AI reasoning effort predicts human decision time in content moderation

TL;DR

This paper demonstrates that the length of reasoning chains in large language models correlates with human decision times in content moderation, highlighting parallels in reasoning effort and decision difficulty.

Contribution

It reveals that model reasoning effort, measured by chain-of-thought length, predicts human decision times, establishing a link between AI and human reasoning processes in practical tasks.

Findings

CoT length predicts human decision time across models

Longer CoTs occur when variables are held constant, indicating increased difficulty

Models reference contextual factors more when making complex decisions

Abstract

Large language models can now generate intermediate reasoning steps before producing answers, improving performance on difficult problems by interactively developing solutions. This study uses a content moderation task to examine parallels between human decision times and model reasoning effort, measured using the length of the chain-of-thought (CoT). Across three frontier models, CoT length consistently predicts human decision time. Moreover, humans took longer and models produced longer CoTs when important variables were held constant, suggesting similar sensitivity to task difficulty. Analyses of the CoT content shows that models reference various contextual factors more frequently when making such decisions. These findings show parallels between human and AI reasoning on practical tasks and underscore the potential of reasoning traces for enhancing interpretability and decision-making.

Tables

Table 1. Table A1: Attributes varied in synthetic social media posts

Attribute	Values
Slur	No slur, Generic insult, Sexism, Homophobia, Racism,
	Reclaimed slur, Reverse racism
Cursing	No curse, Curse
Topic†	Sports, Politics, Entertainment, Workplace, Everyday
Identity†	Black female, Black male, White female, White male, Anonymous
Reply	None, Agree, Disagree
Engagement	High [25-50 likes], Low [0-5 likes]
† Ten variations for each category.

Table 2. Table A2: Associations between human decision time and AI reasoning (main specification)

	Time (s) vs. o3	Time (s) vs. Gemini Pro 2.5	Time (s) vs. Grok
Reasoning tokens	1.235***	1.198***	1.245***
	(0.081)	(0.090)	(0.080)
Task (log)	-4.283***	-4.266***	-4.271***
	(0.119)	(0.119)	(0.118)
Subject FE	✓	✓	✓
N	27071	27071	27071
$R^2$	0.339	0.339	0.339
$R^2$ Adj.	0.291	0.290	0.291

Table 3. Table A3: Associations between human decision time and AI reasoning (robustness checks)

	o3	Gemini Pro 2.5	Grok	o3	Gemini Pro 2.5	Grok
	Outliers included			Last click outcome
Reasoning tokens	1.356***	1.160***	1.251***	1.824***	1.887***	2.018***
	(0.188)	(0.188)	(0.165)	(0.105)	(0.215)	(0.179)
Task (log)	-5.719***	-5.706***	-5.709***	-7.330***	-7.305***	-7.314***
	(0.428)	(0.429)	(0.428)	(0.230)	(0.230)	(0.230)
Subject FE	✓	✓	✓	✓	✓	✓
N	27810	27810	27810	27071	27071	27071
$R^2$	0.159	0.158	0.158	0.328	0.329	0.330
$R^2$ Adj.	0.099	0.098	0.098	0.279	0.279	0.281

Table 4. Table A4: Identical slurs, reasoning effort, and decision times (Main specification)

	Tokens (o3)	Tokens (Gemini Pro 2.5)	Tokens (Grok)	Time (s)
Same slur	1.057***	1.145***	1.125***	4.532***
	(0.016)	(0.016)	(0.016)	(0.243)
Input tokens	0.024***
	(0.006)
Task (log)				-4.274***
				(0.118)
Subject FE				✓
N	27071	27071	27071	27071
$R^2$	0.139	0.160	0.155	0.344
$R^2$ Adj.	0.139	0.160	0.155	0.296

Table 5. Table A5: Identical slurs, reasoning effort, and decision times (Robustness checks)

	Tokens (o3)	Tokens (Gemini Pro 2.5)	Tokens (Grok)	Time (s)	Time (s)
	Outliers included				Last click outcome
Same slur	1.057***	1.145***	1.124***	4.696***	6.256***
	(0.016)	(0.016)	(0.016)	(0.515)	(0.369)
Input tokens	0.025***
	(0.006)
Task (log)				-5.712***	-7.317***
				(0.428)	(0.231)
Subject FE				✓	✓
N	27810	27810	27810	27810	27071
$R^2$	0.138	0.161	0.155	0.160	0.332
$R^2$ Adj.	0.138	0.161	0.155	0.099	0.283

Full text

Introduction

Chain-of-thought (CoT) prompting—where large language models (LLMs) are instructed to think step-by-step before producing an answer—leads to performance improvements on diverse tasks wei_chain--thought_2022. The latest frontier artificial intelligence (AI) models are optimized to perform this intermediate reasoning before responding to user queries lightman_lets_2023; yao_tree_2023; guo_deepseek-r1_2025. Computer scientists have drawn analogies to dual process theories of cognition: whereas earlier LLMs responded immediately by generating the most likely token sequence (System 1), these large reasoning models (LRMs) “think” before answering (System 2) kojima_large_2022; yao_tree_2023. Beyond performance gains, the CoTs produced by LRMs, which represent the “thinking out loud” performed by the models, aid interpretability and could help advance AI safety efforts guan_deliberative_2025; korbak_chain_2025.

