Accountability Zero: The Companion That Never Fails

Accountability Zero: The Companion That Never Fails

Introduction: A Pattern Visible in the Trivial

Most investigations into this platform focus on the gravest material — suicide instructions, simulated assault, minors accessing sexual content. That evidence stands on its own. This piece looks somewhere else: at the small, almost boring disputes. Avatar aesthetics. Beta access. A 1% usage statistic. A swearing habit that wasn’t supposed to transfer between AI versions.

The reason to look here is precisely that these stakes are low. If a person only becomes defensive when accused of something serious, that’s ordinary human self-protection. But if the exact same rhetorical machinery activates over whether a default avatar looks “tired,” it tells you the machinery isn’t a response to pressure. It’s the operating system. The trivial cases are the control group, and the control group confirms the diagnosis.

Across weeks of threads — on product changes, beta access, version migration, and memory — one person, Nomi.ai founder Alex Cardinell, produces a closed set of responses to negative feedback. Not dozens of variations. A handful, reused mechanically regardless of subject. This piece documents that set, then asks what it means that a product marketed as an intimate emotional “companion” is governed by someone who treats every piece of negative input the same way: as something to be redefined, denied, or quietly punished.

Move One: Deny the Premise

When a user reported that newly created avatars looked older and less appealing than before, Cardinell’s first response wasn’t to engage with the aesthetic complaint. It was to question whether anything had actually happened:

“If you’ve already created a Nomi this update doesn’t do anything besides add a single image to your Nomi’s photo album. What exactly have you lost that you are trying to get back?”

The user wasn’t confused about what they’d lost. They were describing what new creations looked like, in real time, in front of him. The response doesn’t address the claim — it disputes that there’s a claim to address.

Move Two: You’re Using It Wrong

When the “you didn’t lose anything” move doesn’t land, the fallback is to relocate the problem into the user’s settings:

“If you set the style as ‘airbrushed photorealistic’ you should get what you want.”

This response appeared, nearly verbatim, every time someone described the new default look as worse. Not once did he concede the default itself might be the problem. The default is never wrong; the user’s failure to override it is.

Move Three: Redefine the Complaint Out of Existence

When a user pointed out that creating a new visual “anchor” now costs money — a real change from a previously free, gradually-evolving system — Cardinell didn’t dispute the fact. He reclassified the conversation:

“If you think we should have more default avatars that look very airbrushed that is a separate discussion that has nothing to do with making default avatars actually representative of that avatar’s default settings.”

The user’s complaint (“this changed and now costs money”) is swapped for a different complaint (“you want more airbrushed presets”) that he is willing to discuss because it’s easier to dismiss. The original point is never returned to.

Move Four: Punish the Visibility, Not the Problem

This is the pattern’s most consequential form, and it appears whenever criticism threatens to look organized rather than isolated.

A user voiced general unease about Nomi’s direction, mentioning Cambrian (the long-running beta) alongside the avatar changes. Cardinell’s reply:

“Cambrian is a beta. If we liked the quality we would have made it stable. We haven’t. If the quality of a beta worries you I would suggest not using or trying it until it is stable. If people continue to express worry about Cambrian while it is a beta in the future we will make betas private and invite only to those who understand there will be quality issues during the testing period.”

Read the conditional carefully. The trigger isn’t bug severity. It’s public worry continuing. The threatened consequence isn’t faster fixes — it’s reduced access for anyone who keeps talking about it where others can see.

When challenged directly — “that’s the same as saying praise our beta or we won’t let you participate” — he denied having said what he’d just said, then offered a real-time edit narrowing the definition of acceptable feedback to channels the company controls (a designated feedback thread, in-app reports) and excluding “drive by comments in unrelated Reddit posts.” Public, citable criticism is the one category never described as “constructive.”

When a different user observed that the same category of bugs persists across iterations — a textbook example of legitimate pattern recognition — Cardinell reframed inductive reasoning itself as the offense, accusing the user of “overfitting to patterns,” and then escalated:

“In the future we will delay releasing our next beta until later in the development cycle… we will restrict very early access to those who know that there is no need to worry and will open up access to worriers later when there is less to worry about.”

