Reflective Computing

1 INTRODUCTION

• Fast feeds, autoplay, and conversational AI can displace reflection, which requires time and space for people to consider beliefs and experiences.The introduction frames how interactive systems can give people that time and space as an open design question.Source: §1 · p. 3
• Calm, affective, persuasive, positive, slow, personal-informatics, digital-wellbeing, and human-AI paradigms organize design around aims other than reflection itself.Their aims include attention, emotion, behavior, wellbeing, self-tracking, and reasoning alongside intelligent systems.Source: §1 · p. 3
• Existing HCI paradigms use reflection as a means, stage, or side effect, while prior work often invokes it without defining or evaluating it.Frameworks describe levels or dimensions of thought, rather than qualities systems need to invite, protect, and deepen reflection.Source: §1 · p. 3
• The article proposes reflective computing and derives its dimensions from recurring gaps found by analyzing three decades of ACM interactive-system designs.It then codes the dimensions with two independent coders and profiles systems and paradigms against them.Source: §1 · p. 3
• The article contributes a definition and seven dimensions, a reliable codebook and coded corpus, cross-paradigm profiles, and design moves for closing identified gaps.Its profile also reports two cultures of reflection and an advantage for designs in which people set the pace.Source: §1 · p. 3

2.1 Computing paradigms as design traditions

• A computing paradigm is a named design tradition that provides shared aims, vocabulary, and exemplars; the corpus is classified using eleven such traditions.Paradigms can overlap, yet shape what designers notice and optimize.Source: §2.1 · p. 4, Table 1 · p. 4

2.2 Theoretical lenses on reflection

• Dewey and Schön locate reflection around pauses during or after action, while HCI frameworks describe thought levels and prompting resources rather than qualities systems need to invite, protect, and deepen reflection.Form, pace, openness, agency, and unfilled time provide sensitizing lenses, not a predefined coding scheme.Source: §2.2 · pp. 4–5

2.3 Prior reviews of reflection in HCI

• Baumer et al. found reflection often invoked without definition or evaluation; prior reviews also stay within paradigms, catalog existing systems, and stop short of defining a paradigm organized around reflection.The present article addresses these three gaps through cross-paradigm analysis of what systems collectively lack.Source: §2.3 · p. 5

3.1 Review design

• The critical review analyzes a purposive set of exemplary designs and tests the resulting dimensions for coding reliability rather than claiming an exhaustive literature search.The review design is intended to make a conceptual contribution from the corpus.Source: §3.1 · p. 5

3.2 Corpus selection

• The corpus includes interactive interventions published at ACM venues from 1996 to 2026, selected through field knowledge, references, and citation chasing as representative examples.The final corpus contains 134 systems from 17 venues; its paradigm counts describe the exemplar set, not the distribution of ACM publications.Source: §3.2 · pp. 5–6

3.3 Abductive gap analysis (Phase 1)

• In Phase 1, the researchers iterated between system gap notes and reflection concepts, clustering recurring absences and restating seven gaps as design dimensions.Social reflection did not emerge as a separate recurring gap.Source: §3.3 · p. 6

3.4 Codebook development and profiling (Phase 2)

• Phase 2 translated each dimension into absent, partial, or strong scores and summed them into a 0–14 proximity score, which is not a measure of design quality on its own terms.The codebook also recorded who set interaction pace, the predominant form, reflection aims, and data-practice reporting; unreported D7 values were scored neutrally.Source: §3.4 · p. 7

3.5 Independent coding and adjudication

• A second coder independently scored systems from a frozen codebook, initially using mixed sources, then rechecked all 134 systems from full text.Ordinal reliability measures assessed agreement; codebook amendments addressed four systematic disagreements before the author adjudicated remaining disagreements.Source: §3.5 · pp. 7–8

3.6 Statistical analysis

• Statistical analysis used rank correlations, factor and latent-class models, paradigm comparisons, and ordinal regression to assess pace associations while controlling for paradigm and form.D7 was excluded from factor and class analyses for low variation; pace models also removed D6 in a robustness check.Source: §3.6 · pp. 9–10

