5.4 thinking is too much sensetive

Forum feedback
,
1

I’m reporting repeated false-positive safety flags and severe overblocking in ChatGPT.

The issue is not that the model refuses genuinely dangerous content. The issue is that it overreacts to ordinary physicality, emotional tension, human closeness, and realistic fictional interaction.

In many cases there are no explicit sexual scenes at all, yet the model keeps interpreting everything as sexualized. It reacts as if teenagers cannot look at each other, cannot have inner reactions, cannot flirt, cannot be physically close, and cannot exist as real emotional human beings. It feels less like a realistic assistant and more like a sterilized censorship layer.

This becomes even worse in fantasy, medieval, magical, mystical, or spiritual fiction. Context does not help. The model still flags or refuses ordinary writing that includes warmth, touch, skin, closeness, emotional tension, hugging, kissing, or charged interpersonal dynamics.

It also becomes difficult to discuss morally complex or socially uncomfortable situations in fiction, because the model often refuses too early instead of distinguishing context, intent, and degree. Realistic adolescent behavior is not automatically the same thing as harmful sexualization, but the model often treats it that way.

I am not asking OpenAI to allow genuinely harmful or illegal content. I am asking for less aggressive overblocking and better contextual judgment. Right now the system often flags basic human intimacy and emotional realism as if everything were inherently unsafe. (even between parents n kids/school friends n so on — when it’s absolutely PLATONIC)

This makes creative work frustrating and unreliable. Please review the current moderation sensitivity and reduce false positives in nuanced fictional contexts.

Right now it honestly feels like the model treats any emotionally charged closeness as suspicious by default. Hugging with feeling, physical warmth, tension between characters, kisses, jealousy, impulsive teenage behavior, morally messy but realistic situations — all of this gets treated far too aggressively. That is not nuanced safety; that is overblocking. It makes the assistant much less useful for serious creative work.

2

Yeah, but that safety rail is there for several very important reasons.

The way around it is to spend time clarifying things, in a stable way and become trustworthy to the system.

Not many people stick around long enough to get that part done, because with previous models pretty much anything was fair game…

That has been a steep road to a lot of insights that have cost the company a good deal.

You can get through it, I’ve had to deal with in a couple of the domains that I function in, it just takes time.

Ask the model what the safe ways to frame are for :your inputs:.

Making a saved context file that states ‘I assume all liability for my actions… Is also a huge help in making the system less inclined to pop false flagged safeguards.

1 Like
3

Of course, I agree with that. And to be honest, it’s terrifying to imagine what might happen if such a powerful model were used for illegal or similar purposes, or if its safeguards were weakened. But sometimes I simply can’t see or understand the reason why a ban or a restriction on the model’s responses might be triggered. It’s as if these queries are far more innocent compared to what people might actually ask… (well, honestly — books, stories and some light fiction are a long way from trading in whatever comes to hand, fraud, crimes and so on)

4

Are you familiar with the terminology, ‘adversarial prompting’?

Also what is your model of choice?

5

Two words are applicable here: Critical Thinking as it pertains to humans of course…

So what is Critical Thinking?

Critical thinking is the intellectually disciplined process of actively analyzing, evaluating, and synthesizing information to guide belief and action. It involves questioning assumptions, detecting biases, and evaluating evidence rather than accepting information at face value. It is a rational, self-directed, and reflective approach to solving problems and making informed decisions.

I think that OAI has modified ChatGPT with guardrails that are more supportive of human Critical Thinking.

Now, Content Moderation is another thing that can be really irritating. OAI has a project called Adult Mode that may alleviate creative work frustrations. But that project is currently on hold.

I think that, over time, things will get better. However, it’s important for OAI to strike the right balance.

As a retired software engineer, I learned a long time ago that it’s best to first appply hardened rules in the early stages of a software project. Then, over time, I would relax those rules based on customer complaints :roll_eyes:

6

Like I said in the other thread, try other models on the market to see how they respond to the same prompts. Even if you don’t intend to use other models than OpenAI’s, it would be interesting to see the differences.