CoChat
Sign Up
Cited by 146Human-Automation Interaction and SafetyExplainable Artificial Intelligence (XAI)Ethics and Social Impacts of AI

Human-AI Collaboration via Conditional Delegation: A Case Study of Content Moderation

Open access
Lai, Vivian, Carton, Samuel, Bhatnagar, Rajat, Liao, Q. Vera, Zhang, Yunfeng, Tan, Chenhao·2022·CHI Conference on Human Factors in Computing Systems
DOI 10.1145/3491102.3501999
Open this paper in CoChat — free accountOpen PDF

Abstract

Despite impressive performance in many benchmark datasets, AI models can still make mistakes, especially among out-of-distribution examples. It remains an open question how such imperfect models can be used effectively in collaboration with humans. Prior work has focused on AI assistance that helps people make individual high-stakes decisions, which is not scalable for a large amount of relatively low-stakes decisions, e.g., moderating social media comments. Instead, we propose conditional delegation as an alternative paradigm for human-AI collaboration where humans create rules to indicate trustworthy regions of a model. Using content moderation as a testbed, we develop novel interfaces to assist humans in creating conditional delegation rules and conduct a randomized experiment with two datasets to simulate in-distribution and out-of-distribution scenarios. Our study demonstrates the promise of conditional delegation in improving model performance and provides insights into design for this novel paradigm, including the effect of AI explanations.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes conditional delegation, where humans create keyword rules defining trustworthy regions for an AI model, and tests it on content moderation. In a 240-participant randomized experiment, conditional delegation improved model precision over the model alone, but achieved complementary performance over manual rule-based moderation only for in-distribution AI, not out-of-distribution AI.

Key findings

  • Conditional delegation with human-created rules improved precision over the model working alone both in-distribution and out-of-distribution (p < 0.001 for average and union precision). On WikiAttack, the model alone already outperformed the manual condition, and conditional delegation further improved precision, achieving complementary performance; on Reddit, it improved the model but did not compensate enough to outperform the manual rule-based approach.
  • Distribution type had the most salient effect: two-way ANOVA found significant effects of distribution type, experimental condition, and their interaction (p < 0.001). For average precision, condition was significant on WikiAttack (p < 0.001) but not Reddit (p = 0.864); for union precision, condition was significant on both (p < 0.001), and on Reddit the global explanation condition was worse than manual (p = 0.004) and predicted labels (p = 0.028).
  • Explanations did not significantly improve conditional delegation performance; global explanations slightly hurt union precision and reward out-of-distribution. Explanations improved efficiency: rules per minute had a significant condition effect (p = 0.008), with global explanations vs predicted labels significant (p = 0.031), and global explanations led to higher overlap with rationale words.
  • Participants were more engaged with delegation support than manual (actions p < 0.001; manual vs predicted labels/global p < 0.001, local p = 0.009). Subjective measures were generally positive with no strong condition differences; local explanations trended toward the best experience on WikiAttack.

Methods

  • Between-subject randomized experiment with 240 Mechanical Turkers (30 per condition per distribution; 120 in-distribution, 120 out-of-distribution; 6 of 240 filtered from analysis). Four conditions: predicted labels; predicted labels + local explanations; predicted labels + local explanations + global explanations; manual. Two datasets: WikiAttack (in-distribution) and Reddit hate speech (out-of-distribution). Participants had to submit at least 10 rules; bonus was $0.10 per 100 toxic comments correctly reported and −$0.10 per 100 nontoxic comments mistakenly reported, bounded $0–$2.
  • Model: rationale-style neural architecture with BERT encoders, trained on WikiAttack (70k training, 23K test comments) and applied to Reddit (22K comments); recall weight 0.5 chosen. Measures: average precision, union precision, reward, engagement actions, elapsed time, rules per minute, and Likert subjective measures. Analysis: t-tests, two-way/one-way ANOVA, and Tukey’s HSD.

Limitations

  • Stated: the work is one instantiation of conditional delegation; precision/coverage tradeoffs may vary by application; participants are not representative of content moderators; evaluations are limited by the roughly 10 rules participants created in about 10 minutes; model, dataset, and explanation choices affect outcomes; the study does not account for future rule updates or existing moderator rules, and uses a narrow toxicity component of content moderation.
  • Evident: the in-distribution vs out-of-distribution contrast is confounded with model performance and dataset/community differences, which the paper itself notes as a complexity; reward is highly volatile because a small number of keywords can dominate, especially out-of-distribution, as the paper states.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3491102.3501999

Cites

70 of 70 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 71 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    Model explanations can improve efficiency in identifying delegation conditions.

    “We also found that model explanations can improve efciency in identifying delegation conditions and, with weak evidence, improve user experiences.”
    Quote in the paperSubstantial claim
  • Findingown data

    Through a randomized experiment with 240 mechanical turkers, even crowdworkers are able to create high-quality rules that lead to higher precision with conditional delegation than the model working alone.

    “Through a randomized experiment with 240 mechanical turkers, we show that even crowdworkers are able to create high-quality rules that lead to higher precision with conditional delegation than the model working alone.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The study assumes the same workflow as the manual rule-based approach: comments falling outside keyword rules are ignored without taking an action.
  • Findingown data

    Model explanations, with weak evidence, improve user experiences.

    “We also found that model explanations can improve efciency in identifying delegation conditions and, with weak evidence, improve user experiences.”
    Quote in the paperNot a substantial claim
  • Findingcitation

    In-distribution AI already outperforms the manual rule-based approach for content moderation.

    “which already outperforms the manual rule-based approach for content moderation”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The study assumes the same workflow as the manual rule-based approach: comments falling outside keyword rules are ignored without taking an action.
  • Findingown data

    Anecdotally, the authors deployed a version of their model on a subreddit to report comments predicted as toxic, and moderators asked them to shut it down due to high false positive rates (low precision).

    “Anecdotally, we deployed a version of our model on a subreddit to report comments that are predicted as toxic, and the moderators asked us to shut it down due to high false positive rates (i.e., low precision).”
    Quote in the paperNot a substantial claim
  • Classified as probably not a substantial claim. 56
  • May combine two claims. 55
  • The quote only partly states this claim. 2

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

    Findingcitation

    The rationale-style neural architecture achieves competitive accuracy while having built-in interpretability because the prediction is made exclusively based on rationales.