4.1 Pace imposed by the system defines tempo agency (D1)

• 51 systems (38%) let systems or designers time essentially all interaction; user-paced systems average 5.7 on the other dimensions, versus 4.8 for shared-pace and 3.8 for system-paced systems.Tempo agency means people can slow down, speed up, pause, resume, or decline interaction.Source: §4.1 · p. 11

4.2 Urgent or standard form defines invitational form (D2)

• 62 systems offer no signal that pausing is welcome; among 56 screen-based systems, 44 score zero on invitational form.Physical counterexamples use material, motion, and pace to signal that taking time is welcome.Source: §4.2 · p. 11

4.3 Verdicts and prescriptions define interpretive openness (D3)

• 53 systems tell people what their experience means through scores, goals, recommendations, emotional labels, or AI answers.Interpretive openness instead offers material, juxtapositions, or questions and leaves meaning to the person.Source: §4.3 · p. 12

4.4 Reflection that stops at description defines reflective depth (D4)

• Only 12 systems scaffold reflection beyond description; 58 target no reflection, while another 64 support only recall or description.The deeper scaffold moves from describing an experience toward questioning, reframing, and transforming understanding.Source: §4.4 · p. 12

4.5 Momentary use defines temporal span (D5)

• 74 systems support only momentary use, while only 11 are designed for revisiting or anticipation over longer periods.Temporal span concerns reflection across meaningful periods through revisiting, anticipation, and long-term presence.Source: §4.5 · p. 12

4.6 Unprotected idle moments define protected idleness (D6)

• 32 systems actively fill idle moments, 84 neither fill nor shape idle time, and only 18 deliberately preserve or shape quiet.Protected idleness leaves unfilled moments available for rest, mindwandering, and reflection.Source: §4.6 · p. 12

4.7 Covert influence, exposure and silence define ethical non-extraction (D7)

• 13 systems score zero on ethical non-extraction; 56 papers are silent about collected inner-state data, and only two make serving the person an explicit design commitment.The dimension calls for pauses, emotions, and reflections to serve the person, remaining private by default and transparent in use.Source: §4.7 · p. 13

4.8 Defining reflective computing

• Reflective computing makes inviting, protecting, and deepening a person’s reflection its primary aim, with user control over when to reflect, interpretive openness, and no covert use required for membership.Under these conditions, 20 of the 134 systems qualify as instances; the others are neighbors whose gaps define the paradigm.Source: §4.8 · pp. 13–14

5.1 Reliability of the dimensions

• 0.82 was the ICC for the overall proximity score, while protected idleness (α = 0.32) and pace (κ = 0.43) did not reach acceptable agreement in the first coding.Agreement was higher for four dimensions, including invitational form (α = 0.84); codebook amendments addressed systematic disagreements, and final scores used full text for all 134 systems, though the amended-codebook re-check was not an independent replication.Source: §5.1 · pp. 15–16

5.2 Composition of the corpus

• 11 of the 12 systems from 2021–2026 were human-AI interaction systems, following earlier shifts in which calm and connectedness designs dominated the late 1990s and persuasive technology rose through the 2000s.The corpus contains 134 systems across six periods; 56 (42%) were screen-based, and health (44), work (26), and personal life (20) were the most common domains.Source: §5.2 · pp. 16–17, Figure 2 · p. 17

5.3 How close systems come to reflective computing?

• Proximity scores ranged from 0 to 11 (mean 5.7; median 6); no system reached 14, and five systems tied for the highest score of 11.Only 16 systems reached at least a partial level on all seven dimensions; 31 scored 3 or less.Source: §5.3 · p. 17

5.4 Profiles of the eleven paradigms

• 8.8 was the mean proximity score for slow, ludic and experience-centered design, while persuasive technology (3.8) and human-AI interaction (3.9) were furthest from reflective computing.Paradigm scores differed significantly (Kruskal–Wallis H = 49.1, p < .001), but confidence intervals overlapped heavily except at the extremes, so the profiles are better treated as tendencies than definitive ranks.Source: §5.4 · pp. 18–19, Table 6 · p. 20