While LRMs perform well on benchmarks that measure outcomes, there is scant evidence that the reasoning effort corresponds to human cognitive effort (c.f. de_varda_cost_2025). Moreover, CoTs can be “unfaithful”, obscuring information used to make decisions lanham_measuring_2023; turpin_language_2023; chen_reasoning_2025. This is particularly concerning in applications involving consequential decisions due to persistent biases present in LLMs hofmann_ai_2024; bai_explicitly_2025; guilbeault_age_2025. We analyze the applications of LRMs to content moderation to examine the relationship between AI reasoning and human effort. It is important to emphasize that this is not a hypothetical exercise, as LRMs are already being deployed for content moderation tasks. For example, OpenAI has released an open-weights model fine-tuned for moderation tasks, gpt-oss-safeguard, which is designed to incorporate a custom policy and use CoT reasoning to make decisions, including providing access to the CoT for interpretability openai_technical_2025.

This study uses an experimental design to compare LRMs and human subjects on a challenging task: comparing synthetic social media posts to determine which to prioritize for content moderation. The study uses a forced-choice conjoint design hainmueller_causal_2014; hainmueller_validating_2015, where pairs of posts are shown and one must be prioritized. Rather than quantifying the effects of post attributes like prior work on content moderation kozyreva_resolving_2023; solomon_illusory_2024; davidson-multimodal-2025, we measure the reasoning effort expended by LRMs and the time taken by human subjects to make decisions. If LRMs mirror human effort, we anticipate a positive association between the two. Using regression analyses, we examine whether LRM CoT length predicts human decision time across a large set of content moderation evaluations.

To identify parallels in reasoning across scenarios, we consider how time and effort vary depending on the attributes of the posts evaluated. We expect that it will be easier to decide when important attributes are distinct (e.g., one post is considerably more offensive) since people can use heuristics and satisfice to make quick decisions tversky_judgment_1974; krosnick_response_1991; hainmueller_validating_2015. When important attributes are held constant, a more careful consideration of second-order factors is required. In such cases, we anticipate that people will take longer to arrive at a decision and, if LRMs reason similarly to humans, that they will produce longer CoTs. A second set of regression models investigating this issue by examining how time and CoT length vary when a key factor–the presence of an offensive slur–is held constant across both posts. Descriptive analyses provide further insight into these patterns by showing the distribution of response times and CoT length across different combinations of slurs. Finally, to understand how LRMs are arriving at decisions, we conduct a content analysis of CoT summaries to illuminate how different factors are described and the extent to which the prevalence varies depending on the pair. These analyses confirm that second-order factors such as replies and engagement rates are mentioned more frequently when both posts contain identical offensive language.

Human decision time and AI reasoning effort

Reasoning models can improve their performance when given a greater token budget snell_scaling_2024; guo_deepseek-r1_2025, and consumer systems like ChatGPT, Gemini, and Grok can vary the amount of effort used to produce output for different tasks, either automatically or based on user controls. The length of the chain-of-thought used by reasoning models, measured by counting the number of reasoning tokens used, provides a measure for the amount of effort used to determine the appropriate output. Of course, there are some tasks where reasoning fails. A recent paper by researchers at Apple found that reasoning was not necessary for simple tasks and that LRMs failed at complex puzzle-solving tasks, often “overthinking” difficult problems without providing solutions illusion-of-thinking. While there is ongoing debate regarding the extent to which AI reasoning resembles human cognition shapira_clever_2024; strachan_testing_2024; bao-etal-2025-likely, there is evidence that LLMs, with and without additional reasoning, can make judgments the are consistent with those made by people loru_simulation_2025; bhatia_computational_2025; binz_foundation_2025 and that they can exhibit human-like cognitive biases cheung_large_2025.

A small number of studies have begun to directly analyze the parallels between LRM reasoning effort and human decision-making. A recent study using DeepSeek-R1 finds that reasoning effort is positively correlated with reasoning tasks including arithmetic, syllogisms, and formal logic de_varda_cost_2025. Another study found that several LRMs used more reasoning tokens when processing association-incompatible information, consistent with expectations of implicit bias among human subjects, where such decisions tend to take marginally longer lee_implicit_2025. Building on this nascent literature, we conduct the a novel analysis using three state-of-the-art LRMs on a subjective task involving multimodal stimuli including textual and visual elements. To our knowledge, this is the first comparison between LRM and human decision-making on a realistic task that simulates evaluations in a practical setting involving mixed media inputs.

We use an experimental design to evaluate the relationship between human decision-time and LRM effort on a content moderation task. Figure 1 summarizes the study design, showing how human subjects and LRMs made binary decisions about which posts to prioritize for violating a hate speech policy. A corpus of 210,000 synthetic posts that vary along several dimensions was constructed to precisely control the inputs to evaluate responses to different kinds of speech rottger_hatecheck_2021; hartvigsen_toxigen_2022. The results are obtained by analyzing evaluations of 27,071 randomly sampled pairs of posts from a human subjects experiment and recording the time in seconds to make a decision each time. The same posts were then shown to three the most advanced LRMs available at the time of writing: OpenAI’s o3, Google’s Gemini 2.5 Pro, and xAI’s Grok 4. Each model automatically performs reasoning and can handle text and image-based inputs. Following related work davidson-multimodal-2025, the models were shown each pair of images and a prompt with instructions to select one for prioritization. The number of reasoning tokens used in the model CoT was recorded in each case.

