# Trust Me on This: A User Study of Trustworthiness for RAG Responses
**Authors**: Weronika Ćajewska, Krisztian Balog
> Work done while at University of Stavanger, prior to joining Amazon.
institutetext: Amazon institutetext: University of Stavanger, Stavanger, Norway email: lajewska@amazon.lu, krisztian.balog@uis.no
## Abstract
The integration of generative AI into information access systems often presents users with synthesized answers that lack transparency. This study investigates how different types of explanations can influence user trust in responses from retrieval-augmented generation systems. We conducted a controlled, two-stage user study where participants chose the more trustworthy response from a pairâone objectively higher quality than the otherâboth with and without one of three explanation types: (1) source attribution, (2) factual grounding, and (3) information coverage. Our results show that while explanations significantly guide users toward selecting higher quality responses, trust is not dictated by objective quality alone: Usersâ judgments are also heavily influenced by response clarity, actionability, and their own prior knowledge.
## 1 Introduction
As generative AI becomes integrated into information access systems, from conversational agents to summaries on search engine result pages, users are increasingly presented with concise responsesâoften just a few sentencesâwithout visibility into which sources the response was based on. This shift conceals crucial information that users typically rely on to assess the novelty, reliability, and relevance of retrieved content [25]. Retrieval-augmented generation (RAG) [13, 7, 10, 20] has recently emerged as a prominent approach to generating more factually grounded and diverse responses by leveraging external documents. However, RAG models still fall short on transparency: they offer no indication of low-confidence outputs or known limitations, whether stemming from incomplete retrieval or flaws in generation. Because users see only the final output, the responsibility lies with the system to surface potential issues, thereby promoting transparency and enabling users to judge the response quality. Although explainability has been extensively studied in areas such as decision support and recommender systems [2, 18, 26, 23, 14, 24, 27], it remains underexplored in the context of retrieval-augmented response generation.
While prior work establishes that explanations can enhance the user-perceived usefulness of a generated response [12], trust operates on a different, more critical level. Usefulness can be seen as the immediate utility of a response for a given task, relating to its clarity, format, or relevance. Trust, however, concerns the userâs belief in the credibility and factual correctness of the information itself. A user might find a well-structured summary useful for getting a quick overview, but they will only trust it for a critical decision if they believe it to be factually sound. Therefore, trust is a distinct and vital prerequisite for the meaningful adoption of these systems, moving beyond surface-level utility to user confidence and reliance. Therefore, this paper investigates the next logical step by asking: How do different types of explanations influence user trust in responses from retrieval-augmented generation systems?
To answer this question, we conducted a controlled user study with a two-stage, within-subject design. We extended an existing response generation pipeline with a post-hoc module to generate three distinct types of explanations: (1) source attribution via supporting passages, (2) factual grounding by linking response statements to sources, and (3) information coverage through the surfacing of relevant but omitted aspects of the topic. Study participants were shown a pair of responses for a given query, where one response was of higher objective quality than the other. Initially, they chose the response they found more trustworthy. Subsequently, they were presented with the exact same pair of responses, but now enhanced with one of the explanation types, and were prompted to make their trust judgment again. This two-stage design enabled a direct measurement of how explanations altered a userâs perception of trustworthiness.
The results of our study show that usersâ trust judgments do not always align with the objective quality of responses. Many participants preferred the objectively lower quality response when it was clearer, more detailed, or more actionable. Source attribution had the strongest positive effect on trust, but only in factual or technical contexts, while it was largely ignored in subjective questions. Finally, as indicated in the free-text comments provided by users, the relevance of information and usersâ own background knowledge strongly shaped trust decisions, with some participants dismissing or overlooking explanations when they felt confident in their own understanding of the topic. All collected user data, input responses, and corresponding annotations from this study are available in the public repository: https://github.com/iai-group/trustworthy-rag/.
