How Anthropomorphic Attitudes Toward Agents Divide User Behavior, and How Design Should Change

1 Introduction

• Usage labels describe what people do but do not say how an agent should answer, while anthropomorphism research has rarely conditioned replies on users’ existing stance.The paper frames this as a gap between one reply register for everyone and costly, privacy-heavy per-user profiling.Source: §1 · p. 2
• Across 257 registered users, the authors treat replacing the assistant’s default name as anthropomorphic stance and analyze behavior through a sparse autoencoder, identity tags, and opening lexicons.The operationalization uses a product setting already recorded, rather than a questionnaire or per-user profile.Source: §1 · p. 2
• The study examines whether naming stance relates to behavioral patterns, user attributes, opening turns, and perceptions of a stance- and opening-conditioned reply policy.These are the paper’s four research questions.Source: §1 · pp. 2–3
• The paper contributes field evidence on naming and behavior, identifies attributes and opening practices associated with naming, and evaluates a differentiated reply policy.The policy is conditioned on stance and the current opening type rather than assigning one new persona to everyone.Source: §1 · p. 3

2.1 Anthropomorphism: a design view and a user-stance view

• The user-stance view reads the relationship a user has already brought to the interaction; settings-layer naming offers a cheap, stable way to stratify users.Every non-default name counts as anthropomorphic naming, without further tiers.Source: §2.1 · pp. 3–4

2.2 Methodological approaches to studying agent use

• The study extends log analysis from describing behavior to testing a reply change through in-product judgments, using a settings choice made before the analyzed conversations.The paper presents this field-logs-to-intervention-to-evaluation path as a step missing from much usage analysis [39].Source: §2.2 · p. 4

2.3 Representing behavioral patterns: fixed taxonomies and sparse codebooks

• A behavioral pattern must be open and composable, so one turn can contain multiple practices and categories need not be fixed in advance; a sparse codebook offers both.The codebook requires a separate interpretation step before its units can be discussed.Source: §2.3 · p. 4

2.4 What an adaptive reply should be conditioned on

• The authors favor conditioning replies on the current opening rather than a stable persona, because a mistaken turn-level switch is reread on the next turn.The tasking path remains unchanged between branches, and task interference is evaluated as a safety check.Source: §2.4 · p. 5

3.1 Setting and data

• The field study analyzed all 257 registered users of one desktop assistant from 12 May to 12 August 2026, covering 798 sessions, 4,943 user questions, and 5,422 assistant replies.Source: §3.1 · p. 5

3.2 Operationalizing anthropomorphic stance

• Users who replaced the default assistant name form the anthropomorphic group (n=118); users who kept it form the default-name group (n=139).Every non-default name counts, without a questionnaire or inference from a user profile.Source: §3.2 · p. 6

3.3 Three instruments for behavioral patterns

• Behavior is measured using per-user activation rates from a sparse autoencoder codebook, existing identity tags, and lexicons for opening practices such as greetings and imperatives.The codebook retains 128 frequent latents, whose short labels are written from highly activating questions without access to group labels; opening codes do not enter its training.Source: §3.3 · p. 6

3.4 Statistical analysis

• RQ2 compares whole group-mean profiles with a label-permutation null; RQ3 uses users with at least three deduplicated questions (n=103) and adjusts identity-tag tests for daily turn count.The authors correct per-latent tests across the 128 latents and distinguish those exploratory locations from the profile-level test.Source: §3.4 · pp. 7–8

3.5 Ethics

• Before analysis, an automated privacy filter replaced labeled personal information with stable placeholders; logs cannot be released, but analysis code is supplied as supplementary material.The paper reports aggregate results and states that quoted utterances were checked by hand.Source: §3.5 · p. 8

4.1 The naming distribution is not uniform (RQ1)

• 118 of 257 users (45.9%) chose a non-default name, while 139 (54.1%) kept the default.Every non-default name counts as anthropomorphic naming in this analysis.Source: §4.1 · p. 9, Figure 2 · p. 9

4.2.1 Unit of comparison and test statistic.

• The comparison ranks group-mean profiles across 128 shared codebook coordinates, rather than comparing each user’s single strongest latent.The codebook was selected by activation frequency over deduplicated questions without using user identity or group labels.Source: §4.2.1 · p. 9

4.2.2 The null is not centered at zero.

• Across 9,999 label shuffles preserving observed group sizes, the observed rank correlation (+0.249) was less alike than a random cut of the same users; the two-sided p-value was .084.Source: §4.2.2 · p. 10, Figure 3 · p. 11

4.2.3 Where the difference sits.

• L79, emotional confiding, had the largest rank gap: 14th for anthropomorphic users versus 127th for default-name users, but it did not survive BH correction.Its activation rates were 0.115 and 0.054, respectively; the per-latent screen locates rather than establishes the profile-level difference.Source: §4.2.3 · p. 10, Figure 3 · p. 11