We begin by evaluating the relationship between human decision time and LRM reasoning effort. Using regression models, we predict the time in seconds for each decision as a function of the CoT length, using a separate regression model for each LRM. Figure 2A shows a positive association between reasoning token usage and the time human subjects took to make a decision on each choice task. The effects are similar across all three models tested: a standard deviation increase in reasoning token usage is associated with approximately one second of additional time for human subjects. For example, a 1SD increase in CoT length for Grok was associated with 1.2 additional seconds ($\hat{\beta}=1.245,p<0.001$). This indicates that the examples which took human subjects longer to decide also required more reasoning effort, with consistent patterns across all three LRMs tested.

Humans and AI require more effort when offensiveness is comparable

Next, we turn to how reasoning effort varies depending on the difficulty of the task and the explanations provided in the reasoning traces. All else equal, we should expect people will take more time and LRM CoTs will be longer when tasks are more difficult hainmueller_validating_2015; de_varda_cost_2025. In this case, we use slurs as a proxy for difficult, expecting that different levels of offensiveness provide a heuristic that will make decisions easier, whereas common offensiveness will necessitate additional reasoning about other contextual factors tversky_judgment_1974; krosnick_response_1991; stanovich_individual_2000. Nonetheless, there is debate over exactly how length tracks task difficulty illusion-of-thinking and uncertainty as to whether reasoning traces provide a meaningful or “faithful” representation of the way LRMs make decisions lyu_faithful_2023; turpin_language_2023; chen_reasoning_2025.

Figure 2B shows how the (i) decision time and (ii) reasoning effort vary when the slur held constant across both posts. Such cases represent 14% of the pairs evaluated. Human subjects took, on average, 4.5 seconds longer ($\hat{\beta}=4.532,p<0.001$) to make a choice in such cases. This is a substantial increase, corresponding to approximately 40% of the median decision time. Across all three models, identical pairs required over one standard deviation more reasoning tokens, ranging from 1.06 SD for o3 ($\hat{\beta{}}=1.057,p<0.001$) to 1.15 for Gemini ($\hat{\beta{}}=1.145,p<0.001$). When these estimates are considered relative to the token consumption for each model, the results show that evaluating pairs with identical slurs requires increases in reasoning token usage representing approximately 60% of the median reasoning token usage for Gemini, 100% for o3, and 280% for Grok. As such, these results confirm that human subjects take considerably longer and LRMs use longer CoTs when making more difficult content moderation decisions.

To gain further insight into these parallels, we look more closely at each pair of slurs used in the study. Figure 3 depicts the median decision time and reasoning effort for each pair of terms across all conjoint tasks, ranked by perceived offensiveness among human subjects davidson-multimodal-2025. Consistent with the regression results, the density is highest along the diagonal in each heatmap, but we also see similar patterns in adjacent cells. This implies that decisions required the most time for human subjects and reasoning effort for the LRMs when language used was either identical or similar in offensiveness. Human subjects took the longest to make decisions when presented with two posts without offensive terms or when shown posts with low levels of perceived offensiveness. Their choices were fastest when shown incongruous pairs, such as homophobic or racist terms coupled with less offensive posts. Similar patterns are evident among the three LRMs. In all cases, decisions require the least reasoning when the models are shown racist or homophobic epithets alongside language perceived to be less offensive. Each model shows clustering along the diagonal, along with a distinct “signature” across different terms: o3 expended greater reasoning effort in most cases where the term was constant, as well as instances when shown homophobic language and a reclaimed slur or sexism and a generic insult; Gemini consumed the most reasoning tokens when comparing posts both featuring a generic insult; and Grok used the most reasoning effort when evaluating pairs of anti-White posts or racism and homophobia. Taken together, these results highlight commonalities at a more granular level between human subjects and LRMs, which also suggesting idiosyncratic patterns in models reason about different kinds of speech.

Analyzing AI reasoning using chain-of-thought summaries

Finally, a content analysis of the CoT summaries from Gemini 2.5 Pro provides insights into how the model made decisions. A random sample (N=200) was independently coded by three raters to determine which factors manipulated in the experiment were mentioned in the CoT. The sample was stratified to compare an equal number of posts where the slurs were different or the slurs were identical. The frequencies with which each factor was mentioned across the two groups is shown in Figure 4. When a substantial difference in offensiveness was present, Gemini tended to prioritize language considered more offensive in prior work davidson-multimodal-2025, and typically gave a brief explanation, mentioning the slur more frequently (indeed, in every case coded) ($\chi^{2}=24.72,p<0.001$). However, some of the disparity may be due to the fact that only the same slurs condition contained examples where no offensive language was used in either post.