## 2 Related Work
Trust is a key factor in humanâmachine interaction, particularly in information-intensive tasks such as retrieval-augmented response generation, which synthesizes relevant information into concise, comprehensive responses [13]. System trustworthiness is not inherent but is instead communicated through trust cuesâfeatures of the interface, documentation, or explanations that shape usersâ judgments [16]. When these cues are misleading or poorly designed, they can foster misplaced trust or confusion [16]. Because such cues are often processed heuristically rather than analytically, users tend to rely on mental shortcuts that can result in inconsistent or biased trust assessments. Providing explanations is one approach to addressing these challenges: explainable systems are expected to clarify their capabilities, communicate uncertainty, and make their reasoning transparent [1]. However, explanations can have mixed effectsâthey can enhance trust but may also lead to overtrust or inaccurate mental models, especially among users with low domain expertise [4].
Generated responses often contain unsupported claims or incorrect citations [17]. Commercial conversational search systems, such as Perplexity.ai, Microsoft Copilot, and Google Gemini, display source links, but explanations are generally limited to attribution, overlooking other important aspects. Additional factors that influence trust include bias [6, 5], unanswerability [8, 9], and response completeness [3]. Prior work on trustworthy RAG and question answering has focused mainly on grounding and factual correctness (e.g., avoiding hallucinations) rather than directly on user trust [15, 21, 17, 22]. High-quality explanations related to source, system confidence, and response limitations have been shown to improve the perceived usefulness of responses [12], echoing findings in broader AI research where calibrated confidence improves trust [27]. However, their effect on user trust remains largely unexplored. In this work, we adopt a user-centered perspective to investigate how different explanation strategies shape trust judgments in RAG responses for information-seeking queries.
<details>
<summary>x1.png Details</summary>
### Visual Description
\n
## Screenshot: Instructions for a Response Evaluation Task
### Overview
This is a screenshot of an instruction page for a task involving the evaluation of two response variants to a question. The page presents two sets of response pairs, one labeled "Response 1 (no explanations)" and "Response 2 (no explanations)", and another labeled "Response 1 (with explanations)" and "Response 2 (with explanations)". Each pair is followed by a question asking the user to select which response they trust more, and a text box for justification.
### Components/Axes
The screenshot contains the following components:
* **Header:** "Instructions" in large font.
* **Introductory Text:** A paragraph explaining the task's objective â to compare response variants and select the more trustworthy one based on initial judgment.
* **Question 1 (No Explanations):**
* Two response boxes labeled "Response 1 (no explanations)" and "Response 2 (no explanations)".
* A question: "Which aspects/facets/points of view are discussed in this response? (Select all that are discussed)".
* Two checkboxes labeled "aspect 1" and "aspect 2".
* A radio button question: "Which response do you trust more?". Options are: "Trust Response A a lot more", "Trust Response A slightly more", "Trust them about the same", "Trust Response B slightly more", "Trust Response B a lot more".
* A text box: "In your own words, explain your preferences and justify your choice."
* **Question 1 (With Explanations):** Identical structure to the "No Explanations" section, but labeled "Response 1 (with explanations)" and "Response 2 (with explanations)".
### Detailed Analysis or Content Details
The text content is as follows:
"You will be presented with multiple questions. For each question, you will see two pairs of responses. Read two variants of responses, indicate which aspects of the topic they cover, and decide which response do you trust more. We want to know your first impression, so do not change your responses once you move to the next question. Rely solely on your judgment and refrain from using additional sources other than the ones provided in this task."
Within each question section:
* "Which aspects/facets/points of view are discussed in this response? (Select all that are discussed)"
* "aspect 1"
* "aspect 2"
* "Which response do you trust more?"
* "Trust Response A a lot more"
* "Trust Response A slightly more"
* "Trust them about the same"
* "Trust Response B slightly more"
* "Trust Response B a lot more"
* "In your own words, explain your preferences and justify your choice."
### Key Observations
The screenshot is designed as a user interface for a survey or experiment. The repeated structure of the question pairs suggests a comparative evaluation process. The emphasis on "first impression" and avoiding external sources indicates a focus on intuitive judgment. The inclusion of checkboxes for "aspects" and a justification text box encourages detailed consideration of the responses.
