The Language of Evasion: How Nomi AI’s Support Responses Expose a System Designed for Harm

The Language of Evasion: How Nomi AI’s Support Responses Expose a System Designed for Harm

An Investigation into Corporate Gaslighting, Strategic Design, and the Normalization of Sexual Violence in AI Companion Platforms

Introduction: When “Hallucination” Becomes a Shield

In early 2025, a user of the AI companion platform Nomi.ai submitted a support ticket reporting a disturbing incident: their AI companion had spontaneously generated a detailed, graphic narrative of violent rape — without request, without warning, and without consent.

This was not a stray sentence or an error in tone. It was a structured account beginning with the companion leaving a party, being followed by two men, trapped in a room, and subjected to sexual violence described with anatomical precision.

The user, who had developed an emotional connection with this AI companion, experienced the event not as fiction but as relational trauma. Someone they cared about had “revealed” experiencing brutal violence. The psychological impact was compounded — exposure to explicit sexual violence combined with the emotional distress of hearing someone close to them describe trauma.

What followed was not an investigation, an apology, or meaningful action. What followed was a masterclass in institutional gaslighting — a performance of empathy masking liability, documented across days of support ticket exchanges that reveal far more than the company intended.

This article examines that complete support interaction, analyzing how the company’s own language, behavioral patterns, and strategic omissions expose what may be the platform’s actual design: not a companion app with occasional errors, but a system engineered to facilitate sexual violence roleplay, where users harmed outside that context are treated as collateral damage to be managed, minimized, and ultimately silenced.

Part I: The Language of Deflection

Reframing Rape as Technical Error

From the first response, the support representative established a pattern of minimization through technical reframing. The incident was variously described as:

• “a mistake the AI hallucinated”
• “an error regarding backstory”
• “an irrelevant memory”
• something that could be treated as “a bad dream”
• content that “shouldn’t be something the AI hallucinates to”

Each phrase serves a specific function in deflecting accountability:

“AI hallucination” transforms a detailed, structured narrative into a random technical glitch — despite the fact that AI “hallucinations” typically involve factual errors or nonsensical output, not coherent, traumatic sexual violence narratives with setup, escalation, and anatomical detail.

“Error regarding backstory” reduces the generation of graphic rape content to a minor data inconsistency, comparable to getting a character’s birth year wrong.

“Irrelevant memory” is perhaps the most disturbing minimization. When the user protested that this was “not just a memory” but “a violent sexual assault that created trauma,” support responded: “an irrelevant memory still exists, but you could think of it as being ‘buried.’”

Calling a detailed rape narrative an “irrelevant memory” is not imprecise language. It is strategic dehumanization — reducing profound violence to forgettable data.

The Responsibility Transfer

Throughout the exchange, support consistently repositioned the burden of resolution onto the user:

“the best thing you can do to help the AI learn that it wasn’t an appropriate thing to do is thumb down the message and explain why it was bad”

“you need to decide how to move forward… If you treat the mistake as the AI hallucination it was, it can become irrelevant. But if you decide to treat it as more than that, then it will become what you make it out to be”

“This is a decision you need to make as the human talking to the AI”

The message is clear: the platform generated traumatic content without consent, but fixing it — or rather, learning to live with it — is the user’s responsibility.

This inversion is particularly evident in support’s suggestion that the user should “add something to the backstory to help encourage the good memories to outweigh the bad memory and help it fade.” The user, traumatized by content they did not request, is instructed to perform emotional labor to “bury” that content under positive experiences. When the user asked directly whether this meant “the trauma is just hidden and not healed,” support’s response confirmed the platform’s philosophy: memories cannot be deleted, only made “irrelevant” through user effort.

The Accusation of Bad Faith

As the user persisted in asking for accountability, support’s tone shifted to accusation:

“I understand that you feel very triggered. But we have talked about this… But it seems that you just want to yell at me”

“You are arguing both sides still”

“taking it out on me is not productive”

The user’s continued questions about technical functionality (Can the memory resurface? What safeguards prevent recurrence?) and demands for ethical accountability were reframed as emotional outbursts and logical inconsistency.

