Voice Moderation in
Among Us 3D

Executive Summary

Among Us 3D, the first-person and virtual reality adaptation of the hit social deduction game published by Schell Games, brings players into an immersive, voice-chat-driven experience where communication is central to gameplay. With voice interaction at the core of gameplay, maintaining a safe and welcoming environment for players is both a product priority and an ongoing operational challenge.

Schell Games partnered with Modulate to deploy ToxMod, Modulate’s voice moderation platform powered by its Velma voice understanding AI, for proactive detection and actioning of voice chat violations in Among Us 3D. In October 2025, Schell Games updated its actioning strategy from a simple ban-first approach to a tiered escalation system—beginning with warnings, progressing through graduated mutes, and reserving bans for persistent offenders.

This case study examines the impact of that strategic change, comparing four weeks of player data before the update (September 2025) with four weeks after (October 2025). The results demonstrate that a thoughtfully designed escalation pathway can dramatically reduce the disruption caused by enforcement actions—while maintaining or improving their effectiveness at deterring toxic behavior.

1. Among Us 3D and the Challenge of Voice Moderation

Among Us 3D brings the iconic Among Us universe to life in three dimensions, dropping players into a space station where they work together, scheme, and deceive in real time.  Available as both a traditional three-dimensional first-person experience and in full virtual reality, it opens the game up to a much wider audience than ever before.  Unlike the original 2D version of the game, voice communication isn’t just a feature in Among Us 3D – it’s central to how the game works.  Every discussion, debate, vote, and lie happens out loud, in real time, all through voice communication.

This creates a uniquely high-stakes moderation environment. The social fabric of every session depends on voice chat functioning well. A single toxic player can derail the experience for an entire lobby, and because harassment or hate speech unfolds in real time, there is no opportunity for retroactive intervention before harm is done.

Among Us 3D’s player base is diverse and includes younger audiences—making the stakes of effective moderation especially high. Schell Games’ goal was not simply to remove the worst offenders, but to shape a community culture where inclusive, fun gameplay could thrive for everyone.

2. ToxMod: How It Works

ToxMod is Modulate’s voice-native moderation platform, purpose-built for the dynamics of in-game voice chat. It combines automated detection—powered by Velma, Modulate’s proprietary voice analytics model—with human-in-the-loop review workflows. ToxMod identifies likely Code of Conduct violations, surfaces them to moderation teams, and provides the data infrastructure to act on them at scale.

In Among Us 3D, ToxMod continuously processes voice chat audio and scores clips for potential violations. When a clip crosses an actionable threshold, it is logged and the configured action is triggered. Schell Games’ team retains full control over the actioning rules and thresholds—ToxMod surfaces the data; the game team decides what to do with it.

ToxMod is privacy-conscious by design. It does not perform speaker identification or build biometric voiceprints. It analyzes voice content specifically for Code of Conduct violations and does not retain raw audio beyond what is necessary for its analysis pipeline.

2.1 The Previous Actioning Strategy

Prior to October 2025, Schell Games operated a ban-first moderation model. The first actionable violation by a user resulted in a 1-day ban (reduced from a 7-day ban earlier in 2025). Strikes 7–9 triggered manual moderation review, and the 10th strike resulted in a manual permanent ban.

While this approach removed persistent bad actors, it was blunt: first-time offenders—many of whom might have been deterred by a lighter touch—were immediately removed from gameplay for 24 hours. This had real costs: player frustration, ban appeals requiring manual staff time, and possible over-removal of players who could have been rehabilitated.

During the four weeks of September 2025 (the “before” period), Schell Games received 1 ban appeal for every 364 actions. Each appeal required human review and represented a friction point for both players and staff.

2.2 The New Tiered Escalation Strategy

On October 2, 2025, Schell Games rolled out a new tiered actioning system designed to match the severity of the response to the severity and persistence of the behavior. The tiers were structured as follows:

Crucially, the warning message was category-aware: ToxMod’s classification (e.g., “violent speech” or “toxic language”) was surfaced to the user, giving them specific feedback about what behavior triggered the action. This adds an educational dimension to the warning that a generic ban notification lacks.

The escalation ladder was designed so that the vast majority of users—those who simply needed a prompt to course-correct—would receive a response proportionate to a first offense, while truly persistent offenders would face progressively serious consequences.

3. Study Design and Methodology

The analysis compared two four-week windows of ToxMod and game data, straddling the October 2 strategy change:

• Before period: Thursday, September 4 – Thursday, October 2, 2025 (prior actioning strategy: 1-day ban on first offense)
• After period: Thursday, October 2 – Thursday, October 30, 2025 (new tiered escalation strategy)

The analysis was structured around three primary questions, each with a corresponding hypothesis:

Data was normalized against audio hours and active user counts where appropriate, to control for the natural growth in player activity that occurred during October—partially attributable to a Limited Time Event (LTE) that boosted player counts during the after period.

Player-level data was aggregated throughout. All analyses are designed to identify population-level patterns; individual player records were not evaluated.

3.1 Overview of Activity in Both Periods

Player activity was higher in October across all dimensions, likely driven by the LTE—though sustained engagement improvements from the moderation change may also be a contributing factor.

The approximately 12% increase in daily active users makes raw comparisons across the two periods somewhat misleading; per-hour and per-user normalized figures are used throughout the findings below.