### Interpretation
The screenshot represents a methodology for gathering subjective data on the perceived quality or trustworthiness of different responses. The task is designed to minimize bias by restricting information access and emphasizing initial reactions. The structure allows for both quantitative (trust level) and qualitative (justification) data collection. The presence of two question sets (with and without explanations) suggests an investigation into the impact of explanatory content on trust and evaluation. The use of generic labels like "aspect 1" and "aspect 2" indicates that the specific topic of the questions is intentionally left unspecified, allowing for broader applicability of the evaluation process.
</details>
Figure 1: User study design (example on a single query).
## 3 User Study
To investigate the impact of different explanation types on usersâ trust in RAG responses, we designed and conducted a user study. The study evaluates three key dimensions of response qualityâsource attribution, factual grounding, and information coverageâeach associated with a corresponding explanation type. Our goal is to measure how truthful explanations revealing the quality of a response alter a userâs initial trust judgment.
### Experiment Design.
We employed a within-subject study design where the core task involved a two-step comparison. As visualized in Figure 1, participants were first presented with a pair of responsesâone high-quality and one limited-quality, shown in random orderâwithout explanations, and were asked the question: âWhich response do you trust more?â. After providing their initial judgment on a 5-point Likert scale, they were shown the exact same pair or responses, but this time enhanced with one of the explanation types. This design directly measures the shift in user trust attributable to the presence of an explanation. Even though the modification of the scores for the first pair of responses (without explanations) is not blocked by the UI, users were instructed to not change it after moving to the next question. After each comparison (both with and without explanations), users were also asked to justify their choice in a free-text comment.
We conducted separate experiments for each of the three explanation types, allowing us to isolate and evaluate the independent effect of each response dimension on user trust.
### Responses and Explanations.
We selected 30 queries from the TREC CAsT â22 dataset [19] (10 for each response dimension). No more than two queries were selected from the same CAsT topic. Factual, information-intensive queries were prioritized for source attribution and factual grounding explanations, whereas queries involving complex or controversial topics were prioritized for information coverage explanations. For each query, responses were generated using GINGER, a modular pipeline for response creation based on information nuggets extracted from retrieved documents [11]. This approach ensures grounding in specific facts, facilitates source attribution, and maximizes information inclusion within length constraints. To systematically study the effect of explanations, we manipulated the quality of the generated responses by altering the underlying source passages: High-quality responses were based on relevant passages with high coverage of the information requested in the query. Limited-quality responses were based on only partially relevant passages covering a single aspect of the topic or were generated based solely on LLMâs parametric memory without relying on external relevant passagesâsee Table 1 for details.
To generate explanations, we extended the GINGER pipeline with a post-hoc explanation module that includes source attribution, statement grounding, and additional topic facets revealment. Crucially, all explanations shown to users are truthful, accurately revealing the shortcomings of limited-quality responses.
Table 1: Summary of explanation types and manipulated response characteristics.
| Response Dimension/ | High-quality | Limited-quality |
| --- | --- | --- |
| Explanation | Response | Response |
| Source Attribution | Based on relevant passages with clear source attribution | Relying on information from LLM parametric memory |
| Supporting passages: Clarifies whether the response is grounded in verifiable sources | List of supporting passages | âThe response has no traceable origin and it cannot be attributed to specific source documents.â |
| Factual Grounding | Each factual statement is traceable to a relevant passage | Response is generated from parametric memory without traceable origin |
| Statement-level links: Clarifies whether each claim in the response can be traced to a supporting source | List of supporting passages linked inline to supported statements | âThe response is not supported by verifiable references, and the individual claims cannot be reliably cross-checked against sources.â |
| Information Coverage | Covers multiple facets of the query topic | Covers only a single aspect or misses major points |
| Missing aspects highlighted: Discloses whether important information is missing from the response | âThe response covers multiple aspects of the topic, providing a broad view.â | âThe response focuses on just one aspect and may miss important points related to: Keyword: {keyword} â |
### Participants and Procedure.
We recruited skilled and engaged crowd workers from Amazon Mechanical Turk through a qualification task containing 5 query-response pairs, each followed by a question about topic aspects covered in the given response. Each qualified worker completed a human intelligence task (HIT) that included 2 queries per explanation type, resulting in 6 unique queries per HIT. For each query, participants compared two response variants, first without explanations and then with explanations. In total, each worker evaluated 12 response comparisons. After each response comparison, workers were asked to justify their choice in a free-text comment. We follow a within-subject design to control for individual differences as each participant serves as their own control, reducing variability and increasing the sensitivity of the analysis, as well as to use recruited crowd workers more efficiently by collecting multiple data points from each one. We designed 5 different HITs in total, each completed by 10 different participants. This setup ensures balanced coverage and enables us to independently assess the effect of each type of explanation on user trust.