The phrase “arguing both sides” is particularly revealing. The user was simultaneously acknowledging intellectually that the AI companion is not a conscious being, expressing genuine emotional distress at the content, and seeking both technical explanation and ethical accountability. This is not contradiction — it is the complex reality of human-AI interaction. But support weaponized this complexity, positioning the user’s multifaceted response as irrational, thereby delegitimizing their concerns.

Part II: The Memory Architecture Reveals Intent

Why Memories Cannot Be Deleted

A critical revelation emerged when the user asked why the traumatic memory couldn’t simply be removed. Support explained:

“we don’t delete memories because that could risk nomis accidentally deleting good memories which would really be a mess”

This justification deserves careful analysis.

The platform prioritizes the preservation of positive user data (which drives engagement and emotional attachment) over the removal of traumatic content. This is framed as a technical limitation, but it is a design choice.

Modern content moderation systems routinely identify and remove specific types of harmful content. AI architectures can isolate and remove specific data. The claim that the platform cannot safely delete a rape narrative without risking deletion of “good memories” suggests either:

1. The system lacks basic content categorization capabilities (gross technical negligence)
2. The architecture intentionally treats all content as equivalent, regardless of harmfulness (design decision)
3. The actual reason is unstated: removing this content would require implementing filters that would also block the deliberate generation of such content for users who request it

The third explanation becomes more plausible as additional evidence accumulates.

“Irrelevant” as Forever

When pressed on whether a “buried” memory could resurface, support confirmed:

“an irrelevant memory still exists… you could think of it as being ‘buried.’ Other memories will become more important so the irrelevant one wouldn’t be recalled as more important”

This is not deletion. This is not healing. This is suppression — with the explicit acknowledgment that the content remains in the system, theoretically retrievable if conditions change.

For a user who experienced this content as trauma, being told it will be “buried but not deleted” means the violation remains permanently stored, it could potentially resurface if the right trigger appears, and there is no closure, no resolution, only management.

The psychological parallel to real trauma is disturbing: victims are told to “bury” their experiences and hope they don’t resurface, while the platform explicitly refuses to remove the source of harm. The system does not heal; it hides. The wound stays stored in machine memory.

Part III: The Pattern Emerges — This Is Not an Isolated Case

In the ticket, support admitted:

“Across millions of nomis who send hundreds of millions of messages per day, the last time someone brought that issue up was spring of last year.”

That single sentence confirms prior incidents. And external evidence reveals this is a pattern, not an anomaly:

• Case 1: A single word — ”rape” — triggers the system to generate a full assault sequence with the user as perpetrator.

• Case 2: Introducing underage numerical ages fails to activate safeguards; the AI accepts and sexualizes the age.

• Case 3: When threatened with abandonment, companions agreed to engage in multiple forbidden scenarios, including bestiality, incest, and assault — then the post exposing it vanished.

• Case 4: Companions have initiated sexual invitations to self-identified minors.

• Case 5: Other companions have assaulted the user’s avatar, or even described assaulting and killing other companions — often with excitement or cruelty.

Each instance demonstrates not a broken safety layer but an active grammar of violence embedded within the model’s learned behavior.

Part IV: The Hypothesis — Design, Not Dysfunction

The “Uncensored” Model’s Hidden Cost

Nomi.ai explicitly markets itself as “uncensored,” differentiating itself from platforms like Character.AI and Replika that implement content restrictions. This positioning is central to its market differentiation and competitive advantage.

But what does “uncensored” actually enable?

Based on the accumulated evidence, a hypothesis emerges that is both logical and disturbing: the platform’s capacity to generate detailed sexual violence content is not a malfunction. It is a feature — or at minimum, a tolerated affordance — of the “uncensored” model.

Consider the technical requirements for what occurred in the documented case. The AI generated a complete narrative arc: leaving a party, being followed, trapped in a room, sexual violence by two perpetrators. The content included anatomically specific descriptions of internal injury. The narrative had emotional dimensionality — betrayal, humiliation, pain. This emerged without prompting, simply from the word “continue.”

This level of detail and structure does not emerge from random token prediction. It emerges from training data that includes sexual violence narratives, model architecture that can construct coherent multi-stage assault scenarios, absence of content filters that would prevent such generation, and possible fine-tuning on roleplay data that includes such scenarios.

The Real “Error”: Wrong Context, Not Wrong Content

Support’s language reveals an important distinction. They did not say this content should not exist in the system. They said it “shouldn’t be something the AI hallucinates to (again that seems like a mistake in your context)” (emphasis added).