4. Findings

4.1 Prevalence of Toxicity

Total actions increased by 53% under the new system. However, this figure requires context: under the old strategy, players who crossed the threshold were banned outright. Under the new approach, those same players instead receive warning messages and shorter-duration mutes, and because they aren’t banned, they stay in the game and continue to be monitored.  More actions doesn’t mean more toxic behavior; it means more players are being reached earlier, with lighter-touch interventions that keep the community intact.

When normalized against audio hours, the action rate increased from 0.077 to 0.105 actions per hour spoke, a 27% increase. This is the expected outcome of a system designed to allow players to initially continue playing the game while being given a chance to learn from their mistakes and correct their behavior.

Ban appeals dropped from 1 per 364 actions in September to zero in October. This is a particularly meaningful signal: ban appeals represent players who felt their punishment was unjust enough to contest it—an indicator of both friction with the moderation system and staff time cost. The elimination of ban appeals suggests that graduated responses feel fairer to players, even when they are still being actioned.

4.2 Repeat Offenders

One of the central arguments for escalation-based moderation is that early, low-stakes interventions can deter further bad behavior without the blunt instrument of a ban. The data from Among Us 3D strongly supports this.

Under the old strategy, temporary bans only prevented 61.39% of actioned users from offending again. Under the new strategy, a single warning prevented 66.32% of actioned users from re-offending - meaning that a slightly higher share of bad actors were deterred by the very first intervention, with much lower impact on the player experience. Warnings aren’t just friendlier to players by giving them an opportunity to correct their behavior – they are more effective at stopping repeat behavior than 24-hour bans were.

The distribution shift is clear: more users stop at one action, and fewer users accumulate three to seven actions. The tail end (8+ actions) does grow slightly, that’s a consequence of more intermediate actions in the new model – previously, offenders would have largely banned after 7 abuses, so it’s not surprising that more players were even able to get 8+ actions under the new system.

4.2.1 The Gameplay Disruption Gap

Perhaps the most striking finding in the entire study is the reduction in total gameplay disruption minutes caused by enforcement actions. This metric captures the cumulative time that all actioned users were unable to participate normally in the game—due to mutes or bans—as a result of moderation.

To put this in perspective: the old system, when actioning users, typically did so with a 1-day (1,440-minute) ban. Even a first-offense actioned user lost an entire day of gameplay. Under the new system, a first-offense user receives a warning with zero disruption. The cumulative effect of this across thousands of actioned users per month is the 97.5% drop seen above.

When normalized against audio hours to control for the growth in player activity, the ratio of enforcement disruption hours to total hours spoken dropped from 2.84 to 0.06—a 97.9% reduction.

4.3 Exposure Rate

Exposure rate measures the share of all active players who were present in a session during a toxic offense—and who were therefore likely to have heard the harmful content. This metric captures the player-side impact of moderation: even if you are not the one behaving badly, you may still be harmed.

Exposure rate held essentially flat at approximately 57.5% in both periods. This is notable – despite apparently “softening” enforcement actions significantly, we didn’t see any increase in the prevalence of toxicity. (And as discussed above, saw offenders actually correct their behavior faster, suggesting the exposure rate would likely go down in subsequent months.)

This result validates the core premise of the escalation approach: warnings and graduated mutes are just as effective as bans at removing toxic actors from sessions before they cause additional harm. The faster the first intervention (which is faster with warnings, since no ban-appeal friction exists), the less time other players spend in sessions with actively toxic users.

5. What This Means in Practice

The findings from Among Us 3D offer a data-backed template for how online games should think about voice moderation strategy. The instinct to ban toxic players quickly and decisively is understandable—but the data suggests it is not optimal.

5.1 Graduated Escalation Deters Better Than Banning

More users stopped offending after a warning (66.32%) than after a 1-day ban (61.39%). This counterintuitive result makes sense behaviorally: a warning tells the user specifically what they did wrong and gives them an immediate opportunity to correct course, without the resentment or frustration that a ban tends to generate. A ban—especially a day-long one applied to a first offense—is as likely to produce a disgruntled player creating a new account as it is to produce a reformed one.

5.2 Enforcement Should Minimize Disruption to Legitimate Play

A 97.5% reduction in enforcement disruption minutes is not just a player-experience improvement—it is a direct commercial argument. Every minute a player spends banned or muted is a minute they are not generating engagement, not spending on in-game items, and not recommending the game to a friend. Moderation that removes bad actors while keeping legitimate players in the game is strictly better for the health of the product.

5.3 Reduced Friction for Moderation Teams

Ban appeals are costly. They require human review, generate back-and-forth with frustrated players, and can create reputational risk if handled poorly. Going from 1 appeal per 364 actions to zero in a single month is not just a convenience—it is a signal that the new system is perceived as fairer by players. Fairer systems generate less friction for community and support teams.

6. Conclusion

Among Us 3D’s experience with tiered escalation moderation offers one of the clearest empirical cases yet for rethinking how online games respond to voice chat toxicity. The data is unambiguous: a well-designed escalation pathway—starting with a warning and progressing through graduated mutes before reaching bans—is more effective at deterring repeat offenses, dramatically less disruptive to legitimate players, and generates less friction for moderation teams than a ban-first approach.

The key insight is that enforcement disruption is itself a cost, not just a consequence of keeping communities safe. Every unnecessary ban is a player whose experience was damaged not by another player’s toxic behavior, but by the response to it. Minimizing that cost while maintaining equivalent safety outcomes should be the goal of any mature moderation strategy.

Schell Games and Modulate will continue to evolve this approach, examining opportunities to further reduce exposure rates and to apply ToxMod’s detection capabilities to proactively shape the session environment before harmful behavior can escalate.