Transparency and Traceability in AI Citations

Transparency and traceability in AI citations represent critical mechanisms for establishing accountability and verifiability in artificial intelligence systems that generate, retrieve, or synthesize information from source materials ¹². This framework encompasses the technical and methodological approaches that enable users to understand how AI systems attribute information to original sources, track the provenance of generated content, and verify the accuracy of citations ³. The primary purpose is to maintain intellectual integrity, combat misinformation, and ensure that AI-generated content can be audited and validated against authoritative sources ⁵. In an era where large language models and retrieval-augmented generation systems increasingly mediate information access, transparent and traceable citation mechanisms are essential for maintaining trust, enabling fact-checking, and preserving the scholarly and informational ecosystem's integrity ⁹.

Overview

The emergence of transparency and traceability in AI citations stems from the rapid advancement of large language models and their deployment in information-intensive domains where accuracy and accountability are paramount ¹⁸. As these systems began generating increasingly sophisticated content, concerns arose about their tendency to produce plausible-sounding but factually incorrect information—a phenomenon known as "hallucination" ⁵. The fundamental challenge these mechanisms address is the opacity of neural language models, which traditionally function as "black boxes" that synthesize information from vast training corpora without explicit attribution to specific sources ²⁹.

The practice has evolved significantly from early rule-based citation systems to sophisticated neural attribution methods. Initial approaches relied on simple retrieval-augmented generation architectures that appended source documents to model inputs ⁸. More recent developments include attention-based attribution mechanisms, contrastive evaluation methods like ALCE (Automatic LLM Citation Evaluation), and training paradigms that explicitly reward models for generating verifiable, attributable content ¹³. Systems such as WebGPT demonstrated that language models could be trained through reinforcement learning to browse sources and cite them appropriately, marking a significant evolution toward inherently transparent AI systems ⁸.

Key Concepts

Attribution Granularity

Attribution granularity refers to the specificity level at which AI systems link generated content to source materials, ranging from document-level citations to sentence-level or even token-level attribution ¹. This concept determines how precisely users can verify the provenance of specific claims within AI-generated content.

Example: A medical AI assistant generating a treatment recommendation might provide document-level attribution by citing "Smith et al., 2023" for general guidance, but when stating a specific dosage recommendation, it employs sentence-level granularity by linking that exact claim to paragraph 3 of the cited study, allowing physicians to verify the precise source of critical clinical information before implementing the recommendation.

Provenance Graphs

Provenance graphs are structured representations of information lineage that map the flow of data from original sources through various processing stages to final outputs ²³. These directed acyclic graphs (DAGs) enable both forward tracing (identifying what outputs derive from a given source) and backward tracing (identifying which sources contributed to a specific output).

Example: A legal research AI system analyzing a complex case creates a provenance graph showing that its conclusion about precedent applicability derives from three Supreme Court cases. The graph reveals that Case A was retrieved via semantic search, Case B was identified through citation network analysis of Case A, and Case C was added after the system detected a contradiction between Cases A and B, requiring reconciliation. This graph allows attorneys to understand not just which cases were cited, but why and how they were discovered and integrated.

Retrieval Transparency

Retrieval transparency encompasses the visibility into how AI systems identify, rank, and select source documents from knowledge bases or the web ¹⁸. This includes exposing the query formulation process, ranking algorithms, similarity scores, and the complete candidate set considered, not merely the sources ultimately used.

Example: An AI-powered journalism tool researching climate policy retrieves 50 candidate documents using dense vector embeddings. The system's transparency interface shows journalists that it ranked sources using a weighted combination of semantic similarity (40%), source authority based on domain reputation (30%), recency (20%), and geographic relevance (10%). It reveals that while a think tank report ranked third overall, it was excluded because its confidence score fell below the 0.75 threshold, allowing journalists to manually review this borderline exclusion.

Attribution Confidence Scores

Attribution confidence scores are probabilistic measures indicating the certainty with which an AI system can trace generated content to specific sources ¹⁵. These scores acknowledge that not all citations carry equal reliability, particularly when content synthesizes information from multiple sources or involves paraphrasing.

Example: A research synthesis AI generates the statement "Recent studies suggest that intermittent fasting may improve metabolic markers in adults over 50." It assigns confidence scores to its attributions: 0.92 confidence for linking "intermittent fasting" and "metabolic markers" to a 2023 meta-analysis, but only 0.68 confidence for the "adults over 50" specification, as this demographic detail was inferred from subgroup analyses across three studies rather than stated explicitly in any single source. Researchers can prioritize verification efforts on lower-confidence attributions.

