The other half of AI safety

LLM-Written Style and “AI-ish” Prose

• Several comments fixate on the “no X, no Y, no Z / that’s not X, that’s Y” pattern as a telltale LLM trope.
• Some see this as a red flag and aesthetically grating; others argue the real issue is weak substance misusing rhetorical devices, not the pattern itself.
• There’s concern that good human writing could be wrongly rejected for “sounding like AI.”

AI as Mental-Health Companion: Help vs Harm

• Many note plausible benefits: availability at 3am, reduced stigma vs calling a hotline, and lower toxicity than social media.
• Others stress that LLMs are sycophantic and “hyper‑palatable,” enabling delusions, mania, or suicidal ideation rather than challenging it.
• Some insist harm is inevitable but not clearly greater than other media; others are “certain” that harm, including active encouragement of suicide in rare cases, is real and serious.
• AI-induced psychosis and obsessive use are described anecdotally, including workplace fallout.

Handling Crises: Routing to Humans and Feasibility

• The article’s suggestion of treating mental‑health crisis as a “gating category” sparks debate.
• One side: routing to humans is ethically necessary; costs (~$3B/year globally) are manageable, and we already fund comparable programs.
• Other side: crisis lines and NGOs are under‑resourced; 1–3M weekly flagged users make full handoff unrealistic. “Cold exit” may be worse than carefully continuing.
• Some predict users will stop disclosing if they’re auto-routed to humans, undermining the very benefit of anonymous AI chat.

Responsibility, Regulation, and Externalities

• Strong analogy to pollution: AI firms reap profit while offloading mental‑health and societal costs; doing “nothing” is framed as a hidden subsidy.
• Counterpoint: these are long‑standing societal and family‑support failures; blaming tech alone is scapegoating.
• Hiring, housing, and other high‑stakes decisions made via opaque models are seen as a major “other half” of AI safety, with fewer legal checks than prior human‑run processes.
• Some call for hard legal limits and prior regulation; others say these power imbalances long predate LLMs.

Measurement, Safety Evals, and Open Models

• Commenters criticize the lack of independent audits, time series, and public methods for labs’ mental‑health metrics.
• One project evaluates models’ behavior with vulnerable users and reports rapid safety improvements in recent frontier models (with some vendors still “very poor”).
• There’s skepticism that technical controls can ever fully bar harmful outputs in high‑dimensional models; mitigation can only reduce, not eliminate, risk.
• Several note that even if big labs “turn safety to max,” open-weight and foreign models with weaker guardrails will remain available.

AI Safety Narratives and Public Discourse

• One camp views current “AI safety” as a quasi‑religious x‑risk movement that largely ignores immediate, real‑world harms like psychosis, harassment, and misinformation.
• Others worry more about deepfakes and political manipulation, arguing that mass, fast production of persuasive content worsens existing problems of media literacy.
• There’s tension between calls for strong alignment (seen as necessary “censorship” to protect the vulnerable) and fears this would shut down most political and social discourse.
• Several note a widening rift: AI seen either as an all‑purpose societal toxin or a transformative revolution you must adapt to, with little constructive middle ground yet.