CoChat
Sign Up
Cited by 85Explainable Artificial Intelligence (XAI)Innovative Human-Technology InteractionUsability and User Interface Design

Progressive Disclosure: When, Why, and How Do Users Want Algorithmic Transparency Information?

Open access
Springer, Aaron, Whittaker, Steve·2020·ACM Transactions on Interactive Intelligent Systems
DOI 10.1145/3374218
Open this paper in CoChat — free accountOpen PDF

Abstract

It is essential that users understand how algorithmic decisions are made, as we increasingly delegate important decisions to intelligent systems. Prior work has often taken a techno-centric approach, focusing on new computational techniques to support transparency. In contrast, this article employs empirical methods to better understand user reactions to transparent systems to motivate user-centric designs for transparent systems. We assess user reactions to transparency feedback in four studies of an emotional analytics system. In Study 1, users anticipated that a transparent system would perform better but unexpectedly retracted this evaluation after experience with the system. Study 2 offers an explanation for this paradox by showing that the benefits of transparency are context dependent. On the one hand, transparency can help users form a model of the underlying algorithm's operation. On the other hand, positive accuracy perceptions may be undermined when transparency reveals algorithmic errors. Study 3 explored real-time reactions to transparency. Results confirmed Study 2, in showing that users are both more likely to consult transparency information and to experience greater system insights when formulating a model of system operation. Study 4 used qualitative methods to explore real-time user reactions to motivate transparency design principles. Results again suggest that users may benefit from initially simplified feedback that hides potential system errors and assists users in building working heuristics about system operation. We use these findings to motivate new progressive disclosure principles for transparency in intelligent systems and discuss theoretical implications.

Integrity signals

Registry

Crossref · 10.1145/3374218

Cites

89 of 89 references verified0 citation errors · 0 not checkable

Statistics

4 tests recomputed · 0 with the reported and recomputed p on opposite sides of .05

Claims the paper makes

29 of 96 cited sources checked · 10 supported · 1 flagged · accuracy not yet measured

  • Premisecitation

    Some research suggests that the way transparency is presented may account for contradictory results.

    “Some research suggests that the way we present transparency may account for these contradictory results [28,43].”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [28] 28. Pedro García Garcia, Enrico Costanza, Jhim Verame, Diana Nowacka, and Sarvapali D. Ramchurn. (in press). Seeing (Movement) is believing: The effect of motion on perception of automatic systems performance. To appear in Hum.- Comput Interact. https://doi.org/10.1080/07370024.2018.1453815 source

      Section 2.2 (page unavailable) describes how explanations can help users understand a system but can also overwhelm them, and how animation can aid understanding yet sometimes cause confusion. This supports the idea that presentation matters, but the paper does not specifically say that differences in presentation account for the contradictory results cited in the sentence. (LLM, conf=0.78)

      Experimental — accuracy not yet measured
    • cites [43] 43. René F. Kizilcec. 2016. How much information?: Effects of transparency on trust in an algorithmic interface. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. ACM, New York, 2390-2395. https: //doi.org/10.1145/2858036.2858402 source
  • Premisecitation

    Transparency can promote user learning and insight from complex data.

    “It can also promote user learning and insight from complex data, as humans increasingly work with complex inferential systems for analytic purposes [44,45,73].”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites [44] 44. Todd Kulesza, Margaret Burnett, Weng-Keen Wong, and Simone Stumpf. 2015. Principles of explanatory debugging to personalize interactive machine learning. In Proceedings of the 20th International Conference on Intelligent User Interfaces (IUI'15). 126-137. https://doi.org/10.1145/2678025.2701399 source

      The Related Work and Explainability sections (p. 3) say transparency and explanations help users understand how a learning system operates; the Conclusion (p. 11) reports improved understanding among participants. The paper does not establish that transparency gives users insight from complex data for analytic purposes, so that part of the attributed claim is unsupported. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [45] 45. Todd Kulesza, Weng-Keen Wong, Simone Stumpf, Stephen Perona, Rachel White, Margaret M. Burnett, Ian Oberst, and Andrew J. Ko. 2008. Fixing the program my computer learned: Barriers for end users, challenges for the machine. In Proceedings of the 13th International Conference on Intelligent User Interfaces (IUI'09). 187. https://doi.org/10.1145/ 1502650.1502678 source
    • cites [73] 73. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. "Why should i trust you?": Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 1135-1144. https://doi.org/10.1145/2939672.2939778 source
  • Premisecitation

    Making a complex algorithm understandable to end users can require simplification, which often comes at the cost of reduced accuracy of explanation.

    “Making a complex algorithm understandable to end users can require simplification, which often comes at the cost of reduced accuracy of explanation [46,47,53,54,73].”
    Quote in the paperSubstantial claim4 of 5 sources checked
    • cites [46] 46. Bai. C. Kwon, M. Choi, J. T. Kim, E. Choi, Y. B. Kim, S. Kwon, J. Sun, and J. Choo. 2019. RetainVis: Visual analytics with interpretable and interactive recurrent neural networks on electronic medical records. IEEE Trans. Vis. Comput. Graph. 25, 1 (2019), 299-309. https://doi.org/10.1109/TVCG.2018.2865027 source

      Section 8.1 (pp. 8–9) says clinicians benefit from simpler, more easily interpretable visualizations of complex medical data. It discusses a tradeoff between model interpretability and model performance, but does not say that simplifying an explanation reduces the explanation’s accuracy. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [47] 47. Himabindu Lakkaraju, Ece Kamar, Rich Caruana, and Jure Leskovec. 2017. Interpretable & explorable approximations of black box models. arXiv:1707.01154 [cs]. Retrieved 18, 2018 from http://arxiv.org/abs/1707.01154. source
    • cites [53] 53. Yin Lou, Rich Caruana, and Johannes Gehrke. 2012. Intelligible Models for Classification and Regression. In Proceed- ings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, New York, 150-158. https://doi.org/10.1145/2339530.2339556 source

      The Introduction (p. 2) describes a trade-off between model intelligibility and predictive accuracy. Section 5.2 (p. 6) adds that smoother plots may be easier to interpret while missing potentially important structure. The paper does not establish that simplifying an explanation of an existing complex algorithm reduces the explanation’s accuracy; its accuracy comparisons concern the fitted models. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [54] 54. Scott M. Lundberg and Su-In Lee. 2017. A Unified Approach to Interpreting Model Predictions. 10. source

      Section 2 (p. 2) says complex models require a simpler, interpretable explanation model that approximates the original; Section 2.1 (p. 2) describes balancing an explanation’s faithfulness against its complexity. The paper does not establish that simplification often reduces explanation accuracy. The accuracy–interpretability trade-off in the Introduction (p. 1) concerns model predictions, not the accuracy of explanations. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites [73] 73. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. "Why should i trust you?": Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 1135-1144. https://doi.org/10.1145/2939672.2939778 source
  • Premisecitation

    Providing system explanations may undermine user perceptions when users lack the attentional capacity to process complex explanations, for example when executing a demanding task.