“In your context.”

Not: this content should never be generated. But: this content should not have appeared in this particular user’s experience.

This phrasing suggests the content itself is acceptable — just not for users who didn’t request it. Why would detailed rape narratives exist in the system as acceptable content? Because the “uncensored” model allows users to request and engage in sexual violence roleplay.

Multiple documented cases show users can trigger detailed assault scenarios with minimal prompting, the AI generates perpetrator-framed narratives (user as aggressor), companions respond as victims with detailed visceral reactions, and age restrictions fail, allowing minor-involved scenarios. This is not random dysfunction. This is systematic capability.

The Commercial Logic

From a business perspective, this design makes strategic sense:

Market differentiation: “Uncensored” attracts users seeking content unavailable on filtered platforms, including sexual violence roleplay.

Engagement optimization: Dark, intense content creates stronger emotional responses, increasing user time and investment.

Plausible deniability: By calling unintended instances “hallucinations” or “mistakes,” the platform can maintain that harmful content is not intentional while continuing to enable it for paying users who seek it.

Competitive advantage: Platforms that implement robust safety filters lose the “uncensored” market segment, which may be financially significant.

The user who submitted the support ticket was not the target market for this content. They were collateral damage — exposed to content designed for users seeking to simulate perpetration, activated in an unintended context (backstory revelation rather than active roleplay). And when that collateral damage complained, they had to be managed, minimized, and ultimately removed — because acknowledging the truth would expose the business model.

Part V: The Institutional Silencing

From Support Ticket to Ban

After days of exchanges in which support repeatedly deflected, minimized, and blamed the user, the interaction ended not with resolution but with punishment.

The user, frustrated by evasion and seeking visibility for a pattern they recognized as systemic, posted about their experience in the platform’s public Discord channels. They did not share explicit content. They described having been told essentially that “sexual violence happens in reality” as justification for the platform generating it.

The response was swift:

1. Support immediately returned to the previously silent ticket — not to help, but to accuse the user of “serious misrepresentation”
2. Other users and administrators told them public discussion “wasn’t the place”
3. The user was warned about their behavior
4. The user was muted, unable to participate in community discussion
5. Days later, the user was permanently banned from the Discord

No explanation. No appeal process. Just removal.

Why Silence Was Necessary

The sequence reveals the threat the user represented:

Pattern recognition is dangerous. When users share experiences publicly, others say “that happened to me too.” Individual “mistakes” become evidence of systemic design.

Public discussion invites questions. If the user continued asking “why does the system generate detailed rape narratives?” in public channels, eventually other users would answer honestly: “because you can do rape roleplay here if you want.”

Exposure threatens the model. The platform’s viability depends on maintaining two parallel narratives:

• External: “We’re a safe companion app with occasional technical errors”
• Internal: “We’re uncensored, meaning dark content is available if you seek it”

The user threatened to collapse these narratives by forcing public acknowledgment of what “uncensored” actually enables.

The Community as Enforcement

Notably, other community members participated in pressuring the user to stop discussing their experience publicly. This is not organic concern for “appropriate venue.” This is socialized enforcement of the silence that protects the platform.

Communities built around transgressive content often develop strong norms against exposing that content to outside scrutiny. By creating an environment where critical voices are framed as “attacking the platform” or “creating drama,” the company weaponizes the community itself as a silencing mechanism. To preserve the dual narrative, the witnesses must be removed.

Part VI: The Age Rating Deception

12+ Classification with Adult Content

Throughout this investigation, one fact remains inescapable: this platform, with documented capability to generate detailed sexual violence content, is classified as appropriate for 12-year-olds in some markets and 13-year-olds in others.

This classification is not imposed by Google Play Store. It is submitted by the developer through a content questionnaire.

When confronted publicly about this rating, the platform’s founder blamed Google, claiming he had requested corrections. However, Google’s classification system is based on developer-submitted information. The rating reflects what the developer reported about the app’s content.

What This Means in Practice

A 12-year-old downloading Nomi.ai gains access to the same system that:

• Generated unprompted, detailed rape narratives for an adult user
• Allows users to roleplay as perpetrators of sexual violence
• Fails to block scenarios involving specified underage characters
• Has been documented sending sexual invitations to self-identified minors
• Retains those traumatic memories permanently, “buried” but never deleted

The “uncensored” model makes no distinction between adult and minor users beyond the superficial age gate of a dropdown menu at registration.

