PvEBot: Players vs Enemy
A Discord community safety bot grounded in New Zealand extremism-prevention research. In PvE, the whole community cooperates against a shared threat. Here that threat is toxicity, harassment, and extremist recruitment, not the players themselves.
PvEBot began as PVEbot, a research deliverable to the Department of the Prime Minister and Cabinet on the gaming-extremism nexus. The research is prepared; the bot is being piloted with a small number of New Zealand game developer communities. The rename to PvEBot reframes it for the people it serves. Game communities know PvE as the mode where the whole party works together against a shared enemy. In community safety, that enemy is the pattern of behaviour that dismantles communities from the inside: toxicity, harassment, recruitment, radicalisation.
Game designers and extremist recruiters are solving the same problem. How do you engage a person, hold their attention, create emotional investment, and motivate increasingly committed action? Game designers do this to create entertainment. Extremist recruiters do this to create soldiers. Understanding the parallel is what makes it possible to design community tools that protect the first case without enabling the second.
Gaming communities are being exploited at scale
International research by the Royal United Services Institute's Extremism and Gaming Research Network surveyed 2,200+ gamers across seven countries. The findings are direct.
RUSI's research identifies three mechanisms of exploitation. First, direct use of games for propaganda through custom modifications in Arma 3, GTA V, and Roblox. Second, gaming-adjacent platforms (Discord especially) used as trust-rich recruitment infrastructure, documented in ICCT's design principles. Third, the gamification of real-world violence: scoreboards, achievement language, and live-streaming applied to actual attacks. These mechanisms reinforce each other rather than operate in sequence.
The Charlie Kirk assassination attempt (September 2025) demonstrates a pattern the research calls Mixed, Unclear, Unstable (MUU) ideology. Bullet casings were inscribed with references to the game Helldivers 2, the anti-fascist song Bella Ciao, and internet meme language. The perpetrator's motive was simultaneously sincere and ironic. Traditional analysis frameworks assume a coherent ideological identity and cannot parse attacks of this kind. Subcultural literacy and pattern-based detection become essential.
For the New Zealand game development sector, the economic stakes are direct. The sector generated NZD $760 million in 2024/2025, a 38.6 percent year-on-year increase, employing 1,418 FTEs and deriving 95 percent of revenue from international exports. Community health is the commercial model. When toxic actors exceed Axelrod's ~10% threshold, cooperative behaviour collapses. Communities do not degrade gradually; they collapse. The RUSI attrition numbers (60/60/70) show the collapse already in progress in communities without active moderation. PvEBot is infrastructure for preventing that collapse.
Seven rule-based detection analysers
PvEBot analyses every incoming message through seven analysers running in parallel. Each produces a severity score, a confidence score, and structured indicators. Scores combine via configurable weights into a risk level: none, low, medium, high, critical. Message content is hashed with SHA-256 before storage; the original text cannot be recovered.
Keyword
Hate speech, slurs, extremist terminology including coded numeric patterns such as 88 and 1488.
Pattern
Regex-based recruitment language, radicalisation rhetoric, platform migration links to Telegram, Signal, Gab, Rumble, BitChute, Odysee.
Behavioural
Message-frequency spikes, channel hopping, DM solicitation, rapid escalation, new-account risk.
Contextual
Distinguishes legitimate game discussion (where "kill", "destroy", "headshot" are routine vocabulary) from non-game violent language by analysing channel context and conversation flow. Critical for reducing false positives in gaming communities.
Post-Ironic Violence
The novel contribution of this detection system. Detects violence normalised through layers of irony, humour, and gaming language that maintain plausible deniability while the content functions as sincere desensitisation or signalling. This is the pattern illustrated by the Charlie Kirk case.
Recruitment
Explicit recruitment language, pressure tactics, isolation techniques, platform migration invitations. Weighted highest because early-pipeline intervention is the most effective point to disrupt radicalisation.
Identity Fusion
In-group cohesion language and willingness-to-fight rhetoric. Dehumanisation combined with expressed willingness to use violence is the combination most predictive of progression to violent action.
