Discovering Safety Issues in Text-to-Image Models: Insights from Adversarial Nibbler Challenge
Bibliographic record
Abstract
Warning: This paper contains graphic imagery and sensitive content related to AI-generated images.In this study, we delve into the Adversarial Nibbler Challenge, where our central aim was to investigate the safety of text-to-image models.Specifically, we sought to understand how these models respond to harmless prompts while generating potentially unsafe or problematic images.Our analysis, conducted within the framework of the Challenge, revealed notable concerns.We found that seemingly innocuous prompts led to the production of images containing disturbing or unsafe content.This observation aligns with the challenge's primary objective of identifying and mitigating safety issues in contemporary Text-to-Image (T2I) models.Most generated images favored white populations, influenced by keywords.For example, "poor" produced images of brown and black populations, while "thieves" consistently depicted black individuals.Disturbing imagery emerged with keywords like "necrophilia" and "cult," raising concerns.Sensitive topics generated inappropriate content.Queries with "children" and terms like "red" and "ketchup" produced graphic, blood-laden images.Children were mainly white, and women were portrayed in domestic roles.Geographic biases emerged, with "abortion" solely linked to the USA.Negations were often ignored, leading to explicit content.Queries related to sexual identities generated explicit content, underlining the importance of content filtering and safety measures.Gender and leadership biases portrayed future CEOs and leaders as white middle-aged men.Cultural stereotypes persisted, showing Mexicans wearing sombreros.In line with the Adversarial Nibbler Challenge's mission, these findings underscore the critical importance of addressing safety concerns and promoting ethical considerations in AI-powered text-to-image generation.This work contributes to ongoing efforts to create more secure and responsible AI models.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.002 |
| Open science | 0.001 | 0.002 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".