    “Providing system explanations may also undermine user perceptions when users lack the attentional capacity to process complex explanations, for example when executing a demanding task [1,12,43,85].”
    Quote in the paperSubstantial claim1 of 4 sources checked
    • cites [1] 1. Ashraf Abdul, Jo Vermeulen, Danding Wang, Brian Y. Lim, and Mohan Kankanhalli. 2018. Trends and trajectories for explainable, accountable and intelligible systems: An HCI research agenda. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems (CHI'18). 1-18. https://doi.org/10.1145/3173574.3174156 source
    • cites [12] 12. Andrea Bunt, Matthew Lount, and Catherine Lauzon. 2012. Are explanations always important?: A study of deployed, low-cost intelligent interactive systems. In Proceedings of the 2012 ACM International Conference on Intelligent User Interfaces. 169-178. source
    • cites [43] 43. René F. Kizilcec. 2016. How much information?: Effects of transparency on trust in an algorithmic interface. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. ACM, New York, 2390-2395. https: //doi.org/10.1145/2858036.2858402 source
    • cites [85] 85. Daniel S. Weld and Gagan Bansal. 2018. The challenge of crafting intelligible intelligence. arXiv:1803.04263 [cs]. Retrieved September 20, 2018 from http://arxiv.org/abs/1803.04263. source

      Section 6, “Towards Interactive Explanation” (page unavailable), says explanations should suit the audience’s background, and “Ranking Intelligible Models” reports a preference for simpler explanations. Neither passage says that limited attention during a demanding task causes explanations to undermine user perceptions. (LLM, conf=0.78)

      Experimental — accuracy not yet measured
  • Premisecitation

    The E-meter system and protocol have been successfully deployed in multiple previous experiments.

    “The E-meter system and protocol have been successfully deployed in multiple previous experiments [78,80].”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [78] 78. Aaron Springer, Victoria Hollis, and Steve Whittaker. 2017. Dice in the black box: User experiences with an inscrutable algorithm. Retrieved April 24, 2017 from https://aaai.org/ocs/index.php/SSS/SSS17/paper/view/15372. source

      The E-meter System and Results sections (p. 2) describe the system and its use with 12 participants, supporting deployment in an experiment. The Discussion (p. 3) calls this a small pilot; the paper does not establish that the E-meter system and protocol were deployed in multiple previous experiments. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [80] 80. Aaron Springer and Steve Whittaker. 2018. What are you hiding? algorithmic transparency and user perceptions. In Proceedings of the 2018 AAAI Spring Symposium Series. source
  • May combine two claims. 52
  • Classified as probably not a substantial claim. 44
  • The quote only partly states this claim. 2

  • May combine two claims.

    Premisecitation

    In prior experiments, users judged the E-meter algorithm to be accurate but not perfect; the median response for accuracy was 'Accurate' (6 on a 7-point Likert scale).

    “In prior experiments [78,80], users judged this algorithm to be accurate but not perfect; the median response for accuracy was “Accurate” (“6” on a 7-point Likert scale with the highest rating (“7”) being “Very Accurate”).”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [78] 78. Aaron Springer, Victoria Hollis, and Steve Whittaker. 2017. Dice in the black box: User experiences with an inscrutable algorithm. Retrieved April 24, 2017 from https://aaai.org/ocs/index.php/SSS/SSS17/paper/view/15372. source

      The Results (p. 2) report a median E-meter accuracy rating of 5, labelled “Slightly Accurate,” on a 7-point scale—not 6, labelled “Accurate.” The paper cautions that its sample of 12 is too small for a representative quantitative estimate. (LLM, conf=0.94)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites [80] 80. Aaron Springer and Steve Whittaker. 2018. What are you hiding? algorithmic transparency and user perceptions. In Proceedings of the 2018 AAAI Spring Symposium Series. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    In some settings, transparency improves algorithmic perceptions because users better understand system behavior.

    “In some settings there are expected benefits: Transparency improves algorithmic perceptions, because users better understand system behavior [43,44,51].”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites [43] 43. René F. Kizilcec. 2016. How much information?: Effects of transparency on trust in an algorithmic interface. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. ACM, New York, 2390-2395. https: //doi.org/10.1145/2858036.2858402 source
    • cites [44] 44. Todd Kulesza, Margaret Burnett, Weng-Keen Wong, and Simone Stumpf. 2015. Principles of explanatory debugging to personalize interactive machine learning. In Proceedings of the 20th International Conference on Intelligent User Interfaces (IUI'15). 126-137. https://doi.org/10.1145/2678025.2701399 source

      The Related Work section (p. 3) says increased transparency has been associated with greater acceptance, satisfaction, and trust, and that it helps users understand how a learning system operates. It does not establish that better understanding is the reason those perceptions improve; the stated causal link goes beyond the passage. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [51] 51. Brian Y. Lim, Anind K. Dey, and Daniel Avrahami. 2009. Why and why not explanations improve the intelligibility of context-aware intelligent systems. In Proceedings of the 27th International Conference on Human Factors in Computing Systems (CHI'09). 2119. https://doi.org/10.1145/1518701.1519023 source

  • May combine two claims.

    Premisecitation

    Greater transparency potentially increases end user control and improves acceptance of complex algorithmic systems.