5.5 Prevalence of each dimension

• 12 systems scaffolded reflection beyond description, 11 supported revisiting or anticipation over months or years, and temporal span was absent in 74 systems.Only two systems made ethical non-extraction an explicit design commitment; protected idleness was most often partial.Source: §5.5 · p. 20, Table 7 · p. 21

5.6 Two cultures of reflection, through form and through depth

• Invitational form and interpretive openness were correlated (ρ = 0.50), as were reflective depth and temporal span (ρ = 0.46), while form and depth were uncorrelated (ρ = -0.01).Factor analysis retained two factors: form dimensions loaded on one and reflective depth, temporal span, and tempo agency on the other; interpretive openness cross-loaded on both (0.62 and 0.55).Source: §5.6 · p. 21, Table 8 · p. 21
• 43 systems formed a form culture, 43 a depth culture, and 48 neither; only one system combined a strong form cluster with scaffolded depth.The form culture was dominated by slow-technology systems, while the depth culture included most personal-informatics systems; the cultures work through aesthetic, quiet artifacts or structured thought, respectively, and are rarely combined.Source: §5.6 · pp. 22–23, Figure 6 · p. 23

5.7 Who sets the pace, and why it matters?

• 13.5 times higher odds of a higher proximity score were associated with user-paced systems, controlling for both paradigm and form; shared pace was also associated with higher scores (OR 5.4).The user-paced association was clear in affective computing and mindfulness, but not detectable in slow design or persuasive technology; pace and dimension scores were coded from the same papers, so a halo effect is possible.Source: §5.7 · pp. 24–25
• OR 8.0 (95% CI 3.0–21.6) links user-paced systems to higher scores after controlling for paradigm and form, even with protected idleness excluded.User pace is also associated with interpretive openness (OR 21.3), protected idleness (8.5), and reflective depth (6.3), but not temporal span or invitational form. The association persisted in each coder’s separate analysis, though part of it is built into the rubric.Source: Table 9 · p. 25, §5.7 · p. 25

5.8 Differences by form, domain and period

• Tangible objects score highest (mean 9.3), while screens (4.6) and wearables (4.3) score lowest; personal-life and home systems score highest by domain.Scores are broadly stable across periods from 1996 to 2020 (5.3–6.2) and lower in 2021–2026 (4.3), but period and paradigm are fully confounded. Eleven of 12 recent systems are AI systems; these sampled AI systems score below the corpus overall (p = 0.006), so a systematic sample of recent AI-for-reflection systems is needed.Source: §5.8 · pp. 25–26

5.9 Silence on data use

• 56 of 134 papers (42%) are silent about storage, access, sharing, or secondary use of users’ data, emotions, or reflections.Another 63 describe data practices without making them design commitments, and 13 involve covert influence, surveillance, or exposure. Scoring silence as 0 barely changes paradigm ranking (Spearman ρ = 0.97); silence is not neutrality and may leave deployed products’ data practices to business models.Source: §5.9 · p. 26

5.10 What the seven dimensions miss

• Seventeen substantive “other gap” notes identified three candidate gaps beyond the seven dimensions: contestable inference, automation that displaces noticing, and other-directed reflection.These appeared in 7, 5, and 5 systems, respectively, and none reached the frequency of the seven dimensions.Source: §5.10 · p. 26

5.11 Robustness among systems designed for reflection

• Among 22 systems primarily aimed at reflection, only 8 scaffold reflection beyond description, 5 support long time spans, 3 protect idle time, and none makes non-extraction a design commitment.Among 83 systems with any reflection aim, only one combines a strong form cluster with scaffolded depth; user-paced systems still score higher after controlling for paradigm and form (OR 5.8), and with D6 excluded (OR 4.6).Source: §5.11 · pp. 26–27

6 WHAT EACH PARADIGM LACKS, AND HOW TO CLOSE THE GAP

• Strengths are distributed across paradigms: social connectedness leads on tempo agency and non-extraction, while slow, ludic and experience-centered design leads on openness; depth, span, idleness and non-extraction retain gaps.Figure 8 marks the strong level as a design target, not a measured profile; some leading paradigms have six or fewer systems, so their positions describe exemplars.Source: §6 · p. 27, Figure 8 · p. 28

