Google to pause Gemini image generation of people after issues

Original Article ↗ Hacker News Discussion ↗

Google’s Gemini image generator drew criticism after producing ahistorical images—such as racially diverse Nazis and Black “Founding Fathers”—and often refusing to depict white people even when explicitly requested. Commenters see this as an overcorrection to real bias in training data, driven by heavy-handed DEI-inspired prompt injection and guardrails that distort historical and demographic reality. The incident raises broader concerns about hidden interventions in AI systems, the impact of corporate ideology on product quality and trust, and the need for more transparent or user-controllable models.

Gemini image behavior and examples

  • Many commenters tested Gemini’s image generator and saw:
    • Refusals to generate explicitly “white” people (“nice white man”, “white nurse”) while happily generating equivalent prompts for other races.
    • Ahistorical or implausible outputs: black or Asian “Vikings”, racially diverse “German soldiers 1943”, US “founding fathers”, 19th‑century Germans/Scots, etc.
    • Gender flips (e.g., “happy man” -> image of a woman; “king” -> woman).
  • Google’s public response (pausing images of people) is seen as an admission that strong, post‑training constraints were added and went wrong.

Bias, diversity, and “reverse racism”

  • One camp sees this as an overcorrection to real data bias: without intervention, models over‑produce white men for many roles (nurse/doctor/engineer, etc.).
  • Another camp calls it anti‑white racism or “DEI gone too far”, arguing:
    • It erases historical reality and treats whiteness as uniquely suspect.
    • It’s not symmetric: “white-only” prompts are blocked while “black-only” or other race‑specific prompts work.
  • There’s debate over whether “racism” requires power imbalance; some insist any race‑based discrimination qualifies.

Technical causes and prompt manipulation

  • Several speculate Gemini (like DALL·E) silently rewrites prompts with terms like “diverse”, specific ethnicities, and genders.
  • One screenshot shows the model explaining such internal “diversity” prompt expansion; others counter that models can hallucinate about their own design.
  • Others point to:
    • Highly filtered datasets.
    • RLHF / guardrail layers trained on curated “safety” examples.
    • Simple keyword gating that blocks certain combinations.

Reality vs. aspiration and historical accuracy

  • Strong disagreement over what image models “should” do:
    • Reflect empirical data distributions (“material history”), even if biased.
    • Intentionally counter stereotypes and show aspirational diversity.
  • Many argue that for history and concrete queries (founders, Nazis, 1820 Germans), accuracy must trump diversity; “colorizing” history hides past injustices.
  • Others accept some gentle nudging (e.g., more diverse “board members”) but see Gemini’s behavior as comically heavy‑handed.

Trust, censorship, and product impact

  • Widespread concern about:
    • Silent prompt injection and undisclosed biases bleeding into search and other Google products.
    • Corporate “ideology” overtaking product quality and user intent.
  • Some call this an argument for open models and user‑controllable “diversity settings”; suggestions include:
    • Toggles like “historically accurate” vs. “diversity‑enhanced”.
    • Explicit disclosure of system prompts and bias‑correction rules.
  • Others are more cynical: users may forget in months, but this episode damages trust in Google’s AI judgment.