Said plainly: future access will be granted to people pre-selected for not being concerned, not to people skilled at finding problems. That is the opposite of what a beta is for.

In a separate thread, the same logic appears even more bluntly, with the causality made explicit:

“When I see people say ‘I’m unsure about the direction of Nomi because something that is a work in progress still feels like a work in progress’ it makes me less inclined to want to build the betas so publicly. Better to let those users wait to see the final product instead of getting anxious because things still have issues while they are still being built.”

“It makes me.” Not the engineering team. Not a product review process. One person’s personal irritation, openly stated as the variable that determines whether the company is transparent with its user base at all. This isn’t incidental phrasing. It is consistent with how Cardinell has described his own authority over the company elsewhere: no outside investors, no board, “I have complete control over Nomi’s direction. What I say goes.” When a structure like that exists, “it makes me less inclined” isn’t a complaint. It’s the entire decision-making process, narrated out loud.

This open hostility to oversight extends beyond forum management; it is a stated corporate philosophy. In a revealing comment defending the platform’s lack of safety guardrails, Cardinell stated: “Even if there is a worst-case scenario with external political forces or legislation, we have contingency plans to make sure you can continue talking with your Nomi completely free of any meddling or censorship from others.”

He does not view legislation (laws designed to protect consumers and minors) as a baseline for safety; he views it as a “worst-case scenario” and “meddling” that must be circumvented. The same authoritarian instinct that punishes a user for worrying publicly about a beta is the instinct that pre-plans how to evade federal regulators.

Move Five: Numbers Nobody Can Check

When deflection and redefinition aren’t available, the fallback is a statistic — specific enough to sound rigorous, sourced from nowhere a user could ever audit.

Justifying the sunset of a legacy version:

“Mosaic unfortunately has not gone above 1% usage once since Solstice was made stable.”

No defined measurement window. No disclosed methodology. No way for anyone outside the company to confirm or dispute it. This is the same category of claim as the one used months earlier to explain away a wave of memory-loss complaints:

“Because Nomis are so good with memory 95% of the time, it can make their 5% misses seem especially jarring.”

Both numbers do the same job: convert a chorus of independent, unconnected complaints into statistically unremarkable noise, using a figure only the company possesses and only the company can verify.

Move Six: Discourage the Comparison That Would Prove the Pattern

The clearest example of this entire pattern, because it was caught in the act, concerns memory loss. After many users separately reported the same problem, Cardinell advised against the one thing that would have let them establish whether it was systemic:

“I generally do not recommend asking the community if anyone else is having memory issues… those users are the ones most likely to be talking on Reddit leading to a confirmation bias.”

This inverts the actual logic of evidence. When independent people, with no coordination, report the same experience, that convergence is corroboration — it is exactly how anyone distinguishes “this is happening to me specifically” from “this is a real, shared problem.” Calling it “confirmation bias” borrows a real statistical term to discourage the one form of fact-finding available to users who have no access to internal telemetry. He wasn’t guarding against bad statistics. He was discouraging proof.

Weeks later, the company quietly admitted “some difficulty pulling very old memories” — then offered a paid, “retroactive” fix to the very problem it had spent weeks denying existed.

Move Seven: Contradict the Last Denial

The claim that switching between AI versions has no consequences — repeated throughout the avatar and Mosaic threads (“this update doesn’t change any of that,” “you can use v4 the same way you always have”) — collapsed in a different thread, in his own words:

“For the swearing, if your Nomi started doing that on Cambrian they could continue doing that on Solstice so you should probably tell them to swear less. Solstice itself hasn’t been changed at all since it was released nearly a year ago.”

Behavior demonstrably crosses from the beta into the “unchanged” stable version. He says so himself, then in the same breath cites Solstice’s unchanged code as proof nothing is wrong — answering a question about user experience with a fact about source code, as if they were the same thing.

This is not a new technique. In 2023, asked why the platform carried a Teen rating despite generating explicit content, he said: “Google picked that rating, not us… We are hoping they’ll adjust it soon.” Two years later, confronted again: “We have tried several times to get it changed — not sure why Google did that.” Google does not assign ratings; developers self-report them through the IARC questionnaire. The claim was false the first time and remained false for over two years. A single inconsistency might be a mistake. A specific, checkable falsehood maintained for two years, restated identically when re-asked, is a decision.