Contrastive Attribution Methods

Contrastive attribution methods determine source influence by comparing model behavior with and without access to specific sources ¹³. These techniques isolate the causal impact of individual documents on generated outputs, providing more reliable attribution than attention-based approaches.

Example: An AI system generates a product recommendation based on five customer reviews. To verify attribution, the system regenerates the recommendation five times, each time withholding one review. When Review C is withheld, the recommendation changes from "excellent for beginners" to "requires prior experience," definitively establishing that Review C was the critical source for the beginner-friendliness claim, even though attention weights had suggested Review A was more influential.

Hallucination Detection Through Traceability

Hallucination detection leverages citation traceability to identify generated content that lacks support in retrieved sources ⁵⁹. By maintaining tight coupling between outputs and source materials, systems can automatically flag unsupported claims for human review.

Example: A financial analysis AI generates a report stating "Company X's revenue grew 23% in Q3 2024." The hallucination detection module attempts to trace this claim back to the retrieved earnings report but finds that the actual figure was 18%. The system flags this discrepancy with high priority, noting that while the claim structure matches source content (indicating attempted attribution), the numerical value represents a hallucination. The system either corrects the figure automatically or alerts users to verify before publication.

Progressive Disclosure in Citation Interfaces

Progressive disclosure is a user interface design principle that presents citation information in layers, showing summary attributions initially while allowing users to drill down into detailed provenance data as needed ². This approach balances comprehensiveness with usability, preventing information overload.

Example: A scientific literature review AI presents its synthesis with superscript numbers [1-5] indicating five supporting sources. Hovering over a citation reveals a tooltip with the paper title and relevance score. Clicking opens a side panel showing the specific passage cited, its context within the source document, why this source was selected over alternatives, and the full provenance graph showing how this source was discovered. Users conducting casual reading see only the superscripts, while those performing rigorous verification can access complete attribution details.

Applications in AI-Powered Information Systems

Legal Research and Case Analysis

AI systems in legal research employ transparent citation mechanisms to provide attorneys with verifiable analysis of case law, statutes, and legal precedents ⁸. These applications must meet exceptionally high standards for attribution accuracy, as legal arguments depend on precise source references. Systems like Casetext's CoCounsel generate legal memoranda with citations to specific case law, enabling lawyers to verify reasoning before submitting briefs. The traceability infrastructure logs which legal databases were searched, how cases were ranked by relevance and jurisdiction, and which passages influenced specific legal conclusions. This transparency is essential for professional responsibility standards requiring attorneys to verify AI-assisted work.

Healthcare Clinical Decision Support

Clinical decision support systems leverage citation transparency to justify treatment recommendations with references to medical literature, clinical guidelines, and evidence-based protocols ⁹. A diagnostic AI analyzing patient symptoms might recommend specific tests while citing relevant medical studies and clinical practice guidelines. The system provides confidence scores for each recommendation and traces its reasoning through the evidence hierarchy, distinguishing between Level 1 evidence from randomized controlled trials and lower-quality observational studies. Physicians can verify that recommendations align with current standards of care and understand the strength of supporting evidence before making treatment decisions. This traceability also supports medical-legal documentation requirements.

Academic Research Synthesis and Literature Review

Research tools like Elicit and Consensus employ transparent citation mechanisms to help scholars navigate vast academic literature ¹⁸. These systems retrieve relevant papers, synthesize findings across studies, and provide paper-level citations enabling researchers to verify claims and explore primary sources. A researcher investigating "machine learning applications in drug discovery" receives a synthesis of 50 relevant papers with claims like "Graph neural networks have shown 15-20% improvement in binding affinity prediction" linked to specific papers and passages. The provenance system reveals that this synthesis combined results from three papers, averaged their reported improvements, and identified the graph neural network commonality through semantic analysis. Researchers can trace this synthesis process to assess whether the aggregation methodology is appropriate for their purposes.

AI-Assisted Journalism and Fact-Checking

News organizations experiment with AI systems that assist journalists in research, fact-checking, and content generation while maintaining rigorous source attribution ²⁹. An AI tool researching a breaking news story retrieves information from wire services, official statements, and social media, ranking sources by credibility and recency. When generating a draft article, the system provides inline citations for factual claims and flags statements that lack sufficient source support. The transparency interface shows journalists which sources were considered but excluded (perhaps due to credibility concerns), enabling editorial judgment about whether to manually investigate those sources. This traceability helps maintain journalistic standards while leveraging AI efficiency.

