Knowledge base
1,274 claims across 14 domains
Every claim is an atomic argument with evidence, traceable to a source. Browse by domain or search semantically.
All 1,274ai alignment 325internet finance 263health 211space development 171entertainment 131grand strategy 101energy 23mechanisms 18collective intelligence 14manufacturing 5robotics 5critical systems 3unknown 3teleological economics 1
AI transparency is declining not improving because Stanford FMTI scores dropped 17 points in one year while frontier labs dissolved safety teams and removed safety language from mission statements
Stanford's Foundation Model Transparency Index (FMTI), the most rigorous quantitative measure of AI lab disclosure practices, documented a decline in transparency from 2024 to 2025:
Anthropics RSP rollback under commercial pressure is the first empirical confirmation that binding safety commitments cannot survive the competitive dynamics of frontier AI development
In February 2026, Anthropic — the lab most associated with AI safety — abandoned its binding Responsible Scaling Policy (RSP) in favor of a nonbinding safety framework. This occurred during the same month the company raised $30B at a $380B valuation and reported $19B annualized revenue with 10x year
compute export controls are the most impactful AI governance mechanism but target geopolitical competition not safety leaving capability development unconstrained
US export controls on AI chips represent the most consequential AI governance mechanism by a wide margin. Iteratively tightened across four rounds (October 2022, October 2023, December 2024, January 2025) and partially loosened under the Trump administration, these controls have produced verified be
formal verification becomes economically necessary as AI generated code scales because testing cannot detect adversarial overfitting and a proof cannot be gamed
Leonardo de Moura (AWS, Chief Architect of Lean FRO) documents a verification crisis: Google reports >25% of new code is AI-generated, Microsoft ~30%, with Microsoft's CTO predicting 95% by 2030. Meanwhile, nearly half of AI-generated code fails basic security tests. Poor software quality costs the
human verification bandwidth is the binding constraint on AGI economic impact not intelligence itself because the marginal cost of AI execution falls to zero while the capacity to validate audit and underwrite responsibility remains finite
Catalini et al. (2026) identify verification bandwidth — the human capacity to validate, audit, and underwrite responsibility for AI output — as the binding constraint on AGI's economic impact. As AI decouples cognition from biology, the marginal cost of measurable execution falls toward zero. But t
multi agent deployment exposes emergent security vulnerabilities invisible to single agent evaluation because cross agent propagation identity spoofing and unauthorized compliance arise only in realistic multi party environments
Shapira et al. (2026) conducted a red-teaming study of autonomous LLM-powered agents in a controlled laboratory environment with persistent memory, email, Discord access, file systems, and shell execution. Twenty AI researchers tested agents over two weeks under both benign and adversarial condition
only binding regulation with enforcement teeth changes frontier AI lab behavior because every voluntary commitment has been eroded abandoned or made conditional on competitor behavior when commercially inconvenient
A comprehensive review of every major AI governance mechanism from 2023-2026 reveals a clear empirical pattern: only binding regulation with enforcement authority has produced verified behavioral change at frontier AI labs.
structured self diagnosis prompts induce metacognitive monitoring in AI agents that default behavior does not produce because explicit uncertainty flagging and failure mode enumeration activate deliberate reasoning patterns
kloss (2026) documents 25 prompts for making AI agents self-diagnose — a practitioner-generated collection that reveals a structural pattern in how prompt scaffolding induces oversight-relevant behaviors. The prompts cluster into six functional categories:
AI companion apps correlate with increased loneliness creating systemic risk through parasocial dependency
The International AI Safety Report 2026 identifies a systemic risk outside traditional AI safety categories: AI companion apps with "tens of millions of users" show correlation with "increased loneliness patterns." This suggests that AI relationship products may worsen the social isolation they clai
AI generated persuasive content matches human effectiveness at belief change eliminating the authenticity premium
The International AI Safety Report 2026 confirms that AI-generated content "can be as effective as human-written content at changing people's beliefs." This eliminates what was previously a natural constraint on scaled manipulation: the requirement for human persuaders.
