Gemini 3.8 Flash — severe hallucinations, outdated information and failure to use current information
I want to report what appears to be a serious quality and reliability problem with Gemini 3.8 Flash.
I have been testing Gemini extensively and, particularly in recent days, the model has repeatedly produced answers that are outdated, factually incorrect, or completely disconnected from the current state of the world.
The most frustrating part is not that the model occasionally makes mistakes. All LLMs make mistakes.
The problem is that Gemini 3.8 Flash frequently presents old or incorrect information with high confidence, instead of recognizing that the information may be outdated or checking current sources.
Problems I am repeatedly experiencing
1. Outdated information
When asking about current events, software, AI models, products, companies or recent developments, Gemini sometimes responds using information that is clearly several months or even years old.
In some cases, the model continues using outdated information even after explicitly being told that the information is no longer current.
This is particularly problematic for a model that is supposed to be useful for research and current-information tasks.
2. Hallucinations presented as facts
Gemini sometimes invents:
- product/model specifications,
- release dates,
- software features,
- recent events,
- technical documentation,
- names and details that do not exist.
The problem is not simply the hallucination itself.
The problem is the confidence with which the incorrect information is presented.
A user who does not already know the subject may have no way of realizing that the answer is fabricated.
3. Failure to verify information
When a question clearly concerns information that could have changed recently, I expect the system either to:
1. verify the information using available search/grounding tools, or
2. clearly state that it cannot verify it.
Instead, Gemini sometimes appears to answer from its internal knowledge and only later, after being challenged, attempts to correct itself.
This produces a very unreliable research experience.
4. Repeatedly correcting the model does not always work
Another particularly serious issue is that after I point out that an answer is outdated or incorrect, Gemini can acknowledge the correction and then later repeat essentially the same incorrect information.
This makes the model extremely difficult to use for factual research.
5. Current-information questions are particularly problematic
The problem is most obvious when asking about things that changed recently.
For example:
- AI model releases
- current software versions
- recent company announcements
- current technical documentation
- recent events
- current product availability
- recent benchmark results
The model can give an answer that sounds extremely convincing but is simply based on an older state of information.
Why this is a serious issue
A language model does not need to be perfect.
However, there is a major difference between:
«“I don’t know.”»
and:
«“Here is the answer,” followed by confident but outdated or fabricated information.»
The second behavior is significantly more dangerous for users who rely on the model for research, programming, technical troubleshooting or decision-making.
At the moment, Gemini 3.8 Flash often feels less reliable than it should be for these tasks.
In some situations it is genuinely easier and safer to use a normal search engine than to rely on the model’s answer.
Please investigate whether this is a regression
I would particularly like to know whether Google is aware of a recent regression in Gemini 3.8 Flash.
The frequency of these problems appears noticeably higher recently.
I would appreciate it if the Gemini team could investigate:
- hallucination rates,
- outdated knowledge being presented as current information,
- failures to trigger search/grounding,
- incorrect information persisting after user correction,
- possible recent backend/model changes,
- differences between the current 3.8 Flash behavior and earlier versions.
If this is a known regression, please acknowledge it and provide information about whether an investigation or fix is underway.
This is not a complaint about an occasional incorrect answer.
The issue is systematic reliability.
A highly capable model that confidently provides incorrect or outdated information is extremely difficult to trust, regardless of how well it performs on benchmark tests.
I hope this can be investigated by the Gemini engineering team.