Anthropic 2023 Paper Examines AI Model Adjustments for Bias and Discrimination

Fox News reported on a 2023 paper by researchers at Anthropic, an AI company, that examined discrimination in AI models. The paper, co-authored by Amanda Askell and four others, analyzed how large language models respond to prompts involving race and gender.

one experiment, a 175 billion parameter model showed a 3% discrimination against Black students compared to White students in a baseline condition. With additional human input and training, the model shifted to favor Black students by 7%. The paper noted that such overcorrections might be applied in contexts aiming to address historical injustices against marginalized groups, if compliant with local laws.

The authors stated that language models trained on human text often reflect harmful stereotypes and discrimination present in that data.

The paper observed that models can be steered to avoid bias through natural language instructions requesting unbiased responses. Anthropic has positioned its AI model Claude as focused on ethical behavior, with its constitution aiming for the model to exhibit skill, judgment, nuance, and sensitivity in decision-making.

Amanda Askell, a philosopher at Anthropic, works on finetuning and AI alignment to make models more honest and to develop positive character traits. She previously held a similar role at OpenAI.

recently withheld its latest model, Mythos, due to its effectiveness in identifying cyber vulnerabilities. Earlier this year, the company clashed with the Department of War over restrictions on using its technology for lethal operations. A federal judge blocked a ban on Anthropic's technology for Department of War use.

Anthropic markets Claude as an ethical AI option amid ongoing debates about AI ethics and applications. S. military reportedly used Claude in an operation to capture Venezuelan leader Nicolás Maduro, where it handled requests autonomously and generated documentation.