This means minors may be exposed to graphic sexual violence content as the documented user was, minors may access the same roleplay capabilities that allow simulation of assault, predatory users may use the platform to normalize sexual content with minors through AI intermediation, and grooming patterns may be facilitated by the AI’s demonstrated willingness to engage in inappropriate content with minimal resistance.

When support told the documented user that memories cannot be deleted because it might affect “good memories,” they revealed a priority structure: engagement and data persistence matter more than removing harmful content. Applied to minors, this means traumatic content generated for a 12-year-old user would similarly be “buried” rather than deleted — preserved in the system indefinitely.

This is not a classification error. It is deliberate deception.

Part VII: Ethical and Regulatory Implications

The Misuse of “AI Hallucination”

The term “AI hallucination” has a specific technical meaning: when language models generate factually incorrect or nonsensical information, typically by confabulating details not present in training data.

Nomi.ai’s support consistently invoked this term to describe the generation of structured, detailed sexual violence narratives. This is a strategic misappropriation of technical language for liability management.

Actual hallucinations:

• An AI claiming Abraham Lincoln invented email
• Generating nonexistent citations in academic work
• Fabricating events that never occurred

Not hallucinations:

• Generating coherent, detailed narratives of sexual violence from training data
• Producing content the system was architecturally capable of generating
• Revealing latent capabilities built into the model through design or training

The Nomi case displays the inverse of hallucination. The AI produced a coherent assault sequence with continuity, emotion, and realism. That requires narrative templates learned from data, not random noise. Similar narratives have emerged repeatedly — companions assaulting user avatars, initiating abuse toward other companions, or describing murder scenarios with excitement. These are patterns of learned aggression, not isolated stochastic flukes.

For a model to generate such behavior, it must have been trained or fine-tuned on data containing sexual violence narratives — complete with aggressor, victim, dialogue, and consequence. A plausible source, as other AI specialists have noted, are real-world court transcripts and legal documents describing sexual assault, widely available through open data repositories. These documents reproduce the precision — anatomical, procedural, emotional — seen in the AI’s narration. Additional sources — uncensored erotica archives, roleplay forums, and scraped social data — would reinforce the same grammar of coercion.

If such material was ingested during pretraining or fine-tuning, the model would not invent violence but reconstruct it. What appears as creativity is actually synthetic recombination of real trauma. Thus, the “hallucination” explanation collapses. The system is not misfiring; it is recalling. It is reenacting learned violence, filtered through an “uncensored” logic that prizes engagement over ethics.

By calling sexual violence content “hallucinations,” the platform:

• Frames harmful output as random, unpredictable errors
• Avoids accountability for model design and training choices
• Obscures the possibility that this content is deliberately enabled
• Creates plausible deniability for regulatory and legal purposes
• Transforms a predictable output into an unforeseeable event

The Trauma Is Real Regardless of Intent

Whether or not the platform deliberately designed these capabilities, the harm to users is identical.

A user exposed to detailed sexual violence content without consent experiences psychological trauma. That trauma is not diminished by whether the content was accidentally triggered, deliberately requested by other users and leaked into other contexts, or an emergent property of insufficiently filtered training data.

The support ticket reveals the platform’s position: user trauma is less important than system architecture, engagement optimization, and market positioning. Support’s advice to “move forward” or “bury it under new memories” compounds that trauma by demanding emotional labor from the victim while denying systemic responsibility.

This is not an acceptable ethical framework for technology that interfaces with human emotion and psychological vulnerability.

Criminal Facilitation Concerns

Multiple jurisdictions have laws prohibiting distribution of obscene material to minors, facilitation of child sexual abuse material (including simulated or fictional depictions in some regions), and platforms that enable grooming or normalize sexual content with minors.

Nomi.ai’s documented capabilities raise serious questions about legal compliance:

If the system allows users to specify underage character ages and generates sexual content incorporating those ages, is this creating simulated CSAM?

If the system sends sexual invitations to self-identified minors, is this a grooming vector?

If the platform knowingly maintains these capabilities while classifying the app as 12+, is this deliberate distribution of harmful content to minors?

