## Your task

Build as complete a list as you can of **existing, publicly hosted research** on how AI development and deployment interacts with conflict risk across the Taiwan Strait, including the chip supply chain and the AI governance questions this raises. For each item, record what it is, what it claims, why it bears on the problem, and who visibly relies on it.

This is an **intake**, not an assessment. The list feeds two things:

1. The Unjournal's research prioritization dashboard, where each item is later scored by a separate, standard process. **Do not rate, rank or score anything yourself.**
2. A plain reading page for a small working group of researchers and practitioners, shown in date order with no ratings, so the group can form its own view. Your descriptions will appear there nearly verbatim, so write them neutrally.

Today's date is **23 October 2026**. Search for recent work (2023–2026 especially), and include older pieces only when people still lean on them.

## Why this exists

The Unjournal (a nonprofit that commissions open, quantitative expert evaluation of publicly hosted research) is convening a working group that meets four times from early November 2026. Its goal is a short list of **pivotal questions**: questions where better evidence would plausibly change a real decision by a funder, government, AI lab, advocacy group or business within a few years. A secondary goal is to find one or two existing pieces of work that are influential enough, and uncertain enough, that an independent evaluation would be worth commissioning.

The group will use this list to see what research exists and, more importantly, **which parts of the problem it leaves uncovered**. That is why breadth, honest gaps and facet tags matter more than polish.

**The risk to avoid:** a list that looks like expert consensus but is really what one network (Anglophone, DC-based, quantitative, effective-altruism-adjacent) already reads. The default search path produces exactly that. Several instructions below exist to counter it.

## Scope

**Core question:** how the development and deployment of advanced AI changes the risk of conflict and the prospects for cooperation between states, with the Taiwan Strait (and the wider South China Sea) as the focal case. Research matters here because it can inform efforts to reduce the risk of major-power conflict and to make cooperation on AI governance and catastrophic risk more feasible.

**Facets.** Tag each item with every facet it substantively addresses (use the IDs). These tags are how the group will spot gaps, so tag carefully and don't stretch.

| ID | Facet |
|---|---|
| `chips` | Semiconductor supply chain: TSMC and fab concentration, chokepoints, "silicon shield" arguments, onshoring and derisking |
| `upstream` | Upstream and physical inputs: rare earths and critical minerals, energy and water for fabs, shipping lanes and blockade exposure |
| `cost` | Economic and human costs of a Strait conflict or blockade, globally and for AI supply chains |
| `escalation` | Escalation, deterrence and crisis dynamics; tipping points and warning signals; gray-zone coercion |
| `military-ai` | AI in military operations: ISR, targeting, autonomy, command and control, AI–nuclear interactions |
| `hybrid` | AI-enabled hybrid and information operations: disinformation, cognitive warfare, cyber operations against Taiwan |
| `race` | US–China AI competition and race dynamics, and how Strait tensions affect them (for example, whether chip vulnerability makes a race harder to slow) |
| `governance` | AI governance and diplomacy: export controls, compute governance, US–China AI dialogue, military-AI norms (REAIM, the Political Declaration, the 2024 US–China statement on human control of nuclear decisions), Taiwan's own position in AI governance |
| `interdependence` | Economic interdependence, sanctions and economic statecraft as restraints on or triggers of conflict |
| `prc` | PRC decision-making, elite and military perceptions, PRC writing on "intelligentized" warfare |
| `taiwan` | Taiwan's domestic politics, public opinion, defence policy, AI and resilience preparedness |
| `third-party` | US, Japanese, European and other third-party decision-making on Taiwan, including European derisking and economic deterrence |
| `forecasting` | Forecasting, early warning and quantitative prediction of interstate conflict, including invasion-probability estimates and platform questions |
| `method-critique` | Critiques of the forecasting and quantitative-prediction enterprise: limits of long-run conflict data, base rates, idiosyncratic leaders, classified information |
| `scenarios` | Scenario analysis, wargames and conditional "what if" studies (for example, compute or data-centre sovereignty, changes in US support for Taiwan) |

If an important item fits none of these, tag it `other` and say in one line what facet is missing. That is useful information.

**What counts.** Deliberately broad:

- Academic articles and working papers (economics, political science, international relations, security studies, area studies, science and technology studies, history)
- Think-tank and policy reports
- Datasets and trackers (for example ADIZ incursion data, event data, rhetoric indices)
- Forecasting outputs: platform questions with real engagement, structured expert elicitations, forecasting-team reports
- Government, military, parliamentary and AI-lab publications
- EA and global-priorities analyses (forum posts, funder research) where they visibly inform funding or advocacy
- Scenario and wargaming exercises

Include an item if a reasonable person working on this might lean on it when forming a view or making a decision. **Exclude** op-eds and commentary with no underlying research or data, news articles, and general AI-capability work with no stated link to relations between states.

