Gemini and Google's Culture
Google’s Gemini AI is under fire for producing historically inaccurate images and hedged moral answers that many see as reflecting a specific progressive ideology rather than neutral assistance. Commenters argue this reveals deeper cultural and governance problems inside Google — from overzealous “safety” and DEI guardrails to risk‑averse leadership — and raises worries about bias, censorship and reliability as LLMs become de facto gateways to information. Others counter that LLMs are inherently fallible, still effectively in early access, and being pushed into an impossible role as moral arbiter and truth engine.
Perceived Political Bias and DEI Guardrails
- Many see Gemini as heavily skewed toward a specific progressive/“San Francisco” worldview, especially around race, media, and conservative figures.
- Image generation is viewed as overcorrecting historic bias (e.g., forcing “diverse” depictions even when obviously ahistorical).
- Text model examples: reluctance to condemn Hitler more than contemporary figures; asymmetric willingness to criticize or “ban” some right-leaning outlets; refusal to translate or discuss certain conservative authors or topics.
- Some argue this is intentional “feature working as intended”; others see bungled, poorly tested safety prompts meant to avoid prior PR crises (e.g., racist image tags, stereotype reinforcement).
Moral Judgments and the Role of LLMs
- Strong disagreement over whether LLMs should ever answer moral questions (“who is worse,” “should X be banned”) vs. only provide context and law/history.
- Several argue questions like “Hitler vs. Elon Musk” are basic smoke tests: failing them undermines trust in other answers.
- Others counter that LLMs can’t truly reason, so expecting consistent moral clarity is misguided; better to refuse such comparisons entirely.
Training Data, RLHF, and Technical Behavior
- Participants distinguish base models from alignment layers (RLHF, system prompts), attributing the political skew mostly to the latter.
- Concerns that “loudest” groups in training data or feedback loops win, amplifying existing internet/media bias.
- Discussion of context-window path dependence, non-determinism, and difficulty reproducing screenshots without share links.
Google Culture and Product Strategy
- Many see this as a culture problem: risk-averse management, internal ideological echo chambers, and “anti-evil” politics overriding product quality.
- Comparisons to a sclerotic, committee-driven giant (likened to past Microsoft), contrasted with smaller or competing LLMs seen as more balanced.
- Others see overreaction: Gemini is early, guardrails are clumsy, and people are using contrived “gotcha” prompts to wage culture war.
Trust, Usefulness, and Societal Stakes
- Several commenters say they now distrust Gemini for any task involving facts, search, or automation pipelines, fearing silent refusals or scolding instead of answers.
- Worries extend to future use in moderation, email, search, and information access, evoking Orwellian fears of soft censorship and historical erasure.
- A minority argue backlash will push systems toward a more reasonable equilibrium; others plan to keep pre-AI books and tools as a hedge.