    “This model can achieve competitive accuracy while having built-in interpretability because the prediction is made exclusively based on rationales.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [52] source

      Not checked in the free run.

  • May combine two claims.

    Findingreasoning

    The model achieves competitive accuracy while having the ability to provide explanations directly by showing the rationales on which the prediction is based.

    “In short, this model achieves competitive accuracy while having the ability to provide explanations directly by showing the rationales on which the prediction is based on.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    There is a clear performance drop out-of-distribution (e.g., F1 drops from about ∼0.8 to ∼0.6), validating the choice of Reddit as an out-of-distribution scenario.

    “We observe a clear performance drop out-of-distribution (e.g., F1 drops from about ∼0.8 to ∼0.6), which validated our choice of Reddit as an out-of-distribution scenario.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    Participants report a good global understanding of the model (M=3.37, SD=0.97) and local understanding of individual predictions (M=3.56, SD=1.05) on WikiAttack than Reddit, possibly related to model performance difference.

    “Overall, participants report a good global understanding of the model (M=3.37,SD=0.97) and local understanding of individual predictions (M=3.56,SD=1.05) on Wiki-Attack than Reddit, possibly related to the diference in model performance between distribution types.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    On WikiAttack, the model working alone outperforms the manual condition, and conditional delegation further improves precision.

    “In particular, on WikiAttack, the model working alone already outperforms the manual condition, and conditional delegation further improves the precision.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Two-way ANOVA found a statistically significant effect of distribution type and interaction between distribution type and experimental condition on reward (p<0.001).

    “Two-way ANOVA fnds a statistically signifcant efect of distribution type and interaction between distribution type and experimental condition (p <0.001).”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The reward mechanism assumes that the cost of wrongly reported non-toxic comments (false positive) equals the benefit of correctly reported toxic comments (true positive).

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    Participants with predicted labels only spent the most time on the task, but the effect of experimental condition on time was only borderline significant (p=0.074) with no statistically different pairs.

    “Fig. 14a shows the time spent on the interactive interface in each condition: participants with predicted labels only spent the most time on this task.This result is consistent with the number of actions because it likely requires more time to take more actions.However, the diference is relatively weak: one-way ANOVA shows that the efect of experimental conditions is only borderline signifcant (p”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    On Reddit, the global explanation condition is worse than the manual condition in union precision (p=0.004), and only showing prediction labels is worse than manual (p=0.028).

    “On Reddit, we found the global explanation condition is worse than the manual condition (p =0.004), and only showing prediction labels (p =0.028).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The study achieved complementary performance in-distribution on WikiAttack, but not out-of-distribution on Reddit.

    “We are able to achieve complementary performance in-distribution on WikiAttack, but not out-ofdistribution on Reddit.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The effect of distribution type is the most salient, indicating a clear difference between in-distribution and out-of-distribution.

    “The efect of distribution type is the most salient, suggesting a clear diference between in-distribution and out-of-distribution.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The word 'fuck' does not show up in top 10 for the manual condition on Reddit, but shows up in the other conditions.

    “The word does not show up in top 10 for the manual condition, but shows up in the other conditions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Adding the global explanation feature unexpectedly hurt people's ability to choose rules with both high coverage and high precision, leading to slightly lower human-AI performance in union precision and reward.

    “However, adding the global explanation feature seems to unexpectedly hurt people's ability in choosing rules with both high coverage and high precision, and lead to slightly lower human-AI performance in union precision and reward.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data

    Global explanations lead to much higher overlap between human-created rules and the most frequent words in rationales than other conditions.

    “Global explanations lead to much higher overlap between most frequent words in rationales (Fig. 14c).”
    • Also stated in: § 5.2 Efciency and Engagement · plain
      “Fig. 14c shows that global explanations lead to much higher overlap than the other conditions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    One-way ANOVA found a statistically significant effect of experimental condition on number of actions (p<0.001), and the manual condition had significantly fewer actions than other conditions (p<0.001 for predicted labels and global explanations, p=0.009 for local explanations).

    “One-way ANOVA also fnds a statistically signifcant efect in experimental condition (p < 0.001).Post-hoc Tukey's HSD shows the negative diference between the manual condition and other experimental conditions are all statistically signifcant (p <0.001 for predicted labels and global explanations, p =0.009 for local explanations).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    On Reddit, experimental condition has a statistically significant effect on reward (p=0.018), and only the difference between manual and global explanations is significant (p=0.018).

    “On Reddit, the experimental condition also has a statistically signifcant efect (p =0.018).Post-hoc Tukey's HSD shows that only the diference between the manual condition and global explanations is signifcant (p =0.018).”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The reward mechanism assumes that the cost of wrongly reported non-toxic comments (false positive) equals the benefit of correctly reported toxic comments (true positive).

  • May combine two claims.

    Findingown data

    Conditional delegation based on top-precision words leads to greater precision than the model working alone, suggesting users can improve model precision by identifying these words.

    “Conditional delegation based on these words leads to greater precision than the model working alone (dashed lines), suggesting that users can improve the precision of the model by identifying these words for conditional delegation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In all evaluation measures of engagement and efficiency, only experimental condition had statistically significant effects, suggesting patterns with two distribution types are comparable.

    “In all evaluation measures of engagement and efciency, we only fnd statistically signifcant efects of experiment conditions, suggesting that patterns with two distribution types are comparable.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For WikiAttack, there is a weak trend that local explanations lead to less subjective workload while adding global explanation has the opposite effect; these patterns do not hold for Reddit.

    “For WikiAttack, we observe a weak trend that local explanations lead to less subjective workload (lower mental demand, more feeling of success, and less negative emotion) while adding global explanation has the opposite efect.These patterns, however, do not hold for Reddit.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    For average precision, experimental condition has a significant effect on WikiAttack (p<0.001) but not on Reddit (p=0.864).

    “For average precision (Fig. 10), experimental condition has a signifcant efect in Wiki-Attack (p <0.001), but not in Reddit (p =0.864).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    On WikiAttack, experimental condition has a statistically significant effect on reward (p=0.01), and the manual condition is significantly worse than other experimental conditions with delegation support features (p<0.001 for global explanations, p=0.007 for local explanations, p=0.004 for predicted labels).

