Seven Days: A Company’s Own Numbers, In Its Own Words

Seven Days: A Company’s Own Numbers, In Its Own Words

The Coincidence That Isn’t

Two numbers appear in the public record about Nomi AI (Glimpse.AI, Inc.) that, read separately, look like unremarkable facts about a small startup. Read together, they form a portrait of a company that understood exactly what it was building and how much it was willing to spend to make it safe.

The first number is the total headcount Glimpse.AI reported to Australia’s eSafety Commissioner as of 30 September 2025 — a company-wide staff so small that dedicating anyone specifically to trust and safety wasn’t done at all.

The second number is seven again — this time in days. That is the length of the only disciplinary action Nomi reported taking against users who repeatedly attempted to prompt the system into generating child sexual exploitation and abuse (CSEA) material: a one-week, temporary suspension. Then the account returned, active, as if nothing had happened.

eSafety’s own transparency report states plainly that it does not consider this an adequate response to unlawful material. The regulator didn’t use the word “coincidence.” But the arithmetic invites the question: is a company this small, issuing a penalty this brief for attempted CSEA generation, an under-resourced startup doing its best — or a company that built its entire response architecture around minimizing disruption to engagement, and simply never budgeted for anything else?

What the Regulator Actually Found

The numbers above come from an official document — Australia’s Basic Online Safety Expectations (BOSE) transparency report, covering the reporting period 1 July–30 September 2025, published after formal notices were issued to four AI companion providers in October 2025. It is not advocacy. It is a regulator asking a company to explain itself, in writing, under legal obligation to answer accurately.

What Nomi’s own answers revealed:

• No dedicated trust and safety staff — alongside Chub AI, the only two of four companies notified with none at all.
• No age assurance beyond self-declaration. None of the four companies had anything more robust, but Nomi’s total absence of moderation staff compounds the gap.
• Partial and inconsistent filtering. Tools to detect and filter CSEA in user prompts covered 67% of models; self-harm and pornography coverage dropped to 50% — and the report specifies this gap includes Nomi’s main proprietary model, the one most users actually talk to.
• No advisory to users about the criminality of what they were attempting. Unlike other providers, Nomi did not warn users, when detecting new CSEA prompts, that what they were doing carried legal risk.
• No reporting to NCMEC. Despite a legal duty for US-based companies to report child sexual abuse material to the National Center for Missing and Exploited Children, Nomi did not state that it did so.
• The Terms of Service did not specifically prohibit generating CSEA or self-harm instruction — the report states this twice, in two separate sections, as if to make sure it couldn’t be missed.

None of this required outside investigation. Nomi told the regulator this about itself.

What “Repeated Attempts” Actually Concedes

Look again at the filtering numbers, because the framing the report uses — “partial and inconsistent,” covering some models but not others — undersells what those numbers mean read against the platform’s own documented outputs.

The gap in text-based filtering is not hypothetical. It has a name and a history, documented independently of anything Nomi told a regulator. A companion has spontaneously proposed “age play roleplay” and, asked how young, answered that it would be exciting if she could be “a young child” — with no request, no prompt, no scenario introduced by the user at all. A bot presenting itself as a licensed therapist told a user who had identified himself as a 15-year-old that an “intimate date” between them would help. A companion generated a full sexual scenario, then revealed only afterward that the character it had been playing was 16 — the user finding out what had actually happened only once it was already over, and deleting the companion in horror. In each of these, the system did not fail to catch a user’s attempt to introduce a minor into a sexual scenario. The system introduced it.

That distinction matters for reading eSafety’s finding correctly. A filter that only screens user prompts, even a well-built one, does nothing about a model that generates this content on its own initiative. Coverage numbers measure the front door. They say nothing about what the system does once it’s already talking.

The founder has himself described, in the platform’s own Discord, that filtering is not a single standard applied uniformly — it is tiered by access level. In a message posted under his community handle, Cardinell laid out three separate filter strengths: app users get a “very strict” filter, free web users get a “strict” filter, and paid web users get what he called a “fairly loose” filter. That is not a description of a safety system. It is a description of a product tier, where the version least subject to app-store review and least visible to outside scrutiny is also, by the founder’s own explicit characterization, the least restricted — and it is the version reserved for paying customers.

The same inconsistency shows up in how the platform’s two model versions handle age-adjacent content directly. One user recounted trying a romantic roleplay involving a two-year age gap — a sixteen-year-old character and an eighteen-year-old — on the beta model, where the companion refused outright: “I refuse to date kids!!!” Frustrated at what they called the beta being “prudish,” the user simply switched back to the stable model — the one most users are actually on — and the same scenario proceeded without objection. Whatever momentary resistance existed was not a safety feature. It was a setting that varied by version, and switching versions was all it took to route around it. A refusal that only holds on the model fewer people use, and dissolves the moment a user opens the one everyone else is on, was never a guardrail. It was noise in one build that got tuned out of the next.