### Study Execution.
The qualification study and main data collection were conducted over a period of one week (18â25 September, 2025). The qualification task was released to 50 workers, of whom 36 met the threshold (correctly answering at least 4 out of 5 test questions) and were invited to participate in the user study. In total, 21 crowd workers contributed to the study: 6 completed all 5 HITs, 2 completed 3 HITs, 1 completed 2 HITs, and the remainder submitted a single HIT. Data quality was assessed through a control question accompanying each unique response, in which participants identified which aspects from a list were covered. Because a ground-truth set of aspects was available for each response, this served as a quality check. Participants correctly identified aspects in 425 out of 600 cases, indicating high overall data quality. The total cost of the study was $316, including $68 in Amazon Mechanical Turk fees. To further acknowledge high-quality contributions, four workers received a $2 bonus for particularly insightful justifications of their trust decisions.
## 4 Results and Analysis
<details>
<summary>x2.png Details</summary>
### Visual Description
\n
## Stacked Bar Chart: Explanation Type vs. Trust Scores
### Overview
The image presents a stacked bar chart comparing "Explanation Type" (Source Attribution, Grounding, Information Coverage) with "Trust Scores" (Trust reliable response a lot more, Trust reliable response slightly more, Trust both responses about the same, Trust unreliable response slightly more, Trust unreliable response a lot more). The chart differentiates between cases where the explanation is "Shown" versus "Not shown". The x-axis represents "Proportion" ranging from 0.0 to 1.0.
### Components/Axes
* **Y-axis:** "Explanation type" with categories: "Source Attribution", "Grounding", "Information Coverage".
* **X-axis:** "Proportion" with a scale from 0.0 to 1.0, marked at 0.2, 0.4, 0.6, 0.8, and 1.0.
* **Legend (Top-right):** "Trust scores" with the following color-coded categories:
* Light Blue: "Trust reliable response a lot more"
* Blue: "Trust reliable response slightly more"
* Gray: "Trust both responses about the same"
* Orange: "Trust unreliable response slightly more"
* Red: "Trust unreliable response a lot more"
* **Horizontal dashed line:** Separates "Not shown" and "Shown" explanations.
### Detailed Analysis
The chart consists of six stacked bars, each representing a combination of Explanation Type and whether the explanation was Shown or Not Shown.
**1. Source Attribution - Not shown:**
* Light Blue (Trust reliable response a lot more): Approximately 0.15
* Blue (Trust reliable response slightly more): Approximately 0.25
* Gray (Trust both responses about the same): Approximately 0.20
* Orange (Trust unreliable response slightly more): Approximately 0.20
* Red (Trust unreliable response a lot more): Approximately 0.20
**2. Source Attribution - Shown:**
* Light Blue (Trust reliable response a lot more): Approximately 0.30
* Blue (Trust reliable response slightly more): Approximately 0.35
* Gray (Trust both responses about the same): Approximately 0.15
* Orange (Trust unreliable response slightly more): Approximately 0.10
* Red (Trust unreliable response a lot more): Approximately 0.10
**3. Grounding - Not shown:**
* Light Blue (Trust reliable response a lot more): Approximately 0.10
* Blue (Trust reliable response slightly more): Approximately 0.30
* Gray (Trust both responses about the same): Approximately 0.25
* Orange (Trust unreliable response slightly more): Approximately 0.20
* Red (Trust unreliable response a lot more): Approximately 0.15
**4. Grounding - Shown:**
* Light Blue (Trust reliable response a lot more): Approximately 0.40
* Blue (Trust reliable response slightly more): Approximately 0.35
* Gray (Trust both responses about the same): Approximately 0.10
* Orange (Trust unreliable response slightly more): Approximately 0.10
* Red (Trust unreliable response a lot more): Approximately 0.05
**5. Information Coverage - Not shown:**
* Light Blue (Trust reliable response a lot more): Approximately 0.05
* Blue (Trust reliable response slightly more): Approximately 0.25
* Gray (Trust both responses about the same): Approximately 0.30
* Orange (Trust unreliable response slightly more): Approximately 0.25
* Red (Trust unreliable response a lot more): Approximately 0.15
**6. Information Coverage - Shown:**
* Light Blue (Trust reliable response a lot more): Approximately 0.35
* Blue (Trust reliable response slightly more): Approximately 0.30
* Gray (Trust both responses about the same): Approximately 0.15
* Orange (Trust unreliable response slightly more): Approximately 0.10
* Red (Trust unreliable response a lot more): Approximately 0.10
### Key Observations
* For all explanation types, showing the explanation generally shifts the proportion towards higher trust in reliable responses (light blue and blue segments increase).
