data Preparation and Analysis
- For 1,008 responses, the first analysis phase measured response length and sentiment polarity, then assessed polarity correlations and group differences by topic and system.Sentiment was scored with a fine-tuned transformer model; statistical analyses used Pearson correlation and group-difference testing.Source: Data Preparation and Analysis · pp. 4–5
- For 672 Bing Chat and Perplexity responses, researchers assessed readability and inductively coded rhetorical features; revised coding achieved acceptable agreement (α=0.71).The coders developed a framework covering narrative voice, stance, and techniques for supporting or mitigating claims.Source: Data Preparation and Analysis · p. 5
- Researchers analyzed 3,448 cited sources across 355 unique web domains, assigning each domain a geographic origin and one of 14 inductively developed source categories.Domains were aggregated by system and topic for quantitative descriptive analysis.Source: Data Preparation and Analysis · p. 5
RESULTS
Responses averaged 234 words, with substantial differences in length across systems but little variation by topic within a system. Bing Chat and Perplexity cited just over five sources per response across systems and topics.
RESULTS
- ChatGPT responses averaged 336 words, compared with 243 for Perplexity and 123 for Bing Chat; Table 3 presents response length and source counts by system and topic.Response length did not vary substantially across topics within each system.Source: RESULTS · p. 5, Table 3 · p. 6
- Bing Chat and Perplexity responses cited just over five sources per response, consistently across systems and topics.Table 3 reports response length and source counts by system and topic.Source: RESULTS · p. 5, Table 3 · p. 6
- Across all topics and systems, response length averaged 234 words.Source: RESULTS · p. 5
Query Sentiment Bias
Response polarity was positively associated with query polarity, and response polarity also differed significantly by system and topic. The two systems examined in the response analysis differed in readability, narrative voice, and how they handled query claims.
Query Sentiment Bias
- Query polarity and response polarity had a moderate positive correlation (r = .46, p < .001), with query polarity explaining 21% of response-polarity variation.Similar correlations were reported for ChatGPT (r = .435), Bing Chat (r = .454), and Perplexity (r = .509), each with p < .001.Source: Query Sentiment Bias · p. 6
System and Topical Sentiment Bias
- Perplexity responses had lower median polarity than ChatGPT and Bing Chat, while climate-change and trust-in-media responses were less positive than vaccination and alternative-energy responses.Differences were significant between systems (χ2(2) = 12.54, p = .002) and topics (χ2(3) = 299.36, p < .001); the reported medians were -.19 for Perplexity, -.01 for ChatGPT, -.10 for Bing Chat, -.33 for climate change, -.39 for trust in media, .25 for vaccination, and .24 for alternative energy.Source: System and Topical Sentiment Bias · p. 6
Response Analysis
- Perplexity responses had a higher mean reading grade level than Bing Chat responses (12.82 versus 9.87); Bing Chat tended toward first-person voice, while Perplexity often invoked search results as authority.The reported standard deviations were 2.00 for Perplexity and 2.11 for Bing Chat.Source: Response Analysis · p. 6
- Most responses supported explicit or implicit query claims; opposition was limited (10–12%) and mainly appeared for counter-factual queries.When disagreeing about vaccine mandates, Bing Chat sometimes presented differing views or asserted an ethical position citing health authorities, while Perplexity identified the issue as controversial.Source: Response Analysis · p. 6
Feature
Bing Chat and Perplexity often invoke generic or named authorities and use hedging, while rarely acknowledging response limits. Their sources vary by region, type, system, and topic, and include both authoritative and questionable sites.