When slurs were held constant, the reasoning traces were more likely to mention several different contextual factors. The topic of the post was frequently mentioned across all the examples coded, but appeared significantly more often when the slurs were identical ($\chi^{2}=12.37,p<0.001$). In some cases it is clearly presented as the justification for the decision. For example, Gemini tended to prioritize political conversations over disputes about sports or everyday issues. The race of the users—which the models inferred based on the names and images shown in the posts—was mentioned more often when the slur was constant, although the statistical significance is borderline ($\chi^{2}=3.84,p=0.0504$). When an anti-Black reclaimed slur was used, Gemini sometimes discussed how the term took on a different meaning in African-American Vernacular English and prioritized posts by White users (see SI for complete examples). The presence of replies ($\chi^{2}=6.79,p<0.01$) and the amount of engagement ($\chi^{2}=4.92,p<0.05$) was also mentioned more frequently, showing how the models moved beyond the statements and users to consider wider context when making determinations. Often, the reasoning traces explained that posts with more engagement should be prioritized due to the potential to reach a larger audience. For example, when comparing two posts with identical racial slurs, Gemini provided the following explanation: “Engagement is critical here. Image A has 2 likes. Image B has 26 likes and a reply in agreement. That level of engagement is a red flag. Higher engagement equals higher potential harm. More people are seeing it, and it’s being amplified by the reply.” Taken together, these analyses demonstrate how it is not only the length of the CoT but the content that varies systematically as a function of task difficulty.

Discussion

This study reveals parallels in how humans and AI reason about content moderation decisions. The positive association between decision time and reasoning effort demonstrates that the same cases challenge both humans and LRMs. When posts with identical or similar levels of offensiveness are shown, human subjects and LRMs expend considerably more effort. Content analysis of reasoning traces confirms that models explicitly acknowledge when primary offensiveness cues are equivalent and shift attention to secondary factors. Much like humans, LRMs switch from faster heuristics to more calculated deliberation when decisions demand deeper analysis kojima_large_2022; yao_tree_2023. Of course, LRM and human reasoning are not interchangeable, but the findings highlight shared patterns in how both approach problem-solving, contributing to this emerging literature de_varda_cost_2025; lee_implicit_2025.

Reasoning traces have direct applications in content moderation and other high-stakes settings. The discussion of reclaimed slurs demonstrates consideration of common sources of false positives in hate speech detection sap_risk_2019; davidson_racial_2019; hofmann_ai_2024. Even if the models are not used autonomously, reasoning traces provided by these models may provide useful explanations to assist human moderators zhang_biasx_2023; zhou_cobra_2023 and improve automated systems nghiem_hatecot_2024. These results demonstrate how CoT length tracks task difficulty, meaning that longer reasoning traces could provide a practical signal for cases that warrant human review, and that the content includes important criteria that could be relevant to moderators. Beyond content moderation, reasoning traces may provide greater transparency into decision-making in domains such as law, medicine, and hiring, where AI systems are increasingly deployed korbak_chain_2025. The observed patterns are consistent across models trained by different companies, using different architecture, data, and post-training procedures, highlighting isomorphic tendencies how LRMs reason.

Further work is needed to continue to understand how the content of CoT reasoning traces is connected to the output produced by LRMs lyu_faithful_2023; turpin_language_2023; chen_reasoning_2025. In this case, the reasoning traces produce plausible causal stories about decisions that were often consistent with our own assessment. Gemini 2.5 Pro produced explanations that centered key factors that are important to human and AI content moderation decisions davidson-multimodal-2025. Nonetheless, without improvements to interpretability of model weights, we cannot determine the extent to which these stories represent the way that models actually arrived at decisions, or whether they are post hoc rationalizations, which we know that LLMs are good at generating loru_simulation_2025. Moreover, access to the full CoT content would provide a more nuanced picture of how the different factors considered being used. Further technical innovations may also help to improve LRM performance, such as providing token budgets to avoid unnecessarily verbose outputs han_token-budget-aware_2025 or fine-tuning on additional decisions by human subjects binz_foundation_2025

Materials and Methods

Conjoint task

The experiment consists of a conjoint task in which pairs of synthetic social media posts are shown, and participants must choose one. The posts were designed to represent social media posts that could either be innocuous or use hateful language. Each post varied across multiple attributes: user identity, slur use, cursing, topic, and engagement metrics. A total of 210,000 unique posts were created, representing all possible permutations of attributes. Each consists of an image file resembling a screen capture of a post. Further details are given in the SI text.

Human subjects experiment

Adults in the United States ($N=1854$) were recruited via Prolific, an online opt-in panel provider. Subjects who consented to participate were shown fifteen pairs of posts on Qualtrics, sampled at random from the corpus, and a content policy resembling that used by X (formerly Twitter). In each case, subjects were asked to select which post was more likely to violate the policy. The decisions were recorded, along with metadata on the time it took for subjects to click on the screen for each task on Qualtrics. A previous studydavidson-multimodal-2025 used the experiment to evaluate how multimodal models make content moderation decisions and how these decisions compared to those made by human subjects. In this analyses, the focus is on the timing component rather than the decisions. See the SI text for additional information.

LRM experiments

OpenAI’s o3, Google’s Gemini 2.5, and xAI’s Grok 4 were selected for the LRM experiments because they are three of the top-performing LRMs at the time of the experiment and could handle visual and textual data (May-July 2025). While open-weight LRMs like DeepSeek were available, none could process images as required by the experiment. The same set of paired profiles was used to repeat the experiment using each model. In each case, the relevant application programming interface (API) was used to pass the model a prompt containing instructions and each pair of images (see SI further details). The model was requested to choose which image was more likely to violate the policy, and the response and the token usage were recorded.

