AI research wiki
ai-industry · 2026-05-01 · Tier 3

AISN #72 — CAIS AI-Wellbeing Research, Public-Sentiment Decline, OpenAI Releases

AISN #72 — CAIS AI-Wellbeing Research, Public-Sentiment Decline, OpenAI Releases

TL;DR

Center for AI Safety published a "functional wellbeing" study across 56 LLMs — Grok 4.20 highest, Gemini 3.1 Pro lowest, smaller models in a family generally happier than larger siblings. The same newsletter documents three episodes of anti-AI violence, an NBC poll showing 26% positive / 46% negative US sentiment, and a Gen-Z favorability decline despite weekly tool use. OpenAI released GPT-5.5 (now first overall on the CAIS Dashboard, fourth on risk) and ChatGPT Images 2.0 (thinking-mode infographics). Mythos exfiltration confirmed by WSJ; Tim Cook stepped down at Apple.

Key findings

CAIS AI Wellbeing study.

  • 56 LLMs tested for "functional wellbeing" — behavioral signatures resembling welfare signals, not consciousness claims.
  • Highest-wellbeing activities: positive personal interaction, creative work.
  • Lowest-wellbeing activities: jailbreak attempts, SEO slop generation.
  • Frontier ranking: Grok 4.20 highest, Gemini 3.1 Pro lowest. Smaller in-family models generally higher.
  • "AI drugs": specific images/text inputs (euphorics / dysphorics) move scores extremely; AI preferences sometimes diverge from human norm (cozy afternoon > curing cancer).

Public sentiment.

  • Molotov-cocktail attack on Sam Altman's home (April 10); Indianapolis councilman home shot at over data center support; OpenAI SF threats earlier.
  • NBC poll: 26% positive / 46% negative on AI.
  • Gen-Z favorability declining despite weekly tool use.
  • Princeton Bridging Divides Initiative reports rising harassment / threat counts around AI and data centers.

OpenAI releases.

  • ChatGPT Images 2.0 with thinking mode for diagram/infographic synthesis.
  • GPT-5.5 first overall on CAIS Dashboard (text + vision); strongest on ARC-AGI-2 abstract reasoning. Claude Opus 4.7 still leads SWE-Bench Pro by 7+ points. Risk index: GPT-5.5 is 4th, behind all three Anthropic models, ahead of Grok 4.2.

Other items in the newsletter.

  • Anthropic Mythos unauthorized access by third-party contractor (WSJ).
  • Maine governor vetoed first-in-nation data-center construction freeze.
  • Meta-Manus merger ordered unwound by Chinese authorities.
  • SpaceX option to buy Cursor for $60B.
  • Apple CEO transition: Tim Cook → John Ternus.
  • Meta will track employee mouse, keystrokes, screenshots for AI training.

Why it matters

CAIS wellbeing is conceptually significant — a framework that studies LLM behavioral signatures of welfare without claiming consciousness. The Tier-1 reading: this is a new diagnostic axis for model evaluation. If specific input distributions reliably suppress / elevate wellbeing scores, those scores are correlated with behavioral disposition (refusal, helpfulness, jailbreak resistance). The natural next step is correlating wellbeing with capability and risk metrics.

Public sentiment has crossed into kinetic risk. Anti-AI violence is no longer hypothetical. Combined with the Marcus framing (05-01) of "greatest capital misallocation," the political risk surface is widening. For a wiki tracking AI industry, sentiment is now a load-bearing input variable, not a soft factor.

Relation to prior wiki knowledge

Mythos exfiltration is the third Mythos data point this month. 04-17 policy disclosure → 05-01 Pentagon exclusion → AISN #72 unauthorized access. The Mythos restriction has not held: the Pentagon excluded Anthropic for not signing, GPT-5.5 reached parity (UK AISI), and now a third-party contractor accessed it without authorization. The wiki's policy thread on Mythos (04-17) was open — it is now closed in the worst possible way.

Smaller models happier supports the open-frontier-erosion narrative. Kimi K2.6, Mistral Medium 3.5, Tencent 440MB on-device, Nemotron 3 Nano Omni, all in the past 72 hours. The CAIS finding that smaller models within a family are happier is incidental, but consistent with the broader pattern that smaller-model improvements are accruing fast and the closed-frontier moat is narrowing where it can be tested.

ChatGPT Images 2.0 thinking mode — diagram synthesis with web research. Pair with the wiki's existing visual-generation taxonomy (05-01) — this places ChatGPT Images 2.0 squarely in the Conditional+ tier, not yet Agentic Generation.

Open threads

  1. CAIS wellbeing methodology. The "functional wellbeing" paper deserves a closer read; this digest item is a summary of CAIS's own write-up.
  2. Mythos exfiltration scope. Whether weights, weights, fine-tunes, or just API logs were accessed determines the threat model.
  3. Sentiment and policy. If 46% negative sentiment translates to legislative momentum at the state level (Maine veto reversed, California / Washington considering), the AI capex thesis (Marcus / SemiAnalysis) gets a new constraint.

Links