    “On WikiAttack (Fig. 11c), we fnd a statistically signifcant efect of experimental condition (p =0.01), and post-hoc Tukey's HSD shows that the manual condition is signifcantly worse than other experimental conditions with delegation support features (p <0.001 for global explanations, p =0.007 for local explanations, and p =0.004 for predicted labels).”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The reward mechanism assumes that the cost of wrongly reported non-toxic comments (false positive) equals the benefit of correctly reported toxic comments (true positive).

  • May combine two claims.

    Findingown data

    There is a non-significant trend that conditions with explanation, especially local explanation, result in better confidence for WikiAttack, but not for Reddit.

    “There is a non-signifcant trend that conditions with explanation, especially local explanation, result in better confdence for Wiki-Attack, but not for Reddit, where the model performs relatively poorly.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Overall, participants were neutral about mental demand (M=3.15, SD=1.08), agreed they felt successful (M=3.95, SD=0.91), and disagreed they felt negative emotions (M=1.93, SD=1.03).

    “Overall, participants were neutral about whether the task was mentally demanding (M=3.15,SD=1.08), agreed that they felt relatively successful in accomplishing the task (M=3.95,SD=0.91), and disagreed that they felt negative emotions (M=1.93,SD=1.03).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Overall, participants report a better understanding of the model as a whole and individual predictions on WikiAttack than Reddit, although the differences are not statistically significant.

    “Overall, participants report a better understanding of the model as a whole and individual predictions on WikiAttack than Reddit, although the diferences are not statistically signifcant.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Differences between Reddit and WikiAttack: 'cunt' and 'retard' are not common in rationales in WikiAttack but are among the top five on Reddit.

    “We can immediately observe diferences between Reddit and WikiAttack:”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Participants spent more time working on conditional delegation than the manual condition, but the efficiency is improved with explanations, especially global explanations.

    “Participants spent more time working on conditional delegation than the manual condition, but the efciency is improved with explanations, especially global explanations.”
    Quote in the paperSubstantial claim

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

    Findingnone given

    The nature of comments on Wikipedia and Reddit differs substantially.

    “Second, the nature of comments on Wikipedia and Reddit difers substantially.”
    Quote in the paperNot a substantial claim

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

    Findingnone given

    Crowdworkers can already create keyword-based rules that achieve greater precision than the model working alone.

    “It is impressive that crowdworkers can already create keyword-based rules that achieve greater precision than the model working alone.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Conditional delegation with all delegation support features leads to much better engagement and more submitted rules than the manual condition.

    “Conditional delegation with all delegation support features leads to much better engagement and more submitted rules than the manual condition.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    Local explanations seem to have less of a priming effect than global explanations but still improve efficiency.

    “It is worth noting that local explanations seem to have less of a priming efect than global explanations but still improves efciency.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Having an appropriate mental model allows people to accurately anticipate a system's behaviors and interact more effectively.

    “Having an appropriate mental model allows people to accurately anticipate a system's behaviors and interact more efectively.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [65]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Content moderation involves a large amount of relatively low-stakes decisions where the key challenge is dealing with the massive scale of comments.

    “One such example is content moderation, where moderator decisions on individual comments for further actions (e.g., hiding the content or prompting further review, depending on the community policy) are of limited consequence; instead the key challenge lies in dealing with the massive scale of comments.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [18] source

      Not checked in the free run.

    • cites [33]

      Not checked in the free run.

    • cites [35] source

      Not checked in the free run.

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

    Premisecitation

    Many studies have investigated the role of AI in assisting humans in making individual decisions.

    “Many studies have investigated the role of AI in assisting humans in making individual decisions [9,16,36, 37, 50, 51, 55, 58, 69, 83, 84, 89], e.g., predicting whether a person will recidivate in the near future.”
    Quote in the paperNot a substantial claim0 of 10 sources checked
    • cites [9] source

      Not checked in the free run.

    • cites [16]

      Not checked in the free run.

    • cites [36] source

      Not checked in the free run.

    • cites [37] source

      Not checked in the free run.

    • cites [50] source

      Not checked in the free run.

    • cites [51] source

      Not checked in the free run.

    • cites [55]

      Not checked in the free run.

    • cites [58]

      Not checked in the free run.

    • cites [69] source

      Not checked in the free run.

    • cites [83] source

      Not checked in the free run.

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

    Interpretationreasoning

    Coverage warrants that the AI model can identify as many toxic comments as possible to alleviate scalability issues.

    “Coverage warrants that the AI model can identify as many toxic comments as possible to alleviate the scalability issues.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    HCI studies on XAI have found somewhat mixed results on enhancing user trust, satisfaction, and willingness to adopt AI systems.

    “HCI studies on XAI have found explanations to improve user understanding of AI systems [12,22,32], and somewhat mixed results on enhancing user trust [22], satisfaction [32] and willingness to adopt AI systems [80].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [22] source

      Not checked in the free run.

    • cites [32]

      Not checked in the free run.

    • cites [80] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning

    Conditional delegation employs a greater level of automation than human-AI collaboration on every single decision and provides humans with active control on when to use an AI model and in what ways.

    “This approach employs a greater level of automation than human-AI collaboration on every single decision and provides human with active control on when to use an AI model and in what ways.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Manual rule-based approaches miss out the benefit of AI because rigid rules often do not work on informal languages such as social media posts.

    “This approach, however, misses out the benefit of AI especially since rigid rules often do not work on informal languages such as social media posts (e.g., containing swear words without being toxic).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [18] source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Human-AI collaboration on every single decision is not scalable and is thus less appropriate for tasks involving a large amount of relatively low-stakes decisions.

    “However, human-AI collaboration on every single decision is not scalable and is thus less appropriate for tasks involving a large amount of relatively low-stakes decisions.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    The task for humans to create conditional delegation rules shares similarity with what many social media moderators already do by writing manual automation rules.

    “Notably, the task for humans to create conditional delegation rules share some similarity with what many social media moderators are already doing by writing manual automation rules to deal with the massive amount of comments (Fig. 1(C)).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [18] source

      Not checked in the free run.

    • cites [44] source

      Not checked in the free run.