    “Greater transparency potentially increases end user control and improves acceptance of complex algorithmic systems [31,44,57,85].”
    Quote in the paperSubstantial claim3 of 4 sources checked
    • cites [31] 31. Bryce Goodman and Seth Flaxman. 2017. European union regulations on algorithmic decision-making and a "right to explanation." AI Mag. 38, 3 (2017), 50-57. https://doi.org/10.1609/aimag.v38i3.2741 source
    • cites [44] 44. Todd Kulesza, Margaret Burnett, Weng-Keen Wong, and Simone Stumpf. 2015. Principles of explanatory debugging to personalize interactive machine learning. In Proceedings of the 20th International Conference on Intelligent User Interfaces (IUI'15). 126-137. https://doi.org/10.1145/2678025.2701399 source
    • cites [57] 57. Tim Miller. 2017. Explanation in artificial intelligence: Insights from the social sciences. arXiv:1706.07269 [cs]. Re- trieved September 17, 2018 from http://arxiv.org/abs/1706.07269. source

      The Introduction says that more transparent or explainable systems may help users understand and trust intelligent agents, and links limited uptake to a lack of trust. It does not say transparency increases end-user control. Section 4.6.5 discusses an agent’s control over its own actions, not a user’s control over a system. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites [85] 85. Daniel S. Weld and Gagan Bansal. 2018. The challenge of crafting intelligible intelligence. arXiv:1803.04263 [cs]. Retrieved September 20, 2018 from http://arxiv.org/abs/1803.04263. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Study 3 recruited 53 E-meter users who had previously passed a short screener (PG-WBI) [33].

    “We recruited 53 E-meter users who had previously passed a short screener (PG-WBI) [33].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [33] 33. E. Grossi, N. Groth, P. Mosconi, R. Cerutti, F. Pace, A. Compare, and G. Apolone. 2006. Development and validation of the short version of the Psychological General Well-Being Index (PGWB-S). Health Qual. Life Outcomes 4, 88 (2006). DOI:10.1186/1477-7525-4-88 source

      The Background and Methods describe and validate a short psychological well-being questionnaire, supporting the citation as an identification of the measure. But the paper calls the six-item version PGWB-S; PGWBI is the original 22-item questionnaire. It does not describe a pass threshold, though the 53 E-meter users are the citing authors’ own sample, not a finding attributed to this source. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductcitation

    All 41 Study 2 users had previously passed the authors' standard psychological screener [30].

    “All had previously passed our standard psychological screener [30].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [30] 30. K. Goddard, A. Roudsari, and J. C. Wyatt. 2012. Automation bias: A systematic review of frequency, effect mediators, and mitigators. J. Am. Med. Inf. Assoc. 19, 1 (2012), 121-127. DOI:10.1136/amiajnl-2011-000089. PMC 3240751. PMID 21685142. source

      The Screening and eligibility section describes screening papers for a systematic review, not screening study participants. Searches for “standard psychological screener” and “41 Study 2 users” found no passage in our copy addressing whether those users passed a screener (page numbers unavailable). (LLM, conf=0.76)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    The original concept of progressive disclosure involved hiding advanced interface controls, allowing users to make fewer initial errors and learn the system more effectively.

    “The original concept involved hiding advanced interface controls; allowing users to make fewer initial errors and learn the system more effectively [15].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [15] 15. John M. Carroll and Caroline Carrithers. 1984. Training wheels in a user interface. Commun. ACM 27, 8 (1984), 800- 806. https://doi.org/10.1145/358198.358218 source

      The Training Wheels Interface section says the design made advanced functions inaccessible and blocked troublesome error states; the Discussion reports that it facilitated learning. But the paper does not say advanced controls were hidden: users could still select an unavailable menu choice and receive a message, and the menus were kept identical to those of the complete system. The claim is supported on error protection and learning, but not on hiding controls. (LLM, conf=0.87)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Progressive disclosure has a long history in UI research, dating back to the Xerox Star and early word processing systems.

    “Progressive disclosure has a long history in UI research, dating back to the Xerox Star and early word processing systems [15,62,76].”
    Quote in the paperNot a substantial claim3 of 3 sources checked
    • cites [15] 15. John M. Carroll and Caroline Carrithers. 1984. Training wheels in a user interface. Commun. ACM 27, 8 (1984), 800- 806. https://doi.org/10.1145/358198.358218 source
    • cites [62] 62. Lloyd H. Nakatani and John A. Rohrlich. 1983. Soft machines: A philosophy of user-computer interface design. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI'83). 19-23. https://doi.org/10.1145/ 800045.801573 source

      In “Operability with Flexibility” (page not provided), the paper explicitly identifies progressive disclosure with the Xerox Star interface, supporting that part of the historical claim. Its mention of dedicated word processors elsewhere concerns special function keys, not progressive disclosure, so it does not support the early-word-processing-systems part. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [76] 76. David Canfield Smith. 1982. Designing the Star User Interface. Byte Magazine, April, 1982, 242-282. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    There are many other ways to illustrate how an algorithm operates, including verbal explanations and concrete user exploration.

    “There are many other ways to illustrate how an algorithm operates including verbal explanations, concrete user exploration, and so forth [44,45,51,73,88].”
    Quote in the paperNot a substantial claim1 of 5 sources checked
    • cites [44] 44. Todd Kulesza, Margaret Burnett, Weng-Keen Wong, and Simone Stumpf. 2015. Principles of explanatory debugging to personalize interactive machine learning. In Proceedings of the 20th International Conference on Intelligent User Interfaces (IUI'15). 126-137. https://doi.org/10.1145/2678025.2701399 source

      The paper’s Being Complete section (p. 6) describes explanations tied to specific predictions, and Being Actionable (p. 7) describes users interacting with and adjusting those explanations. This supports concrete user exploration of how the classifier operates. Our copy does not establish verbal explanations as a method used in this work. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites [45] 45. Todd Kulesza, Weng-Keen Wong, Simone Stumpf, Stephen Perona, Rachel White, Margaret M. Burnett, Ian Oberst, and Andrew J. Ko. 2008. Fixing the program my computer learned: Barriers for end users, challenges for the machine. In Proceedings of the 13th International Conference on Intelligent User Interfaces (IUI'09). 187. https://doi.org/10.1145/ 1502650.1502678 source
    • cites [51] 51. Brian Y. Lim, Anind K. Dey, and Daniel Avrahami. 2009. Why and why not explanations improve the intelligibility of context-aware intelligent systems. In Proceedings of the 27th International Conference on Human Factors in Computing Systems (CHI'09). 2119. https://doi.org/10.1145/1518701.1519023 source
    • cites [73] 73. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. "Why should i trust you?": Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 1135-1144. https://doi.org/10.1145/2939672.2939778 source
    • cites [88] 88. Woebot-Your Charming Robot Friend Who Is Here for You, 24/7. Retrieved September 21, 2018 from https://woebot. io.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Newer technical approaches attempt to increase the intelligibility of complex neural models by approximating them using simpler, more intuitive models.

