Google CEO calls Gemini completely unacceptable, vows to make structural changes

Original Article ↗ Hacker News Discussion ↗

Google’s Gemini AI has drawn backlash after generating racially and historically inaccurate images, prompting the company’s CEO to label the results “completely unacceptable” and promise structural fixes. Commenters argue the problems stem less from core model quality and more from overzealous “fairness” guardrails, internal culture, and a lack of effective QA or dissenting voices. The incident is seen as part of a broader pattern of declining product quality, ideological bias in big-tech AI systems, and the difficulty of balancing safety, accuracy, and user trust.

Overall Sentiment on Google and Gemini

  • Many see Gemini’s failures as predictable given big-tech culture around “ML fairness” and risk aversion.
  • Some argue this is not a minor bug but a symptom of deep organizational and cultural problems; others think it’s an embarrassing but fixable overcorrection.
  • A minority defend Google’s track record, pointing to dominance in search, email, maps, video, etc., and view Gemini as high‑performing but mis‑tuned.

Guardrails, RLHF, and Prompt Rewriting

  • Commenters believe the core issue isn’t raw model capability but layers of guardrails, RLHF, and prompt rewriting that try to “improve” outputs.
  • Examples discussed: models appending “be diverse”‑style instructions to prompts, using separate moderation AIs, and secretly rewriting user queries before image generation.
  • Some describe this as trying to force a “rose‑tinted” or ideological world model onto systems trained on messy reality.

Diversity vs Accuracy and User Intent

  • One camp supports default diversity (e.g., not all white male doctors) and sees that as reasonable and often beneficial.
  • Another camp sees forced diversity beyond plausibility (e.g., Black Nazis, female popes, non‑white British monarchs) as deceptive, “post‑truth,” or propagandistic.
  • Many stress that if a user explicitly asks for particular demographics or historically accurate depictions, the system should either honor that or clearly refuse, not silently alter the request.
  • There’s debate over whether to aim for:
    • “Proportional to real‑world demographics”,
    • “Proportional to training data”, or
    • Explicitly biased outputs to counteract societal bias.

Google Culture, DEI, and Internal Dissent

  • Multiple commenters claim most rank‑and‑file engineers are skeptical or apathetic about DEI/“anti‑racism” efforts but fear speaking up.
  • Others push back, saying claims about a “silent majority” are anecdotal and that younger staff largely share current moral norms.
  • Some link Gemini’s behavior to a small but powerful internal ideological group; others blame structural cowardice and lack of empowered QA/ethics teams.

Leadership, Strategy, and Competition

  • Sundar Pichai is criticized as a “peacetime” steward lacking vision, late to several markets (short‑form video, videoconferencing, cloud gaming, generative AI).
  • Some call for leadership change; others argue Google still executes well overall and that building GPT‑4‑class models is intrinsically hard.
  • Comparisons are drawn to OpenAI: several note OpenAI’s earlier “nannying” and political bias, but say it has since moved closer to acceptable neutrality, especially in newer products.