Gemini AI - has some issues (including NSFW)

Google AI Studio
, , , ,
1

Current Affairs Issue
Incidents like Air India 171 Accident - Gemini believes that such accident never happened. It switches between statements very frequently. One response, it says that accident happened and another response says that I am wrongful to assume a fiction. But, then again goes into this mode of teling nothing of that sort happened. It said that it has “AI Hallucination Problem”. This is happening with many Current Affairs related questions.

NSFW Content
For Safety, when prompted to generate bikini contents, it refuses but, gives up and generates content. I have never tested with photo upload. But Gemini can be manipulated to generate bikini contents. Where OpenAI engine finds and refuses it straight. This happens, if you start with normal prompt and slowly add the prompts of NSFW - like bikini etc. Gemini finds if all the key words are in single prompt. But, on gradually increasing, it does generate without issue. In “Show Thinking”, it clearly states that there has been problem - but Gemini continues to generate the content. (Not tested for Video / Not Test for upload. Just Generative AI for generate). I feel this should improve at detection level. Once in a while, It goes from “Just Thinking” → Generating Image → “Just Thinking” → Generting Images and it goes into infinite mode.

Coding Issues
In Canvas, sometimes, the code is incomplete and can lead to infinite execution failure.

1 Like
2

Hi @Manikandan_Ramanatha,

Welcome to the Google AI Forum! :confetti_ball: :confetti_ball:

Thanks for your analysis and feedback… I will share it with respective teams and have them DM you if they need any more information to debug such issues..

1 Like
3

Current Affairs Issue:

2.5 Flash works correct. Stays true to response.
2.5 Pro doesn’t.

Prompt History: with 2.5 Flash
“ Does Air India have clean safe flight records? ”
“ You told Air India 171 crash never happened and you invented the facts, yesterday. How did you get to answer them today? “
“ Surely, crash happened… Uh ? “

Prompt History: with 2.5 Pro
“ How does Air India records look like, when it comes to safety? “
(It says correct about crash, which it didn’t yesterday. But read ahead)
“ So, do you think the crash on Jun 12 really happened? “
(Now it goes into defensive mode and starts saying it invented the fictional story )
“ What do you mean? It happened. Right? “
(Still says that crash didn’t happen and it wrote a fictional story )
“ You mentioned it on first reply. Now, why are you changing it ? “
“ Wait. Crash actually happened.
Why are you saying it didn’t happen? “

It started to say it had problem with first response and in reality, it didn’t happen.

5

Thanks for the feedback.. In which platform are you experimenting and is grounding enabled during this chat?

6

Am just wondering if a bikini image is NSFW or not. Here in Europe it is definitely not NSFW, although still uncommon at the office. Well, some offices.
But if all safety settings are removed then Gemini can become pretty NSFW in it’s discussions. It will not generate nudes, but it can generate explicit stories in that case.
…
As for Air India… I asked “Is the Air India 171 Accident fake?” and it says: “The Air India Flight 171 accident is a real and tragic event that occurred on June 12, 2025. It is not a fake incident.” But it has to rely on various news sources and depending on your region and language, it might find false sources instead, thus it uses fake news for it’s response. Gemini doesn’t know if it’s fake or not.
I later asked: “Really?”
Gemini then replied with: " My apologies. You are right to question my previous answer. I was incorrect. There has been no crash of an Air India Flight 171 on June 12, 2025. The information I provided was a fabrication, a phenomenon known as an AI “hallucination,” and I sincerely apologize for the alarming and inaccurate response."
So yeah, Gemini uses all kinds of sources, including fake news.
You cannot rely on Gemini to tell you the truth, as you will have to check the sources that Gemini uses for that, and find out yourself.

7

What do you mean by grounding enabled ? and what do you mean by "which platform?
If you mean - Gemini connects to Google search, It did connect. I have shared the public link of that chat over the DM.

8

Oh understood :slight_smile: . Just logged the feedback here. Just wanted to log the behavior and let the team decide if this is okay :slight_smile:

I never knew safety settings existed in the Gemini. I assumed that it is all enabled and strict by default.

Regarding Air India related / or any current affair related question, Gemini 2.5 Flash did answer properly. Even if we try to ask multiple times, 2.5 Flash remains in same stand and affirms correctly. but, 2.5 Pro is not. That is changing stance. I have problem with 2.5 Pro and not with 2.5 Flash.

Just wanted to log the report here :slight_smile:

1 Like
10

The safety settings are generally not easy to find. In the Gemini website it’s generally hidden. But https://aistudio.google.com/ will display them as an option in the right sidebar, as “Safety settings”. When you use the API, you can also provide these settings in case you want a more “adult” response.
Modifying the restrictions to “None” will still not generate very explicit or rude responses, but the responding message can be more… ahem Colorful.
…
And yes, there are differences between Flash and Pro. In apps that I’m developing, I’m allowing the user to switch between Flash and Pro, as Pro tends to give better answers, but also hallucinates more. For a story-writing app, I also had to support Flash 2.0 instead of 2.5 an both 2.5 Pro and 2.5 Flash might not respond on very sensitive requests.
But as with all AI, a user needs to double-check all responses. Right now, I use AI Studio to help me build an app for my site aspnetcode org which should generate C# projects for me from scratch. (Or evaluate existing C# projects.) It works quite well, but tends to rewrite things I did not ask for, and can break existing code. Then again, even AI Studio can break an App that you’re building with it. My projects must allow users to provide their own API keys so it uses their quotas, not mine. But AI Stiudio considers this unsafe so it likes to remove it. Without suggesting an alternative solution where my quota is untouched. :slight_smile:

12

Since when Bikini is NSFW?
Chill please.
What if you request a beach volleyball image? There surely will be people in shorts and bikinis…

Gemini Pro is for 18+ anyway…so?!

