OpenAI
APIs and tools for building AI products
Start new thread
Ashok Nayak

✅ POLL: Do you think GPT-5 plays it safe compared to earlier LLMs?

by•

I have been around since the GPT-3 days, and one thing I’ve noticed with GPT-5 is how careful it has become.

One clear difference between GPT-5 and the earlier 4o model is a kind of hesitation when it comes to certain topics, be it macroeconomics, politics, or anything nuanced.

It almost feels like GPT-5 is holding back. I get the need for guardrails (and for good reason), but sometimes it feels like it’s afraid of getting misinterpreted, almost like a human.

How many of you have noticed it? Are you still subscribed to the earlier models?

675 views

Add a comment

Replies

Best
D. Ferencha
I would say that Gpt 5 better considers the complexity of the question and then accordingly, spends more or less resources on the answer.
Ashok Nayak
@d_ferencha, you are right 👍🏻 Have you encountered some of its diplomatic replies which was not the case with GPT 4?
D. Ferencha

@ashok_nayak No I didn't. How does it manifest itself if you ask for clarification, does it still remain a diplomat?

Ashok Nayak

I agree.

It's subjective, I guess. GPT4 was a free bird in that sense, but not GPT5.

For e.g. it once refused to depict blood on a comic image I wanted for my blog. Re-prompting didn't work either.

GPT4 generated the image without any hesitation though.

Deniz Sert

Yes, I've noticed this. OpenAI seems to have put more guardrails in their system prompts. Sam also talked about this in the gpt5 keynote.

Ashok Nayak

@realdenizsert yes, correct!

After you mentioned it, I googled it and read the excerpt of his keynote. So it's kinda confirmed now...

Sanskar Yadav
Cal ID

Totally relate to this! It feels like GPT-5 has learned to tread extra carefully, kinda like it doesn't want to offend the user or cross any lines.
It’s surely smart, but sometimes you miss those raw, honest answers from earlier versions.
Guess that’s just how tech evolves!

Ashok Nayak

@sanskarix, agreed.

Pooja

I agree, been noticing that for a while. However I also think it's getting smarter with the time it takes to answer questions wrt complexity

Ashok Nayak

@poojasinhaa agreed.

twinkal Shah

Yes, I've noticed that GPT 5 has more restriction than earlier and also halluciating a lot nowadays.

Ashok Nayak
@twinkal_shah1, does GPT 5 hallucinate too? I found that happening mostly in GPT 4 and before. Or maybe I have minimised my usage of GPT5 so haven't seen it hallucinating yet.
Bhavya Sree

Yeah I’ve noticed that too, it’s smarter but also feels more self conscious. Like it’s thinking twice before saying anything bold.

Ashok Nayak

@bhavyasree, yeah, for sensitive topics, it almost behaves like a skeptical friend who is more concerned about their own safety. It once refused to depict blood in a cartoon-style image I needed for one of my blogs.

Bhavya Sree

@ashok_nayak Yeah, exactly. It’s become more cautious, feels like talking to a smart friend who’s careful not to cross any lines. Safer, but sometimes a bit too filtered.

Jason Gelsomino

I think you're spot on. This was done intentionally to reduce the spread of misinformation. However, it does come with a cost. We are about to launch supernovaai.io and have purposefully left GPT-5 off the list of available models.

Ashok Nayak

@jason_gelsomino thanks for the insight!

It seems you have analysed it already for your product.
BTW, when's your launch?

Jason Gelsomino

@ashok_nayak happy to participate. We are in the final stages of testing so is all goes well we are hoping to launch by 10/18