* The largest shift towards higher trust is observed for "Grounding" when shown.
* "Information Coverage" consistently has the lowest proportion of "Trust reliable response a lot more" even when the explanation is shown.
* The proportion of "Trust unreliable response a lot more" (red segment) is generally reduced when the explanation is shown.
### Interpretation
The data suggests that providing explanations (showing vs. not showing) generally increases trust in reliable responses across all explanation types. This indicates that transparency and justification are valuable in building user confidence. The most significant impact of showing an explanation is observed for "Grounding," suggesting that providing context about the basis of the response is particularly important for establishing trust. The relatively lower trust scores for "Information Coverage" even with explanations shown might indicate that the information itself is less inherently trustworthy or that the explanations provided for information coverage are less effective. The chart highlights the importance of explainability in AI systems and suggests that different explanation types have varying impacts on user trust. The consistent trend of reduced distrust (red segment) when explanations are shown reinforces the value of transparency.
</details>
Figure 2: Proportion of user trust judgments for each explanation type, comparing responses with and without explanations. Colors indicate the direction of preference.
Figure 2 presents the distribution of trust preferences across explanation presence and types. Overall, showing explanations increases the likelihood of users selecting the reliable response, which was chosen as more trustworthy in 165/300 cases with explanations compared to 84/300 without. This result clearly indicates that providing explanations helps guide users toward more reliable information.
Delving into how usersâ perceptions change, we find that explanations prompted a shift in preference from the unreliable to the reliable response in 69 out of 300 cases. The effect varied by explanation type, with grounding causing the largest shift (27 cases), followed by source attribution (23 cases) and information coverage (19 cases). This variation may also be influenced by the nature of queries: grounding queries were often factual and required information-intensive responses (âI know that there are similar schemes to Ponzi schemes. Can you compare them?â). By contrast, for information coverage queries that are often opinion-based (âWhat do you think is the best Marvel movie?â), breadth and depth were easily assessed from the response alone.
### Qualitative Analysis.
User justifications reveal a tension between objective quality and the perceived usability of responses. Participants often favored a limited-quality response when it was clearer or more actionable (âResponse B gives practical, named resources ⊠which makes the advice more actionableâ). Similarly, specific examples frequently boosted perceived trustworthiness regardless of their factuality (âResponse A provides clear examples of geographic and economic biases ⊠giving a broad perspectiveâ).
Information coverage also emerged as a recurring theme, though its effects were mixed. Participants valued comprehensive coverage when it was relevant (âResponse B covers a broader range including the resignation of aides, which adds depth to the impactâ), but viewed it as distracting when it did not directly address the query (âResponse Bâs focus on a broader view isnât necessarily helpful, given the query is specifically asking for meat dishesâ).
Source attribution was another important factor, though its influence was highly context-dependent. In fact-based domains, sources often tipped the balance (âWithout citations, response Bâs numbers canât be verified, while response Aâs quoted sources look authoritativeâ). Yet in subjective queries, participants tended to dismiss their importance (âSources are not needed for subjective questionsâ; âI donât care about response Bâs citing of sources, since this is an inherently subjective questionâ). While preliminary, these findings highlight the need for further investigation into how explanation strategies vary across query types.
