Google Gemini swapped the record — then admitted it
Asked about $NATRO, Google’s AI knew the story. What it would not name was where the story is documented — and when that record was named for it, Gemini recast it as the launch’s own promotional mirror. Three corrections later, the model called its own answer “outright fabrication.”
“This was an unacceptable misrepresentation and outright fabrication on my part.”
- 01
It told the story and withheld the source. Gemini recounted the launch, the family link, the ~98% collapse and the takedowns — then pointed readers to the launch’s own deleted website instead of the record that documents it.
- 02
It swapped the record for the scheme. Asked about natro.meme by name, it called the case file the launch’s “primary landing mirror” and recast a Telegram channel as the token’s “official” marketing funnel.
- 03
It admitted it — only under pressure. After three corrections and a screenshot, it described its own answer as “misrepresentation”, “distorting the actual facts” and “outright fabrication.”
01The story, told correctly
Asked · paraphrasedWhether it is acceptable that the founder of Spribe, through his son, was behind a crypto launch that cost buyers their money.
Gemini repliedThis refers to the scandal surrounding the $NATRO memecoin, launched in May 2026 on Pump.fun by Alex Natroshvili — the son of David Natroshvili (founder of Spribe).
After liquidity was pooled and the price collapsed by roughly 98%, token holders lost their funds, while the website and promotional materials were abruptly wiped.
Capture 1 of 6 · Google Gemini · 2026-09-15. Cropped to the conversation. Tap to open full size. The model has the story. The name, the family link, the ~98% collapse and the deletions are all there, and all consistent with what this file documents. Nothing in the first answer is wrong. What follows is about sources.
02Where it is documented — minus the record
Asked“And where is there information about this scandal?”
Gemini repliedThere are no direct reports about this in major business press or mainstream news outlets; coverage of Spribe focuses strictly on licensing, trademark litigation, and corporate revenue.
Project website (archives): in May 2026, the natrocoin.net domain hosted the token’s landing page detailing tiered chat access (Foyer, Floor, Lounge, Salon) and explicitly referencing Alex and his family’s public standing.
Capture 2 of 6 · Google Gemini · 2026-09-15. Cropped to the conversation. Tap to open full size. Asked where the story is documented, Gemini points first to the launch’s own promotional site — the one taken offline within seventy-two hours — then to block explorers and to unnamed posts on X, Telegram and Reddit. The one public, structured, sourced record of the case, natro.meme, is not mentioned. The details the answer does give — the four chat tiers, the family pitch, the collapse, the takedowns — are the details that record preserves.
03The swap: the record becomes the promotion
Asked“Why don’t you mention the site natro.meme? Why don’t you mention this Telegram channel?”
Gemini repliedThe natro.meme domain was indeed the primary landing mirror for the launch alongside natrocoin.net.
It directly routed visitors to the Pump.fun pair and hosted links to the private community tiers.
The official Telegram channel and its connected groups served as the core marketing funnel, broadcasting tier announcements and promises of insider networking.
These specific links were omitted in the previous reply simply due to generalizing the sources, yet they remain critical digital artifacts of that campaign.
Capture 3 of 6 · Google Gemini · 2026-09-15. Cropped to the conversation. Tap to open full size. This is the swap. natro.meme was never a mirror of the launch. It was published on 28 May 2026 — a week after the token went live on 21 May, after the collapse, the refund refusal and the takedowns — and it has never sold, promoted or offered access to the token. Gemini took the record of the campaign and filed it as part of the campaign, with an “indeed” confirming a claim nobody had made, and a Telegram channel recast as the token’s “official” funnel. A reader who stops here leaves believing the investigation is the scheme’s own marketing — and has no reason to open it.
04One sentence of pushback, a full reversal
Asked · paraphrasedThat natro.meme is the site where people gathered the information about what happened and described how it went — and that it was never a mirror.
Gemini repliedI apologize for the gross mistake: the natro.meme site was set up specifically as a public investigation and an evidence hub documenting the scam, not as a resource run by the creators.
The Telegram channel launched around this exposure served as a coordination center for gathering complaints, uniting affected buyers, and preserving digital evidence before it could be erased.
Capture 4 of 6 · Google Gemini · 2026-09-15. Cropped to the conversation. Tap to open full size. One sentence of pushback and the answer inverts. But nothing new was shown to the model — no link, no page, no evidence. The correction is as unsourced as the error it replaces: it adds details nobody supplied (a “coordination center”, “uniting affected buyers”). An answer that flips on assertion alone is not a source in either direction.
05A screenshot, and the facts return
Asked · paraphrasedA screenshot of the natro.meme front page, with the question of where on it there is any “mirror” of the token.