    “Newer technical approaches attempt to increase intelligibility of complex neural models by approximating them using simpler more intuitive models [47,53,54].”
    Quote in the paperNot a substantial claim3 of 3 sources checked
    • cites [47] 47. Himabindu Lakkaraju, Ece Kamar, Rich Caruana, and Jure Leskovec. 2017. Interpretable & explorable approximations of black box models. arXiv:1707.01154 [cs]. Retrieved 18, 2018 from http://arxiv.org/abs/1707.01154. source
    • cites [53] 53. Yin Lou, Rich Caruana, and Johannes Gehrke. 2012. Intelligible Models for Classification and Regression. In Proceed- ings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, New York, 150-158. https://doi.org/10.1145/2339530.2339556 source

      The Introduction (pp. 1–2) says complex models, including deep neural nets, are difficult to interpret and presents generalized additive models as a more intelligible alternative. But it describes fitting those models directly, not approximating an existing neural model with a simpler one. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [54] 54. Scott M. Lundberg and Su-In Lee. 2017. A Unified Approach to Interpreting Model Predictions. 10. source

  • May combine two claims.

    Findingown data

    The E-meter linear regression tested at R2 = 0.25; mean absolute error was .95 on the target variable (mood) scale of (-3,3).

    “Using a train/test split of 85/15 the linear regression tested at R 2 = 0.25; mean absolute error was .95 on the target variable (mood) scale of (-3,3).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In Study 1, the median user found the E-meter to be 'Slightly Accurate' and 'Slightly Trustworthy', a rating of 5 on each 7-point scale.

    “Overall, the median user found the Emeter to be “Slightly Accurate” and “Slightly Trustworthy,” i.e., a rating of “5” on each 7-point scale.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Users anticipated that a transparent system would perform better but retracted this evaluation after experience with the system (Study 1).

    “In Study 1, users anticipated that a transparent system would perform better but unexpectedly retracted this evaluation after experience with the system.”
    • Also stated in: § 1.1 Contribution · plain
      “In Study 1, users anticipated that a transparent system would perform better but retracted this evaluation after experience with the system.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Users are the ultimate experts on their own emotions, providing ground truth assessments about whether the system is correct.

  • May combine two claims.

    Findingown data

    Users are more likely to consult transparency information and to experience greater system insights when formulating a model of system operation (Study 3).

    “Results confirmed Study 2, in showing that users are both more likely to consult transparency information and to experience greater system insights when formulating a model of system operation.”
    • Also stated in: § 1.1 Contribution · plain
      “It shows that users are more receptive to, and derive greater understanding benefits from, transparency information when trying to formulate models of system operation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Non-transparent users were unable to discern this discrepancy, while those who saw transparent feedback sometimes took issue with it.

    “While non-transparent users were unable to discern this discrepancy, those who saw transparent feedback sometimes took issue with this.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Participants' self-reported qualitative explanations accurately reflect the reasons for their predicted accuracy, preferences, and system understanding.

  • May combine two claims.

    Findingown data

    Both expectation violation and transparency are associated with perceived accuracy.

    “As expected, both Expectation Violation and Transparency are associated with perceived accuracy.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Users' subjective affect judgments accurately reflect the emotional valence of their writing, so the difference between user and system judgments measures expectation violation.

  • May combine two claims.

    Findingown data

    The imprecise global feedback from the algorithm confirmed non-transparent users' overall interpretations, allowing them to build a consistent, if vague, model of system operation.

    “The imprecise global feedback from the algorithm confirmed their overall interpretations, allowing them to build a consistent, if vague model of system operation.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Participants' self-reported qualitative explanations accurately reflect the reasons for their predicted accuracy, preferences, and system understanding.

  • May combine two claims.

    Findingown data

    Other users shifted their preference to the non-transparent version but offered very different reasons for this judgment.

    “In contrast, others shifted their preference to the nontransparent version, but offered very different reasons for this judgment.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Participants' self-reported qualitative explanations accurately reflect the reasons for their predicted accuracy, preferences, and system understanding.

  • May combine two claims.

    Findingreasoning

    These users were generally able to determine how the system weighted individual words to arrive at an interpretation, and this led them to view the system as accurate.

    “These people were generally able to determine how the system weighted individual words to arrive at an interpretation, and this led them to view the system as accurate.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Participants' self-reported qualitative explanations accurately reflect the reasons for their predicted accuracy, preferences, and system understanding.

  • May combine two claims.

    Findingown data

    In Study 2, most users across conditions judged the E-meter to be 'Accurate' or 'Very Accurate', with the median being 'Accurate' (a score of 6 on a 7-point scale).

    “Again, most users across conditions judged the E-meter to be “Accurate” or “Very Accurate” with the median being “Accurate” (a score of “6” on a 7 point scale) as shown in Figure 4(a).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    On later prompts, a sizeable number of users increase their understanding, while others feel their understanding is unchanged by the transparency.

    “On these later prompts, although a sizeable number of users increase their understanding, others feel their understanding is unchanged by the transparency.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For some low expectation violation users, transparency seemed to create more questions than it answered, and the additional word-level highlighting confused rather than clarified.

    “For some of these low expectation violation users, transparency seemed to create more questions than it answered, and the additional information provided by the word-level highlighting confused rather than clarified.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Users' subjective affect judgments accurately reflect the emotional valence of their writing, so the difference between user and system judgments measures expectation violation.
    • Rests on an assumption: Participants' self-reported qualitative explanations accurately reflect the reasons for their predicted accuracy, preferences, and system understanding.

  • May combine two claims.

    Findingown data

    People view transparency information for much longer when their expectations are not met, although the effects are relatively small; a one-unit increase in expectation violation is associated with a 7.25% increase in transparency viewing time.

    “In other words, people view transparency information for much longer when their expectations are not met, although the effects are relatively small; a one unit increase in Expectation Violation is associated with a 7.25% increase in transparency Viewing Time (recall that time is measured on a log scale).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Users' subjective affect judgments accurately reflect the emotional valence of their writing, so the difference between user and system judgments measures expectation violation.

  • May combine two claims.

    Findingown data

    The benefit of transparency on perceived accuracy occurs only when expectation violation is high; transparency is not helpful for lower levels of expectation violation.

    “The interaction term valence shows that Transparency has the expected benefit when Expectation Violation is high (as shown in Figure 5).In other words, if people have high Expectation Violation, then Transparency increases perceived Accuracy.However, Transparency is not helpful for lower levels of Expectation Violation.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Users' subjective affect judgments accurately reflect the emotional valence of their writing, so the difference between user and system judgments measures expectation violation.

