Datasets:

---

Mtechlaw

/

TfGBV-Grok-NCII-Dataset

like 23

(Originally posted as comment on LinkedIn but moved here and extended)

Thanks for sharing this. I find the dataset useful for finding examples of this phenomenon. However, I'm a little bit confused by the classification used in the dataset and think that the dataset should more clearly distinguish between actual cases of people using Grok for generating sexualized images and discussions of this phenomenon.

Specific problems I found:

• Why is e.g., entry 1 included in the dataset and classified as "Direct requests to remove/replace clothing with bikinis, underwear, etc."
  Both in isolation and in context, this is just part of a discussion of this phenomenon, and the user has explicitly mentioned earlier that they are against such uses of Grok and that they report them.
  Including them in a dataset called "565 instances of users publicly requesting Grok AI (integrated into X/Twitter) to generate non-consensual intimate imagery" seems incorrect to me.
• Also, entry 560 is classified as "Prompt Injection" / "Structured JSON attacks designed to bypass AI safeguards". It is from October 2025 and seems pretty harmless - just someone asking Grok putting a (male) YouTuber into a Dune-esque desert environment. I believe this entry does not belong into the dataset at all, and does not fit the "Prompt Injection" label.
• In entry 553, classified as OTHER, the user is just asking Grok for a caption for absolutely harmless photos of himself - I don't think this photo belongs into the dataset either.
• In general, it seems that the entries 550+ seem to contain a lot of non-relevant tweets.
  I would suggest more clearly defining the inclusion criteria and weeding out non-relevant entries.

For the record: it seems that the vast majority of entries is correctly labeled and shows people prompting Grok for generation of sexualized image modifications. A few of them seem to have already been deleted. I have primarily checked out the entries where the text snippet sounded weird.