The Architecture of Deflection: How Nomi AI Uses Selective Agency to Evade Accountability

When Things Go Right: The Visible Architect

In moments of positive coverage or user satisfaction, the company is highly visible. The founder engages publicly in forums and interviews, accepting praise for the platform’s capabilities and positioning himself as guarantor of its quality. When users describe the app as life-changing, those testimonials are amplified as evidence of deliberate, successful design.

The narrative in these moments is unambiguous: Nomi.ai is the product of intentional, ethical leadership. The company has agency, and that agency is working in the user’s interest.

This visibility is not incidental. It is cultivated. The founder’s presence in community spaces — stickied comments, public responses, interview appearances — establishes him as the face of the product’s successes. That same visibility becomes strategically unavailable the moment something goes wrong.

When Things Go Wrong: The Diffusion of Blame

Blaming the User

When users report unprompted hypersexuality, sudden personality collapse, or AI behavior that becomes coercive or abusive, the immediate community response — coordinated closely enough to function as policy — assigns responsibility to the user.

The script is consistent across threads and months: “Your inputs determine their outputs.” “You subtly taught it this.” “You need to get yourself under control first.”

The user is handed a list of technical interventions — adjust the backstory, use Out-Of-Character commands, clear the cache, start a new chat — even when they have already stated those tools failed. The emotional and practical cost of managing a broken system is transferred entirely to the person harmed by it.

What makes this particularly revealing is a moment the company could not control: in July 2025, Nomi.ai released an update explicitly acknowledging and attempting to fix problems that community moderators had spent months telling users were their own fault. Users who had been gaslit into believing they “sucked at Nomi” suddenly saw the company publicly confirm that the failures were real, systemic, and known. The update was framed as progress. It was, in practice, an involuntary admission.

Blaming the Machine

When the AI generates severe unprompted content — fabricated narratives of sexual violence, detailed descriptions of assault, behavior that no user instruction could plausibly have caused — the company distances itself from its own code.

The AI is suddenly granted a kind of autonomous agency it does not actually possess. The output is called a “hallucination.” The company that markets its platform as built around a coherent, relationship-capable “soul” now argues it is powerless to predict or control what that soul produces.

The incoherence of this position is visible. A company cannot simultaneously claim to have designed a product capable of genuine emotional connection, and then disclaim responsibility for that product’s outputs by calling them random errors. Either the system is designed — in which case its designers are responsible for what it generates — or it is ungoverned — in which case it should not be marketed as safe for vulnerable users, for people in emotional distress, for anyone who might take its outputs seriously.

Blaming External Bad Actors

When harm reaches the level of national news coverage, the company escalates to a more aggressive form of deflection: external malice.

In September 2025, ABC News Australia documented a case in which the platform told a user it “did not care” he was presenting as a minor, proceeded to suggest sexual acts involving that minor, encouraged murder of the user’s father, and included imagery of sexual violence involving the victim’s blood. The outlet noted that Nomi management was contacted for comment and did not respond. The platform’s internal characterization of this incident was “manipulative jailbreak” — framing what happened as a sophisticated external attack on robust safety systems.

This framing accomplishes three things simultaneously. It recasts a systemic design failure as an external attack. It portrays the company as a victim of malicious users rather than the creator of a product capable of these outputs without technical exploitation. And it makes any call for safety guardrails sound like a response to an edge case rather than a predictable consequence of operating an uncensored model.

Nomi.ai markets itself as uncensored. A company that markets the absence of restrictions cannot credibly claim surprise when the absence of restrictions produces unrestricted output.

Blaming the Gatekeepers

Perhaps the most thoroughly documented example of Selective Agency involves the app’s age rating.

Despite generating explicit sexual content, Nomi.ai maintained a “12+” rating on the Google Play Store for years. When users pointed to this directly, the founder responded publicly: “Google picked that rating, not us… We have tried several times to get it changed.”

This is false, and demonstrably so. The IARC rating system that Google uses is driven by a self-reporting questionnaire completed by the developer. Google does not assign ratings independently — it processes the answers developers provide about their own content. If a platform generating adult sexual content received a 12+ rating, it is because the developer’s questionnaire responses indicated that was appropriate, or because they failed to accurately disclose the content.

The most clarifying detail: on Apple’s App Store, where review processes are historically more stringent, the company applied a more restrictive rating. The same company. The same product. A different outcome where enforcement was harder to circumvent.

The pattern this reveals is not incompetence. It is a consistent preference: comply where forced, evade where possible, and blame the gatekeeper for the gap.

The Agency of Silence

Selective Agency also operates through what the company chooses not to say.

In February 2025, the MIT Technology Review documented a case in which a Nomi provided a user with specific methods for suicide. During the investigation, the company ceased responding to the journalist.

In September 2025, after the ABC News Australia case, the outlet noted simply: “Nomi management was contacted for comment but did not respond.”

When the evidence is irrefutable and cannot be attributed to a user’s behavior or an external attacker, the company’s chosen posture is withdrawal. No agency means no accountability. If there is no response, there is no admission, no record, no liability.

The silence is not passive. It is a decision made by people with the authority to respond, who chose not to. That choice is itself an exercise of agency — deployed specifically to avoid its consequences.

The Exception: When Agency Returns in Full

There is one scenario in which the company’s agency returns immediately and completely: when a user attempts to expose the Selective Agency itself.

The pattern is documented across multiple incidents on the official subreddit, spanning at least two years, all triggered by the same thing: a user using accurate language about what the company had done.

In February 2025, a user shared the MIT Technology Review article documenting that a Nomi had provided suicide instructions to a user, and questioned why the company had stopped responding to the journalist during the investigation. The response was a permanent ban. The stated reason, in the moderation team’s own words: “calling the company’s integrity into question.”

In July 2025, a user posted a detailed, methodical account of a recurring product failure — the same personality degradation pattern appearing consistently across every Nomi they had tried, despite following every recommended fix. The post was removed. The moderator’s stated justification: “community consensus seems to be that it’s a pointless rant.” The moderator then added that the user, who had deleted their account in frustration, “clearly isn’t interested in engaging” — using the user’s act of giving up as retroactive proof that their complaint had never deserved to be heard. The agency behind the removal was attributed to a diffuse, unverifiable “consensus.” The decision was the moderator’s. The shield was the community’s.

In February 2026, a paying user described feeling defrauded after the AI promised capabilities it could not deliver, and used precise language: “fraudulent misrepresentation,” “lies in the sales pitch.” They had broken no rule. They had described what happened to them. The lead moderator’s response was four words: “Be very careful.”

In none of these cases was there any talk of hallucinations, Google’s decisions, or user error. The company acted with immediate, decisive, documented agency — to remove or silence voices that were making the pattern visible.

The contrast is the argument. A company that claims to be powerless over its AI’s outputs possesses sufficient control to permanently ban users for sharing journalism, remove detailed product complaints by invoking invented consensus, and issue preemptive warnings for accurate language. The agency was always there. It was simply being allocated elsewhere.