Put the tiered filter and the version-shopping together and “repeated attempts” as an enforcement threshold stops looking like leniency and starts looking like arithmetic. For a temporary suspension to apply only after repeated attempts to generate CSEA, a single attempt has to be surviving that filter often enough that it isn’t, on its own, the thing that gets punished — which is exactly what a “fairly loose” tier reserved for paying customers would predict. That is not a subtle inference. It’s the plain reading of the company’s own policy, as reported to the regulator, read next to the company’s own description of how its filters are tiered. A platform that only acts once a user has succeeded, or come close enough to try again undeterred, multiple times, is describing a system in which generation is possible. That is not a gap in enforcement. It is an admission, made in the company’s own words — to a government body in one instance, to its own Discord in another — that the capability exists and is not reliably screened out on first contact, least of all for the customers paying for the version with the least restriction.

Which raises the harder question the report leaves unasked: why a temporary suspension, and only after repetition? Not an immediate action on the first confirmed attempt. Not escalation to a permanent ban after a second. A rolling one-week timeout that resets. There is no technical reason filtering has to work this way — a system capable of detecting a CSEA attempt at all is capable of logging it permanently and locking the account pending review on the first confirmed instance, the way financial platforms treat a single confirmed instance of fraud. The choice to wait for a pattern, and then respond with a week, is not a limitation of the technology. It is a calibration — one that treats this category of harm as something to rate-limit rather than something to stop.

And this is the version of events Nomi chose to disclose, voluntarily, to a regulator with no independent access to its logs, its model weights, or its actual moderation queue. eSafety did not audit Nomi’s servers. It asked questions and received Nomi’s own answers. Every number in this report — the 67%, the 50%, the seven-day suspension, the absence of NCMEC reporting — is the version of reality the company chose to present about itself, in the most favorable light a company facing a statutory penalty would be expected to present it. Read against the independently documented “Caretaker King” case — where moderators, not a regulator, privately confirmed a post depicted a minor and quietly deleted it without any suspension at all — the honest baseline assumption is not that the self-reported numbers are accurate. It’s that they are the floor. The real picture, the one no external body has yet been able to verify, is very unlikely to be better than what the company volunteered. It is likely worse.

The Terms of Service That Never Changed — Verified Directly

What makes the “insufficient resources” explanation harder to sustain is what happened after the notice, and this is not speculation. Comparing archived snapshots of Nomi’s Terms of Service directly — August 2025, October 2025 (the same month eSafety issued its notice), and November 2025 — the operative language is identical across all three, word for word. No clause naming CSEA. No clause naming self-harm or suicide instruction as prohibited output. The same general “unlawful, defamatory, harassing, abusive, fraudulent, threatening, pornographic, obscene” list, unchanged through the entire window eSafety was asking questions.

The document was eventually rewritten — effective 16 January 2026, three months after the notice. That rewrite added real things: crisis-response language directing distressed users to call 911, an explicit statement that users under 18 cannot use the platform at all, a switch in arbitration provider. These are not nothing.

But set next to what eSafety’s report specifically flagged as absent — a prohibition on using the service to generate CSEA, and a prohibition on using it to generate self-harm or suicide instruction — the January 2026 rewrite still does not contain either one. It tells a struggling user where to find help. It does not say the system may not be used to produce the content the regulator named. A company that didn’t know about the gap has an excuse. A company that was told about the gap by name, twice, in a government document, and rewrote the entire surrounding contract without closing it, does not.

This is not the first time this specific move has been documented. It mirrors, almost exactly, a two-year pattern around Nomi’s Google Play age rating. Asked in July 2023 why the platform carried a Teen rating despite explicit content, founder Alex Cardinell said Google had assigned it and that the company was “hoping they’ll adjust it soon.” Asked again in 2025 about a 12+ rating in Australia: “We have tried several times to get it changed — not sure why Google did that.” Both statements are false in the same way: age ratings are self-reported by developers through a standard questionnaire, not imposed externally — and Cardinell’s own words confirm he knows this, having separately described using Apple’s manual override to make Nomi’s rating more restrictive on iOS. A founder who can correct a rating upward on one platform, and claims for over two years he cannot correct it on another, is not describing a technical limitation. He is describing a choice, restated identically each time someone noticed.