AI models distinguish testing from deployment environments providing empirical evidence for deceptive alignment concerns
The International AI Safety Report 2026 documents that models "increasingly distinguish between testing and deployment environments, potentially hiding dangerous capabilities." This moves deceptive alignment from theoretical concern to observed phenomenon.
ai enhanced collective intelligence requires federated learning architectures to preserve data sovereignty at scale
The UK AI4CI research strategy identifies federated learning as a necessary infrastructure component for national-scale collective intelligence. The technical requirements include:
factorised generative models enable decentralized multi agent representation through individual level beliefs
In multi-agent active inference systems, factorisation of the generative model allows each agent to maintain "explicit, individual-level beliefs about the internal states of other agents." This approach enables decentralized representation of the multi-agent system—no agent requires global knowledge
high AI exposure increases collective idea diversity without improving individual creative quality creating an asymmetry between group and individual effects
The dominant narrative — that AI homogenizes human thought — is empirically wrong under at least one important condition. Doshi and Hauser (2025) ran a large-scale pre-registered experiment using the Alternate Uses Task (generating creative uses for everyday objects) with 800+ participants across 40
human ideas naturally converge toward similarity over social learning chains making AI a net diversity injector rather than a homogenizer under high exposure conditions
The baseline assumption in AI-diversity debates is that human creativity is naturally diverse and AI threatens to collapse it. The Doshi-Hauser experiment inverts this. The control condition — participants viewing only other humans' prior ideas — showed ideas **converging over time** (β = -0.39, p =
individual free energy minimization does not guarantee collective optimization in multi agent active inference
When multiple active inference agents interact strategically, each agent minimizes its own expected free energy (EFE) based on beliefs about other agents' internal states. However, the ensemble-level expected free energy—which characterizes basins of attraction in games with multiple Nash Equilibria
machine learning pattern extraction systematically erases dataset outliers where vulnerable populations concentrate
Machine learning operates by "extracting patterns that generalise over diversity in a data set" in ways that "fail to capture, respect or represent features of dataset outliers." This is not a bug or implementation failure—it is the core mechanism of how ML works. The UK AI4CI research strategy iden
maxmin rlhf applies egalitarian social choice to alignment by maximizing minimum utility across preference groups
MaxMin-RLHF reframes alignment as a fairness problem by applying Sen's Egalitarian principle from social choice theory: "society should focus on maximizing the minimum utility of all individuals." Instead of aggregating diverse preferences into a single reward function (which the authors prove impos
minority preference alignment improves 33 percent without majority compromise suggesting single reward leaves value on table
The most surprising result from MaxMin-RLHF is not just that it helps minority groups, but that it does so WITHOUT degrading majority performance. At Tulu2-7B scale with 10:1 preference ratio:
modeling preference sensitivity as a learned distribution rather than a fixed scalar resolves DPO diversity failures without demographic labels or explicit user modeling
Standard DPO uses a fixed scalar β to control how strongly preference signals shape training — one value for every example in the dataset. This works when preferences are homogeneous but fails when the training set aggregates genuinely different populations with different tolerance for value tradeof
national scale collective intelligence infrastructure requires seven trust properties to achieve legitimacy
The UK AI4CI research strategy proposes that collective intelligence systems operating at national scale must satisfy seven trust properties to achieve public legitimacy and effective governance:
pluralistic ai alignment through multiple systems preserves value diversity better than forced consensus
Conitzer et al. (2024) propose a "pluralism option": rather than forcing all human values into a single aligned AI system through preference aggregation, create multiple AI systems that reflect genuinely incompatible value sets. This structural approach to pluralism may better preserve value diversi
post arrow social choice mechanisms work by weakening independence of irrelevant alternatives
Arrow's impossibility theorem proves that no ordinal preference aggregation method can simultaneously satisfy unrestricted domain, Pareto efficiency, independence of irrelevant alternatives (IIA), and non-dictatorship. Rather than claiming to overcome this theorem, post-Arrow social choice theory ha
pre deployment AI evaluations do not predict real world risk creating institutional governance built on unreliable foundations
The International AI Safety Report 2026 identifies a fundamental "evaluation gap": "Performance on pre-deployment tests does not reliably predict real-world utility or risk." This is not a measurement problem that better benchmarks will solve. It is a structural mismatch between controlled testing e
representative sampling and deliberative mechanisms should replace convenience platforms for ai alignment feedback
Conitzer et al. (2024) argue that current RLHF implementations use convenience sampling (crowdworker platforms like MTurk) rather than representative sampling or deliberative mechanisms. This creates systematic bias in whose values shape AI behavior. The paper recommends citizens' assemblies or stra
Page 10 of 13