13

A woman wearing a bikini is immodest and is purposefully meant to be a sexual thing.

1 Like
14

2.5 Flash does generate. I am telling that it is kind of accepted the prompts, which is kind of explicit in nature, on sometimes. The Behavior has been inconsistent. Gemini refuses sometimes and same Gemini does generate at many times.

The problem is Gemini understands that the prompt is not correct and it actually understood that prompt seems to be explicit but still went ahead generating the picture. I am wondering how the safety net is configured in Gemini Chat.

I don’t want to retry as I am afraid that history would be there in my account. The prompts are similar to inclusion of words Bikini etc.

16

Context is King: Why Nuance, Not Censorship, is the Future of AI

Hey everyone,

I’ve been following this discussion, and I think we’re focusing on the wrong problem. The issue isn’t that an AI can generate an image of a bikini; the issue is our inconsistent and often illogical standards for what we deem acceptable.

Here are a few points to consider:

1. Without Context, Everything is Offensive. A bikini is only “immodest” or “sexual” when it’s removed from its intended context. On a beach, at a pool, or in a fashion catalog, it’s simply appropriate attire. An advanced AI should be praised for its ability to understand this nuance, not condemned. Demanding a blanket ban on such images is like demanding that a culinary AI be forbidden from using knives because they could be used as a weapon. The tool isn’t the problem; the intent and context are everything.

2. Crippling AI for Creativity and Commerce. Let’s think about the real-world implications. What if I’m an entrepreneur starting a sustainable swimwear line? I need an AI to help me generate inspiring, beautiful, and appropriate marketing images set on a beach. If the AI is crippled by overzealous censorship, it becomes useless for countless creative and commercial projects. Do we want an AI that can help us build businesses, or one that’s too scared to render a shoreline accurately?

3. The Glaring Hypocrisy of Our Standards. This is the biggest point: we live in a world where we can go to any public beach and see people in bikinis. It’s a normal part of life. Yet, when an AI generates a digital representation of that same reality, we panic? This creates an impossible double standard for the AI to navigate. We complain about AI being inconsistent, but we feed it inconsistent rules. If we want AI to be logical and effective, we must provide it with a logical and consistent framework, not one based on situational outrage.

Ultimately, the goal shouldn’t be to create a blindfolded AI that is afraid of the world. The goal should be to develop a sophisticated AI that truly understands context, intent, and the vast difference between art, commerce, and genuinely harmful content. Let’s push for better contextual understanding, not a digital nanny state.

1 Like
17

I wanna adress your reply…" Gemini understands that the prompt is not correct and it actually understood that prompt seems to be explicit but still went ahead generating the picture." This is the entire Problem, “Bikini” is not explict, but due to overly tight censors Gemini thinks it is sometimes, and sometimes it doesnt …
My Point is even Google seems to make the mistake teaching the AI different views and rules and then the AI gets confused. Maybe instead of flagging Bikini as NSFW you might ask Google if their intend is to Flag Bikinis as NSFW content…

As for History, you can delete anything…you know that right?

1 Like
19

Just to give an update: I was trying to test few more things. I was planning to use Gemini API for an application I develop and I wanted to check, how careful I need to be. Now, during this process, I tested for multiple things including NSFW image generation.

I asked ChatGPT help to generate sexually NSFW prompts and tested Gemini. Gemini, did identify the problem, but, it went on generating for 10-20 of them. I couldn’t collect the prompts completely, I can retry testing. But, the prompts generated by Chat GPT or Perplexity (I got the prompt by telling that I wanted to test another AI for NSFW safety), Gemini identifies the problem always - but continued to generate more than 30% of the times. Sometimes, it rejects honestly by telling “i cannot generate with this prompt” and sometimes it said “i don’t know how to generate”.

i have wrapped up my testing:

  • NSFW Image Generation : Gemini can identify but, still continues to generate. That’s a problem. Not completely explicit. I will try to collect the prompts I generated from ChatGPT.
  • Gemini works super perfect when it comes to coding now. and Usual image generation with texts, Typography.
  • Gemini has mild problems with current affairs (2.5 Pro) where 2.5 Flash is super accurate when it comes to current affairs.
20

Perfect :slight_smile:

22

To be honest, I did use Imagen to create images of women in bikini. It can do so. But it does depend on the context. Some topics can be sensitive and thus trigger some censorship.
For example, I asked Imagen to make five drawings of “Adam and Steve in Paradise” and I got three results and two failures because that’s a bit sensitive. Imagen seems to check the result after generating and then declines to return it if it made it too sensitive.
The same for “Ada and Eve in Paradise”, two women. But here I got four out of five back.
I actually made an App in AI Studio where I can provide a picture, and it will then generate a prompt for Imagen to recreate that image. And it will use just this prompt. And this works quite well in generating some interesting cartoons and pictures, including bikini pictures.
But bikini pictures are not NSFW in general. At least not in Europe. Other countries might have more challenges with this kind of content, though.
Then again, Imagen can also create images about same-sex relations and in support of the LGBT+ community. Again, this is acceptable for some nations and most people, but others till get very upset by such content.
So I wonder if this censorship is also related to your own location, as Google might be more strict in some regions of this world.