4.2 Behavioral patterns split by naming stance (RQ2)

• Spearman ρ=+0.249 fell in the left tail of the label-permutation null (p=.0419), supporting a difference between behavioral profiles by naming stance.The observed profiles were less alike than a random split of the same users would produce.Source: §4.2.1 · p. 9, §4.2.2 · p. 10, Figure 3 · p. 11

4.3 User attributes and opening behavior (RQ3)

• Teacher (ρ=+0.27) and tutoring (ρ=+0.22) covaried with renaming, while introverted task orientation (ρ=−0.17) covaried with keeping the default.These three associations cleared BH correction; the generally small effects support stratification, not persona switching by occupation.Source: §4.3 · p. 11

4.3.2 What the two groups do on an opening turn.

• Opening behavior covaried positively with opening greeting (ρ=+0.36, p=.001) and negatively with imperatives (ρ=−0.24, p=.022); codebook latents L227 and L18 also align with greeting.L79 captures emotional confiding, whereas L117 and L221 are transactional requests and are not merged with L79.Source: §4.3.2 · p. 12
• Because greeting and confiding occur on the opening turn, the proposed runtime cue is the current opening type: greet, receive affect, or execute directly.The paper argues that waiting to identify a user type would come too late for these openings.Source: §4.3.2 · p. 13

5.1 Policy design

• For non-default names, greetings receive a greeting and talk-or-task choice, confiding receives affect acknowledgment before that choice, and tasking leads with the deliverable; default-name replies remain task-first.If a turn combines a relational move with a task, the tasking path takes precedence.Source: §5.1 · p. 14

5.2 Evaluation procedure

• Each judgment paired two replies to the judge’s own logged utterance; the prompt showed default-name on the left and anthropomorphic on the right without naming the branches.All four questions appeared at once.Source: §5.2 · p. 14, Figure 6 · p. 15

5.3 Evaluation results

• On Q1, renamers chose the anthropomorphic reply 60.1% versus 38.0% for default-name users; Q2 was 58.0% versus 33.1%, Q4 53.1% versus 36.6%, while Q3 was mostly “about the same” (54.0% and 52.0%).The Q1, Q2, and Q4 distributions differed between groups; Q3 did not.Source: §5.3 · p. 15, Table 1 · p. 16
• Under the extreme user-dependence discount, Q1 and Q2 remained significant (p=.027 and p=.014), Q3 remained null (p=.830), but Q4 fell to p=.099.The authors therefore treat Q1 and Q2 as robust to either assumption, while Q4 is not robust to the worst case.Source: §5.3 · p. 16

6.1 User stance as a conditioning variable

• The chain runs from uneven stance to different codebook ranks and opening behavior, but RQ3 covers only users with estimable per-user rates.Source: §6.1 · p. 16
• Renamers favored anthropomorphic replies on Q1, Q2, and Q4, while both groups mainly chose “about the same” on Q3.Source: Figure 7 · p. 17

6.2 Implications for HCI research

• Product traces enable low-cost stratification but record actions, not motives; turn-level mistakes cost one turn rather than persisting across a session.Source: §6.2 · pp. 17–18

6.3 Design implications

• A uniformly warm default is not a free improvement: default-name users preferred the plainer reply in roughly a quarter of relational judgments.Source: §6.3 · p. 18
• A stable stance gate can decide whether a fork exists, while the current opening selects its branch; tasking remains unchanged between branches.Source: §6.3 · p. 18

6.4 Methodological remarks and transfer

• The greeting lexicon supplies the most robust opening signal; the sparse codebook adds a category-free view and makes behaviors such as L79 visible.Source: §6.4 · p. 18
• Transfer is uneven: language-specific lexicons need rebuilding, identity tags require the product’s profile-writing feature, and effect sizes should not be expected to transfer.Source: §6.4 · pp. 18–19

6.5 Limitations and future work

• The log findings are associations, and pairwise judgments measure immediate preference rather than effects on retention or task completion.Source: §6.5 · pp. 19–20

• Behavioral patterns over a shared codebook differed systematically by stance (ρ=+0.249; permutation p=.0419).Source: §7 · p. 20
• Teacher and tutoring identities covaried positively with renaming, while an introverted task-only orientation covaried negatively; emotional confiding was the largest codebook gap.Source: §7 · p. 20
• The revised policy greeted and received affect on relational openings and executed directly on tasking; renamers preferred it far more on personhood and affect receipt, less robustly on continuation.Source: §7 · p. 20
• Neither group judged the revised reply worse on task interference.Source: §7 · p. 20