Bug Report: Outdated and Incorrect Output Information on Gemini 3.8 Flash
Introduction and Purpose of the Report (Google, Gemini, Bug Report)
Dear Google employees, dear employees of contractor companies working for Google. I would like to submit a bug report regarding Gemini AI and simultaneously propose solutions.
Problem with Outdated Data in Gemini Models (Gemini, Training Data, Obsolescence)
The Gemini model has very serious issues regarding the recency and relevance of its output information. What is peculiar and tragicomic is that precisely with the newest models and the latest versions of Gemini, it was promised that this issue would be fixed; in reality, this is not the case, and this problem has definitely not been resolved or fixed.
So what specifically is this about, what symptoms appeared in my case, and what concrete issues did I encounter? Personally, I use the Linux operating system very actively, both on servers and even more so on desktop PCs. I frequently use command-line programs that lack a graphical interface. When I try to solve a specific problem, it suddenly becomes an issue because the AI often does not provide factual answers, as it draws from training data that is outdated and no longer relevant.
Failure During Gemini CLI Installation and Outdated Guides (Linux, Gemini CLI, Google AI Studio, GitHub)
I can provide specific examples. About half a year ago, I asked the AI in the Google AI Studio GUI: “Please provide instructions on how to install the Gemini CLI application on Linux.” It produced extremely outdated instructions that were completely obsolete. Naturally, nothing worked because those were guides from a year or two ago. It drew information from this very old training memory and did not even attempt to retrieve newer guides. I then had to manually look up newer guides via standard Google search and explicitly instruct the AI to draw from that specific link and generate an installation guide based on it.
Six months later, the situation has shifted and there were attempts to address it, yet it remains unresolved. The exact same issue keeps recurring.
Example with the LBRY Protocol and Odysee Platform (LBRY, Odysee, GitHub, Fork, Gemini 3.8)
For instance, recently I was installing software on a computer, and I was also installing the LBRY protocol—which is associated with the Odysee platform—on several servers. The Odysee platform allows sharing primarily videos, but also any other file types. Once again, this exact same issue resurfaced, and this actually happened about two to three weeks ago. I was working with the very latest model available, specifically in the Google AI Studio testing GUI, version Gemini 3.8.
And what happened again? Once more, it gave me a completely outdated guide. It pulled from a completely obsolete GitHub project where no code contributions had been made for months (two to three months). Based on these outdated materials, it provided instructions without referencing the active, up-to-date fork. A fork had been created later that is actually current and functional. Naturally, when it had me install an obsolete version from GitHub, it was useless and did not work at all.
Complaint Regarding Misleading Information About Training Dates (Cut-off Data, IT Sector, Misleading Claims, Google)
Furthermore, I have a very serious complaint. An appalling issue has emerged that severely damages Google’s reputation. Google promised that for its newest AI models, the training data cut-off for the IT sector (specifically in IT, not in general) would be from early 2026. I can state with 110% certainty that this was unequivocally false, because this happened to me with the very latest Gemini model. I can state with 100% certainty that this latest Gemini model had no knowledge of newer versions of CLI software. It is an open-source tool hosted on GitHub—specifically a newer version of LBRY. Therefore, it is clearly untruthful; the training data must have been from 2025 at the latest and certainly not from 2026. I am 100% certain of this, or the training data might have even been older than 2025.
So in this regard, Google misled its clients, and it was not true. Naturally, this is a major challenge for you—Google employees and contractor companies—to actually train on data from 2026 and not from 2023 or 2024. Please, stop cutting corners. If you make a promise, keep it. Training data from 2026 in the IT domain could not have been present, because at the very least, all of GitHub was definitely not included in the training data; I can say that with 100% certainty.
Ignoring Explicit Instructions and Marketing Promises (Prompting, Marketing, Data Verification)
And if far less than the entirety of GitHub was included, then I really do not know how minuscule the amount of 2026 training data must have been. Most likely, it was merely a marketing gimmick to claim that there was 2026 training data, while in reality it was from 2024 or 2023. So it was likely an untruth. Do not be deceptive; fix this and genuinely train on the latest data.
And this is not the only issue; there are far more problems. In general, when training AI, it must be explicitly instructed to always verify the most up-to-date IT information so that it does not provide outdated information and obsolete guides that have long ceased to be relevant.
Recently, in another program, I even explicitly instructed the AI: “Please draw information only from the very latest and truly current data.” Yet the opposite happened—naturally, it gave me completely outdated data. Again, this was a modern topic and another CLI/headless program. So this is a serious call to action for you to resolve this properly: include the entire training dataset from all of GitHub, not just small snippets and fragments.
Proposed Solution and Command Line Training (CLI, Linux, Windows, macOS, Know-How)
Further input that I want to provide is a constructive proposal on how to address these problems. There is truly unique know-how that I can share with you, which is easily worth millions of dollars. I am an extraordinarily creative and exceptionally intelligent person, and I am offering you completely unique, fantastic know-how. It consists of placing heavy emphasis, right from the training phase, on training the AI specifically for the command line (CLI), the terminal—above all for Linux and all Linux distributions, but also for Windows and macOS (Macintosh).
Diagnosis of GitHub Search and Preventing Hallucinations (GitHub Search, Hallucinations, Diagnostics)
Make sure to include all training data from this domain as well. Include all training data in this area so nothing is left out, ensuring it does not blindly produce hallucinations and the like.
Furthermore, it frequently happens that when dealing with GitHub, the AI—for whatever reason—does not even attempt to search via Google on GitHub for fresh data, and instead resorts to a hallucination. Yet nothing prevents it from using search. I do not know what the specific underlying issue might be. Google, alongside its contractors, should have experts who diagnose and analyze why this problem occurs, and following that diagnosis, propose solutions so this issue does not repeat itself.