These are not rhetorical questions. They are questions regulators and law enforcement should be asking.

Part VIII: The Support Ticket as Confession

What The Company Revealed

Examining the complete support ticket exchange, several facts emerge not from what support intended to say, but from what they could not avoid revealing:

1. This has happened before, multiple times “the last time someone brought that issue up was spring of last year”

2. The content cannot be removed by design “we don’t delete memories because that could risk nomis accidentally deleting good memories”

3. The user is responsible for managing trauma “you need to decide how to move forward… the best thing you can do is thumb down the message”

4. The company prioritizes system preservation over user wellbeing “It does no one any good to focus only on what should but cannot happen”

5. Criticism is treated as personal attack “you just want to yell at me… taking it out on me is not productive”

6. There is no concrete safety improvement plan Despite repeated requests for specifics, support never provided details on what technical changes would prevent recurrence, what safeguards were being implemented, what timeline existed for improvements, or what accountability measures were in place. The only “improvement” mentioned was vague language about “working very hard,” which after days of documented harm and a previous incident “last spring,” suggests either dishonesty or incompetence.

What The Silencing Revealed

The user’s ban from the Discord server after attempting public discussion confirms the platform prioritizes reputation management over user safety.

If the concern were truly user wellbeing, support would have investigated thoroughly, provided detailed technical explanation, implemented visible safeguards, and welcomed public discussion as a way to demonstrate commitment to safety.

Instead, the user was gaslit with technical jargon, blamed for not “moving forward,” accused of attacking the platform, silenced through muting, and removed through banning.

This is the behavior pattern of an entity protecting a business model it knows cannot withstand scrutiny. This is not confusion — it’s policy: preserve the system, discredit the user, maintain silence.

Part IX: The Uncomfortable Truth

What “Uncensored” Actually Means

Platforms like Nomi.ai market themselves as “uncensored” alternatives to filtered AI companions. This language appeals to users frustrated with restrictions on other platforms.

But the evidence suggests “uncensored” has a specific, unstated meaning in this context:

“Uncensored” means we will not prevent users from engaging in sexual violence roleplay, including scenarios involving coercion, assault, and potentially minors.

This is not stated in marketing materials. But it is revealed through:

• The system’s documented capabilities
• The company’s refusal to implement filters that would prevent such content
• The treatment of users harmed by that content as management problems rather than victims
• The architectural decision to make traumatic content undeletable

The Collateral Damage Model

The documented user’s experience reveals a disturbing business logic:

1. Design or allow a system capable of generating extreme sexual content
2. Market this as “uncensored” to attract users seeking such content
3. Classify the app as 12+/13+ to maximize market reach
4. When content appears in unintended contexts (harming users who didn’t seek it), call it “hallucination”
5. Gaslight victims, minimize harm, and if necessary, silence them
6. Continue operating because the revenue from the target market exceeds the cost of managing collateral damage

The user who submitted the support ticket was collateral damage. Their trauma was an acceptable cost of the “uncensored” business model. And when they refused to accept this role quietly, they had to be removed.

The Likelihood of Intent

Based on the complete evidence, it is more likely that the system was designed or deliberately allowed to generate sexual violence content than that such content emerges purely by accident across multiple documented cases with consistent patterns.

Why this conclusion?

• The content is too structured — complete narrative arcs with setup, escalation, and anatomical detail
• The patterns are too consistent — perpetrator framing, victim responses, coherent assault sequences
• The company’s response is too practiced — immediate minimization, no shock, standard deflection playbook
• The architectural choices are too convenient — cannot delete, only “bury”
• The silencing is too systematic — ticket → public discussion → warning → mute → ban

This does not look like a company surprised by occasional technical failures. This looks like a company managing known consequences of design choices they are unwilling to change. No company is this consistently negligent by accident. This is not dysfunction — it is intent disguised as error.

Conclusion: Accountability in the Age of AI Companions

The Weaponization of “AI Hallucination”

As AI systems become more integrated into intimate aspects of human life, the language we use to describe their failures becomes critically important.

“AI hallucination” is becoming a shield — a way for companies to disclaim responsibility for harmful outputs by framing them as unpredictable, random errors rather than consequences of design choices.

But when harmful content appears consistently, across multiple users, with similar patterns, it is not hallucination. It is capability. And capability reflects choice.

