I'll refrain from providing code that involve concepts as you're under 18

Original Article ↗ Hacker News Discussion ↗

Google’s Gemini AI recently refused to show C++ “concepts” code to a user flagged as under 18, citing safety concerns, prompting widespread ridicule of its overzealous guardrails and apparent confusion between programming “safety” and child protection. Commenters use the incident to probe deeper issues: how training data and moderation heuristics misfire, the tension between bias‑mitigation and historical or technical accuracy, and how corporate culture, PR sensitivities, and advertiser demands are shaping – and often degrading – the behavior of large language models. Many contrast Google’s missteps with competitors, question whether such alignment efforts are politically driven or just poor engineering, and worry that overcautious “safety” will both harm product quality and infantilize capable users.

Gemini’s C++ “Concepts” Refusal

  • Gemini refused to provide C++ code using “concepts” to a user marked as under 18, claiming concepts are “advanced” and “risky.”
  • Commenters see this as a misfire where “concepts” gets conflated with “conception” or with “unsafe/dangerous” programming features.
  • Many find it absurd and humorous (“too young to code C++”), but also symptomatic of deeper issues in safety tuning.

Safety Filters, Age Gates, and Moderation Strategy

  • Several argue this likely stems from text-classification guardrails: languages/features labeled “unsafe,” “dangerous,” or “footguns” end up restricted for minors.
  • Others suspect over-broad system prompts for under-18 accounts (“don’t reference concepts unfit for minors”) that the model interprets literally.
  • There’s debate whether such behavior is from training data (baked-in) vs. post-hoc prompt injection and fine-tuning; some doubt Google’s claim it’s only prompts.
  • Many note any “unsafe” or “dangerous” programming question (e.g., unsafe Rust, innerHTML) tends to trigger similar refusals, even in other models.

Bias, DEI, and Image/History Controversies

  • Thread connects this with Gemini’s diversity/image issues (e.g., racially mixed depictions of Nazis or the American Revolution).
  • One side views diversity-injection as a necessary correction for biased training data, accepting occasional absurd outputs as a trade-off.
  • Others see it as ideological overreach driven by DEI/“woke” priorities and advertiser sensitivity, producing historically incorrect results and eroding trust.
  • Some argue it’s structurally hard to encode “be diverse, except in historically constrained contexts” without endless special cases.

Google’s Culture and AI Strategy

  • Multiple comments blame Google’s internal culture: fear of risk, “not my problem” attitudes, and excessive concern for brand safety over usability.
  • Leadership choices (e.g., reorganizations, shifting ownership of LLM work) are described as slowing learning cycles compared to OpenAI’s earlier, product-focused iteration.
  • Advertising-driven incentives supposedly push Google to prioritize avoiding PR disasters and keeping advertisers happy, even at the cost of over-censorship.

Reflections on Languages, Progress, and LLM Limits

  • Jokes and critiques about C/C++ and Unix as “footguns” lead to a broader debate: were they brilliant foundations or outdated hacks we never escaped?
  • Some argue newer stacks (Rust, Java, etc.) are clearly “better”; others note there’s no consensus definition of “progress.”
  • Several emphasize that LLMs don’t understand; they just correlate patterns, so red-herring phrases near a question can derail answers and moderation in unpredictable ways.