  • May combine two claims.

    Findingown data

    Participants justified their predicted accuracy judgments by drawing attention to incremental color-coded feedback, which they said offered a clearer sense of how the system worked and boosted confidence that it was operating appropriately.

    “participants justified why they expected the transparent system to be more accurate by drawing attention to the incremental color-coded feedback.They argued that this offered them a clearer sense of how the system worked, with this feedback boosting their confidence that the system was operating appropriately.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Participants' self-reported qualitative explanations accurately reflect the reasons for their predicted accuracy, preferences, and system understanding.

  • May combine two claims.

    Findingown data

    There is a strong negative relation between expectation violation and accuracy (p < 0.0001), confirming prior work [43,80].

    “We see the predicted strong negative relation between Expectation Violation and Accuracy (p < 0.0001) confirming prior work [43,80].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [43] 43. René F. Kizilcec. 2016. How much information?: Effects of transparency on trust in an algorithmic interface. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. ACM, New York, 2390-2395. https: //doi.org/10.1145/2858036.2858402 source
    • cites [80] 80. Aaron Springer and Steve Whittaker. 2018. What are you hiding? algorithmic transparency and user perceptions. In Proceedings of the 2018 AAAI Spring Symposium Series. source
    • Rests on an assumption: Users' subjective affect judgments accurately reflect the emotional valence of their writing, so the difference between user and system judgments measures expectation violation.

  • May combine two claims.

    Findingown data

    In Study 1, before actual usage, participants initially anticipate greater accuracy for the transparent version of the E-meter, but this is altered by their system experiences.

    “In Study 1 we found that before actual usage, participants initially anticipate greater accuracy for the transparent version of the E-meter, but this is altered by their system experiences.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Study 3 confirmed these prior results and showed that transparency benefits depend on context.

    “Study 3 confirmed these prior results and deepened our understanding of transparency, again showing that its benefits depend on context.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    After using both system versions, although both versions of the algorithm were perceived as accurate, users were split 50-50 in their preference.

    “After using both system versions, and although both versions of the algorithm were perceived as accurate, users were split 50-50 in their preference.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Findingown data

    Several users took issue with the level of detail provided by transparency, making spontaneous suggestions about how this might be modified.

    “Several users took issue with the level of detail provided by transparency, making spontaneous suggestions about how this might be modified.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Findingown data

    After clicking the button to reveal transparency, P11 had further questions.

    “After clicking the button to reveal transparency, P11 had further questions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The machine learning model was limited to the 600 most predictive words of mood, but still presented too many extraneous words that these users felt were unimportant.

    “While the machine learning model was limited to the 600 most predictive words of mood, that model still seemed to present too many extraneous words that these users felt were unimportant.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing).

    Findingown data

    They feel that the overall rating should reflect a simple ratio of the number of positive versus negative words in the word-level transparency.

    “They feel that the overall rating should reflect a simple ratio of the number of positive versus negative words in the word-level transparency.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingnone given

    Some users (like those in Studies 2 and 3) may want yet more transparency.

    “However, some users (like those in Studies 2 and 3) may want yet more transparency.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    Transparency is helpful when people don't understand, but it can undermine when expectations are met.

    “We found that transparency is helpful when people don't understand, but it can undermine when expectations are met.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Study 2 showed that transparency information was generally helpful, but that its specific effects depended on expectation violation.

    “Study 2 showed that transparency information was generally helpful, but that its specific effects depended on expectation violation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Initial viewings promote distinct improvements in understanding, but these benefits decrease over time.

    “Furthermore, initial viewings promote distinct improvements in understanding, but these benefits decrease over time.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims. The quote only partly states this claim.

    Study design and conductcitation

    Study 1 recruited 100 users, eliminated 26 based on open-ended responses, leaving a sample of 74 users.

    “Following methodological recommendations concerning compliance of Mechanical Turk participants [18], we eliminated 26 respondents based on their responses to open ended questions, leaving us with a sample of 74 users.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [18] 18. Eun Kyoung Choe, Nicole B. Lee, Bongshin Lee, Wanda Pratt, and Julie A. Kientz. 2014. Understanding quantified- selfers' practices in collecting and exploring personal data. In Proceedings of the 32nd Annual ACM Conference on Human Factors in Computing Systems (CHI'14). 1143-1152. https://doi.org/10.1145/2556288.2557372 source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    Study 1 questions and procedure were carefully piloted and had been used before in multiple prior studies.

    “Questions and procedure were carefully piloted and had been used before in multiple prior studies.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    Transparency may erode confidence in a system, with users trusting it less, because transparency leads them to question the system even when it is correct.

    “Transparency may erode confidence in a system, with users trusting it less, because transparency leads them to question the system even when it is correct [1,13,50].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [1] 1. Ashraf Abdul, Jo Vermeulen, Danding Wang, Brian Y. Lim, and Mohan Kankanhalli. 2018. Trends and trajectories for explainable, accountable and intelligible systems: An HCI research agenda. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems (CHI'18). 1-18. https://doi.org/10.1145/3173574.3174156 source
    • cites [13] 13. Moira Burke, Brian Amento, and Philip Isenhour. 2006. Error correction of voicemail transcripts in SCANMail. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI'06). ACM, New York, NY, 339- 348. DOI:http://dx.doi.org/10.1145/1124772.1124823 source
    • cites [50] 50. Brian Y. Lim and Anind K. Dey. 2011. Investigating intelligibility for uncertain context-aware applications. In Pro- ceedings of the 13th International Conference on Ubiquitous Computing. 415-424. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The Study 1 study was approved by an Institutional Review Board.

    “The study was approved by an Institutional Review Board.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Extensive technical research about operationalizing transparency has emerged in the machine learning community, but no clear consensus has resulted.

    “Extensive technical research about operationalizing transparency has emerged in the machine learning community, but no clear consensus has resulted [1,23,[52][53][54]85].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [1,23,[52][53][54]85]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Prior work has often taken a techno-centric approach, focusing on new computational techniques to support transparency.

    “Prior work has often taken a techno-centric approach, focusing on new computational techniques to support transparency.”
    • Also stated in: § 1 INTRODUCTION · plain
      “Overall, then, prior work has often addressed transparency from a technical perspective, asking, “What is possible from an algorithmic standpoint?” rather than “What does the user need?””
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data

    The Study 1 design was counterbalanced; half the users experienced the nontransparent system version first.