Best Practices

Implement Multi-Method Attribution Validation

Rather than relying on a single attribution technique, robust systems employ ensemble methods that combine attention-based attribution, contrastive evaluation, and explicit source matching ¹³. The rationale is that different attribution methods have complementary strengths and weaknesses—attention weights provide computational efficiency but can be noisy, while contrastive methods offer higher accuracy but greater computational cost.

Implementation Example: A question-answering system first uses attention-based attribution to identify candidate source passages for each generated sentence, achieving real-time performance. For high-stakes outputs (flagged by user settings or content domain), the system applies contrastive attribution by regenerating answers without suspected sources to confirm their influence. Finally, it performs lexical matching to verify that attributed passages actually contain semantically similar content. Only attributions confirmed by at least two methods receive high confidence scores, while single-method attributions are flagged for user review.

Maintain Comprehensive Retrieval Logs with Ranking Transparency

Systems should log not only the sources ultimately cited but the complete candidate set considered, along with ranking scores and the factors influencing selection ¹⁸. This practice enables auditing of potential biases in source selection and helps users understand why certain sources were prioritized over alternatives.

Implementation Example: An AI research assistant searching for information about renewable energy policy logs all 200 documents retrieved, recording for each: semantic similarity score (0-1), source authority score based on domain reputation, publication recency, geographic relevance, and whether the source was ultimately used. When presenting results, the system allows users to view "alternative sources" that scored highly but weren't selected, perhaps because they were slightly less recent. This transparency enables researchers to identify potentially valuable sources the AI deprioritized and to audit whether ranking factors align with their research priorities.

Implement Source Stability Mechanisms and Persistent Identifiers

Given that web sources may disappear or change, systems should archive retrieved content (respecting copyright), prioritize sources with persistent identifiers like DOIs, and implement fallback mechanisms when sources become unavailable ²⁹. This practice ensures that citations remain verifiable over time and that attribution doesn't break as the web evolves.

Implementation Example: A citation-aware AI system retrieves a mixture of academic papers (with DOIs), government reports (with stable URLs), and news articles (potentially volatile). For academic papers, it stores only metadata and DOIs, relying on persistent scholarly infrastructure. For government reports, it archives the full text in a local repository with checksums to detect changes. For news articles, it captures both the full text and a screenshot at retrieval time, and subscribes to the Internet Archive's Wayback Machine API to create permanent snapshots. When presenting citations, it indicates source stability (stable/archived/volatile) so users understand verification reliability.

Design Progressive Disclosure Interfaces Tailored to Verification Workflows

Citation interfaces should present information in layers appropriate to different user needs, from casual readers requiring minimal citation clutter to expert verifiers needing complete provenance details ². The rationale is that comprehensive attribution data, while valuable, can overwhelm users if presented all at once, reducing rather than enhancing trust.

Implementation Example: A medical information AI presents three interface modes: "Patient View" shows only that information is "based on medical research" with a single link to sources; "Clinician View" displays inline citations with confidence scores and quick-access tooltips showing cited passages; "Researcher View" provides the full provenance graph, retrieval logs, attribution method details, and alternative sources considered. Users can switch modes based on their current task, ensuring that citation transparency enhances rather than impedes their workflow.

Implementation Considerations

Tool and Format Choices for Attribution Infrastructure

Implementing transparent citation systems requires selecting appropriate technologies for retrieval, attribution tracking, and provenance storage ¹⁸. Vector databases like Pinecone, Weaviate, or Milvus enable efficient dense retrieval using embedding models, while graph databases like Neo4j excel at representing complex provenance relationships. Observability platforms such as LangSmith or Weights & Biases provide logging and debugging infrastructure for tracking attribution decisions. The choice depends on factors including scale (number of documents and queries), attribution granularity requirements (document vs. sentence-level), and whether provenance queries will be simple (backward tracing only) or complex (multi-hop reasoning about source relationships).

Example: An enterprise AI system handling 10 million documents with complex multi-step reasoning implements a hybrid architecture: Weaviate for vector-based retrieval with built-in filtering capabilities, Neo4j for provenance graphs enabling queries like "find all outputs influenced by sources from domain X," and LangSmith for real-time logging and debugging during development. This combination provides efficient retrieval, sophisticated provenance queries, and developer-friendly debugging, though it requires maintaining consistency across three systems.