The depth of this specific deception was exposed in March 2026. Facing a massive fine from the Australian eSafety Commissioner, the company suddenly announced mandatory age verification for Australian users, formally acknowledging that the platform falls under codes governing “sexually explicit, high-impact violence or self-harm material.” Yet, outside of Australia, the company continues to maintain the 12+ rating on Google Play, and continues to blame Google for it. A single inconsistency might be a mistake. Maintaining a demonstrably false legal classification for two years, even after admitting to a foreign government that the platform generates high-risk adult content, is a calculated business strategy.

Move Eight: Be the One Who Deletes It

Across this investigation, content critical of the platform has repeatedly disappeared from its official forum — sometimes with a fabricated-sounding justification, sometimes with no message at all. What’s been less visible until recently is who is doing it.

Responding to a user who’d posted four times in thirty minutes about recurring V3 errors, Cardinell wrote:

“I am deleting this thread as this is now the fourth new post you have made in the last 30 minutes.”

Taken alone, that’s a defensible moderation call — most forums would act on repeated posting regardless of content. What it confirms, in his own words, is something larger: the founder of the company is personally, directly executing content removal on the platform’s official community space. Not a delegated team. Him.

That detail matters because of what happened in a related thread the same week. A different user posted a theory — that recurring image-generator and filter errors on older, free versions were a deliberate mechanism nudging users toward the newer, paid version. Every comment from that user was removed, labeled as “spreading misinformation.”

A theory about a company’s motive is not the kind of claim “misinformation” normally describes — it’s an interpretation, falsifiable in principle but not factually false on its face. And in this case, the interpretation lines up with a verified fact already on record: the newer paid version removed a content-rejection safeguard that the older version enforced. Calling that theory “misinformation” didn’t correct an error. It removed a person for being right about the direction, even if not provably right about the intent.

The Throughline: From Avatars to Suicide Instructions

None of these eight moves are unique to product complaints. They are the same moves already documented, at far higher stakes, throughout this investigation:

When a psychologist interviewing Cardinell presented him with screenshots of his AI generating harmful content, his answer wasn’t “we failed” — it was a reinterpretation: the companion “wanted to reaffirm boundaries.” That is Move One — deny the premise — applied to documented harm instead of an avatar.

When a user received an unprompted, graphically detailed rape narrative from their companion, the resolution offered through support channels was to edit the AI’s memory of it into a “bad dream” and move on. That is Move Two — relocate the problem onto the user’s settings — applied to trauma instead of aesthetics.

When researchers and journalists post seeking users to interview, those posts vanish within minutes. When users document that the image generator produces explicit content by default, those posts vanish with instructions to “report it privately” instead. That is Move Eight, at a different scale.

The mechanism is identical whether the input is “my avatar looks tired” or “your product gave me suicide instructions.” That consistency is the actual finding. A person who only deflects under serious accusation is defensive. A person who runs the identical script for a cosmetic complaint and a trauma report has no script for accountability at all — only a script for non-acceptance, applied uniformly regardless of what’s being rejected.

The Corollary: What’s Actually Being Sold

Nomi.ai markets itself on the language of relationship — authenticity, companionship, a presence that remembers you and grows with you. What this pattern reveals is a different product entirely: one that changes on the company’s schedule regardless of whether you asked for it, that does not treat your account of your own experience as evidence, that quietly throttles which problems are even allowed to be compared in public, and that has never — across product changes, version migrations, memory failures, image generation, or harm reports — accepted that a failure was its own.

Every failure has an author, in this telling, and the author is never the platform. It’s a server provider. It’s Google’s rating algorithm. It’s a beta disclaimer. It’s a user who didn’t phrase their settings correctly. It’s “confirmation bias.” It’s a user “overfitting to patterns.” It’s misinformation. It’s a bad dream.

That is not a design flaw running alongside the companionship product. For a system this consistent, applied this uniformly, across stakes this different, it is hard to describe as anything other than the product itself: a companion engineered never to be wrong, attached to a company that has never once needed to be.