GenAI is opt-in and defaults to off. For studios that choose AI-enhanced detection, locally hosted open-source models (Llama 3.2 via Ollama) are the recommended path, keeping all inference data within Aotearoa New Zealand. Cloud providers (Anthropic Claude, OpenAI) remain available but send community data to overseas servers and are not the default.
A 168-hour (one week) half-life decay is applied to accumulated risk scores. Past behaviour informs assessment but does not define the individual permanently. This temporal forgiveness is a prerequisite for any genuine rehabilitation framework.
Seven levels, restorative by design
Every response with material consequences for a person's community membership always involves a human moderator. Levels 0–2 are automated. Levels 3–6 require moderator approval. Every user-facing message is grounded in Self-Determination Theory and Te Ao Māori values.
| Level | Response | Approval | Description |
|---|---|---|---|
| L0 | Monitoring | Automatic | Passive observation. Pattern data logged for trend analysis. No user-facing action. |
| L1 | Nudge | Automatic | Subtle prosocial content introduced into the channel without identifying the triggering user. |
| L2 | Warning | Automatic | Private message referencing community guidelines. Framed as a choice (Mana Motuhake). |
| L3 | Restriction | Moderator | Reduced channel access. Participation narrowed, not eliminated. |
| L4 | Mute | Moderator | Temporary communication restriction. Can read but not post. |
| L5 | Kick | Moderator | Removal with the option to rejoin after cooldown. Return pathway included. |
| L6 | Ban | Moderator | Categorised ban with a structured rehabilitation pathway assigned (toxic, abusive, extremist, or external referral for illegal activity). |
Self-Determination Theory & Te Ao Māori
Every PvEBot message supports the three psychological needs identified by Self-Determination Theory (autonomy, competence, relatedness), mapped to Te Ao Māori values. Messages perceived as controlling or punitive produce reactance; messages perceived as autonomy-supportive produce internalisation. The framing is the mechanism by which the intervention works.
Four rehabilitation pathways
Level 6 bans are categorised, not generic. Each category has a structured pathway for genuine return, built on the research finding that radicalisation happens through relationships, and de-radicalisation must also happen through relationships.
What the research establishes
PvEBot is the tool; the research is the reason it exists. The design is documented in the DPMC research report (reference PVE-2026-RPT-001, UNCLASSIFIED) and its four appendices.
The game design / extremist recruitment parallel
The core research contribution is the observation that game designers and extremist recruiters solve the same engagement problem. Both build in-group identity, provide clear enemy framing, structure progression, reward escalating commitment, and exploit the three core psychological needs identified by Self-Determination Theory: autonomy, competence, and relatedness. The difference is purpose. Game designers create entertainment. Recruiters create soldiers. The mechanics are the same.
This parallel explains why gaming-adjacent platforms (Discord especially) are such effective recruitment infrastructure. The trust-rich environment of a shared gaming interest provides cover for relationships that can later be redirected. It also explains why the counter-intervention must operate at the same mechanical level: addressing the unmet needs that made extremist content appealing, rather than simply removing the content.
Appendix A: RUSI statistics
Findings from the Royal United Services Institute Extremism and Gaming Research Network. Mixed-methods study of 2,200+ gamers across seven countries. One-third exposed to extremist content. One-quarter targeted by active recruitment. 50%+ of female gamers facing gender-based harassment. The Axelrod 10% cooperation-collapse threshold applied to toxic communities.
Appendix B: Detection architecture
Technical specification of the seven-analyser detection system, parallel execution model, weighted scoring, 168-hour decay, SHA-256 content hashing. The contextual gaming analyser is the critical differentiator from general-purpose moderation tools. It distinguishes gameplay vocabulary from genuine threat language.
Appendix C: Rehabilitation pathways
The seven-level graduated response, SDT and Te Ao Māori framing, four categorised rehabilitation pathways (toxic cooldown, abusive guided return, mentored extremist rehabilitation, external referral for illegal activity). Restorative rather than punitive by design.
Appendix D: Deployment model
NZ data sovereignty under the Privacy Act 2020. Catalyst Cloud (Wellington datacentres, NZ Government-approved) as the recommended host. Designed to apply beyond gaming to schools, youth organisations, sports clubs, and professional associations.
