Suggested Pages

100 pages the wiki should add, ranked by priority. Priority is based on how often the topic is mentioned across existing pages and its importance to AI safety coverage. Sourced from the Feb 2026 gap analysis.

100 of 100 results

				Why
100	AI Governance	concept	183	Mentioned on 183 pages. Sub-topics exist (compute governance, etc.) but no umbrella page.
99	Reinforcement Learning	concept	81	Mentioned on 81 pages. Underpins RLHF, reward modeling, and alignment methods — no standalone page.
98	Retrieval-Augmented Generation (RAG)	concept	404	Referenced on 404 pages — most-mentioned concept without a page.
97	GPT-4	capability	193	Central reference point for frontier capabilities, mentioned on 193 pages.
96	Claude (Model Family)	capability	186	Anthropic's flagship model, mentioned on 186 pages, no standalone page.
95	Frontier Model (Concept)	concept	161	The concept of 'frontier model' is used on 161 pages with no definition page.
94	Training Data	concept	111	Fundamental topic (curation, bias, consent, copyright) — 111 page mentions.
93	Safety Evaluations	response	106	Referenced on 106 pages; evals are how labs demonstrate safety.
92	Misalignment Potential	ai-transition-model	110	110 EntityLinks to this transition-model factor, no page exists.
91	Civilizational Competence	ai-transition-model	104	104 EntityLinks to this transition-model factor, no page exists.
90	Model Evaluation (Methodology)	response	55	Referenced on 55 pages. Capability evals, dangerous-capability evals, eval science — no methodology page.
89	Fine-Tuning	concept	89	Key technique for adapting models; safety implications of open fine-tuning.
88	Gemini (Google DeepMind)	capability	88	Google's frontier model family, mentioned on 88 pages.
87	Llama (Meta)	capability	82	Most widely used open-weights model. 82 page mentions.
86	DeepSeek	organization	71	Chinese frontier lab (R1, V3) — changed compute-efficiency assumptions globally.
85	GPT-5 / Next-Gen OpenAI	capability	67	Frequently referenced as next capability milestone, 67 page mentions.
84	Transformer Architecture	concept	65	The architecture underlying all frontier models; no explainer page.
83	Jailbreaking & Prompt Injection	risk	81	Primary attack vector against deployed LLMs. Covers direct jailbreaks + indirect injection in tool-use contexts.
82	Multimodal AI	capability	49	Vision/audio models have distinct safety challenges. 49 mentions.
81	Training Runs & Compute Cost	concept	47	Economics of training — cost, duration, environmental impact. 47 mentions.
80	Foundation Model (Concept)	concept	46	Distinct from 'frontier model' — the general category. 46 mentions.
79	Hallucination	risk	41	Most user-visible AI failure mode. 41 mentions, no dedicated page.
78	AI Chips & Hardware	concept	39	GPU/TPU/custom silicon — hardware is a key governance lever. 39 mentions.
77	Semiconductor Industry	concept	35	Supply chain chokepoints (TSMC, ASML). 35 mentions.
76	Grok (xAI)	capability	33	xAI's model. 33 mentions; no entity or page.
75	Embeddings & Vector Search	concept	33	How models represent knowledge. 33 mentions.
74	AI Incidents Database	incidents	32	Only 2 incident pages exist. Need a comprehensive tracker.
73	Benchmarks & Leaderboards	concept	28	How capabilities are measured; gaming and limitations. 28 mentions.
72	Mistral AI	organization	27	Leading European frontier lab. Important for EU AI Act context.
71	DPO & RLHF Alternatives	response	27	DPO, IPO, KTO — alternatives to RLHF for alignment. 27 mentions.
70	Transition Turbulence	ai-transition-model	26	26 EntityLinks to this transition-model factor, no page.
69	Synthetic Data	concept	24	Self-play and synthetic training data — model collapse risk. 24 mentions.
68	Model Weights (Security & Access)	concept	40	Referenced on 40 pages. Weight theft, open release decisions, proliferation risk — no standalone page.
67	Pre-Training	concept	21	The initial training phase. Distinct safety considerations from fine-tuning.
66	Knowledge Distillation	concept	20	Compressing large models; safety properties may not transfer. 20 mentions.
65	Post-Training (RLHF, Safety)	concept	20	Where safety alignment happens in practice. 20 mentions.
64	Content Provenance & C2PA	response	19	Technical countermeasure to deepfakes and AI content. 19 mentions.
63	Misuse Potential	ai-transition-model	18	18 EntityLinks to this transition-model factor, no page.
62	AI Watermarking	response	18	SynthID, text watermarks — detection of AI-generated content.
61	Data Annotation & AI Labor	concept	18	Ghost work, RLHF annotators, labor conditions. 18 mentions.
60	Intelligence Explosion	concept	17	Core AI safety concept (Good, Bostrom). 17 mentions, no page.
59	Voice Cloning	risk	17	Fraud, impersonation, consent issues. 17 mentions.
58	Model Cards & Documentation	response	17	Standard disclosure format for AI models. 17 mentions.
57	Context Windows	concept	17	Key capability dimension (4k to 1M+) with safety implications.
56	TSMC	organization	16	Single point of failure for advanced chips. 16 mentions.
55	Epoch AI	organization	16	16 EntityLinks to this org — key data source for AI trends.