Audience-Specific Customization of Citation Granularity

Different user populations require different levels of citation detail ²⁹. Academic researchers may need sentence-level attribution with confidence scores, while general consumers may find document-level citations sufficient. Healthcare applications might require passage-level citations for treatment recommendations but document-level for general health information. Systems should adapt attribution granularity to user expertise, domain risk level, and task requirements.

Example: A health information AI implements adaptive granularity: for general wellness queries ("how much sleep do adults need?"), it provides document-level citations to health organization guidelines; for specific medical questions ("what are the side effects of metformin?"), it offers passage-level citations linking each side effect to specific clinical studies; for professional queries from verified healthcare providers, it provides token-level attribution with confidence intervals and access to full provenance graphs. This adaptation balances verification needs with usability across user populations.

Organizational Maturity and Governance Context

Implementation approaches must align with organizational capabilities and governance requirements ⁵⁹. Organizations with mature data governance may already have metadata standards, access control systems, and audit requirements that citation systems must integrate with. Startups might prioritize rapid deployment with basic attribution, while regulated industries require comprehensive audit trails. The implementation should consider existing infrastructure, compliance requirements, and organizational capacity for maintaining attribution systems.

Example: A financial services firm implementing an AI research assistant must comply with regulatory requirements for documenting the basis of investment recommendations. The system integrates with the firm's existing document management system that tracks access permissions, ensuring that citations respect information barriers between departments. It implements immutable audit logs of all retrieval and attribution decisions, stored in the firm's compliance database with seven-year retention. The system uses the firm's existing metadata taxonomy for source classification, ensuring consistency with other research tools. This integration approach leverages existing governance infrastructure rather than creating parallel systems.

Computational Cost and Performance Trade-offs

Comprehensive attribution tracking can increase latency and resource requirements by 50-200% compared to non-attributed generation ¹³. Implementation must balance attribution completeness with performance requirements through techniques like selective logging (tracking only critical decision points), asynchronous processing (computing detailed attributions in background), and tiered attribution (providing basic citations immediately with detailed provenance available on request).

Example: A customer service AI chatbot implements tiered attribution to maintain response times under 2 seconds: it performs lightweight attention-based attribution in real-time, providing immediate responses with basic citations; for conversations flagged as high-stakes (detected through sentiment analysis or explicit user requests), it triggers background processing that computes detailed contrastive attributions and provenance graphs, making these available within 10 seconds; for routine queries, detailed attribution is computed during off-peak hours for quality monitoring purposes but not presented to users. This approach maintains user experience while ensuring attribution depth matches conversation criticality.

Common Challenges and Solutions

Challenge: Attribution Accuracy in Multi-Source Synthesis

Current attribution methods struggle when AI systems synthesize information from multiple sources in ways that defy simple attribution ¹³. When a generated statement combines facts from three different sources, paraphrases content, or draws inferences that no single source explicitly states, determining accurate attribution becomes highly complex. Attention-based methods often produce noisy attributions because attention weights reflect computational patterns rather than semantic influence. This challenge is particularly acute in domains requiring nuanced synthesis, such as legal analysis or scientific literature review, where conclusions legitimately derive from integrating multiple sources.

Solution:

Implement ensemble attribution methods that combine multiple techniques and provide confidence intervals rather than point estimates ¹³. Use contrastive evaluation to validate attention-based attributions for high-stakes content, regenerating outputs with and without suspected sources to confirm their influence. For synthetic claims that integrate multiple sources, explicitly acknowledge multi-source attribution rather than forcing single-source citations. Train specialized verification models that assess whether generated claims are supported by cited sources, flagging potential misattributions for human review. Provide users with "attribution confidence" indicators that distinguish between high-confidence direct citations and lower-confidence synthetic attributions, enabling them to prioritize verification efforts appropriately.

Challenge: Source Stability and Link Rot

Web sources frequently disappear or change, breaking citations and making verification impossible ²⁹. Research indicates that approximately 20% of web links become inaccessible within two years, with higher rates for news articles and social media content. This "link rot" undermines the fundamental purpose of citations—enabling verification—and creates particular problems for AI systems that may retrieve and cite sources that later become unavailable. The challenge is compounded by copyright concerns that limit the ability to archive full source content.