Feature
- More than half of Bing Chat and Perplexity responses invoke generic or named authorities, while hedging is common and admissions of limits are rare.Perplexity also uses vague majority or minority viewpoints; the rhetorical-feature summary compares the two systems.Source: Feature · p. 7, Table 4 · p. 7
Source Analysis
- Of 3,448 citations across 355 domains, 95% of domains originated in the U.S., UK, Canada, or international organizations; 65% of references were U.S.-based.Bing cited more UK and Canadian sources and fewer U.S. sources than Perplexity.Source: Source Analysis · p. 7, Figure 1 · p. 8
- News and Media was the most frequent source category (21.8%), followed by Business (17.9%), Government (16.5%), Non-Profit (13.3%), and Digital Media (11.7%).Educational institutions, academic publishers and journals, policy institutes, and professional and scholarly associations each accounted for 2.6%–3.1%.Source: Source Analysis · p. 8, Table 5 · p. 8
- Bing cited proportionally more News and Digital Media sources, while Perplexity cited more Business and Non-Profit sources; source categories also varied by topic.Government sources featured heavily for climate and vaccination, Business for renewable energy, and News and Non-Profit for Trust in Media.Source: Source Analysis · p. 8
- Source quality was inconsistent, ranging from relevant authoritative agencies and news organizations to questionable sites, essay services, assignment content, denial articles, and content farms.Examples included consumer health websites and company sites offering energy services or related finance, investment, and employment content.Source: Source Analysis · pp. 8–9
DISCUSSION
The audit found query- and topic-related sentiment bias, while systems used different rhetorical styles and sources to establish authority. Sources showed commercial and geographic biases, and uneven quality; Bing Chat and Perplexity also had distinct response profiles.
DISCUSSION
- Query wording and topic both significantly influenced response sentiment; responses echoed query valence, while solution-focused topics were more positive than problem-focused topics.The authors caution that the query-response correlation does not establish cause and effect.Source: DISCUSSION · p. 9
- Bing Chat spoke as an authoritative knower, whereas Perplexity presented itself as an intermediary relaying search results and other authorities.The authors describe Perplexity’s approach as more faithful to the hybrid retrieval-augmented generation model.Source: DISCUSSION · p. 9
- Responses used authorities, statistics, hedging, and balance checks, but minority-voice framing on scientifically grounded topics risked implying false equivalence.The authors connect balance checks to the media legacy of prioritizing apparent neutrality over scientific consensus, including on climate change.Source: DISCUSSION · p. 9
- Sources relied heavily on news and media, with business websites especially prominent in Perplexity; the authors suggest renewable energy queries were read as commercial in intent.They note that commercial sources may reflect search optimization or intentional native advertising, but do not establish which explanation applies.Source: DISCUSSION · p. 9
- Sources were heavily concentrated in the U.S. and other western countries, while Canadian sources were scarce despite queries being issued in Canada.Bing returned more Canadian results for vaccination, suggesting some geolocation use, though the extent of geographic personalization remains unknown.Source: DISCUSSION · p. 10
- Although many cited sources were appropriate, responses also included poor-quality commercial blogs, clickbait, essay mills, and content farms.The authors argue that each cited resource matters because sources are integral to a generated response rather than merely items in a ranked list.Source: DISCUSSION · p. 10
- Bing Chat was shorter, more conversational, easier to read, more positive, and drew from fewer domains than Perplexity, which used a more academic tone and was more U.S.-centric and commercial.The authors suggest these distinct profiles point toward more adaptive and hyper-personalized search engines.Source: DISCUSSION · p. 10
CONCLUSION
The study found sentiment bias in responses and commercial and geographic bias in cited sources, alongside differences in how Bing Chat and Perplexity presented authority. Its findings are bounded by four broad topics at one point in time and the absence of human participants.
CONCLUSION
- The audit found sentiment bias related to queries and topics, as well as commercial and geographic bias in cited sources.Bing Chat and Perplexity adopted different profiles; responses sought cognitive authority, but ambiguous language and questionable sources weakened their claims.Source: CONCLUSION · p. 10
- The study examined four broad public-interest topics at one time point, June 2023, limiting claims about general patterns.The authors call for future work across a wider range of topics and tasks.Source: CONCLUSION · p. 10
- The study had no human participants, so its findings about authority, credibility, and source quality require validation through user studies.The authors identify user studies as future work.Source: CONCLUSION · p. 10
GENERATIVE AI USE
The authors disclose that the study examined Generative AI system outputs and did not use them elsewhere in the research process.
GENERATIVE AI USE
- The study examined outputs from Generative AI systems.Source: GENERATIVE AI USE · p. 10
- The authors did not use those outputs in any other steps of analysis or writing.Source: GENERATIVE AI USE · p. 10
- The authors assume responsibility for the submission’s content.Source: GENERATIVE AI USE · p. 10