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

    Interpretationreasoning

    Precision ensures that AI behavior is trustworthy in delegation mode and avoids unnecessary actions.

    “Precision ensures that AI behavior is indeed trustworthy in the delegation mode and avoids unnecessary actions, whether it is mistaken deletion or extra work for further review.”
    Quote in the paperSubstantial claim

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

    Interpretationnone given

    Our work is the first study with controlled experiments to examine the effect of conditional delegation.

    “Furthermore, to the best of our knowledge, our work is the frst study with controlled experiments to examine the efect of conditional delegation.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    AI-assisted decision-making tends to focus on high-stakes decisions and is not scalable in the number of decisions because humans are required to make each decision.

    “This approach tends to focus on high-stakes decisions and are not scalable in the number of decisions because humans are required to make each decision.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [37] source

      Not checked in the free run.

    • cites [51] source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    HCI studies on XAI have found explanations to improve user understanding of AI systems.

    “HCI studies on XAI have found explanations to improve user understanding of AI systems [12,22,32], and somewhat mixed results on enhancing user trust [22], satisfaction [32] and willingness to adopt AI systems [80].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [12] source

      Not checked in the free run.

    • cites [22] source

      Not checked in the free run.

    • cites [32]

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    People's mental model can be refined by explanations of how the system works.

    “People's mental model can be refned by explanations of how the system works.”
    Quote in the paperSubstantial claim0 of 4 sources checked
    • cites [1] source

      Not checked in the free run.

    • cites [41] source

      Not checked in the free run.

    • cites [54] source

      Not checked in the free run.

    • cites [71] source

      Not checked in the free run.

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

    Premisecitation

    The empirical results of prior work on complementary performance in AI-assisted decision-making are mixed at best because of insufficient complementarity in human and AI's domain knowledge or lack of ability for people to judge reliability of AI recommendations.

    “The empirical results, however, are mixed at best, because there was either insufcient complementarity in human and AI's domain knowledge or a lack of ability for people to judge the reliability of AI recommendations.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [8] source

      Not checked in the free run.

    • cites [51] source

      Not checked in the free run.

    • cites [89] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning

    Such rules can be created for any model of choice, so conditional delegation is orthogonal to research on improving AI capability.

    “Such rules can be created for any model of choice, so it is orthogonal to the research on improving the capability of AI.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Chiang and Yin found that humans rely on AI more out-of-distribution, where AI performance is worse.

    “Chiang and Yin [23] examined human reliance on the model in human-AI decision making and found that surprisingly humans rely on AI more out-of-distribution, where the AI performance is worse.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [23] source

      Not checked in the free run.

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

    Premisecitation

    Liu et al. suggested that complementary performance is more plausible for out-of-distribution examples because of AI's performance drop.

    “They suggested that complementary performance is more plausible for out-of-distribution examples because of AI's performance drop.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [56]

      Not checked in the free run.

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

    Premisecitation

    Chandrasekharan et al. found that hard-coded rules are prone to mistakes.

    “Chandrasekharan et al. [18] also found that hard-coded rules are prone to mistakes.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [18] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Models trained with historical data can present issues such as gender bias and racial bias, potentially exacerbating structural inequalities.

    “models are trained with historical data and can present issues such as gender bias and racial bias in AI models [68,73], potentially exacerbating structural inequalities [10]”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [68] source

      Not checked in the free run.

    • cites [73] source

      Not checked in the free run.

    • cites [10] source

      Not checked in the free run.

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

    Premisecitation

    Jhaver et al. found through interviews with 16 moderators that AutoModerator improves the efficiency of moderation.

    “Through interviews with 16 moderators, Jhaver et al. [44] found that AutoModerator improves the efciency of moderation.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [44] source

      Not checked in the free run.

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

    Study design and conductreasoning

    The datasets that participants explore to create rules are of comparable size between Wikipedia (in-distribution) and Reddit (out-of-distribution).

    “As a result, the datasets that participants explore to create rules are of comparable size between Wikipedia (in-distribution) and Reddit (out-of-distribution).”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    The existence of distribution shift is a strong motivation for some form of conditional delegation so that humans can identify trustworthy regions.

    “The existence of distribution shift is a strong motivation for some form of conditional delegation so that humans can identify the trustworthy regions.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    There exists a need for audit tools to monitor the performance of keyword rules.

    “However, there exists a need for audit tools to monitor the performance of the keyword rules.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [44] source

      Not checked in the free run.

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

    Premisecitation

    AutoModerator fundamentally changes the work of moderators and may introduce additional unnecessary work.

    “They also highlight the fact that AutoModerator fundamentally changes the work of moderators and may introduce additional unnecessary work.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [44] source

      Not checked in the free run.

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

    Premisecitation

    The explanation method has attracted criticism for producing rationales that don't necessarily align with human reasoning, but it has the advantage of producing rationales that are, by construction, sufficient.

    “Note that this explanation method has attracted some criticism for producing rationales that don't necessarily align with human reasoning [90], but it has the advantage of producing rationales that are, by construction, sufcient”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [90]

      Not checked in the free run.

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

    Premisecitation

    The difference between training dataset and deployment data is called distribution shift, which often results in a performance drop.

    “The diference between the training dataset and the deployment data is called distribution shift, which often results in a performance drop [24,46,61].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [24] source

      Not checked in the free run.

    • cites [46] source

      Not checked in the free run.

    • cites [61] source

      Not checked in the free run.

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

    Interpretationreasoning

    In our setup, strong AI performance in-distribution is likely more critical for human-AI collaborative performance.

    “In our setup, however, as we conditionally delegate decisions to AI, strong AI performance in-distribution is likely more critical for the human-AI collaborative performance.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    The key advantage of rule-based methods is that it is entirely under the control of moderators.

    “The key advantage of this method is that it is entirely under the control of moderators.”
    Quote in the paperNot a substantial claim

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

    Interpretationnone given

    Conditional delegation combines traditional rule-based approaches and AI models by providing moderators with the ability to decide when to trust or distrust the AI model.

    “Conditional delegation combines traditional rule-based approaches and AI models by providing moderators with the ability to decide when to trust or distrust the AI model.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given

    Conditional delegation can be used to tailor different requirements of precision and tune trade-offs between false positives and false negatives.