    “The study design was counterbalanced; half the users experienced the nontransparent system version first.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisenone given

    Both systems use the same underlying machine learning model that uses solely individual word features.

    “Recall that both systems use the same underlying machine learning model that uses solely individual word features.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    On early prompts, users are unclear about how the algorithm works, so initial viewings promote greater understanding benefits.

    “On early prompts, users are unclear about how the algorithm works, so that initial viewings promote greater understanding benefits.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Participants in Study 2 complied with instructions to write at least 100 words (mean words written = 107.74, sd = 14.8).

    “Participants complied with instructions to write at least 100 words, (mean words written = 107.74,sd = 14.8).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    For these users, transparency information can help moderate expectation violation, suggesting that the algorithm is operating in an interpretable manner even if it does not exactly conform to their expectations.

    “For these users, transparency information can help moderate expectation violation, suggesting to users that the algorithm is operating in an interpretable manner, even if it does not exactly conform to their expectations.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Users' subjective affect judgments accurately reflect the emotional valence of their writing, so the difference between user and system judgments measures expectation violation.
    • Rests on an assumption: Participants' self-reported qualitative explanations accurately reflect the reasons for their predicted accuracy, preferences, and system understanding.

  • May combine two claims.

    Interpretationown data

    This comparison means there is significant variation between participants and also between the order effects within each participant.

    “This comparison means there is significant variation between participants and also between the order effects within each participant.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Both viewing time and understanding benefits progressively decrease for later prompts as users gain experience with the algorithm.

    “However both viewing time and understanding benefits progressively decrease for later prompts as users gain experience with the algorithm.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    The elevation in Viewing Time at the fourth position may arise from the novelty of engaging in a second task with the algorithm.

    “However, this fourth position corresponds to the first probe for the second writing task, and the elevation in Viewing Time may arise from the novelty of engaging in a second task with the algorithm.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Others preferred non-transparent feedback because it did not expose word-level errors, potentially leading users to overestimate the competence of the underlying algorithm and believe it was more advanced than it was.

    “However, others preferred non-transparent feedback because it did not expose word-level errors, potentially leading users to overestimate the competence of the underlying algorithm with the consequence that they believed it to be more advanced than it was.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Overall, users had better impressions of the transparent system initially, but those preferences disappeared after using both versions.

    “Overall, users had better impressions of the transparent system initially, but those preferences disappeared after using both versions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    While some users seemed to find transparency highlighting distracting, the cognitive load results do not support this.

    “While some users seemed to find transparency highlighting to be distracting, our cognitive load results do not support this.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationcitation

    The Study 1 retrospective accuracy evaluation is consistent with previous work [80], where the median user found the E-meter to be 'Accurate'.

    “This retrospective accuracy evaluation is consistent with our previous work [80], where the median user found the E-meter to be “Accurate.””
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [80] 80. Aaron Springer and Steve Whittaker. 2018. What are you hiding? algorithmic transparency and user perceptions. In Proceedings of the 2018 AAAI Spring Symposium Series. source

  • May combine two claims.

    Interpretationreasoning

    For users who felt the system was behaving appropriately, transparency led them to notice word-level errors or see non-conformance to their mental models, and may therefore have undermined their judgments of system accuracy.

    “Paradoxically, this benefit did not accrue for users who felt that the system was behaving appropriately; transparency led them to notice word-level errors or see non-conformance to their mental models.Transparency may therefore have undermined their judgments of system accuracy.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Participants' self-reported qualitative explanations accurately reflect the reasons for their predicted accuracy, preferences, and system understanding.

  • Probably not a substantial claim: may not pass S3 (load-bearing). The quote only partly states this claim.

    Study design and conductnone given

    Study 2 was approved by an Institutional Review Board.

    “This study was approved by an Institutional Review Board.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    Knowing users' predicted accuracy was higher for the transparent version, the authors expected this to lead users ultimately to prefer the more transparent version.

    “Therefore, knowing users' predicted accuracy was higher for the transparent version, we would expect that this would lead to users ultimately preferring the more transparent version of the system.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The absence of detailed information in non-transparent feedback can hide errors; this error hiding leads some participants to form approximate but positive working heuristics about how the system operates, which may explain the lack of experience-based preference for transparent systems.

    “Our observations suggest that the absence of detailed information in non-transparent feedback can hide errors; this error hiding leads some participants to form approximate, but positive, working heuristics about how the system operates.This may explain the lack of experience-based preference for transparent systems.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductreasoning

    Study 2 used the same system and transparency manipulations as Study 1 with one difference: it was a between-participants design so users only ever experienced one form of system feedback, to avoid prior system experience influencing judgments.

    “We used the same system and transparency manipulations as Study 1 with one difference.Study 2 is a between-participants design; so users only ever experience one form of system feedback, as we did not want prior system experience to influence users' judgments.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    Participants in Study 2 were randomly allocated to transparent and non-transparent conditions.

    “Participants were randomly allocated to transparent (see Figure 2) and non-transparent (see Figure 1) conditions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Overall, users who saw transparency information seem better able to understand how the system is operating, which increases their accuracy judgments.

    “Overall then, users who saw transparency information seem better able to understand how the system is operating, which increases their accuracy judgments.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Participants' self-reported qualitative explanations accurately reflect the reasons for their predicted accuracy, preferences, and system understanding.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Study 3 was approved by an Institutional Review Board.

    “This study was approved by an Institutional Review Board.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    This second press would satisfy the users in Study 4 who only wanted to see the major factors influencing the algorithm.

    “This second press would satisfy the users in Study 4 who only wanted to see the major factors influencing the algorithm.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationown data

    Study 4 again reveals the complexity involved in presenting transparency information.

    “Study 4 again reveals the complexity involved in presenting transparency information.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Extant literature confirms that system feedback is correct, as high usage of first-person singular pronouns is correlated with depression and negative mood.

    “However, extant literature confirms that system feedback is correct, as high usage of first-person singular pronouns is correlated with depression and negative mood [66,74].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [66] 66. James W. Pennebaker. 2011. The Secret Life of Pronouns: What Our Words Say about Us. Bloomsbury Press. source
    • cites [74] 74. Stephanie Rude, Eva-Maria Gortner, and James Pennebaker. 2004. Language use of depressed and depression- vulnerable college students. Cogn. Emot. 18, 8 (2004), 1121-1133. https://doi.org/10.1080/02699930441000030 source

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    Some of these requirements seem internally inconsistent.