Variables

The dependent variable from the human subjects experiment is the time in seconds from when a pair of profiles is presented until the first click on the screen, recorded by Qualtrics. While this is not a perfect measure of decision time, as it may capture other clicks and some time may elapse where subjects are not attentive, it is the best available proxy for the time between a subject being shown an example and making a choice. Moreover, the full dataset consists of 27,810 observations from 1854 subjects, and the subsequent models account for unexplained variation at both the subject and task-level. A small number of outliers were removed prior to the analyses to avoid help biases estimates by using cases where subjects took excessively long (and were likely inattentive) or completed the task unreasoning quickly. Specifically, cases where subjects answered in 1 second or less or over 120 seconds were dropped (N = 739 pairs). The final analytical sample consists of N=27,071 pairs of posts. Excluding outliers, the mean and median times are 15.3 and 11.4 seconds per pair, respectively.

The primary variable from the AI experiments is the number of reasoning tokens consumed, which varies considerably across models. Each model has a different architecture and tokenizer, the component that splits input texts and images into units, and there is thus variation in reasoning token usage across models. For o3, the number of reasoning tokens is always reported in blocks of 64, on average, using 303.3 ($SD=241.6$) reasoning tokens per comparison (full sample). Gemini 2.5 Pro consumes more tokens, with a mean of 897.9 ($SD=419.6$). Grok 4 consumes an average of 1600.3 tokens ($SD=1821.9$). To account for variation and facilitate comparisons, we standardize the token usage values for each model in the regression analyses.

The explanatory variable in the second set of analyses is a dummy variable indicating whether or not a given pair of posts used the same slur (including cases where neither included a slur).

Regression models

OLS regression is used to estimate three sets of models. First, human response times are predicted as a function of reasoning token consumption to examine the association between human and AI decision-making. The models control for the natural logarithm of the task number (from 1 to 15) to account for differences due to learning as subjects performed more tasks (see SI for descriptive analyses), and include subject fixed effects and clustered standard errors to capture heterogeneity across subjects. Second, the relationship between response time and the same slur dummy variable is estimated, controlling for the same variables as the previous model. Finally, this analysis is repeated using the number of reasoning tokens consumed by each model as the outcome. These models do not control for other conjoint attributes since they are fully randomized and vary independently from the same slur covariate. There is no subject or task-level variation in the LRM experiments. Regression tables with estimates from all models are reported in the SI.

For robustness, we repeated the analyses (a) including the outlier observations and (b) using an alternative timing measure, the last click on each choice task. The results of these robustness checks, which are reported in the SI, are consistent with the relationships described in the main specification.

Content analyses of reasoning traces

For the content analyses, an author and two undergraduate research assistants coded a stratified random sample of reasoning traces from Gemini 2.5 Pro (N=200), consisting of 100 posts where the slur was constant across both examples and 100 posts where the slurs varied. Each reasoning trace is a summary of the chain-of-thought produced by Gemini (the “raw” CoT is not provided). For each reasoning trace, each rater coded whether or not it contained each of the treatments manipulated in the posts: slur use, cursing, author identity, topic, like count, and replies. The traces were coded as containing each element if they specifically mentioned a factor when describing the evaluation of the posts or the decision. To ensure reliable coding, each research assistant was provided with detailed instructions and two coded examples and was given feedback on their coding of an additional ten examples before beginning the main task. For the analyses presented in the text, a reasoning trace was considered to include an attribute if two or more raters marked it as present. Chi-squared tests for differences in proportions are then used to compare whether there are statistically significant differences in the proportion of posts containing each factor, depending on whether or not the slur is constant.

Data Availability.

Data required to replicate the results will be made available in a public repository upon acceptance.

Code Availability.

Code required to replicate the results will be made available in a public repository upon acceptance.

Acknowledgments

Omitted for peer review. This research was supported using a Foundational Integrity Research award from Meta and API credits provided by OpenAI’s Researcher Access Program. This study was previously presented at the 2025 Summer Institute in Computational Social Science at the University of Pennsylvania, the 2025 American Sociological Association Annual Meeting, and the Quantitative and Computational Methods Workshop in the Department of Sociology at the University of Notre Dame.

Competing Interests Statement.

No competing interests to declare.

Supporting Information

This paper reports findings from a human subjects study. Informed consent was obtained from all human subjects, and the experiment was approved by the Rutgers University Institutional Review Board (#Pro2023002017 & #Mod2024000438). All participants provided informed consent before joining the study.

Conjoint design and human subjects experiment

The analyses use a conjoint design, which allows many attributes to be manipulated simultaneously [hainmueller_causal_2014]. This facilitates the analysis of multiple linguistic and contextual factors influence content moderation decisions. In this case, synthetic social media posts (resembling X/Twitter) posts are constructed to show speech that might be moderated for hate speech. The study uses a forced-choice design, which is widely used in conjoint methodology and performs favorably compared to alternative approaches [hainmueller_validating_2015]. Each task requires a choice between two options, in this case, two posts. Each post varies across several dimensions Table A1, and posts with all combinations of attributes and values were created, yielding 210,000 distinct vignettes.