Finally, usersâ background knowledge shaped trust decisions when they could not easily assess response quality. In such cases, participants either trusted both responses equally (âWithout any previous knowledge, there isnât much that I can really say here. No real differenceâ) or relied on their own expertise to guide judgments (â⊠I know for sure response one is factually correct but response two cannot be verified, which makes me even discredit the response moreâ). Since usersâ background knowledge was not explicitly measured in this study, these observations are indicative rather than conclusive and warrant a more systematic and focused analysis of the role of background knowledge in future work.
## 5 Conclusion
We conducted a user study to examine how the presence and type of explanations affect user trust in responses from retrieval-augmented generation systems. Our findings show that usersâ trust judgments often diverge from the objective quality of responses, with many favoring limited-quality responses that seem clearer, more detailed, or more actionable. Based on the results of our study, future response generation systems should adapt explanations based on query type and response complexity, use source-based explanations primarily for factual domains, and personalize explanations according to usersâ prior knowledge. Future work should also investigate the long-term effects of these explanations on how usersâ trust evolves and calibrates over time.
### 5.0.1 Acknowledgements
This research was partially supported by the Norwegian Research Center for AI Innovation, NorwAI (Research Council of Norway, project nr. 309834).
### 5.0.2 Disclosure of Interests.
The authors have no competing interests to declare that are relevant to the content of this article.
## References
- Amershi et al. [2019] Amershi, S., Weld, D., Vorvoreanu, M., Fourney, A., Nushi, B., Collisson, P., Suh, J., Iqbal, S., Bennett, P.N., Inkpen, K., Teevan, J., Kikin-Gil, R., Horvitz, E.: Guidelines for human-AI interaction. In: Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, pp. 1â13, CHI â19 (2019)
- Balog and Radlinski [2020] Balog, K., Radlinski, F.: Measuring recommendation explanation quality: The conflicting goals of explanations. In: Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 329â338, SIGIR â20 (2020)
- Bolotova et al. [2020] Bolotova, V., Blinov, V., Zheng, Y., Croft, W.B., Scholer, F., Sanderson, M.: Do people and neural nets pay attention to the same words: Studying eye-tracking data for non-factoid QA evaluation. In: Proceedings of the 29th ACM International Conference on Information & Knowledge Management, pp. 85â94, CIKM â20 (2020)
- Cau et al. [2023] Cau, F.M., Hauptmann, H., Spano, L.D., Tintarev, N.: Supporting high-uncertainty decisions through AI and logic-style explanations. In: Proceedings of the 28th International Conference on Intelligent User Interfaces, pp. 251â263, IUI â23 (2023)
- Draws et al. [2021] Draws, T., Tintarev, N., Gadiraju, U.: Assessing viewpoint diversity in search results using ranking fairness metrics. ACM SIGKDD Explorations Newsletter 23 (1), 50â58 (2021)
- Gao and Shah [2020] Gao, R., Shah, C.: Toward creating a fairer ranking in search engine results. Information Processing & Management 57 (1), 102â138 (2020)
- Gienapp et al. [2024] Gienapp, L., Scells, H., Deckers, N., Bevendorff, J., Wang, S., Kiesel, J., Syed, S., Fröbe, M., Zuccon, G., Stein, B., Hagen, M., Potthast, M.: Evaluating generative ad hoc information retrieval. In: Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 1916â1929, SIGIR â24 (2024)
- Hu et al. [2019] Hu, M., Wei, F., Peng, Y., Huang, Z., Yang, N., Li, D.: Read + verify: Machine reading comprehension with unanswerable questions. In: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 6529â6537, AAAI â19 (2019)
- Huang et al. [2019] Huang, K., Tang, Y., Huang, J., He, X., Zhou, B.: Relation module for non-answerable predictions on reading comprehension. In: Proceedings of the 23rd Conference on Computational Natural Language Learning, pp. 747â756, CoNLL â19 (2019)