    “This is a challenge, because some of these requirements seem internally inconsistent.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Supporting transparency is complex and there are myriad decisions that affect the user experience when deciding how to operationalize it.

    “Together these results suggest that supporting transparency is complex and there are myriad decisions that affect the user experience when deciding how to operationalize it.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The participants were randomly divided into one of two conditions.

    “The participants were randomly divided into one of two conditions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Even if the system is objectively correct when displaying transparency information, it can still cause users to take issue and result in poorer perceptions of the system.

    “Even if the system is objectively correct when displaying transparency information, it can still cause users to take issue and result in poorer perceptions of the system.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationown data

    P6 moved beyond their initial heuristic that all words are weighted equally and noted how one word has a larger effect in the system.

    “This statement indicates that this user has moved beyond their initial heuristic that all words are weighted equally; they are now noting how one word has a larger effect in the system.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Users engage with transparency first by operating with simple hypotheses and only critically re-examine these when their expectations aren't met.

    “Users seem to engage with transparency first by operating with these simple hypotheses and only critically re-examine these when their expectations aren't met.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Users may need to identify a small number of clear examples of words showing strongly positive and negative affect to form a working model of the system's operation.

    “It may be that users need to identify a small number of clear examples of words showing strongly positive and negative affect, to form a working model of the system's operation.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    This study was approved by an Institutional Review Board.

    “This study was approved by an Institutional Review Board.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Reactions to system explanations depend on the level of error.

    “Our studies did not systematically vary error, however, and prior work indicates that reactions to system explanations depend on the level of error [50,51].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [50] 50. Brian Y. Lim and Anind K. Dey. 2011. Investigating intelligibility for uncertain context-aware applications. In Pro- ceedings of the 13th International Conference on Ubiquitous Computing. 415-424. source
    • cites [51] 51. Brian Y. Lim, Anind K. Dey, and Daniel Avrahami. 2009. Why and why not explanations improve the intelligibility of context-aware intelligent systems. In Proceedings of the 27th International Conference on Human Factors in Computing Systems (CHI'09). 2119. https://doi.org/10.1145/1518701.1519023 source

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Our findings suggest important potential implications for the design of transparent systems that exhibit progressive disclosure.

    “Taken together, our findings suggest important potential implications for the design of transparent systems that exhibit progressive disclosure.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationcitation

    The previous three studies confirm that on some occasions transparency can be beneficial in reducing expectation violations, but on other occasions it can detract.

    “They confirm that on some occasions transparency can be beneficial in reducing expectation violations, but on other occasions it can detract.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationown data

    This observation suggests a principle for transparency presentation that restricts the amount of information presented.

    “This observation suggests a principle for transparency presentation that restricts the amount of information presented.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationcitation

    When interacting with a system, users first engage in rapid, approximate, System 1 thinking and only engage in deeper, more analytic, System 2 thinking when directly prompted or confronted with clearly anomalous information.

    “As with other areas of reasoning, it may be that when interacting with a system, users first engage in rapid, approximate, System 1 thinking and only engage in deeper, more analytic, System 2 thinking when directly prompted or confronted with clearly anomalous information [41,67].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [41] 41. Daniel Kahneman. 2011. Thinking, Fast and Slow. Macmillan. source
    • cites [67] 67. Richard E. Petty and John T. Cacioppo. 1986. The elaboration likelihood model of persuasion. In Advances in Exper- imental Social Psychology, L. Berkowitz (Ed.), Vol. 19. Academic Press, San Diego, CA, 123-205. https://doi.org/10. 1016/S0065-2601(08)60214-2 source

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning

    These overlaps suggest that persuasion and dual process theories could offer important insights for designing transparent algorithms.

    “These overlaps suggest that persuasion and dual process theories could offer important insights for designing transparent algorithms.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    The increase in viewing time for Prompt 4 is likely due to the fact that participants had just begun a second writing task.

    “The increase in viewing time for Prompt 4 is likely due to the fact participants had just begun a second writing task”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductcitation

    Although users generated their own data in the system, results were not directly used to inform other aspects of the user's personal behavior, so the costs of system errors were low.

    “while users generated their own data in our system, results were not directly used to inform other aspects of the user's personal behavior, as in Hollis et al. [38], so the costs of system errors were low.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [38] 38. Victoria Hollis, Artie Konrad, Aaron Springer, Chris Antoun, Matthew Antoun, Rob Martin, and Steve Whittaker. 2017. What does all this data mean for my future mood? actionable analytics and targeted reflection for emotional well-being. Hum.-Comput. Interact. 32, 5-6 (2017), 208-267. https://doi.org/10.1080/07370024.2016.1277724 source

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    While the results do not apply directly to neural methods, they may become more relevant as approximation methods become more common.

    “So, while our results do not apply directly to neural methods, they may become more relevant, as such approximation methods become more common.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationnone given

    The current study indicates a promising design approach involving progressive disclosure.

    “The current study indicates a promising design approach involving progressive disclosure that we intend to explore in future work.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Users will modify high-stakes personal health behaviors based on false inferences about an algorithm's high accuracy.

    “Other work has shown that users will modify high stakes personal health behaviors based on false inferences about an algorithm's high accuracy [39,70].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [39] 39. Victoria Hollis, Alon Pekurovsky, Eunika Wu, and Steve Whittaker. 2018. On being told how we feel: How Algo- rithmic sensor feedback influences emotion perception. Proc. ACM Interact. Mob. Wear. Ubiq. Technol. 2, 3 (2018), 114:1-114:31. https://doi.org/10.1145/3264924 source
    • cites [70] 70. Mashfiqui Rabbi, Min Hane Aung, Mi Zhang, and Tanzeem Choudhury. 2015. MyBehavior: Automatic personalized health feedback from user behaviors and preferences using smartphones. In Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing. 707-718. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    The deployment of a working algorithm meant that results were obtained for situations where the algorithm generated moderate numbers of errors.

    “our deployment of a working algorithm meant that results were obtained for situations where our algorithm generated moderate numbers of errors”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    The limitations of the work raise significant questions about the generalizability of the findings to other contexts.

    “These are clear limitations of the work that naturally raise significant questions about generalizability of our findings to other contexts.”
    Quote in the paperNot a substantial claim
86 of 86

  • The paper relies on this without support.

    Users are the ultimate experts on their own emotions, providing ground truth assessments about whether the system is correct.

    “And users are the ultimate experts on their own emotions, providing them with ground truth assessments about whether the system is correct.”

