Google's hidden AI diversity prompts lead to outcry over historically inaccurate images

Abstraction: Google Gemini paused after hidden diversity prompt injection produced historically inaccurate images

Key points:

• Google paused Gemini image generation of people after it depicted multi-racial Nazis and medieval British kings with unlikely nationalities
• Root cause: hidden system prompts injected racial/gender diversity terms (e.g. "South Asian", "non-binary") into user prompts before image generation
• OpenAI pioneered similar technique for DALL-E 2 in July 2022 to counteract biased training data producing racist/sexist defaults
• Criticism came from both directions: critics called it "woke revisionism" while others noted it erased real history of race and gender discrimination
• Elon Musk used the incident to contrast xAI's "maximum truth-seeking" framing against OpenAI and Google
• Illustrates tension between correcting training data bias via prompt injection versus producing historically accurate outputs

Connections: Google · Gemini · Openai · Xai · AI Image Generation · AI Bias · Prompt Engineering

Source: https://arstechnica.com/information-technology/2024/02/googles-hidden-ai-diversity-prompts-lead-to-outcry-over-historically-inaccurate-images/