A review of Doan & Seering, “The Design Space for Online Restorative Justice Tools: A Case Study with ApoloBot” — CHI ’25.
For decades, restorative justice has offered a powerful alternative to retributive systems in our schools, courtrooms, and neighbourhoods. Its strength lies in its deeply human-centred process: facilitated dialogue, the vulnerable sharing of harm and need, and the co-creation of a path toward repair. But what happens when we attempt to translate this profoundly interpersonal practice into code and deploy it in the chaotic, often anonymous, digital spaces where a growing share of our conflicts now occur?
This is no longer a hypothetical question. A new study by Bich Ngoc (Rubi) Doan and Joseph Seering of KAIST (the Korea Advanced Institute of Science & Technology), presented at the 2025 CHI Conference on Human Factors in Computing Systems in Yokohama, takes a pragmatic step into this frontier. Their tool, ApoloBot, is a Discord bot designed to facilitate apologies after interpersonal harm. The full paper is open-access on arXiv at https://arxiv.org/abs/2502.18861; the official ACM publication is available via DOI https://doi.org/10.1145/3706598.3713598. The findings provide a sobering yet hopeful map of the terrain — essential reading for anyone interested in the future of restorative practice online.
The scale of the problem the researchers address is significant. They cite an Ipsos survey indicating nearly 60% of online users have experienced some form of harm, while over 40% of those harmed chose not to report it. For volunteer moderators — the unpaid labour holding much of the social internet together — the available toolkit is almost entirely punitive. Content removal, temporary mutes, permanent bans: efficient at scale, but as the authors observe, this carceral architecture often leaves those harmed feeling unsupported and those who caused harm confused and unreformed. The relational damage goes unaddressed.
This is precisely the gap restorative justice was designed to fill. Drawing on the now-classic framing offered by Howard Zehr, harm is understood not as rule-breaking to be punished, but as a “violation of people and relationships” that creates obligations to make things right. The question Doan and Seering pose is whether software can plausibly carry any part of that work.
The Design: Punitive Logic, Restorative Door
ApoloBot was built as a “technological probe” — not a final product, but a deliberate provocation to surface the design space. Technically, it runs on Node.js with MongoDB and Heroku, using the discord.js library: a stack any reasonably competent solo developer could reproduce. Its design elegantly hybridises familiar punitive logic with core restorative principles, layering a possibility of repair on top of the mute command moderators already know.
A moderator initiates the process with a /apolomute command, specifying an offender, a victim, a mute duration, and a reason. This action both enforces a traditional sanction and opens a door to repair. The bot then creates separate private chat threads for each party. The person harmed is told the offender has been sanctioned and given a choice: request an apology, or let the punishment stand. This voluntary step is foundational, honouring victim autonomy — the principle that no one harmed should be conscripted into another’s healing. If they opt in, they write their request. The offender, in their own thread, is given a parallel choice: write an apology to potentially have their mute lifted, or decline. If they comply, a human moderator reviews the response before it ever reaches the other party, acting as a gatekeeper to filter out the insincere, deflective, or weaponised. If the apology is accepted, the offender is released. At any point, a “no” or a timeout reverts the process to a standard punitive mute. There is no forced forgiveness — only an opportunity for it.
A brief note on language: the paper uses victim and offender throughout, mirroring the field’s older vocabulary. Contemporary restorative practice increasingly favours person harmed and person who caused harm — language that resists fixing identity to a single act. I retain the paper’s terms below when referencing its findings, but the reframe matters, and any future tool of this kind should consider whether its prompts encode permanent labels or transitional roles.
The Method: Probe, Deploy, Reflect
With ApoloBot as their conversation starter, the researchers ran a three-phase study. Sixteen experienced Discord moderators were interviewed in onboarding sessions to gather context on their existing practices and reactions to the tool. Six of them then deployed ApoloBot in their real communities across a four-week field study, with weekly check-ins. Exit interviews followed. The resulting data offers a richly textured picture of where restorative tools can take root online — and the stark conditions where they will likely fail.
The Where and When: Finding Fertile Ground
The study’s most critical insight is that a restorative tool is not a universal fix. Its viability is deeply conditional. Moderators who found the greatest value in ApoloBot came from communities that were, in essence, relationally pre-disposed. Servers built around content creators, mental health support, or creative arts already possessed the cultural infrastructure for restoration. As one moderator noted, such a tool can “align better with their brand,” transforming conflict resolution into an asset that builds community cohesion rather than just suppressing disruption. In these spaces, an apology isn’t “cringe”; it’s a valued act that reinforces group norms around empathy and growth.
This stands in stark contrast to the “impersonal” spaces where the tool was deemed irrelevant — technical support servers, for example, where members appear briefly to get an answer and leave with no social investment. More tellingly, ApoloBot was actively rejected in highly toxic, large-scale communities, particularly around competitive gaming, where “trash talk” is the norm. Here, moderators feared the tool would be weaponised, with apologies becoming targets for mockery and vectors for further harm. One described the cultural barrier bluntly: the mindset is “it’s the norm to be toxic here,” making the offer of a restorative process not a gift but a joke. The lesson is vital: a restorative tool cannot be parachuted into a retributive culture. It can only amplify a community’s existing, latent values. (This echoes a long-standing finding from face-to-face practice — programs imposed without cultural buy-in collapse into compliance theatre.)