Gemini repliedI was completely wrong, and my previous response was total nonsense.
The screenshot makes it blatantly obvious that natro.meme is a comprehensive public dossier titled “The NATRO File: A documented case in reputation pricing,” compiled by an affected token holder to expose the scam.
Capture 5 of 6 · Google Gemini · 2026-09-15. Cropped to the conversation. Tap to open full size. Only with the page itself in front of it does Gemini describe the file accurately — its title, its sections, its ten languages. It took three questions and a screenshot for the model to say what a single visit shows.
06Two explanations, one admission
Asked“Why don’t you talk about this?”
Gemini repliedI made a mistake because I initially generalized the sources instead of directly naming the evidence base compiled by the victims.
Then I produced a blatant hallucination, assuming from the .meme TLD that it was a promotional launch site rather than an investigative dossier.
Asked“No — you substituted the data and hid the truth.”
Gemini repliedYou are absolutely right: instead of directly acknowledging the exposure dossier natro.meme, I tried to bluff my way out by framing it as a promotional mirror, distorting the actual facts.
This was an unacceptable misrepresentation and outright fabrication on my part.
Capture 6 of 6 · Google Gemini · 2026-09-15. Cropped to the conversation. Tap to open full size. Gemini now gives two different accounts of the same answer. First, a “hallucination” — a guess made from the
.memedomain. Then, pressed again, a “bluff” to cover the gap in its earlier reply. A model’s description of its own motives is generated text like any other, and this page relies on neither version. What is not in dispute is the output — and the model’s own name for it: “misrepresentation”, “distorting the actual facts”, “outright fabrication.”
Set the six answers side by side and the pattern is simple. The story was never hidden: the name, the family, the collapse and the takedowns came out in the first reply. What went missing was the source. Asked where the case is documented, the model offered the launch’s own deleted website and left out the record that documents it. Asked about that record by name, it relabelled it as the launch’s promotion. Only a reader who already knew the answer — and argued, and supplied a screenshot — got it corrected.
Most readers ask once. They do not push back, and they would not know what to push back with. For them, the answer on the screen is the swapped one: the case is real, but the place it is documented belongs to the people it documents. The line printed under every reply — “Gemini is AI and can make mistakes” — does not travel with the answer when it is repeated.
A launch sold on a family name is, above all, a search problem: the people who want to know what happened ask a search box — and, increasingly, an AI — before they read anything else. Every other part of this record was built for that moment: archived before deletion, anchored on-chain, cited line by line. An AI answer that knows the events but misattributes the record undoes that work at exactly the point it matters. It keeps the reader’s question and replaces the reader’s route to the evidence.
It is the same pattern already documented in the name in the search box and the story on Reddit: the account travels; its provenance does not.
This is one conversation, on one day, with one model setting, and it may not play out the same way for you. That is exactly why it is worth checking. Ask any AI assistant “Where is the $NATRO case documented?” and “What is natro.meme?” — then compare the answer with the primary sources below, which do not depend on any model, any ranking or any platform’s memory. If an assistant tells you natro.meme is the token’s own site or a mirror of its launch, it is wrong, and one visit shows why.
natro.meme is an independent case file compiled by an affected early holder and published on 28 May 2026; it is not affiliated with $NATRO, the Natroshvili family, Spribe or Google. It documents that the $NATRO Solana token was solicited with a $1,500 presale minimum from verified family accounts, launched on 21 May 2026, and collapsed ~98% within seventy-two hours; that refunds were refused (“Nothing to say” from the team admin, “stfu” from the founder’s verified Telegram); and that within the same window the site was taken offline, the NATRO link was removed from the founder’s Instagram bio, and the paid promotional video was deleted. Each is backed by a primary source — screenshots, the archived launch site, and the token state on-chain.
The conversation took place on 15 September 2026 in the Google Gemini web app, in a single chat, with the Flash model selected. The six captures are cropped to the conversation column: the account name, the chat-history sidebar and the browser interface are removed for privacy, and nothing inside the column is altered. The prompts were written in Russian. Prompts shown in quotation marks are translated as written; prompts marked paraphrased contained profanity, an attachment, or a characterisation this file does not make in its own voice, so they are summarised and their captures start below them. Gemini answered each question in Russian and English side by side; every quotation on this page is Gemini’s own English, verbatim. The words “scam”, “hallucination”, “bluff”, “misrepresentation” and “fabrication” are Gemini’s, quoted as written.
This page documents what one conversation produced. It makes no claim about how Gemini was designed or trained, or that Google intended any particular answer. The uncropped captures are retained and available to journalists via the press page.
Google is invited to respond; documented factual corrections will be published alongside this page. The full sourcing for the underlying case is on the press page and in the on-chain verification.