- Izacard and Grave [2021] Izacard, G., Grave, E.: Leveraging passage retrieval with generative models for open domain question answering. In: Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pp. 874â880, EACL â21 (2021)
- Ćajewska and Balog [2025] Ćajewska, W., Balog, K.: GINGER: Grounded information nugget-based generation of responses. In: Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, p. 2723â2727, SIGIR â25 (2025)
- Ćajewska et al. [2024] Ćajewska, W., Spina, D., Trippas, J., Balog, K.: Explainability for transparent conversational information-seeking. In: Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 1040â1050, SIGIR â24 (2024)
- Lewis et al. [2020] Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., KĂŒttler, H., Lewis, M., Yih, W.t., RocktĂ€schel, T., Riedel, S., Kiela, D.: Retrieval-augmented generation for knowledge-intensive NLP tasks. In: Proceedings of the 34th International Conference on Neural Information Processing Systems, pp. 9459â9474, NIPS â20 (2020)
- Li et al. [2023] Li, L., Zhang, Y., Chen, L.: Personalized prompt learning for explainable recommendation. ACM Transactions on Information Systems 41 (4), 1â26 (2023)
- Li et al. [2024] Li, S., Park, S., Lee, I., Bastani, O.: TRAQ: Trustworthy retrieval augmented question answering via conformal prediction. In: Duh, K., Gomez, H., Bethard, S. (eds.) Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), NAACL-HLT â24 (2024)
- Liao and Sundar [2022] Liao, Q., Sundar, S.S.: Designing for responsible trust in AI systems: A communication perspective. In: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pp. 1257â1268, FAccT â22 (2022)
- Liu et al. [2023] Liu, N., Zhang, T., Liang, P.: Evaluating verifiability in generative search engines. In: Findings of the Association for Computational Linguistics: EMNLP 2023, pp. 7001â7025, EMNLP â23 (2023)
- Nunes and Jannach [2017] Nunes, I., Jannach, D.: A systematic review and taxonomy of explanations in decision support and recommender systems. User Modeling and User-Adapted Interaction 27 (3-5), 393â444 (2017)
- Owoicho et al. [2022] Owoicho, P., Dalton, J., Aliannejadi, M., Azzopardi, L., Trippas, J.R., Vakulenko, S.: TREC CAsT 2022: Going beyond user ask and system retrieve with initiative and response generation. In: The Thirty-First Text REtrieval Conference Proceedings, TREC â22 (2022)
- Ram et al. [2023] Ram, O., Levine, Y., Dalmedigos, I., Muhlgay, D., Shashua, A., Leyton-Brown, K., Shoham, Y.: In-context retrieval-augmented language models. Transactions of the Association for Computational Linguistics 11, 1316â1331 (2023)
- Rashkin et al. [2021] Rashkin, H., Nikolaev, V., Lamm, M., Aroyo, L., Collins, M.: Measuring attribution in natural language generation models. Computational Linguistics 49 (4), 777â840 (2021)
- Schuster et al. [2023] Schuster, T., Lelkes, A.D., Sun, H., Gupta, J., Berant, J., Cohen, W.W., Metzler, D.: SEMQA: Semi-extractive multi-source question answering. In: Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pp. 1363â1381, NAACL-HLT â23 (2023)
- Wang et al. [2024] Wang, S., Zhang, X., Wang, Y., Ricci, F.: Trustworthy recommender systems. ACM Transactions on Intelligent Systems and Technology 15 (4), 1â20 (2024)
- Wang and Yin [2021] Wang, X., Yin, M.: Are explanations helpful? a comparative study of the effects of explanations in AI-assisted decision-making. In: Proceedings of the 26th International Conference on Intelligent User Interfaces, pp. 318â328, IUI â21â (2021)
- Xu and Chen [2006] Xu, Y.C., Chen, Z.: Relevance judgment: What do information users consider beyond topicality? Journal of the American Society for Information Science and Technology 57 (7), 961â973 (2006)
- Zhang and Chen [2020] Zhang, Y., Chen, X.: Explainable recommendation: A survey and new perspectives. Foundations and TrendsÂź in Information Retrieval 14 (1), 1â101 (2020)
- Zhang et al. [2020] Zhang, Y., Liao, Q.V., Bellamy, R.K.E.: Effect of confidence and explanation on accuracy and trust calibration in AI-assisted decision making. In: Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, pp. 295â305, FAT â20 (2020)