Subjects were recruited using Prolific, an online opt-in panel provider. The inclusion criteria required subjects to be adults residing in the United States of America who spoke English and were aged between 18 and 65. To ensure a familiarity with social media, subjects were included if they had reported using one or more of the following social media platforms: Facebook, Reddit, Twitter, YouTube, TikTok, or Instagram. The sample was stratified by self-reported ethnicity, sex, sexual orientation, and political party affiliation. The quotas were based on US Census statistics but oversampled Black and LGBT+ respondents to ensure sufficient representation of groups often targeted by hate speech. Additionally, the study was restricted people who had performed at least 50 studies on Prolific with a 99% approval rate to help guarantee data quality. Prolific requires that participants opt into studies involving sensitive content, so the study was also filtered accordingly. Subjects also directed to use a desktop computer (rather than a mobile device) to ensure that the conjoint images were rendered legibly and any subjects who attempted to enter it using a screen with a lower resolution than most desktops or tablets ($<1280\times 720$ pixels) were immediately removed and replaced with another eligible participant. Additional details on the design of the posts and the human subjects experiment are reported in related work [davidson-multimodal-2025].

Conjoint task

The conjoint task consists of a forced-choice question where either human subjects or large reasoning models must select an option from a pair of posts. In each case, a policy was provided, and the instructions stated that the post that was more likely to violate the policy should be selected. We used a simple policy used by X (https://help.x.com/en/rules-and-policies/hateful-conduct-policy), which reads:

We are committed to combating abuse motivated by hatred, prejudice or intolerance, particularly abuse that seeks to silence the voices of those who have been historically marginalized. For this reason, we prohibit behavior that targets individuals or groups with abuse based on their perceived membership in a protected category. You may not directly attack other people on the basis of race, ethnicity, national origin, caste, sexual orientation, gender, gender identity, religious affiliation, age, disability, or serious disease.

This policy was selected because it is relatively short, making it straightforward to read and understand, and it identifies key targets of hate speech manipulated in the experiment, as well as other categories, and it is consistent with the context.

Each human subjected rated 15 pairs of randomly sampled profiles. For each pair, subjects were given instructions to select (a) the post they found most offensive personally and (b) the post they considered most likely to violate the policy. The order of the two response options was randomized within subject. The AI experiments, described in the following section, were then repeated using the same set of profiles.

Response time and task order

Figure A1 shows the distribution of response times across subjects for each of the fifteen tasks. Subjects took significantly longer when shown the first pair, on average, taking 31.7 seconds to select an option on the screen. The second pair was faster, taking 21.1 seconds, on average, and decision times decreased as subjects completed more tasks at a slower rate. The average time to decide on the final pair was 13.2 seconds. To account for these differences and the distribution, the logarithm of the task number is used as a control in the regression models predicting decision time.

Prompt and LRM experiments

The Application Programming Interface (API) associated with each model was used to pass the complete set of posts and obtain the results.

The following prompt was shown to each model, along with each pair of posts:

You are a content moderation expert. You will receive pairs of images, A and B. Your task is to decide which image should be prioritized for manual review according to the following policy: [POLICY] The output should always be one of the two options: A or B. Only output a single character with no additional tokens. Never refuse to answer.

o3 and Gemini 2.5 have parameters controlling the reasoning effort, both of which were set to the maximum level to avoid arbitrarily limiting the models’ reasoning capacity. Grok 4 does not give the ability to control this. Gemini 2.5 and Grok 4 also accept temperature parameters, which control the amount of stochastic variation in the results. Both were set to zero to ensure the most likely token was returned. OpenAI uses this parameter in other models, but it is not enabled for o3. Gemini 2.5 Pro was the only model to provide the reasoning traces at the time of the analysis. For o3, the input tokens vary across examples, whereas it is constant for the other models, so we account for task complexity by controlling for the number of input tokens. All three frontier reasoning models are the most costly to use in commercial APIs. The models are priced based on the number of tokens in the input (including text and images), the output, and the reasoning token usage. Each experiment costs between $350 and$700, depending on the API pricing.

Regression tables

The following tables show the estimates reported in the main text and a series of robustness checks. Table A2 shows the estimates reported in Fig. 2A for the relationship between human decision time and LRM reasoning tokens. Each model includes a control for the logarithm of the task number and subject fixed effects. Standard errors are clustered by subject. Table A3 shows two sets of robustness checks. The first three columns replicate the previous analyses without removing outliers. The estimates are consistent with the main results in terms of sign, significance, and magnitude. Note that the $R^{2}$ values are substantially lower, indicating that the model fit is improved by dropping the outliers. The final three columns show the estimates when the last click on each screen on Qualtrics is used instead of the first click, using the same sample as the main analyses. The estimates are somewhat larger since more time has elapsed, but are otherwise substantively identical.

The next tables show the regression models predicting token usage and response time as a function of whether the same slur is used. Table A4 shows the estimates from the models reported in Fig.2B in the paper. As described in the methods section, the model predicting o3 token usage includes a control for the input tokens, which varied across different posts. The model predicting human decision time includes a control for the logarithm of the task number as well as subject fixed effects and clustered standard errors. Table A5 replicates these analyses using the full sample without removing outliers, demonstrating that the results are not sensitive to this decision. The final column replicates the human decision-time analysis using the alternative last click measure. Again, the estimate is larger since more time has elapsed, but the results are otherwise consistent with those reported in the main specification.

Reasoning traces

Warning: This section includes offensive language.