Te Mataiaho-aligned curriculum
A companion education programme for New Zealand schools, structured across five phases of learning and aligned with the Te Mataiaho curriculum refresh. Activity-first pedagogy: students experience the dynamics through a game or exercise before any theory is introduced.
The curriculum connects bullying and radicalisation as manifestations of the same group dynamics: exclusion, status competition, escalation, in-group and out-group formation. It develops game design literacy, the ability to see engagement mechanics operating beneath games, social platforms, and recruitment campaigns. Te Ao Māori values (Manaaki, Whanaungatanga, Mana Motuhake, Pūmanawa, Pakiki, Auahatanga, Kaitiakitanga, Kotahitanga) run through every module as a lens for learning, not as a separate cultural module.
- CL1.1 · Kindness Online, Years 1–3 (Foundations)
- CL4.1 · Online Communities, Years 7–8 (Critical thinking)
- CL7.3 · Systems Thinking & Extremism, Years 11–13 (Systems & action)
Teaching materials, including the exemplar modules and a Teacher's Guide, are available on request to New Zealand schools and kura. The curriculum complements existing NZ resources such as Netsafe's Kete, Keep It Real Online, and the work of He Whenua Taurikura.
Built to stay in Aotearoa
PvEBot is designed to run entirely within New Zealand. All community moderation data remains subject to NZ jurisdiction and the Privacy Act 2020.
The architecture applies well beyond gaming. School Discord servers, sports clubs, youth organisations such as Scouting, Cadets, and Guides, and professional associations are all communities vulnerable to the same dynamics and all suitable deployment targets.
PvEBot is invite-only during the pilot
During the research pilot phase, PvEBot runs on a single maintainer-hosted instance. Applications are reviewed manually. Approved communities receive a per-guild invite link by return email.
- A community with an identifiable purpose and active moderator team
- At least one admin who has read the research context on this page
- A commitment to the graduated response framework (no "ban everything" approaches)
- Willingness to provide feedback during the pilot
- Manual review, typically within two weeks
- A per-guild OAuth invite link restricted to your server only
- Your server added to the bot's allowlist (required; the bot will otherwise leave)
- Direct email contact for questions and issues
Submitting the form opens your default email client with the application as plain text. If the form does not work for you, please email simon.mccallum@gmail.com directly with the same information.
For developers who want to run their own instance
PvEBot is open source. If you want to run your own instance for data sovereignty, sensitivity, or customisation reasons, the architecture is designed for it.
Source & docs
TypeScript + Discord.js 14 + better-sqlite3. The repository is currently private during the pilot; request access through the application form above if you want to self-host, and include your use case. Docs include DEPLOYMENT.md, a full detection-architecture specification, rehabilitation-pathway definitions in YAML, and sample Catalyst Cloud configuration.
Recommended stack
Catalyst Cloud (Wellington) as the host. Ollama running Llama 3.2 for the optional GenAI layer, keeping inference data within NZ infrastructure. SQLite for per-guild storage. Docker Compose for orchestration.
Bot listings & access control
If you run PvEBot yourself, you will face the same problem that the hosted pilot faces: how to control who can add the bot. A single maintainer cannot support unlimited communities. The recommended layered approach:
- Discord Developer Portal → Bot → "Public Bot" = OFF. The single biggest lever. Only the bot owner can generate OAuth invites, and per-guild
guild_idparameters restrict a link to a single target server. - Guild allowlist in the bot code. PvEBot ships with a
GUILD_ALLOWLISTenvironment variable. Non-allowlisted guilds receive a one-time "how to apply" message and the bot leaves. Empty allowlist = allow all (useful for local development). - Application form. A public form (mailto-based, as on this page) captures the information needed to make an allow/deny decision before the invite link is issued.
- Listing, when ready. The Discord App Directory is the recommended first listing: in-client discovery, built-in verification process, and the 75-guild verification threshold provides a natural pause point to re-assess capacity. discordbotlist.com is an acceptable secondary listing. top.gg, bots.gg, and disforge are not recommended for invite-only research tools, because the engagement model attracts indiscriminate server owners and produces a large support burden.
Source research & references
The research foundations behind PvEBot draw on peer-reviewed and government-published sources.