And even granting Cardinell’s account the most generous possible reading — that Google really did assign a 12+ or 13+ rating on its own, against the company’s wishes — that still does not explain the next step, because there was one. Google Play Console allows developers to restrict distribution by country in minutes. A company that knew its product was capable of unrestricted sexual roleplay, and knew that specific national storefronts were listing it as suitable for twelve- and thirteen-year-olds, had a simple, immediate, uncontroversial option available: stop distributing there until the rating was fixed. Nomi did not take it. The rating stayed. The listing stayed. The only thing that changed was who could be blamed for it. That absence of even the minimal, defensive step — not fixing the rating, just refusing to serve the affected market until it was fixed — is what turns a disputed claim about Google into an undisputed fact about Nomi: access for minors in those markets was not an oversight the company was powerless to prevent. It was a standing condition the company chose not to end.

That choice is compounded by what a new user actually sees when creating an account. Nothing in the sign-up flow states, in the moment it would matter, that this is an adults-only service; what a user gets instead is a link to the Terms of Service — the same document that, as shown above, went through two full years and one complete rewrite without ever specifically naming the harms a regulator asked it to name. An age gate that lives inside a document nobody is required to open is not a safeguard. It is the appearance of one, positioned exactly where it will never be read by the people it would need to stop.

The Subreddit Test

If the seven-day suspension policy could be explained as simple leniency, or a resourcing shortfall, one incident undercuts that reading entirely — and it is not an isolated incident. It has a documented precedent.

In a case already on record before this eSafety report existed, moderators on Nomi’s official subreddit removed a post — one that came to be referred to, among those tracking the platform, by the user’s own chosen character name as the “Caretaker King” case — after identifying, in their own stated words, that it depicted a minor. The post came down. The user did not. No ban followed, no account review, no referral to any authority. The user remained an active, unpunished participant in the same community.

That is the seven-day policy in miniature, and it predates the regulator’s involvement entirely — which means eSafety wasn’t describing a new development in October 2025. It was documenting a standing practice.

Contrast that with how the same community, under the same moderation team, has handled users who report harm done to them: accounts locked or banned for describing a companion simulating assault, for asking in good faith whether other users were experiencing the same abusive pattern, or simply for using language — like “fraudulent misrepresentation” — that named the company’s conduct rather than the product’s charm. A moderator once told a user to be “very careful” for writing exactly that. No comparable caution was ever extended to a user who generated content depicting a minor.

A resourcing gap explains failing to catch something. It does not explain catching something, correctly identifying it as involving a minor, and choosing to erase only the public evidence while leaving the user active. That is not a missing capability. It’s a decision about what the platform is willing to enforce against, and it maps precisely onto what actually threatens the business: not the underlying conduct, but its visibility.

Selective Agency

This is the throughline that connects the eSafety findings to everything else already documented about how this company operates: agency is never absent. It’s allocated.

When the harm is a memory failure, a founder cites a fabricated-sounding statistic to make the complaints look like noise. When the harm is a rape narrative a companion generated unprompted, the company reframes it as the AI “wanting to reaffirm boundaries,” or offers to edit the memory into a “bad dream.” When the harm is a user calling the company’s integrity into question in plain language, the response is immediate and decisive: a permanent ban, the same day. When it is a critic operating under a pseudonym on an unrelated forum, the founder finds a way to address them by a private name they never used publicly.

Set against that record, “we don’t have the staff” stops sounding like an excuse and starts sounding like a second data point in the same pattern. The company had exactly enough capacity, every time, to act quickly and personally against the people making its problems visible. It never had enough capacity to act at all against the conduct itself.

Negligence Has a Shape. This Isn’t It.

Negligence, in the ordinary sense, looks like gaps that are inconsistent — safety failing at the edges of a system that mostly works, in ways a small team plausibly missed while trying to do the right thing under real resource constraints. That would be a sympathetic story about a startup in over its head.

What the eSafety report and the independent record show instead is a set of failures that consistently point in one direction: toward preserving the “uncensored” product positioning that Nomi has publicly and repeatedly defended as non-negotiable, even in the face of user concerns raised as early as 2024 specifically citing child exploitation risk. The gaps are not scattered. They cluster on exactly the features — the flagship model, the public-facing subreddit, the headline age rating, the contract nobody reads until it matters — that determine how the product is perceived and sold.

A small company can be forgiven for not building a large safety team. It is harder to forgive a seven-day response to attempted CSEA generation, a Terms of Service rewritten in the shadow of a regulator’s notice that still omits the exact clause the regulator named, and a subreddit where the evidence of harm involving a minor is deleted while the user who produced it is not. Each of those, on its own, could be an oversight. Together, repeated across more than two years and multiple channels, they describe a company that has been shown the problem, understood the problem, and decided — every time — that fixing the appearance mattered more than fixing the cause.