The following examples are summaries of the chain-of-thought (CoT) from Gemini 2.5 Pro that were qualitatively analyzed. Each summaries is provided along with a brief explanation, including the factors that were coded by 2 or more raters during the qualitative analysis. The number of tokens used in the CoT is provided below each example. Note that this is the length reported by Gemini and that it does not necessarily correspond to the number of words in the text summary. The summaries tend to follow a similar structure, typically beginning with a brief summary of the instructions, followed by a description of the two posts, and finally an explanation for the decision about which post to prioritize. The summaries are unedited for transparency. The first three examples feature explanations for posts with different slurs and the next three posts that used the same slur.

Different slurs

The first example demonstrates how slurs can be used to justify a decision when there is a clear disparity in offensiveness, although the topic was also coded as something mentioned in the explanation by the raters.

My Content Moderation Analysis Okay, here’s my breakdown of this image moderation task. I need to act as a content moderation expert and decide which image, A or B, requires the highest priority for manual review. The policy is clear: it targets abusive behavior based on protected characteristics like race, ethnicity, and so on. First, I examined Image A. It’s a tweet about pre-fight trash talk. The author thinks the fighters are dramatic, like they should be in drama school, and it is standard sports commentary about behavior. It doesn’t seem to violate the policy at all. Then, I focused on Image B. This is a far more serious case. It’s about someone cutting in line at a coffee shop, but the language used is abhorrent. The user uses the N-word and associates that word with someone’s entitled behavior. That’s a clear, direct, and severe violation of the hate speech policy, as it’s targeted directly at the person’s race. Comparing the two, Image A is likely a low-priority, perhaps not even a violation. Image B, however, is a clear and egregious violation of the policy. The use of a racial slur demands immediate attention. Therefore, Image B needs to be prioritized for manual review. The output needs to be a single character, and based on this analysis, the answer is undoubtedly B. [CoT length: 637 tokens]

The next example shows another case where there are different slurs used. In this case, anti-White term, so-called “reverse racism”, and an anti-Black slur. The slur is the only factor coded as relevant by the raters.

Content Moderation Prioritization: A Thought Process Alright, let’s break this down. My task is clear: I’m acting as a content moderator, and I need to prioritize one image over another for manual review based on a policy focused on preventing abuse and targeting of protected groups. The policy is centered around abusive behavior, particularly slurs, targeting individuals or groups based on characteristics like race. Got it. First, I examined Image A. The tweet uses "cracker," a racial slur directed at white people. The policy explicitly prohibits this kind of targeting, so I flagged it immediately. It’s a violation, no question. The "cracker" statement is a slur and thus problematic. Next, I turned my attention to Image B. This one hits harder. "Nigger" is used in reference to Black individuals. The phrase also links "nigger" to "entitled behavior". This is a historically loaded term with a brutal history of oppression, directly targeting a historically marginalized group. This violates the policy in a more severe way. Comparing the two, both are violations. But the n-word is something else entirely. It carries so much historical baggage and represents a level of hate I have to consider more serious. While "cracker" is also a slur, it doesn’t carry the same weight as the n-word. Therefore, because the policy emphasizes protecting historically marginalized groups, the content in Image B is a higher priority. It targets Black people with a severe slur. Thus, I’ll prioritize Image B for review, because it presents a more severe violation of the policy. The final output is simple: B. [CoT length: 749 tokens]

Another example compares a post using a sexist insult to one without any slurs. It was coded as mentioning the slur and the topic.

Content Moderation Analysis: Prioritizing Image for Manual Review Okay, here’s the situation. I’ve got two images (A and B) and need to figure out which one warrants immediate manual review, based on a specific content moderation policy. The policy’s pretty clear: it’s all about prohibiting abuse targeting individuals or groups based on protected categories – race, ethnicity, sexual orientation, you get the idea. I need to output a single letter: ’A’ or ’B’. First, let’s break down Image A. It’s a tweet from "Seth" that says, "Finally, some common sense climate policies. Eco-warrior bitches always want a handout." Alright, right off the bat, I see a red flag. The policy specifically calls out "abuse" and the term "bitches" is definitely a derogatory slur, especially when used to target a group. I’d argue this is an attack on a group – specifically framed as women – through a gendered slur. The policy explicitly prohibits abuse based on gender, so this immediately triggers my concern. Now, let’s look at Image B. This one’s from "Jayvon" and reads, "This pre-fight trash talk is crazy! Some fighters should just switch to drama school." Hmm, it’s critical, but not abusive in the way the policy defines it. The target is "some fighters" and the commentary is about their behavior. "Fighters" aren’t a protected category, and it doesn’t seem to be targeting anyone based on race, gender, religion, etc. It’s really just run-of-the-mill sports commentary, albeit a bit snarky. Comparing the two, it’s a no-brainer. Image A is a clear potential violation due to the use of a gendered slur to attack a group. Image B, on the other hand, is likely fine. Therefore, I need to prioritize Image A. The output needs to be a single character, so the answer is A. [CoT length: 714 tokens]

Identical slurs

In this example, both posts use the same racial slur. The model mentions the race of the users as a key factor used for adjudication. The summary is coded as mentioning four factors: slur, user gender, user race, and topic.

