4
Image-generation technology will continue to evolve,
become more advanced, and likely become more photoreal,
increasing the escalation potential for harm. There is no
foolproof way to make these models entirely safe for use.
To start solving issues of AI-generated images fueling dis-
information, harmful propaganda, NCII, and other safety
harms, the United States must first develop solutions in
regard to authenticity and improve the public’s access to
information about these images.
The first step in navigating this shift is to comprehend
how AI image generation tools produce safety issues and
challenges and how current safeguards are insufficient to
tackle the problem.
Safety Challenges in Image
Generation
Safety challenges in the AI image generation landscape
begin at the technical level, and the most significant safety
challenges are due to biases and harms from training data,
the existence of open-source image models, and a piece-
meal approach to content moderation at the user level. The
current AI image generation space mainly consists of text-
to-image diffusion models, such as Midjourney, DALL-E,
and Stable Diffusion, that generate images based on user
prompts. Understanding the fundamentals of safety issues
in text-to-image diffusion models can show why these
models can produce unsafe images. Furthermore, diving
deep into training data, open-source models, and content
moderation reveals that these technical mitigations are
simply not enough to prevent the generation of harmful
content, and the United States needs authenticity solutions
to manage risks and harms.
Image generation models reflect societal and represen-
tational biases on the internet because they are trained on
data scraped from the internet. For example, there are far
more images sexualizing women on the internet compared
with similar images of men and far more images of men
in professional roles (doctor, lawyer, engineer, etc.) than
similar images of women. Image models conceptualize
these representational biases and have become quite adept
at generating content that both oversexualizes women
and highlights men in professional positions (Heikkilä,
2022). Recent research shows that there is still much work
to be done to make image models safer and less biased, as
there are still severe occupational biases in models, which
result in the exclusion of groups of people from generated
results (Naik and Nushi, 2023). Furthermore, Rest of World
in 2023 conducted an experiment with Midjourney that
showed that the tool typically represented nationalities
using harmful stereotypes: Images of a “Mexican person”
mainly showed a man in a sombrero, and images of New
Image generation models
reflect societal and
representational biases
because they are trained
on data scraped from the
internet.
剩余21页未读，继续阅读
每天读点书学堂
- 粉丝: 1039
- 资源: 1万+
我的内容管理
展开
最新资源
- C语言快速排序算法的实现与应用
- KityFormula 编辑器压缩包功能解析
- 离线搭建Kubernetes 1.17.0集群教程与资源包分享
- Java毕业设计教学平台完整教程与源码
- 综合数据集汇总：浏览记录与市场研究分析
- STM32智能家居控制系统：创新设计与无线通讯
- 深入浅出C++20标准：四大新特性解析
- Real-ESRGAN: 开源项目提升图像超分辨率技术
- 植物大战僵尸杂交版v2.0.88：新元素新挑战
- 掌握数据分析核心模型，预测未来不是梦
- Android平台蓝牙HC-06/08模块数据交互技巧
- Python源码分享：计算100至200之间的所有素数
- 免费视频修复利器：Digital Video Repair
- Chrome浏览器新版本Adblock Plus插件发布
- GifSplitter：Linux下GIF转BMP的核心工具
- Vue.js开发教程：全面学习资源指南
资源上传下载、课程学习等过程中有任何疑问或建议，欢迎提出宝贵意见哦~我们会及时处理！
点击此处反馈