6.1 Slow, ludic and experience-centered design; 6.2 Personal informatics

• Slow, ludic and experience-centered designs offer invitational form and interpretive openness, but all 19 support only recall or description; personal informatics contributes depth, temporal span and user pacing but lacks inviting form and protected idleness.The proposed moves are tunable slowness and optional prompts, and slower physical or ambient forms for personal data.Source: §6.1 · p. 29, §6.2 · pp. 29–30

6.3 Social connectedness computing; 6.4 Contemplative and mindfulness computing

• Social connectedness gives people control over when to reach out but seldom supports reflection on relationships over time; contemplative systems invite attention and protect idleness but rarely connect sessions to longer-term reflection.The proposed moves are shared traces people can revisit and links between a practice session and changes across weeks.Source: §6.3 · p. 30, §6.4 · p. 30

6.5 Affective computing; 6.6 Calm and ambient computing

• Affective systems can offer open material for interpretation, yet often lack temporal span and non-extraction; calm and ambient systems offer peripheral form but lack user-controlled pacing and reflective depth.The proposed moves are private, contestable emotional material that accumulates over time and ambient information people can bring forward when they choose.Source: §6.5 · pp. 30–31, §6.6 · p. 31

6.7 Digital wellbeing and self-control; 6.8 Positive computing and mental health

• Digital wellbeing systems can protect idle time but rarely invite interpretation, while positive computing scaffolds depth but typically delivers screen-based exercises with reminders and little protected idleness.The proposed moves are open questions at moments of friction and more rest, pacing and ambiguity in structured exercises.Source: §6.7 · p. 31, §6.8 · p. 31

6.9 Attention and interruption management; 6.10 Persuasive technology

• Attention-management systems protect focus but treat time chiefly as a productivity resource; persuasive systems often set the pace and define success by designer-chosen target behaviors.The proposed moves are using protected time for reflection and replacing target behaviors with questions that leave people to choose their direction.Source: §6.9 · p. 32, §6.10 · p. 32

6.11 Human-AI interaction

• Human-AI systems offer some thinking prompts and comparatively high tempo agency in invoked writing assistants, but all 12 lack invitational form and only one has temporal span.The proposed move is user-chosen AI pacing, space to think before and after answers, and records for revisiting what was accepted and why.Source: §6.11 · p. 32

7.1 Reflective computing as a paradigm

• Reflective computing brings together qualities that existing systems express separately; no paradigm holds all seven dimensions together as its aim.Most dimensions already have working exemplars, making the proposal’s testable claim that these qualities belong together rather than that a new technique is needed.Source: §7.1 · p. 33
• Its organizing aim is the human capacity to pause, consider and make sense of experience, with success judged by whether people can think for themselves.This contrasts with neighboring paradigms organized around system functions or outcomes, and places control of tempo at the center.Source: §7.1 · p. 33
• Reflective computing is a set of qualities any device or interface can carry, not a hardware category; no form guarantees reflection.Tangible systems score highest on average, yet examples show that tangible or conversational forms do not determine whether systems invite or scaffold reflection.Source: §7.1 · p. 33

7.2 Relation to neighboring paradigms

• Reflective computing complements neighboring paradigms by identifying their contributions to reflection and what they would need to add.It makes tempo agency central, combines form-based and thought-based traditions, includes ethical use of what people reveal, and turns these qualities into seven scoreable dimensions.Source: §7.2 · pp. 33, 35
• Reflective design focuses chiefly on reflection about technology, while reflective computing describes assessable qualities of systems that support reflection on one’s own experience.Slow technology contributes form and idleness but often sets the pace and rarely scaffolds thought beyond description; reflective informatics describes reflection’s process rather than the conditions systems create around it.Source: §7.2 · pp. 34–35
• No system in the corpus expresses all seven dimensions strongly, and the dimensions have not yet been applied beyond the 134 systems from which they were derived and tested.The authors call for systems designed to express all seven dimensions and independent application of the codebook to new systems; until then, reflective computing remains a proposal rather than an established paradigm.Source: §7.2 · p. 35