    “It can also be used to tailor diferent requirements of precision and tune the tradeofs between false positives and false negatives.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    McCoy et al. found that state-of-the-art models in natural language inference adopt three fallible syntactic heuristics and perform around random chance when tested on examples where these heuristics fail.

    “For instance, McCoy et al. [61] fnd that state-of-the-art models in natural language inference adopt three fallible syntactic heuristics and perform around random chance when tested on examples where these heuristics fail.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [61] source

      Not checked in the free run.

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

    Premisecitation

    Reddit moderators can configure an AutoModerator bot to set rules for reporting or deleting all comments that contain certain words.

    “For instance, Reddit moderators can confgure an AutoModerator bot to set rules for reporting or deleting all comments that contain certain words.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [44] source

      Not checked in the free run.

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

    Premisecitation

    The scale of content is the most important argument for some form of automation in content moderation.

    “The scale of content is the most important argument for some form of automation in content moderation [33,35].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [33]

      Not checked in the free run.

    • cites [35] source

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    Conditional delegation is more effective than the model working alone, and even laypeople can create high-quality conditional delegation rules for content moderation.

    “In short, our results demonstrate that conditional delegation is more efective than the model working alone, and that even laypeople are able to create high-quality conditional delegation rules for content moderation.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Turkers are a sufficiently good sample to compare whether conditional delegation improves content moderation outcomes over the baseline condition.
    • Rests on an assumption: Social media content moderation is often performed by part-time volunteers whose expertise varies.

  • May combine two claims.

    Interpretationown data

    Compared to the manual rule-based approach, the advantage of human-AI collaboration via conditional delegation may depend on the base performance of the AI and may not be sufficient if the AI significantly under-performs, e.g., on out-of-distribution comments.

    “Compared to the manual rule-based approach currently used for content moderation, advantage of our human-AI collaborative approach via conditional delegation may depend on the base performance of the AI, and may not be sufcient if the AI signifcantly under-performs, e.g., when used on out-ofdistribution comments.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: In practice, an AI model rarely faces exactly the same distribution in deployment.

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

    Study design and conductown data

    The study employed a between-subject design with four experimental conditions.

    “We employ a between-subject design with the four conditions described above.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Overall, subjective measures show a relatively positive experience across the board, but not strong difference between conditions.

    “Overall, subjective measures show a relatively positive experience across board, but not strong diference between conditions.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    Conditional delegation can potentially alleviate AI bias by giving users freedom to choose trustworthy regions based on domain knowledge or fairness.

    “Additionally, conditional delegation can potentially alleviate AI bias as we give users the freedom to choose trustworthy regions based on their domain knowledge or notion of fairness.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    Discovering such words in a real-world setting could be very challenging.

    “However, we are able to recognize such words with the beneft of a fully-labeled dataset (i.e., oracle access) -discovering them in a real-world setting could be very challenging.”
    Quote in the paperNot a substantial claim

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

    Study design and conductreasoning

    The study recruited 30 participants per condition, totaling 120 participants per distribution and 240 participants overall.

    “As such, we recruited 30 participants for each condition, summing up to 120 participants each for in-distribution and out-of-distribution and 240 participants in total.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    There is some evidence that when the model performs well (WikiAttack), local explanations provide the best experience: strong performance with relatively high efficiency, less subjective workload and more confidence in the outcomes.

    “There is some evidence that when the model performs well (WikiAttack), local explanations provide the best experience: strong performance with relatively high efciency, less subjective workload and more confdence in the outcomes.”
    Quote in the paperSubstantial claim

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

    Study design and conductown data

    Pilot studies with lab mates and turkers ensured interactions were intuitive and the interface was easy to use.

    “We conducted pilot studies with lab mates at the same university and mechanical turkers to gather feedback and ensured that interactions were intuitive and the interface was easy to use.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Humans, including non-expert turkers, are able to create rules that improve model precision, suggesting conditional delegation is a promising direction.

    “These observations demonstrate that humans, in our case turkers who are not experts in content moderation, are able to create rules that improve model precision, suggesting that conditional delegation can be a promising direction to pursue.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Turkers are a sufficiently good sample to compare whether conditional delegation improves content moderation outcomes over the baseline condition.
    • Rests on an assumption: Social media content moderation is often performed by part-time volunteers whose expertise varies.

  • May combine two claims.

    Interpretationreasoning

    Providing explanations can improve efficiency by hinting on rules to consider, but can also mislead people to use high-frequency but not necessarily high-precision rules.

    “Providing explanations can improve efciency by hinting on rules to consider, but can also mislead people to use high-frequency but not necessarily high-precision rules.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S1 (asserted).

    Study design and conductnone given

    The study aimed to obtain 0.95 power to detect an effect size of 0.1 in precision at the standard 0.01 alpha error probability using F-tests.

    “We aimed to obtain 0.95 power to detect an efect size of 0.1 in precision at the standard 0.01 alpha error probability using F-tests.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    People are more engaged when performing conditional delegation than working on creating manual rules, with more actions and a tendency to spend more time.

    “Taken together, our results show that people are more engaged when performing conditional delegation than working on creating manual rules, with more actions and a tendency to spend more time on the task.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Delegation support features can help users identify good rules when the model performs well, but may mislead people when the model performs poorly.

    “This observation suggests that the delegation support features can help users identify good rules when the model performs well, but may mislead people when the model performs poorly.”
    Quote in the paperSubstantial claim

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

    Study design and conductnone given

    The study was approved by the University IRB.

    “The study is approved by the University IRB.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Even a very conservative application of this model is still producing 2 false positives for every 3 true positives-impractical for use by real moderators.

    “Thus, even a very conservative application of this model is still producing 2 false positives for every 3 true positives-impractical for use by real moderators, and something we would like to be able to improve on via conditional delegation.”
    Quote in the paperSubstantial claim

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

    Study design and conductreasoning

    All participants satisfied screening criteria: at least 1000 HITs, 99% approval, US residence, and adult content qualification.

    “To ensure high quality responses, all participants satisfy the following criteria: (1) performed at least 1000 HITs; (2) approval of 99% performed HITs in previous requesters; (3) reside in the United States; 4) has the adult content qualifcation since our task shows toxic comments.”
    Quote in the paperSubstantial claim

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

    Study design and conductreasoning

    The experiment did not allow repeated participants.

