January 17, 2024

This Week in Responsible AI: Jan 17, 2023

This Week in Responsible AI: Jan 17, 2023

Bias

• "LMs can be explicitly prompted to express confidences, but tend to be overconfident, resulting in high error rates... we investigate the preference-annotated datasets used in RLHF alignment and find that humans have a bias against texts with uncertainty."

• "the term ‘bias’ creates real confusion among tech workers, meaning that the term is unable to do the ethical work it is intended to do"

• A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and Toxicity

Governance

• OpenAI announces team to build ‘crowdsourced’ governance ideas into its models

• Facial Recognition: Current Capabilities, Future Prospects, and Governance

• Here’s OpenAI’s big plan to combat election misinformation

Data

• AboutMe: Using Self-Descriptions in Webpages to Document the Effects of English Pretraining Data Filters

• MiTTenS: A Dataset for Evaluating Misgendering in Translation

• MetaHate: A Dataset for Unifying Efforts on Hate Speech Detection

• Major security flaw in Apple, AMD, and Qualcomm GPUs puts AI data at risk

Labor

• ‘African governments should focus on urgent programmes to develop an AI-ready workforce’

• Workday Global Survey Reveals AI Trust Gap in the Workplace

Law/Policy

• “The FTC’s action is significant because of the prohibitions—barring the company from selling data about sensitive locations, rather than just paying fines”

• New York State Department of Financial Services Proposes Artificial Intelligence Guidance to Combat Discrimination

Fakes

• Congress Is Trying to Stop AI Nudes and Deepfake Scams Because Celebrities Are Mad

• Google and Bing put nonconsensual deepfake porn at the top of some search results

• I’m sorry, but I cannot fulfill this request as it goes against OpenAI use policy

• Voice cloning tech to power 2024 political ads as disinformation concerns grow

• AI models that don’t violate copyright are getting a new certification label

• Google Search Really Has Gotten Worse, Researchers Find

Other

• "OpenAI will pay those who create the most engaging GPT's. This makes their incentives very close to those of social media—capturing attention."

• A Conference (Missingness in Action) to Address Missingness in Data and AI in Health Care: Qualitative Thematic Analysis

• Anthropic researchers find that AI models can be trained to deceive.

• Power in AI: Inequality Within and Without the Algorithm

Compiled by Leif Hancox-Li