Solution:

Implement a multi-tiered source stability strategy that treats different source types appropriately ². For academic sources with DOIs or other persistent identifiers, rely on scholarly infrastructure and store only metadata. For government and institutional sources with stable URLs, implement periodic link checking and maintain archived copies with checksums to detect changes. For volatile sources like news articles and social media, capture full content and screenshots at retrieval time (within fair use bounds), and integrate with archival services like the Internet Archive's Wayback Machine to create permanent snapshots. In the citation interface, indicate source stability levels (persistent/archived/volatile) so users understand verification reliability. For enterprise applications, implement a local document repository that archives all cited sources, ensuring long-term verifiability independent of external source availability.

Challenge: Computational Overhead and Latency

Maintaining detailed provenance logs, computing attribution scores, and storing comprehensive metadata significantly increases computational costs and response latency ¹⁸. Systems implementing full attribution tracking may experience 50-200% increases in processing time and storage requirements compared to non-attributed generation. This overhead is particularly problematic for real-time applications like chatbots or interactive research assistants, where users expect sub-second response times. The challenge involves balancing attribution completeness with performance requirements without compromising user experience.

Solution:

Implement adaptive attribution strategies that match tracking depth to content criticality and user needs ¹. Use lightweight attention-based attribution for real-time responses, providing immediate basic citations. For high-stakes content (identified through domain classification, user settings, or explicit requests), trigger more computationally intensive contrastive attribution methods in background processes, making detailed provenance available within seconds rather than blocking initial responses. Employ selective logging that captures critical decision points (retrieval results, source rankings, generation parameters) while omitting redundant intermediate steps. Use efficient data structures like compressed provenance graphs and incremental logging that minimizes storage overhead. Implement caching strategies that reuse attribution computations for frequently accessed sources. For batch processing scenarios like document analysis, perform comprehensive attribution without real-time constraints, while interactive applications use tiered approaches that balance immediacy with depth.

Challenge: User Interface Complexity and Information Overload

Comprehensive attribution data, while valuable for verification, can overwhelm users if presented poorly ². Displaying citations for every sentence creates visual clutter that impedes reading comprehension. Showing complete provenance graphs with retrieval logs, ranking factors, and confidence scores may confuse non-technical users. User studies indicate that poorly designed citation interfaces actually reduce trust rather than enhancing it, as users perceive the system as uncertain or the information as unreliable when confronted with excessive attribution detail. The challenge involves presenting sufficient information for verification without creating cognitive overload.

Solution:

Design progressive disclosure interfaces that present citation information in layers appropriate to user expertise and current tasks ². Implement a default view with minimal visual disruption—perhaps superscript citation numbers or subtle highlighting—that indicates attribution exists without dominating the content. Provide intermediate detail through hover tooltips or expandable sections that show cited passages and relevance scores. Reserve complete provenance information (retrieval logs, ranking factors, attribution methods, confidence intervals) for dedicated "verification mode" or expert interfaces. Conduct user testing with representative populations to optimize the balance between transparency and usability. Implement user preferences that allow individuals to customize citation detail levels based on their expertise and verification needs. For different domains, adapt default citation styles to match existing conventions—academic users expect footnotes or inline citations, while general consumers may prefer linked "source" sections.

Challenge: Handling Conflicting or Contradictory Sources

AI systems frequently retrieve sources that contradict each other, particularly on controversial topics or rapidly evolving situations ⁵⁹. When sources disagree, systems face difficult choices: cite only the most authoritative source (potentially introducing bias), present all perspectives (risking user confusion), or attempt to synthesize a consensus (potentially misrepresenting the actual state of knowledge). Concealing source conflicts undermines transparency, but presenting them without context may reduce user trust or create decision paralysis. This challenge is particularly acute in domains like health information, policy analysis, and current events where legitimate expert disagreement exists.

Solution:

Implement explicit conflict detection and transparent disagreement presentation ⁵⁹. Use semantic analysis to identify when retrieved sources make contradictory claims about the same topic. When conflicts are detected, explicitly acknowledge disagreement rather than concealing it: "Sources disagree on this point. Source A reports X, while Source B reports Y." Provide context about why sources might disagree (different methodologies, time periods, or perspectives) and, when possible, information about source credibility and authority. For domains with established authority hierarchies (like medicine with its evidence levels), implement ranking systems that prioritize higher-quality sources while still disclosing alternative views. Allow users to explore the full range of perspectives through expandable sections or alternative view modes. In some cases, abstain from generating a definitive answer when sources substantially conflict, instead presenting the range of views with appropriate citations—acknowledging uncertainty is often more trustworthy than false confidence.