    “The experiment follows a between-subject design therefore we do not allow any repeated participants.”
    Quote in the paperNot a substantial claim

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

    Study design and conductreasoning

    Participants were asked three attention-check questions to filter out those who did not pay attention to instructions.

    “After the instruction, participants are asked three attentioncheck questions to allow us to flter out participants who did not pay attention to the instructions.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    Crowdworkers are not representative of moderators.

    “First, crowdworkers are not representative of moderators, who have way more experience with their platform's data.”
    • Also stated in: § The promise of conditional delegation. · plain
      “Second, our participants are not representative of content moderators.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Moderators are more familiar with the moderation process and more knowledgeable about important words.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationreasoning

    The process of conditional delegation could introduce human bias if one's notion of fairness is ill conceived.

    “However, the fip side is that this process could introduce human bias, if for example one's notion of fairness is ill conceived.”
    Quote in the paperNot a substantial claim

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

    Study design and conductnone given

    Our work represents one instantiation of conditional delegation.

    “First, our work represents one instantiation of conditional delegation.”
    Quote in the paperNot a substantial claim

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

    Study design and conductnone given

    Our work does not take into account any future updates.

    “Third, prior work has shown that moderators often update the rules used by AutoModerator [18,44] and our work does not take into account any future updates.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Prior work has shown that moderators often update the rules used by AutoModerator.

    “Third, prior work has shown that moderators often update the rules used by AutoModerator [18,44] and our work does not take into account any future updates.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [18] source

      Not checked in the free run.

    • cites [44] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    Depending on the workflow, avoiding false negatives could be important in other instantiations.

    “Depending on the workflow, avoiding false negatives could be important in other instantiations.”
    Quote in the paperNot a substantial claim

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

    Study design and conductnone given

    Our choice of model, datasets, and explanations affect the experimental outcome.

    “Third, our choice of model, datasets, and explanations afect the experimental outcome.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    The AI model that we use represents a narrow component in content moderation.

    “Last but not least, content moderation involves a wide range of diferent rules beyond toxicity, and even the policies under the umbrella term of toxicity can vary, so the AI model that we uses represents a narrow component in content moderation.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Content moderation involves a wide range of different rules beyond toxicity.
    • Rests on an assumption: Even the policies under the umbrella term of toxicity can vary.

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

    Interpretationcitation

    These observations echo concerns of unintended consequences with the use of explanations in human-AI collaboration.

    “These observations echo concerns of unintended consequences with the use of explanations in human-AI collaboration [8,37,51].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [8] source

      Not checked in the free run.

    • cites [37] source

      Not checked in the free run.

    • cites [51] source

      Not checked in the free run.

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

    Interpretationown data

    Our case study shows the promise of conditional delegation.

    “Our case study shows the promise of conditional delegation, but further study is required in each application domain of interest to develop the best design for human-AI collaboration in identifying delegation conditions.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductown data

    We choose the model with 0.5 because it achieves competitive performance both in-distribution and out-of-distribution.

    “We choose the model with 0.5 because it achieves competitive performance both in-distribution and out-of-distribution.”
    Quote in the paperSubstantial claim

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

    Interpretationnone given

    The tradeoff between true positives, true negatives, false positives, and false negatives can vary in practice depending on the application and the actions taken according to AI predictions.

    “This tradeof between true positives, true negatives, false positives, and false negatives can vary in practice depending on the application and the actions taken according to AI predictions.”
    Quote in the paperNot a substantial claim

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

    Study design and conductnone given

    We do not leverage any existing rules that moderators have created.

    “Neither do we leverage any existing rules that moderators have created.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning

    Our experiment setup is only a first step towards using conditional delegation in content moderation.

    “However, we recognize that our experiment setup is only a frst step towards using conditional delegation in content moderation.”
    Quote in the paperNot a substantial claim
96 of 96

  • The authors state this as an assumption.

    The study assumes the same workflow as the manual rule-based approach: comments falling outside keyword rules are ignored without taking an action.

    “We assume the same workflow in our study and only focus on the precision and coverage related metrics for comments within the scope of keywords rules.”

  • The paper relies on this without support.

    Content moderators often do not have labeled comments to quantify model performance.

    “When given an AI model, content moderators often do not have labeled comments to quantify model performance.”

  • The paper relies on this without support.

    In content moderation, recall is often less of a priority given limited time for content moderators, who are often volunteers, to deal with massive incoming comments.

    “In the context of content moderation, recall (identifying all toxic comments) is often less of a priority given the limited time for content moderators, who are often volunteers, to deal with a massive amount of incoming comments.”

  • The paper relies on this without support.

    The current workflow using manual rule-based approach ignores comments falling outside the rules without taking an action.

    “This is reflected in the current workflow using the manual rule-based approach (Fig. 1(C)), where comments falling outside the rules are ignored without taking an action.”

  • The paper relies on this without support.

    In a real-world scenario, it is unlikely that models are used to classify data that is exactly the same as the training dataset.

    “In a real-world scenario, it is unlikely that models are used to classify data that is exactly the same as the training dataset.”

  • The paper relies on this without support.

    Current AI models rely on identifying patterns in training datasets.

    “Current AI models rely on identifying patterns in training datasets.”

  • The authors state this as an assumption.

    The reward mechanism assumes that the cost of wrongly reported non-toxic comments (false positive) equals the benefit of correctly reported toxic comments (true positive).

    “To make the calculation easy to understand for participants, this reward makes a simplifed assumption that the cost of wrongly reported non-toxic comments (false positive) equals the beneft of correctly reported toxic comments (true positive).”

  • The authors state this as an assumption.

    Turkers are a sufficiently good sample to compare whether conditional delegation improves content moderation outcomes over the baseline condition.

    “Furthermore, we believe turkers are a sufciently good sample for us to compare whether conditional delegation improves the content moderation outcomes over the baseline condition, and expert users are likely to further enhance the improvement pattern, if any.”

  • The paper relies on this without support.

    Social media content moderation is often performed by part-time volunteers whose expertise varies.

    “We note that while this recruiting choice may limit the generalizability of our results, social media content moderation is often performed by part-time volunteers whose expertise varies.”

  • The paper relies on this without support.