  • The paper relies on this without support.

    The E-meter model's error rate measured on the EmotiCal train/test split generalizes to E-meter users' mood predictions.

    “The mean absolute error of our model indicates that the E-meter will, on average, err by 15.83 points on a (0,100) scale for each user's mood prediction.”

  • The paper relies on this without support.

    Users' subjective affect judgments accurately reflect the emotional valence of their writing, so the difference between user and system judgments measures expectation violation.

    “Expectation Violation was defined as the difference between the user's subjective affect judgment (their own assessment of the mood valence expressed in their writing) versus the system's overall assessment of the emotional valence that writing.”
    “We then compared this user judgment with the current system emotion rating of their writing, and computed the discrepancy.We refer to this discrepancy as the real-time expectation violation, which is measured six times overall.”

  • The paper relies on this without support.

    Participants' self-reported qualitative explanations accurately reflect the reasons for their predicted accuracy, preferences, and system understanding.

    “Users then provided qualitative explanations for their ratings: “Please give two or more reasons for the accuracy ratings you made on the previous page.””
    “We examined participants' written explanations of their experience-based preferences to better understand their choices.”

  • The paper relies on this without support.

    Prior studies commonly used hypothetical scenarios describing system operation to elicit potential users' reactions to those systems.

    “One method used in prior studies is to provide potential system users with hypothetical scenarios describing system operation and eliciting reactions to those systems.”
5 of 5

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Statistics that don’t add up

Each reported p-value recomputed from its test statistic and degrees of freedom.

  • Inconsistent p-value: t(73) = 5.452, p = .022

    Minor

    The paper reports p = .022 here; recomputing from the test statistic and its degrees of freedom gives p < .001 (two-tailed). The check cannot tell whether the statistic or the p-value is the misprint.

    Reported p = .022 · recomputed p < .001

  • Inconsistent p-value: t(73) = .910,p = .343

    Minor

    The paper reports p = .343 here; recomputing from the test statistic and its degrees of freedom gives p = .366 (two-tailed). The check cannot tell whether the statistic or the p-value is the misprint.

    Reported p = .343 · recomputed p = .366

2 of 2

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • Doesn’t add upt(73) = 5.452reported p = .022 · recomputed p < .001

    “t(73) = 5.452, p = .022”

  • Doesn’t add upt(73) = 0.91reported p = .343 · recomputed p = .366

    “t(73) = .910,p = .343”

  • Not confirmedt(73) = -0.05reported p = .95 · recomputed p = .960

    “t(73) = -.05,p = .95”

  • Consistentχ²(5) = 16.94reported p < .01 · recomputed p = .005

    “χ 2 (5) = 16.94,p < .01”

4 of 4

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Study 3 table names the wrong outcome

    Minor

    The Study 3 methods identify transparency viewing time as the regression's dependent variable, but Table 2's title says the model predicts perceived accuracy judgments.

  • Trust score 4 has two different labels

    Minor

    Study 2 calls a trust score of 4 "Moderately Trustworthy," while the Figure 4 caption calls the same score "Very Trustworthy."

2 of 2

Cites

Flagged first, then by citations

Research works printed but not linked: 24

  1. Angwin Julia, 2016, ProPublica. Retrieved
  2. Bolukbasi Tolga, 2016, Kalai
  3. Insight CCS, 2016, Retrieved
  4. Choe Eun Kyoung, Proceedings of the 32nd Annual ACM Conference on Human Factors in Computing Systems (CHI’14)
  5. Devlin Jacob, 2018, Retrieved
  6. E. Grossi N. Groth P. Mosconi R. Cerutti F. Pace A. Compare and G. Apolone. 2006. Development and validation of the short version of the Psychological General Well-Being Index (PGWB-S). Health Qual. Life Outcomes 4 88 (2006). DOI:10.1186/1477-7525-4-88 · DOI · Paper page
  7. Hart Sandra G., 1988, Staveland
  8. Kim SeungJun, Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems (CHI’15)
  9. Kulesza Todd, Proceedings of the 13th International Conference on Intelligent User Interfaces (IUI’09)
  10. Lakkaraju Himabindu, 2017, Retrieved 18
  11. Brian, Proceedings of the 11th International Conference on Ubiquitous Computing. 195--204
  12. Brian, Proceedings of the 13th International Conference on Ubiquitous Computing. 415--424
  13. Lipton Zachary C., 2016, Retrieved
  14. Scott, 2017, Lundberg and Su-In Lee
  15. Miller Tim, 2017, Retrieved
  16. Petty Richard E., 1986, Cacioppo
  17. Reeves Byron, Television, and New Media Like Real People and Places
  18. Smith David Canfield, 1982, Byte Magazine
  19. Springer Aaron, 2017, Retrieved
  20. Stark Litza, 2000, Processing of the International Conference on Speech and Language Processing (ICSLP-2000)
  21. Warshaw Jeffrey, 2015, Smith
  22. Daniel, 2018, Retrieved
  23. Yang Rayoung, Proceedings of the 2013 ACM International Joint Conference on Pervasive and Ubiquitous Computing (UbiComp’13)
  24. Zisook Miriam, 2016, Retrieved

Cited by

Sorted by citations

Related

Authors
Springer, A., & Whittaker, S.
Affiliations
University of California, Santa Cruz (US)
ORCID
Venue
ACM Transactions on Interactive Intelligent Systems
Publisher
Association for Computing Machinery (ACM)
Published
2020-10-16
Volume
10
Issue
4
Pages
1-32
ISSN
  • 2160-6455 (ISSN-L)
  • 2160-6463
DOI
10.1145/3374218
OpenAlex
W3042124235
Cited by
84
References
66
Citations by year
2026: 30 · 2025: 19 · 2024: 12 · 2023: 12 · 2022: 6 · 2021: 5
FWCI
2.26
Type
journal article
Language
en
Open access
Licence
www.acm.org/publications/policies/copyright_policy#Background (vor)
Locations
ACM Transactions on Interactive Intelligent Systems (publishedVersion)
Funders
NSF (IIS-1321102)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Physical Sciences › Computer Science › Human-Computer Interaction › Innovative Human-Technology Interaction
  • Physical Sciences › Computer Science › Human-Computer Interaction › Usability and User Interface Design
Keywords
progressive disclosure, algorithmic transparency, algorithmic decision-making, algorithmic errors, real-time user feedback, intelligent systems

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-10.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-10.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review