54	Open Weights	concept	16	Distinct from 'open source' — weights-only release model.
53	Attention Mechanism	concept	14	Core transformer component. 14 mentions.
52	Capability Overhang	concept	13	When existing hardware can run much more capable models. 13 mentions.
51	Test-Time Compute & Reasoning	capability	11	o1/o3/R1 inference-scaling paradigm — changes safety assumptions.
50	Chinchilla Scaling	concept	11	Compute-optimal training. 11 mentions.
49	Hugging Face	organization	10	Central hub for open-weights models and datasets.
48	Knowledge Graphs for AI	concept	10	Structured knowledge + LLMs. 10 mentions.
47	Alignment Tax	concept	9	Cost of making models safe vs. capable. Key policy concept.
46	Image & Video Generation	capability	16	Diffusion models, DALL-E, Midjourney, Sora — architecture and safety issues.
45	AI Auditing	response	9	Third-party safety audits. Emerging profession.
44	Data Poisoning	risk	8	Supply-chain attack on training data. Distinct from adversarial examples.
43	Brain Emulation	concept	8	Whole brain emulation as alternative path to AGI. 8 mentions.
42	Algorithmic Bias	risk	6	6 dangling EntityLinks. Needs entity + page.
41	Model Collapse	risk	7	Training on AI-generated data degrades quality. Emerging research.
40	AI Consciousness & Moral Status	concept	7	Sentience, moral patienthood, digital minds. 7 mentions.
39	AI Technical Standards	response	38	Referenced on 38 pages. ISO, NIST, IEEE frameworks — how standards interact with regulation.
38	Chain-of-Thought Reasoning	concept	6	Prompting technique that elicits reasoning. Safety implications.
37	Function Calling & Tool Use	capability	6	Agentic capability — models invoking APIs. Security concerns.
36	Differential Privacy	response	6	Mathematical privacy guarantees for training data.
35	Regulatory Arbitrage	risk	5	5 dangling EntityLinks. Companies choosing least-regulated jurisdictions.
34	AI Liability & Legal Frameworks	response	5	Who pays when AI causes harm? Foundational governance question.
33	NVIDIA	organization	4	4 dangling EntityLinks. Dominant AI chip supplier.
32	Compliance Costs	concept	4	4 dangling EntityLinks. Cost of regulation for AI companies.
31	Chinese AI Ecosystem	concept	4	Baidu, Alibaba, Tencent, ByteDance — different safety norms.
30	Reward Modeling	response	3	Positive framing of reward specification. Complements reward-hacking page.
29	Model Merging & Weight Manipulation	risk	3	Open-source technique to combine or modify model capabilities.
28	AI Supply Chain	concept	3	End-to-end: data, compute, training, deployment. Chokepoints.
27	Post-Deployment Monitoring	response	2	Runtime safety monitoring. Most safety work is pre-deployment.
26	Federated Learning	concept	2	Privacy-preserving training across distributed data.
25	AI Energy & Environmental Impact	concept	2	Data center power, water use, carbon footprint of training.
24	Compute Governance Tracking	metric	2	Are compute thresholds actually enforced? No tracking page.
23	Foundation Model Commoditization	model	2	Pricing collapse changes lab safety incentives.
22	In-Context Learning	concept	2	How LLMs learn from prompts. Safety implications for elicitation.
21	AI-Enabled Scientific Fraud	risk	2	Paper mills, fabricated data, fake peer reviews.
20	Speculative Decoding	concept	1	Inference optimization affecting deployment safety properties.
19	Shane Legg	researcher	—	DeepMind co-founder. Entity exists, no page.
18	Nate Soares	researcher	—	MIRI Executive Director. Entity exists, no page.
17	Beth Barnes	researcher	—	Founded METR (Model Evaluation & Threat Research). Entity exists, no page.
16	Gary Marcus	researcher	—	Prominent AI critic and public commentator. Entity exists, no page.
15	Ian Hogarth	researcher	—	Chair of UK AI Safety Institute. Entity exists, no page.
14	Buck Shlegeris	researcher	—	CEO of Redwood Research. Entity exists, no page.
13	Elizabeth Kelly	researcher	—	Director of US AI Safety Institute. Entity exists, no page.
12	ARC Evaluations	organization	—	Entity exists (arc-evals), no page. Key eval org.
11	Redwood Research	organization	—	Entity 'redwood' exists separately from redwood-research page — may need merge or separate page.
10	Homomorphic Encryption for AI	concept	1	Privacy-preserving inference. Niche but growing.
9	Tokenization	concept	3	How text becomes model input. Affects multilingual safety.
8	Deepfake Detection	response	3	Technical countermeasure to deepfakes. Detection arms race.
7	AI Copyright & Fair Use	concept	3	Training data rights, output ownership. Active litigation.
6	Catastrophic Forgetting	concept	2	Models lose capabilities during fine-tuning. Safety implications.
5	Mixture of Experts	concept	2	Architecture used by GPT-4, Mixtral. Efficiency vs. safety.
4	AI Labor Displacement (Empirical)	metric	1	Tracking actual job impacts as of 2026.
3	Red-Teaming-as-a-Service	response	1	Commercial red-teaming offerings and effectiveness.
2	Continual Learning	concept	1	Models that learn after deployment. Safety of ongoing adaptation.
1	AI Military & Intelligence Applications	concept	1	Beyond autonomous weapons — broader military AI use.