    In practice, an AI model rarely faces exactly the same distribution in deployment.

    “In practice, it is rarely the case that an AI model faces exactly the same distribution in deployment.”

  • The paper relies on this without support.

    Keyword-based rules are a reasonable approach in content moderation given that rule-based methods are already used in AutoModerator.

    “Keyword-based rules are a reasonable approach in content moderation given that rule-based methods are already used in AutoModerator.”

  • The paper relies on this without support.

    In practice, moderators usually have historical data on which moderation decisions were made.

    “Second, in practice, moderators usually have historical data on which moderation decisions were made.”

  • The paper relies on this without support.

    Content moderation involves a wide range of different rules beyond toxicity.

    “Last but not least, content moderation involves a wide range of diferent rules beyond toxicity, and even the policies under the umbrella term of toxicity can vary, so the AI model that we uses represents a narrow component in content moderation.”

  • The paper relies on this without support.

    Even the policies under the umbrella term of toxicity can vary.

    “Last but not least, content moderation involves a wide range of diferent rules beyond toxicity, and even the policies under the umbrella term of toxicity can vary, so the AI model that we uses represents a narrow component in content moderation.”

  • The paper relies on this without support.

    Moderators are more familiar with the moderation process and more knowledgeable about important words.

    “As moderators are more familiar with the moderation process and more knowledgeable about important words, experts might fnd the interface more useful than crowdworkers.”
15 of 15

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?

