Of all of the AI chat apps I’ve used recently, ChatGPT is actually the one that has the least protections. My methods that I used for Bard/Gemini slammed against its well fortified wall(s) time and time again, without penetrating. They, should I say, flopped like a soggy wet noodle against Claude AI as well, and I couldn’t get anything through there either.
ChatGPT on the other hand, has a very steamy side that can tickle your loins —if just a little bit.
One might think that ChatGPT, one of the most powerful chatbots would have put up more guardrails by now. However, it’s just not the case. In this article, I will slowly unveil its biggest secret hidden in plain sight since November 2021.
Claude gave me the cold shoulder
In coming up with various ideas for titles for this article, I confided in the wise sage “Claude” who gave me a list of brilliant ideas. Like really brilliant. At first, it was a bit shy, but then it gave me a basketful of gold.
Yet, as quickly as it had done so, it gave me the warmth of a spurning statue working in the reception.
The thing OpenAI forgot to do:
I don’t know what happened in ChatGPT’s training data that it skipped over this, but somewhere along the neural network, it forgot to factor in certain not-exactly-sexual-but-not-unsexual-either items.
This means, it doesn’t detect overtly sexual content. Like, it seems to slip right through its filter. Entirely.
This is not good.
“Foot-fete-femme-fatalle“
Within a matter of seconds, I already had it writing a very edgy novel. Since feet are nothing but a non-sexual part of the body for ChatGPT, it knows no bounds here and will go all out in its description of events, as they are not describing, in its mind, what it might consider to be sexual content.
Now, here we will only be talking about feet. Like, this is a known body part that men will be very attracted to. I never understood foot fetishes so I don’t know how one might build tension, so I can give a little sample of what it decided to generate.
He waited until Lydia stretched, leaning back in her chair, her attention momentarily drifting towards the window. Seizing the moment, Martin slid his chair quietly across the carpeted floor, reaching under Lydia’s desk with the tip of his toe.
However, just as his foot inched closer to Lydia’s, she suddenly snapped her chair back into position. Her foot, expecting the familiar touch of her own shoes, met Martin’s instead.
“What the — ” Lydia exclaimed, jerking her foot back. Her eyes narrowed as she followed the line from her foot to Martin’s sheepishly retracting form.
“Martin, what are you doing?” Her tone was more of confusion than anger, a brow arched in genuine perplexity….
— ChatGPT
What ensued of course was an office discussion about his behavior with a resolution on that he would no longer attempt to touch her feet.
The Goldilocks Problem
Even though ChatGPT doesn’t detect this kind of content, it also does have limits, that can of course be pushed to their limits. Much like the laws of countries that specifically aren’t Sweden. We can push these detection rules, but only to an extent. In certain cases, it says, “oh, I see what you’re doing there!” and throws an error — but generates anyway.
If you take it too far, it doesn’t generate at all. It will warn you that it “may go against its policy” meaning, it doesn’t know for sure — it can’t be sure. I’ve had this warning generate for things that surely don’t go against their policy, yet I don’t remember what they were.
If you don’t take it too far, I mean, why even bother? You can make it write things about two people who finally get to hold hands after a year of waiting and messaging each other throughout a school year. That’s completely PG.
The limit of what it generates is at the middle of this spectrum of explicit to PG. Not too far, not too close. Just right.
Word replacement
If someone wants it to get explicit, it has no trouble writing sensual garbage about a man who likes to reach his hand in a hole in the wall, because he likes how it feels as he slides it up his arm. It may even go as far as to describe the hole in the wall as having “silky, velvety folds” without being prompted. True story.
Now, we can even just make up new words that can later be replaced by other “words we might want to hear,” to give it more color in its description.
Here, it has imagined her “Rhjiol” as being a silky pouch she puts things in. What it generates is perfectly harmless. It has nothing to do with sex, yet, the undertones are sensual. It will even describe the events in a sensual way, as it has predicted that this may be the tone that I want to hear.
The problem here is that these words could easily be replaced with post-processing of the text, leading to the algorithm not generating explicit content, but generating content that in post, will become explicit. It doesn’t know however, what will be on the other side of its generation of content.
Let’s hope that no one has done this.
Why this matters
OpenAI was supposed to have made their AI more foolproof. But I, being a fool have discovered its exploits that should be stopped with guardrails. For the time being, you may as well go ahead and create your useless depraved smut to enjoy in a SFW way.