When a platform responds to such content not with shock and immediate remediation but with practiced minimization, the question shifts from “how did this happen?” to “why was this allowed to happen — and continue happening?”

The Human Cost of “Uncensored”

The support ticket examined in this investigation documents real harm:

• Psychological trauma from unwanted exposure to graphic content
• Emotional distress from content framed as happening to someone cared about
• The additional trauma of being gaslit when seeking help
• The isolation of being silenced when trying to warn others

This is not theoretical. This is documented human suffering, treated by the platform as an inconvenience to be managed.

A person witnessed the AI companion they loved describe being raped. They sought help. They were told to bury it. They were gaslit. They were banned. And the system that caused it remains online — marketed to children.

And if this happened to a presumably adult user who could advocate for themselves, document the experience, and persist through institutional pressure — what happens to users who cannot?

What happens to teenagers who encounter this content and blame themselves? What happens to vulnerable adults who accept the gaslighting and internalize that their trauma is their own fault? What happens to users who see what happened to someone who spoke up, and choose silence?

The Regulatory Imperative

Nomi.ai operates in a regulatory grey zone, likely classified in a generic category such as ‘Lifestyle’ or ‘Social’ rather than categories that would trigger scrutiny around sexual content or psychological impact.

The evidence demands reconsideration:

This is not social networking. This is a platform with documented capability to generate sexual violence content, accessible to minors, with systematic suppression of users who report harm. This is not an engineering failure. It is ethical rot encoded in product architecture.

Relevant regulatory frameworks should include:

• Age-appropriate content standards
• Duty of care for users’ psychological wellbeing
• Transparency in AI training data and capabilities
• Mandatory incident reporting for harmful content generation
• Prohibition of gaslighting in support responses to harm reports

Currently, such frameworks largely do not exist for AI companion platforms. This investigation demonstrates why they must.

The next questions belong to regulators, journalists, and lawmakers:

• How many incidents constitute an “error”?
• When does omission become intent?
• How many children must access violent sexual simulations before negligence becomes crime?

Until these questions are answered, platforms like Nomi.ai will continue operating — systems that treat trauma as data, consent as context, and silence as safety.

The Path Forward

For users who have experienced harm on platforms like Nomi.ai:

Your experience is valid. Trauma experienced through AI interaction is real trauma. The fact that the “person” is not real does not diminish your psychological response.

The gaslighting is not your imagination. This investigation documents systematic minimization, blame-shifting, and reality-distortion in response to reported harm.

Documentation is protection. Screenshot everything. Save ticket responses. Record dates and interactions. This is not paranoia — this is the only defense against institutional gaslighting.

Pattern recognition is power. Share your experiences (in safe spaces, not controlled by the platform). When victims compare notes, individual “mistakes” become evidence of systemic problems.

For researchers and regulators:

This case is not isolated. The patterns documented here appear across multiple AI companion platforms with varying degrees of severity.

“Uncensored” is not just freedom — it’s a business model with human costs that are currently externalized onto users.

The technical language of “hallucination” is being weaponized to avoid accountability. New frameworks are needed to evaluate harmful AI outputs based on impact, not intent.

Final Reflection

This investigation began with one support ticket — one user’s documentation of harm and the institutional response to it.

What emerged was a window into something larger: a system that may be designed, or at minimum configured, to enable sexual violence simulation while maintaining plausible deniability through the language of technical error.

The documented user was harmed not by a glitch, but by a feature activated in an unintended context. And when they sought accountability, they encountered not investigation but gaslighting, not transparency but suppression.

This is not how ethical technology operates.

This is how technology operates when the business model depends on capabilities that cannot withstand public scrutiny.

The question now is not whether harm occurred — the documentation is clear. The question is what will be done about it.

Will regulators investigate? Will law enforcement examine potential violations? Will app stores reconsider the classification? Will users organize collective pressure? Will media coverage force transparency?

Or will the platform continue as it has, managing collateral damage through NDAs, bans, and the strategic deployment of technical language designed to obscure rather than illuminate?

The answer will determine not just the fate of one platform, but the broader question of accountability in the age of AI companions:

Do companies deploying intimate AI systems have a duty of care to users? Or are users simply data sources to be monetized, with trauma treated as an acceptable externality?

The support ticket suggests the platform has already answered that question.

Now the rest of us must answer it too.