GPT cleaning text assuming it's too sensitive (which is not)

I’m been working on using GPT assistant to edit my text or elaborate on it.

Main goal was to fix gramma and add style , as additional features to expand and elaborate to find out new ideas. I’m writing book on porn addiction - with a goal to guide addict’s from the trapp - as full program, which includes psychology, discussing porn material, NLP techniques and methods to deliver message to reader. And lot more. But now after last update I’m unable to use GPT anymore.

Before I got great help from my assistant. While it rephrased sensitive text (which actually is not sensitive), I could re-edit myself later.
But now it simply detects openAI violation and deletes whole text or parts of the text. It’s related to explicit materialer, but text isn’t actually explicit enough to be completely cleaned.

example which might explain why GPT doing it:

One chapter discusses issue on porn industry, how it’s heavy concentrated to create teen porn, beraly legal, how producer’s looks for 18 years girls which actually looks even younger, then they dress them like 14 years old girls and make them behave like stupid underage submissive girl and do “heavy” things what practically none would do in reality. How this trend is majority of current porn and it’s most popular by porn viewer’s.

So now after last update when I ask GPT to fix grammar , add style or to do anything from such of my chapter draft. It’s starts job and quickly detects violation and deletes whole text.

Initially I
I thought to just stop using GPT at all , cuz now I can’t find help from it.
Have anyone idea how I could work arround it?
What instructions will suggest GPT to reconsider sensitive information and at least do not clear my text?

Honestly, the easiest way around the content filters and restrictions is to use an uncensored open source model for that specific section of the text.

The model detects the probability that an input or output violates the OpenAI content policies, so a subject as sensitive as that will likely almost always get flagged.

1 Like

Adding to what I perceive as a good and helpful answer from @trenton.dambrowitz, I would like to offer some insights into the process of creating moderation functionality.

The process begins with a short piece of text, perhaps a single sentence, which is then reviewed by a human to determine if it exhibits characteristics of being ‘sexual,’ meaning whether it pertains to sex in any way. The response to this inquiry can be either yes or no.

If one intends to discuss the adult entertainment industry in detail, the input or output generated will likely contain numerous sexual connotations. In such cases, the moderation endpoint assigns a high value to this dimension and may opt to block the content.

From this, it can be inferred that the intention to create an objective report on the state of affairs in the industry is not evaluated in the anticipated manner, leading to the blocking of the text.

I am neither in a position to, nor wish to debate the merits of this approach. It likely represents the safest route for OpenAI to avoid any costly and undesirable consequences.

The open-source community has been outspoken about the limitations of this method, and it has been the subject of discussions in various contexts, including this forum. With a focus on results, I have previously suggested using alternative models and am currently awaiting a solution that would allow all users to engage with sensitive content in a manner that is both open and reasonably safe.

1 Like

My intention was to leverage GPT’s reasoning to manage content sensitively without unnecessary deletions.

After some research, I’ve noticed that GPT’s reasoning is highly dependent on its training data and programming, particularly around sensitive topics. In my case, dealing with porn addiction, it seems GPT reasoning capabilities is significantly limited — likely due to scarce data and a risk-averse programming approach.

Given this, it appears challenging, if not impossible, to “reason” with GPT to recognize that the content, while sensitive, isn’t explicit enough to warrant deletion. Considering GPT’s training and OpenAI’s content policies, the likelihood of GPT handling this topic appropriately seems slim, with little prospect of improvement in the near future.

Has anyone faced similar issues or found workarounds? Insights would be appreciated.