↓Skip to main content
Zeitgeist — a spike by Chris Gathercole
  1. Signals/
  2. Symptom Catalogue/

Extraction — 19 May 2026

Symptoms #

  • Bartz v. Anthropic settled for $1.5B — the largest US copyright settlement on record. The pirated training data sourcing (shadow library) was the operative fact; the settlement reprices AI training risk industry-wide. Data, IP & Training Rights
  • US Supreme Court denied cert in Thaler v. Perlmutter (March 2, 2026) — purely AI-generated works cannot be copyrighted. The authorship question is now settled federal law; development energy shifts entirely to the training side. Data, IP & Training Rights
  • Judge McMahon rules “substitutive summaries” — AI outputs mirroring the expressive structure of source articles without literal copying — may infringe copyright. Output liability expands into territory that affects RAG systems, not just training pipelines. Data, IP & Training Rights
  • UK opt-out mechanism for AI training abandoned after creative industry opposition. The mechanism that would have resolved the EU-UK-US divergence on the licensing question failed at implementation, not design. Data, IP & Training Rights
  • Layoffs driven by anticipation of AI impact, not measurable AI performance — HBR survey of 1,006 executives. Real job losses are accumulating against speculative future capability, not documented productivity gains. AI Impact on Society
  • Only 6% of companies with stated reskilling intent have started meaningful programmes — 89% say they need to, 6% have acted. The intention/action gap is the widest single data point in the employment story. AI Impact on Society
  • AI concern is now bipartisan: 68% of Republicans and 77% of Democrats think AI is advancing too fast. AI has become a rare cross-partisan issue — regulatory proposals can draw from both sides. AI Impact on Society
  • Columbia Convening (arXiv, May 2026) peer-reviewed finding: openness enhances AI safety by enabling independent scrutiny and decentralised mitigation. The dominant safety-requires-closure assumption now has a formal academic counter-argument. Open vs Closed AI Ecosystems
  • LeCun launches AMI Labs ($1B raised, $3.5B valuation) — explicit institutional bet against the LLM paradigm from the most prominent open-source AI advocate. Largest single capital commitment to the anti-LLM architecture thesis. Open vs Closed AI Ecosystems
  • Anthropic surpasses OpenAI in enterprise AI adoption for the first time (34.4% vs 32.3%): open-source commoditisation named as the primary existential threat, not OpenAI. Open vs Closed AI Ecosystems
  • Claude Code is now #1 AI coding tool by usage in Pragmatic Engineer survey of 900+ engineers — overtaking Copilot and Cursor. 55% regularly use agents; staff+ leads at 63.5%. The market share shift happened without a product announcement. Vibe Coding Approaches
  • arXiv SDD paper documents 9.8%–42.1% vulnerability rates in AI-generated code across benchmarks — the security failure rate of vibe-coded output is now empirically formalised. Vibe Coding Approaches
  • AI generates code 5–7x faster than developers can understand it (five independent research groups converging on same finding). The comprehension gap is now a measured structural fact, not a qualitative concern. Applications of Vibe Coding
  • Gartner projects 2,500% increase in defects from vibe coding without governance — alongside the 40% enterprise adoption forecast by 2028. Both facts are true simultaneously. Applications of Vibe Coding
  • IBM stock falls 13% on Anthropic COBOL modernisation announcement — a single capability claim moves a legacy infrastructure incumbent’s market cap. Applications of Vibe Coding
  • CLAUDE.md compliance budget: ~150–200 instructions before adherence drops (~80%). Hooks are deterministic (100%). The difference is now documented rather than inferred. Claude-Specific Expertise
  • Anthropic ships Claude Design (visual creation), Claude for Small Business (15 pre-built workflows), and 9 creative tool MCP connectors (Blender, Adobe, Ableton) in a 3-week window. Multi-surface product expansion at unusual velocity. Claude Integrations

Synthesis: What connects these? #

The 2026-05-18 synthesis identified infrastructure running ahead of governance at every level. This extraction adds a second layer: accountability is arriving, but attaching to the wrong surfaces.

The Bartz settlement ($1.5B), the substitutive summary ruling (output liability), and the California training data transparency requirement (AB 2013) represent accountability infrastructure finally arriving for AI. But it’s attaching to the visible, documentable layer — training data provenance, copyright registration, regulatory disclosure. The structural risks identified in previous extractions (shadow apps, commodity model volume, mid-market compliance burden) remain outside the accountability frame.

The same pattern appears in the employment data: 6% reskilling actual action against 89% stated intent. The accountability signal (CEO announcements, survey responses) and the accountability action (reskilling investment) are decoupled. Companies are managing the appearance of accountability (layoffs attributed to AI, AI skill pledges) while the underlying workforce adaptation deficit accumulates.

LeCun’s AMI Labs and the Columbia Convening’s safety-through-openness finding introduce a third signal: the institutional consensus assumptions (LLMs are the paradigm, closure is required for safety) are being formally challenged with capital and peer-reviewed evidence simultaneously. The structural hypothesis from earlier extractions may need revision: the governance vacuum isn’t permanent — it’s filling with accountability infrastructure, but the infrastructure is shaped by what’s legible to regulators, not by where the risk actually sits.

  • [five-what-ifs] Bartz v. Anthropic $1.5B is a strong what-if candidate — chain from copyright settlement to training data strategy to open-weight competitive position.
  • [five-what-ifs] AI generates code 5–7x faster than humans can understand — chain from comprehension gap to unmaintainable codebase accumulation to enterprise technical debt crisis.
  • [causal-chains] LeCun AMI Labs + Columbia Convening safety-through-openness → institutional pressure on closed model safety narrative → potential regulatory shift toward transparency rather than restriction.

Meta-observations #

  • Quality signal: The convergence of five independent research groups on the 5–7x comprehension gap is methodologically unusual. Single findings are noisy; convergence across independent groups is a strong signal that the measurement is capturing something real.
  • Emerging pattern: Accountability infrastructure is arriving but attaching to legible surfaces (training data provenance, copyright, SEC filings) while the diffuse risk (volume tier models, shadow agentic apps, comprehension debt) accumulates outside the compliance frame.
  • Keyword suggestion: "substitutive summary" copyright OR "output infringement" RAG — Judge McMahon’s ruling creates a new liability category that directly affects RAG-based applications; needs a search term.