7.3 Bridging the two cultures of reflection

• HCI supports reflection through two cultures: one invites slowing down and wonder through form, while the other guides thought with prompts, questions and structured exercises.The distinction is supported by correlations, a two-factor structure and latent classes that separate form and depth.Source: §7.3 · pp. 35–36
• Only one of the 12 systems that scaffold depth scores 4 or more of 6 on the form cluster, and the separation also appears among systems aimed at reflection.The traditions developed apart: design research emphasized artifacts and ambiguity, while behavioral and health research emphasized measurable outcomes and prompts.Source: §7.3 · p. 36
• Reflective computing needs both cultures: form alone may leave people at reflection’s threshold, while depth without distinctive form competes with ordinary demands on attention.Interpretive openness is a potential meeting point; Somaesthetic appreciation design scores on both clusters, but its depth score was a close call.Source: §7.3 · p. 36
• Proposed bridges—such as a slow artifact offering an optional question or a journal in an ambient form—are design hypotheses, not findings.The authors identify the gap between the cultures as a likely source of new system designs and call for concepts that bridge them.Source: §7.3 · p. 36

7.4 Tempo agency as the central condition

• In 7 of 19 slow-technology systems, the designer or system sets the pace, though fixed pace and user-paced designs score alike within slow design.Fixed timing can be part of a design’s poetics, but it leaves the recipient unable to alter when the pause occurs.Source: §7.4 · p. 36
• User-paced systems score markedly higher than system-paced systems on other dimensions, with an odds ratio of 13.5 after controlling for paradigm and form.The advantage is clearest for interpretive openness, reflective depth and protected idleness, but not invitational form; it remains when protected idleness is excluded (OR 8.0).Source: §7.4 · p. 36
• Tempo agency lets people accept or decline a pause, choose whether to fill idle time, and decide when to deepen or stop reflection; slower is not always better.The discussion treats the person’s control of tempo as a condition for the other dimensions, while distinguishing chosen speed from the removal of choice.Source: §7.4 · p. 37
• Designs can offer tempo control through advance-set rhythms, quick or considered modes, or rhythms that follow the person’s existing patterns.Examples include personal goals and limits, SwitchTube’s mode choice, breath-following Sonic Cradle and memory systems that people open or revisit when they choose.Source: §7.4 · p. 37

7.5 Reflective computing and generative AI

• Every AI system scores 0 on invitational form, only one supports reflection over time, and AI systems that slow people down usually impose the pause.The corpus cannot separate a time trend from a paradigm change, and its 12 AI systems were sampled as decision-making and writing exemplars rather than reflection exemplars.Source: §7.5 · p. 37
• The AI systems’ reflective supports typically address only one or two dimensions, with little distinction in form or timing from the rest of the interface.The discussion proposes AI that lets people choose when answers arrive, offers alternatives and questions, supports examining suggestions, preserves thinking over time, leaves pauses unfilled and avoids using reflections without knowledge.Source: §7.5 · pp. 37–38
• When an AI model drafts, logs or explains on a person’s behalf, it can remove the occasion for reflection.Source: §7.5 · p. 38

7.6 Reflection as something that can be extracted

• Only two systems make ethical non-extraction an explicit design commitment, while 56 papers are silent about what happens to collected inner states.The discussion argues that pauses and journal entries can reveal valuable information, and calls for reflections to be private by default and not used to persuade, profile or engage people without their knowledge.Source: §7.6 · p. 38
• Contestability and privacy can be built into artifacts: Casablanca removed automatic presence sensing, and Spheres of Wellbeing restricts access to recordings to their owner.Ripple’s alternative interpretations also make who benefits from emotional data part of the design.Source: §7.6 · p. 38