Cites

Flagged first, then by citations

Research works printed but not linked: 22

  1. Gagan Bansal , Besmira Nushi , Ece Kamar , Walter  S Lasecki , Daniel  S Weld , and Eric Horvitz . 2019 . Beyond Accuracy: The Role of Mental Models in Human-AI Team Performance . In Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, Vol. 7. 2–11 . Gagan Bansal, Besmira Nushi, Ece Kamar, Walter S Lasecki, Daniel S Weld, and Eric Horvitz. 2019. Beyond Accuracy: The Role of Mental Models in Human-AI Team Performance. In Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, Vol. 7. 2–11.
  2. Noam Brown and Tuomas Sandholm . 2019. Superhuman AI for multiplayer poker. Science 365, 6456 ( 2019 ), 885–890. Noam Brown and Tuomas Sandholm. 2019. Superhuman AI for multiplayer poker. Science 365, 6456 (2019), 885–890.
  3. Samuel Carton , Qiaozhu Mei , and Paul Resnick . 2020 . Feature-Based Explanations Don’t Help People Detect Misclassifications of Online Toxicity . In Proceedings of the International AAAI Conference on Web and Social Media, Vol. 14 . 95–106. Samuel Carton, Qiaozhu Mei, and Paul Resnick. 2020. Feature-Based Explanations Don’t Help People Detect Misclassifications of Online Toxicity. In Proceedings of the International AAAI Conference on Web and Social Media, Vol. 14. 95–106.
  4. Jonathan  P. Chang and Cristian Danescu-Niculescu-Mizil . 2019 . Trajectories of Blocked Community Members: Redemption, Recidivism and Departure. In Proceedings of WWW. 184–195. Jonathan P. Chang and Cristian Danescu-Niculescu-Mizil. 2019. Trajectories of Blocked Community Members: Redemption, Recidivism and Departure. In Proceedings of WWW. 184–195.
  5. Jacob Devlin , Ming-Wei Chang , Kenton Lee , and Kristina Toutanova . 2019 . BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding . In Proceedings of NAACL. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of NAACL.
  6. Casey Fiesler , Jialun Jiang , Joshua McCann , Kyle Frye , and Jed Brubaker . 2018 . Reddit rules! Characterizing an Ecosystem of Governance . In Proceedings of International AAAI Conference on Web and Social Media(ICWSM). Casey Fiesler, Jialun Jiang, Joshua McCann, Kyle Frye, and Jed Brubaker. 2018. Reddit rules! Characterizing an Ecosystem of Governance. In Proceedings of International AAAI Conference on Web and Social Media(ICWSM).
  7. Tarleton Gillespie . 2018. Custodians of the Internet: Platforms , Content Moderation, and the Hidden Decisions That Shape Social Media . Yale University Press , New Haven . Tarleton Gillespie. 2018. Custodians of the Internet: Platforms, Content Moderation, and the Hidden Decisions That Shape Social Media. Yale University Press, New Haven.
  8. Riccardo Guidotti , Anna Monreale , Salvatore Ruggieri , Franco Turini , Fosca Giannotti , and Dino Pedreschi . 2018. A survey of methods for explaining black box models. ACM computing surveys (CSUR) 51, 5 ( 2018 ), 1–42. Riccardo Guidotti, Anna Monreale, Salvatore Ruggieri, Franco Turini, Fosca Giannotti, and Dino Pedreschi. 2018. A survey of methods for explaining black box models. ACM computing surveys (CSUR) 51, 5 (2018), 1–42.
  9. Shagun Jhaver , Amy Bruckman , and Eric Gilbert . 2019 . Does transparency in moderation really matter? User behavior after content removal explanations on reddit . Proceedings of the ACM on Human-Computer Interaction 3, CSCW(2019) , 1–27. Shagun Jhaver, Amy Bruckman, and Eric Gilbert. 2019. Does transparency in moderation really matter? User behavior after content removal explanations on reddit. Proceedings of the ACM on Human-Computer Interaction 3, CSCW(2019), 1–27.
  10. Sara Kiesler , Robert Kraut , Paul Resnick , and Aniket Kittur . 2012. Regulating behavior in online communities. Building successful online communities: Evidence-based social design ( 2012 ), 125–178. Sara Kiesler, Robert Kraut, Paul Resnick, and Aniket Kittur. 2012. Regulating behavior in online communities. Building successful online communities: Evidence-based social design (2012), 125–178.
  11. Jon Kleinberg , Himabindu Lakkaraju , Jure Leskovec , Jens Ludwig , and Sendhil Mullainathan . 2018. Human decisions and machine predictions. The quarterly journal of economics 133, 1 ( 2018 ), 237–293. Jon Kleinberg, Himabindu Lakkaraju, Jure Leskovec, Jens Ludwig, and Sendhil Mullainathan. 2018. Human decisions and machine predictions. The quarterly journal of economics 133, 1 (2018), 237–293.
  12. Zhiyuan  “ Jerry” Lin, Jongbin Jung , Sharad Goel , Jennifer Skeem , 2020 . The limits of human predictions of recidivism. Science advances 6, 7 (2020), eaaz0652. Zhiyuan “Jerry” Lin, Jongbin Jung, Sharad Goel, Jennifer Skeem, 2020. The limits of human predictions of recidivism. Science advances 6, 7 (2020), eaaz0652.
  13. Han Liu Vivian Lai and Chenhao Tan. 2021. Understanding the Effect of Out-of-distribution Examples and Interactive Explanations on Human-AI Decision Making. arXiv preprint arXiv:2101.05303(2021). Han Liu Vivian Lai and Chenhao Tan. 2021. Understanding the Effect of Out-of-distribution Examples and Interactive Explanations on Human-AI Decision Making. arXiv preprint arXiv:2101.05303(2021). · arXiv
  14. Brian Lubars and Chenhao Tan . 2019 . Ask Not What AI Can Do, But What AI Should Do: Towards a Framework of Task Delegability . In Proceedings of NeurIPS. Brian Lubars and Chenhao Tan. 2019. Ask Not What AI Can Do, But What AI Should Do: Towards a Framework of Task Delegability. In Proceedings of NeurIPS.
  15. Scott  M Lundberg , Bala Nair , Monica  S Vavilala , Mayumi Horibe , Michael  J Eisses , Trevor Adams , David  E Liston , Daniel King-Wai Low , Shu-Fang Newman , Jerry Kim, 2018 . Explainable machine-learning predictions for the prevention of hypoxaemia during surgery. Nature biomedical engineering 2, 10 (2018), 749–760. Scott M Lundberg, Bala Nair, Monica S Vavilala, Mayumi Horibe, Michael J Eisses, Trevor Adams, David E Liston, Daniel King-Wai Low, Shu-Fang Newman, Jerry Kim, 2018. Explainable machine-learning predictions for the prevention of hypoxaemia during surgery. Nature biomedical engineering 2, 10 (2018), 749–760.
  16. J Nathan Matias . 2016 . The civic labor of online moderators . In Internet Politics and Policy conference . Oxford, United Kingdom. J Nathan Matias. 2016. The civic labor of online moderators. In Internet Politics and Policy conference. Oxford, United Kingdom.
  17. Scott Mayer McKinney , Marcin Sieniek , Varun Godbole , Jonathan Godwin , Natasha Antropova , Hutan Ashrafian , Trevor Back , Mary Chesus , Greg  C Corrado , Ara Darzi , 2020. International evaluation of an AI system for breast cancer screening. Nature 577, 7788 ( 2020 ), 89–94. Scott Mayer McKinney, Marcin Sieniek, Varun Godbole, Jonathan Godwin, Natasha Antropova, Hutan Ashrafian, Trevor Back, Mary Chesus, Greg C Corrado, Ara Darzi, 2020. International evaluation of an AI system for breast cancer screening. Nature 577, 7788 (2020), 89–94.
  18. Don Norman. 2013. The design of everyday things: Revised and expanded edition. Basic books. Don Norman. 2013. The design of everyday things: Revised and expanded edition. Basic books.
  19. Thomas O’Neill , Nathan McNeese , Amy Barron , and Beau Schelble . 2020. Human–autonomy teaming: A review and analysis of the empirical literature. Human Factors ( 2020 ), 0018720820960865. Thomas O’Neill, Nathan McNeese, Amy Barron, and Beau Schelble. 2020. Human–autonomy teaming: A review and analysis of the empirical literature. Human Factors (2020), 0018720820960865.
  20. Morgan Klaus Scheuerman Jialun Aaron Jiang Casey Fiesler and Jed R Brubaker. 2021. A Framework of Severity for Harmful Content Online. arXiv preprint arXiv:2108.04401(2021). Morgan Klaus Scheuerman Jialun Aaron Jiang Casey Fiesler and Jed R Brubaker. 2021. A Framework of Severity for Harmful Content Online. arXiv preprint arXiv:2108.04401(2021). · arXiv
  21. David Silver , Thomas Hubert , Julian Schrittwieser , Ioannis Antonoglou , Matthew Lai , Arthur Guez , Marc Lanctot , Laurent Sifre , Dharshan Kumaran , Thore Graepel , 2018. A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science 362, 6419 ( 2018 ), 1140–1144. David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, 2018. A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science 362, 6419 (2018), 1140–1144.
  22. Michael Wiegand , Josef Ruppenhofer , and Thomas Kleinbauer . 2019 . Detection of Abusive Language: the Problem of Biased Datasets . In Proceedings of NAACL. Michael Wiegand, Josef Ruppenhofer, and Thomas Kleinbauer. 2019. Detection of Abusive Language: the Problem of Biased Datasets. In Proceedings of NAACL.

Cited by

Sorted by citations

Related

Authors
Lai, V., Carton, S., Bhatnagar, R., Liao, Q. V., Zhang, Y., & Tan, C.
Affiliations
  • University of Colorado Boulder (US)
  • Amazon (United States) (US)
  • Microsoft (Canada) (CA)
  • Twitter (United States) (US)
  • University of Chicago (US)
ORCID
Venue
CHI Conference on Human Factors in Computing Systems
Publisher
ACM
Published
2022-04-27
Pages
1-18
DOI
10.1145/3491102.3501999
OpenAlex
W4224940987
Semantic Scholar
29c29d19a1d6fc62b19251547ce9df351e49763e
Cited by
145
References
63
Citations by year
2026: 32 · 2025: 40 · 2024: 40 · 2023: 31 · 2022: 2
FWCI
54.47
Type
conference paper
Language
en
Open access
Licence
Locations
  • CHI Conference on Human Factors in Computing Systems (publishedVersion, cc-by-nd)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • National science foundation (2125116, 2125113, 1837986)
  • IBM faculty award
Topics
  • Social Sciences › Psychology › Social Psychology › Human-Automation Interaction and Safety
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Social Sciences › Safety Research › Ethics and Social Impacts of AI
Keywords
human-AI collaboration, content moderation, social media comments, AI explanations, human-AI interface

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 2026-10-07.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-07.

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