The “sweet spot” for application also mirrors a principle long understood by in-person practitioners: appropriate social distance. The paper finds that apologies were most effective when a relationship existed but was not yet deeply entrenched — newer members who had enough bond to value restoration, but not so much history that the conflict was intractable. At one extreme, complete strangers had little motivation to invest in apology. At the other, close friends would resolve issues privately, and entrenched rivals with “bad blood” would simply refuse. The bot was most powerful in mending the fragile ties of a nascent community, preventing a single incident from cementing a permanent grudge. One moderator who used ApoloBot successfully observed a tangible shift: after a normal mute, the two parties would “become non-friendly,” but after a bot-facilitated apology, they would return to being “normal to each other again since they got all the emotions out.”
The How: Process Before Outcome
Perhaps the most significant finding for practitioners is the question it raises about what “success” looks like. A punitive system has a clear metric: number of bans. What is the metric for a restorative bot?
Doan and Seering argue persuasively that it cannot simply be the rate of completed apologies. The voluntary nature of the process means dropouts are a feature, not just a bug. A person harmed may request an apology and find the offender’s response insufficient. An offender may simply lack the skills to articulate genuine remorse. In these cases, was the tool a failure? The researchers suggest a restorative reframe: success can be defined by the availability of the choice itself. The act of being asked, “What do you need to feel safe and whole again?” is an intervention in itself, even if the answer is “a full punishment.” This reframes the tool’s apparent “inefficiency” in the short term — mediating an apology takes far longer than clicking a ban — as a long-term investment in community health. The goal is to prevent the recurrence of harm and the festering of resentment that purely punitive systems quietly accumulate.
This connects directly to a core challenge the study identifies: equipping human moderators to be effective facilitators. A bot can structure a process, but it cannot exercise the discernment needed to evaluate an apology or the empathy to guide a participant through a difficult moment. The tool appealed most to moderation teams that were already conversationally oriented — those with a “ladder of escalation” that prioritised interaction and education before action. Teams relying purely on a rigid, rule-based approach found the tool’s inherent need for subjective judgement to be a liability, perceived as a source of inconsistency or unfairness. A clear design implication follows: the tool is not the facilitator. It is a scaffold for the facilitator. Future designs must focus on empowering the human mediator — better dashboards, contextual tips, guidance on writing feedback that helps an offender craft a more meaningful apology. The goal is not to automate a restorative justice conference but to make its core loop — request, response, review — manageable at a scale and in a text-based medium that fully manual practice cannot achieve.
The Uncanny Valley of Justice
The study also surfaces a uniquely digital-age challenge: the perception of the tool itself. One deployment was reportedly derailed by community backlash rooted in profound distrust of bots. Members, conditioned by experience with spammy and opaque automated systems, believed ApoloBot was an AI tool harvesting their private conversations. That the bot’s interactions were scripted and its data collection minimal didn’t matter; the perception was instantly toxic.
This presents what we might call an uncanny valley of justice. A purely human-mediated process is authentic. A purely automated punishment is efficient and impersonal — and users can make a certain peace with that, the way they accept traffic enforcement cameras. But a bot offering a path to reconciliation can feel duplicitous: a machine feigning care, and thereby insulting the very humanity restorative justice seeks to honour.
The researchers’ suggested fix is transparency: radical, up-front explanation of not just the “how” but the “why” of the restorative philosophy beneath the tool. More provocatively, they suggest the bot’s communication might need to be “humanised” — framing prompts not as commands from a faceless system, but as a channel for the human moderator’s intent, perhaps signed in their voice.
Here I’d offer a friendly tension with the authors. “Humanising” the bot risks the wrong kind of dishonesty — a chatbot that pretends to feel. The deeper move may be the opposite: transparently mechanical tools that visibly hand off to humans at every emotional juncture. The bot’s job is to handle the choreography; the human’s job is to handle the soul. Making that handoff legible to participants — this is a human reading your message now — may matter more than softening the bot’s tone. We do not want a digital uncanny valley between code and care; we want a clearly marked bridge.
What This Means for Practice
ApoloBot is not the future of online restorative justice. It is, as the researchers know, a flashlight in a dark room, revealing the contours of what that future could be. For our community, its findings offer both an invitation and a warning.
The invitation is to bring our hard-won understanding of needs, dialogue, and non-punitive accountability into digital spaces starved for these things. Three opportunities the study illuminates:
First, restorative tools belong where relational infrastructure already exists. Our task is to partner with mental-health communities, creative collectives, learning servers, faith-based spaces — the digital equivalents of the schools and faith communities where in-person restorative practice has historically taken root. We work with the grain of latent values, not against it.
Second, the metric for success must be the offer, not the completion. Just as a victim-offender dialogue program is not judged solely by signed agreements, an online tool is not judged solely by accepted apologies. We measure dignity offered, not transactions closed.
Third, the human facilitator remains the load-bearing element. Any investment in tooling must be matched by investment in training the moderators who use it — and platforms that take this seriously will need to begin treating moderation not as unpaid janitorial labour but as skilled relational work.
The warning is that our practice cannot be copy-pasted into code. It must be carefully cultivated in fertile community soil, designed to empower human facilitators, and communicated with a radical transparency that overcomes deep-seated distrust of algorithmic mediation. The opportunity is vast, but the work requires us to blend the timeless principles of our practice with a sophisticated understanding of the new, intricate social ecosystems where harm now lives.
The frontier is open. The map is sketched. The next moves are ours.
Source
Doan, B. N. (Rubi), & Seering, J. (2025). The Design Space for Online Restorative Justice Tools: A Case Study with ApoloBot. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25), Yokohama, Japan. Association for Computing Machinery, New York, NY, USA. Article 694, 1–19.
ACM (paywalled): https://doi.org/10.1145/3706598.3713598 Open-access preprint: https://arxiv.org/abs/2502.18861