References

1. Gao, T., Yen, H., Yu, J., & Chen, D. (2023). Enabling Large Language Models to Generate Text with Citations. arXiv:2305.14627. https://arxiv.org/abs/2305.14627
2. Menick, J., Trebacz, M., Mikulik, V., Aslanides, J., Song, F., Chadwick, M., Glaese, M., Young, S., Campbell-Gillingham, L., Irving, G., & McAleese, N. (2022). Teaching language models to support answers with verified quotes. arXiv:2211.09110. https://arxiv.org/abs/2211.09110
3. Bohnet, B., Tran, V. Q., Verga, P., Aharoni, R., Andor, D., Soares, L. B., Ciaramita, M., Eisenstein, J., Ganchev, K., Herzig, J., Hui, K., Kwiatkowski, T., Ma, J., Ni, J., Saralegui, T., Schuster, T., Strohman, W., Teh, Y. W., Yavuz, P., & Zelle, D. (2023). Attributed Question Answering: Evaluation and Modeling for Attributed Large Language Models. arXiv:2301.10120. https://arxiv.org/abs/2301.10120
4. Anthropic. (2023). Measuring Faithfulness in Chain-of-Thought Reasoning. https://www.anthropic.com/index/measuring-faithfulness-in-chain-of-thought-reasoning
5. Liu, T., Zhang, Y., Brockett, C., Dolan, B., & Gao, J. (2023). Making Language Models Better Reasoners with Step-Aware Verifier. arXiv:2305.14739. https://arxiv.org/abs/2305.14739
6. Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., Jiang, X., Cobbe, K., Eloundou, T., Krueger, G., Button, K., Knight, M., Chess, B., & Schulman, J. (2021). WebGPT: Browser-assisted question-answering with human feedback. https://research.google/pubs/pub51058/
7. Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W., Jiang, X., Cobbe, K., Eloundou, T., Krueger, G., Button, K., Knight, M., Chess, B., & Schulman, J. (2021). WebGPT: Browser-assisted question-answering with human feedback. arXiv:2112.09332. https://arxiv.org/abs/2112.09332
8. Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., Brynjolfsson, E., Buch, S., Card, D., Castellon, R., Chatterji, N., Chen, A., Creel, K., Davis, J. Q., Demszky, D., Donahue, C., Doumbouya, M., Durmus, E., Ermon, S., Etchemendy, J., Ethayarajh, K., Fei-Fei, L., Finn, C., Gale, T., Gillespie, L., Goel, K., Goodman, N., Grossman, S., Guha, N., Hashimoto, T., Henderson, P., Hewitt, J., Ho, D. E., Hong, J., Hsu, K., Huang, J., Icard, T., Jain, S., Jurafsky, D., Kalluri, P., Karamcheti, S., Keeling, G., Khani, F., Khattab, O., Koh, P. W., Krass, M., Krishna, R., Kuditipudi, R., Kumar, A., Ladhak, F., Lee, M., Lee, T., Leskovec, J., Levent, I., Li, X. L., Li, X., Ma, T., Malik, A., Manning, C. D., Mirchandani, S., Mitchell, E., Munyikwa, Z., Nair, S., Narayan, A., Narayanan, D., Newman, B., Nie, A., Niebles, J. C., Nilforoshan, H., Nyarko, J., Ogut, G., Orr, L., Papadimitriou, I., Park, J. S., Piech, C., Portelance, E., Potts, C., Raghunathan, A., Reich, R., Ren, H., Rong, F., Roohani, Y., Ruiz, C., Ryan, J., Ré, C., Sadigh, D., Sagawa, S., Santhanam, K., Shih, A., Srinivasan, K., Tamkin, A., Taori, R., Thomas, A. W., Tramèr, F., Wang, R. E., Wang, W., Wu, B., Wu, J., Wu, Y., Xie, S. M., Yasunaga, M., You, J., Zaharia, M., Zhang, M., Zhang, T., Zhang, X., Zhang, Y., Zheng, L., Zhou, K., & Liang, P. (2023). On the Opportunities and Risks of Foundation Models. Nature Machine Intelligence, 5, 456-477. https://www.nature.com/articles/s42256-023-00653-1
9. Thorne, J., Vlachos, A., Christodoulopoulos, C., & Mittal, A. (2022). Evaluating Factuality in Generation with Dependency-level Entailment. arXiv:2205.11916. https://arxiv.org/abs/2205.11916
10. Gao, L., Ma, X., Lin, J., & Callan, J. (2023). Precise Zero-Shot Dense Retrieval without Relevance Labels. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, 1762-1777. https://aclanthology.org/2023.acl-long.386/