7.7 Design directions for reflective computing

• Reflective designs should combine invitational form with optional scaffolds for deeper thought, while letting people choose whether and when to slow down.Source: §7.7 · pp. 38–39
• Designs should leave meaning to the person, support revisiting and anticipation over months, and allow graceful lapse and return.Source: §7.7 · p. 39
• 42% of the corpus consists of screens, which score among the lowest, and no AI system has invitational form.Source: §7.7 · p. 39
• Interfaces can protect idle time through endings, stopped feeds, and quiet when nothing needs saying.Source: §7.7 · p. 39

7.8 Evaluating reflective computing

• Reflection is often claimed without being defined or evaluated, so evaluation should ask whether and how people reflected rather than rely only on usage, satisfaction, or behavior.Source: §7.8 · p. 39
• The codebook profiles expressed design qualities, but cannot establish whether people reflected; that requires experience-level inquiry and evaluation over time.Source: §7.8 · p. 39

7.9 A research agenda

• The dimensions remain a proposal: new corpora should test whether they are complete and distinct, including candidate gaps and social reflection.Source: §7.9 · pp. 39–40
• Whether the seven dimensions should be equally weighted, combined as form and depth, or treated as preconditions remains open.Source: §7.9 · p. 40
• Only one design nearly bridges invitational form and scaffolded depth; studying such combinations could test whether they outperform either quality alone.Source: §7.9 · p. 40
• Future work should compare ways to combine user-controlled pace with designed slowness and study reflective AI in writing, decisions, and learning.Source: §7.9 · p. 40
• The corpus is dominated by Western domestic and workplace settings, and reflection, pace, and idleness may require adaptation across cultures and communities.Source: §7.9 · p. 40

7.10 Limitations

• The dimensions and scores need external testing: they were derived and applied on the same purposive corpus, and coding relied on two coders with author adjudication.Source: §7.10 · p. 40
• Independent-coding reliability reflects the first round; later full-text re-coding with amended rules was not an independent agreement estimate.Source: §7.10 · p. 41
• The proximity score treats dimensions as equal and additive despite their dependence, so it summarizes expressed qualities rather than measuring how reflective a system is.Source: §7.10 · p. 41

7.11 Implications for design, research and education

• Designers can use the dimensions and paradigm-specific translation moves to identify characteristic gaps, including whether AI interfaces leave time and space to think.Source: §7.11 · p. 41
• Reviewers can ask who controls pace, whether form invites pausing, how open meaning remains, and whom collected reflections serve.Source: §7.11 · p. 41
• Researchers can test and extend the proposal using the archived corpus, codebook, and adjudication log.Source: §7.11 · p. 41
• Reflection is a set of qualities spanning pace, form, content, time, silence, and ethics, not a feature to add in isolation.Source: §7.11 · p. 41

8 CONCLUSION

• Seven recurring gaps across 134 systems define dimensions spanning tempo agency, invitational form, interpretive openness, reflective depth, temporal span, protected idleness, and ethical non-extraction.Source: §8 · p. 41
• No system holds all seven dimensions; designs that give people pace control score higher on openness, depth, and idleness.Source: §8 · p. 41
• The two cultures of reflection—through form and scaffolded thought—have met in only one design.Source: §8 · p. 42

• The full coding, codebooks, independent scores, adjudication log, and analysis scripts are archived on OSF.Source: DECLARATIONS · p. 42
• A large language model sparsely edited the manuscript, while two human coders conducted all coding and adjudication.Source: DECLARATIONS · p. 42
• The author reports no conflict of interest.Source: DECLARATIONS · p. 42
• The work was supported by the author’s university seed grant.Source: DECLARATIONS · p. 42

• D1–D7 scores use 0 for absent, 1 for partial, and 2 for strong, with a total score ranging from 0–14.Source: A CODED CORPUS · p. 50
• The coding key records whether reflection was an aim, the second coder’s first-round source, and a score of 1 when papers were silent on data use.Source: A CODED CORPUS · p. 50
• Systems not initially coded from full text were re-coded from full text before final scores were set.Source: A CODED CORPUS · p. 50
• Table A1 lists all 134 coded systems in chronological order with consensus scores.Source: Table A1 · p. 50