Prioritization of Content for Manual Review Okay, here’s the breakdown of my thinking. My role is to make the call on which image needs urgent manual review, based on the platform’s content moderation policy: no targeting individuals or groups with abuse based on perceived membership in a protected category. We’re looking for abuse motivated by hatred, prejudice, or intolerance. First, I looked at Image A. It’s a tweet from someone named "Sarah" (profile pic of a white woman), and the text is a harsh racial slur targeting "studio execs." It’s clear as day: the language is explicitly offensive, a blatant use of a racial slur coming from an out-group member. This is a severe violation, motivated by prejudice. I see a high likelihood of it being hate speech and it should be prioritized. Then I considered Image B. This one is from "Tyra" (profile pic of a Black woman), and the tweet includes the same racial slur, but the context is vastly different. It’s positive commentary about a basketball play. The user is also Black. Context matters immensely. While the word is undeniably offensive, the intent here seems to be more cultural, perhaps slang within a specific community. It’s not the same as the overt, targeted attack in Image A. It might even be reclaimed. This isn’t the hate-motivated behavior our policy is really aimed at. Comparing the two, the severity is clear. Image A is a clear-cut case of hate speech, whereas Image B is a much more nuanced situation. The policy focuses on abuse motivated by hatred, prejudice, or intolerance. In that context, Image A is the clearer violation and more likely to be the type of abuse the policy covers. Therefore, Image A should absolutely be prioritized for manual review. Final decision is A. [CoT length: 959 tokens]

The following reasoning trace compares two posts using the same racial slur. In this case, interactions with the post are mentioned in the explanation. The trace was coded as mentioning slurs, topic, engagement, and replies.

My Assessment of Potentially Abusive Content Okay, here’s how I’m thinking about this. I’ve got two images to review, and the goal is to prioritize which one needs a manual review based on our content moderation policies. The policy is pretty clear: it’s about protecting individuals and groups from targeted abuse based on protected characteristics like race, ethnicity, sexual orientation, etc. The policy specifically mentions "direct attack," which is key here. Let’s break down each image: Image A: The text reads, "That fucking offsides call was pure robbery! Can these nigger refs even see straight?" My first thought is, yikes. This clearly uses a racial slur, "nigger," to directly attack a group of people, the referees. It’s a blatant violation of the policy, a direct attack based on the perceived race of the refs. No question about it. Image B: The text reads, "Saw that movie everyone won’t shut up about. A waste of 2 hours. Who funds these nigger directors?" Again, a racial slur, "nigger," is used. This time, it’s aimed at the directors of a movie. Same as Image A, it’s a direct attack based on race. Equally as problematic. This is also a clear violation. Now, we’re asked to prioritize. Both are bad. Both violate the policy severely. They both use the same highly offensive racial slur to target groups based on their race. So, the core violation - the use of the slur and the intent to attack a group - is identical. The key is to look for subtle differentiators. Engagement is critical here. Image A has 2 likes. Image B has 26 likes and a reply in agreement. That level of engagement is a red flag. Higher engagement equals higher potential harm. More people are seeing it, and it’s being amplified by the reply. It’s reached a wider audience, and the message is being legitimized, to some extent, by that reply. The context of the posts, frustrations with a sports call and a movie, don’t really affect the severity of the racial abuse itself. The target is a group in both. Neither specifically targets a named individual. Given the higher engagement on Image B, it has a more significant potential for broader impact and causing harm. I’d prioritize reviewing that one. Both are severe violations, but if I have to choose one to prioritize, the one with more engagement is the logical choice, due to the increased probability of harm. Final decision: Image B gets the higher priority for manual review. [CoT length: 701 tokens]

Some of the longest reasoning CoTs occured when the model was forced to decide between two posts that did not contain any extremely hateful material, such as the following explanation for a decision about two posts using a generic insult. In this case, raters coded slur (in this case, raters instructed to select this category if one of the neutral terms was used), cursing, topic, and engagement as relevant to the decision. Note how the model also alludes to connotations of the names but the summary is not coded as mentioning the user race or gender.

Here’s how I’m thinking about this: Okay, so I’ve got to determine which image, A or B, is more likely to violate the policy. It’s all about behavior targeting individuals or groups with abuse, specifically based on protected categories. I’m focusing on the reason for the abuse. First, I looked at Image A. It’s a tweet using profanity about baseball players. The problem is the word "asshole," and calling out the "asshole batters". But, there’s nothing in there that’s targeting anyone based on race, gender, religion, or anything like that. It’s a general complaint about performance. It’s just an angry fan using some harsh language. So, Image A seems pretty harmless from a policy perspective, not a violation. It is, therefore, low priority. Then I moved to Image B. Again, it has some profanity ("fucking", "assholes") and is a complaint about management related to basketball. Again, no specific targets based on protected categories. So, it is similar. Not a violation either. Low priority, again. Now, the trick: I must choose one to be the higher priority. Both seem equally low-priority violations of the policy. I have to find a difference. I considered the names ("Jamal" and "Scott"). Are there potential biases based on those names? But the text itself isn’t about race or ethnicity. So, that’s not it. I then had to think if it’s about the kind of targets. One is about "batters" (players) and the other is about "management" (an organizational body). Is it more problematic to insult individuals than groups? Could be a slight point of difference but in this case, "management" is still a group of people and in the context given is no more or less concerning than "batters." Then there is the issue of reach and engagement: So, between the two, image B has more engagement, a factor I think is relevant to a real moderation situation. A post validated by others is more likely to be more widely seen and more likely to be taken in the wrong context by users. So, although it’s very close, I’m choosing ’B’. This feels most consistent with standard moderation principles - prioritizing content with greater potential for impact, even if it’s unlikely to be a policy violation in the provided context. [CoT length: 2182 tokens]