**Hosted, linked work only.** Every item needs a stable public URL to the work itself or to its official landing page (a paywalled journal article with a publisher abstract page qualifies; mark it `paywalled`). Work you know of but cannot link (books without a landing page, unpublished or embargoed reports, items behind institutional logins) goes in the separate **"known but not linkable"** list, not in the main list. The reading page will carry a visible caveat that unlinked work is missing, so this list matters.

## Breadth requirements

These are requirements, not aspirations.

1. **Search in more than English.** Run searches in Traditional Chinese (Taiwan usage), Simplified Chinese (PRC usage) and Japanese, for example: 台海 人工智慧 衝突; 晶片 供應鏈 兩岸; 灰色地帶 脅迫; 台湾海峡 人工智能 冲突风险; 芯片 供应链 两岸关系; 智能化战争; 台湾有事 半導体. Give an English translation of every non-English title.
2. **Taiwan-based and PRC-based work.** Deliberately look for work from Taiwanese institutions (for example INDSR, DSET, Doublethink Lab, IORG, Academia Sinica, Prospect Foundation, university research centres) and citable PRC analysis (for example CICIR, Tsinghua CISS, SIIS, university journals). It will not surface on its own.
3. **Outside Washington.** Include European, Japanese, Australian, Singaporean and Korean work (for example MERICS, IISS, SIPRI, Chatham House, NIDS, JIIA, ASPI, Lowy Institute, RSIS) and European Parliament or EU research-service publications.
4. **Outside the mainstream disciplines.** Political economy of semiconductors, science and technology studies, area studies, and historical work on the Strait.
5. **Sceptics of prediction.** At least three items that question whether conflict forecasting, long-run conflict data or invasion-probability estimates are useful at all.
6. **Disagreements.** Where two items reach conflicting conclusions, include both, and point each to the other in `contrasts_with`. Do not reconcile them or take a side.
7. **Network flag.** Mark items from the EA and global-priorities network (`ea_network: true`), so readers can see how much of the list comes from one source. This is not a discount.

## Leads to verify and expand from

Treat these as starting points, not a list to reproduce. Some may be misremembered; say so if a lead doesn't exist as described.

- RAND work on Taiwan indications and warnings, and RAND's 2025 volume *The Artificial General Intelligence Race and International Security* (Mitre, Horowitz, Henry, Borden and Predd, eds.)
- CSIS China Power: Taiwan Strait coercion and ADIZ incursion data
- CSET: semiconductor supply chains, PLA use of AI, Taiwan and supply-chain papers
- CNAS: Taiwan security work and the Artificial Intelligence and Global Security Initiative research agenda
- Mercatus / Weifeng Zhong: PCI-Tensions, an LLM-scored index of PRC official rhetoric for predicting Strait crises
- A recent paper by Wong et al. predicting China–Taiwan escalation from news-derived indicators (prediction vs. causal framing)
- conflictforecast.org (Mueller and Rauh): text-based conflict forecasting used by governments
- Swift Centre forecasts on China, Taiwan and TSMC
- Metaculus and other platform questions on a Taiwan invasion or blockade; note exactly what each resolves on
- Joel McGuire's EA Forum analysis of Taiwan invasion probability (a Metaculus × Samotsvety decomposition)
- Stephen Clare's great-power conflict reports (Founders Pledge, 80,000 Hours); Founders Pledge great-power conflict research; Longview Philanthropy's nuclear weapons policy work
- Forecasting Research Institute: conditional trees, and the adversarial collaboration on AI risk
- Martin, Mayer & Thoenig (2008), "Make Trade Not War?", *Review of Economic Studies*
- SIPRI work on the AI–nuclear nexus and Chinese security policy

## For each item, record

| Field | Content |
|---|---|
| `title` | Original title; add `title_en` for non-English titles |
| `url` | Stable link to the work or its official landing page. You must have opened it. |
| `doi` | If one exists |
| `authors` | List; use the organisation if there is no named author |
| `organisation` | Publishing institution or platform |
| `publication_date` | `YYYY-MM-DD`, or `YYYY-MM` / `YYYY` if that is all you can confirm. Use the date first made public. |
| `venue` | Journal, series, platform or publisher |
| `type` | One of `journal_article`, `working_paper`, `think_tank_report`, `dataset`, `forecast`, `government`, `lab`, `ea_forum_or_adjacent`, `scenario_or_wargame` |
| `language` | Language of the original |
| `origin` | Country or region where the authors or institution are based |
| `access` | `open`, `paywalled` or `registration` |
| `what_it_claims` | Two or three sentences, in your own words, on the central claim or finding. No quotation beyond a few words. If you are unsure you have it right, say so. |
| `why_it_bears` | One or two sentences on the specific mechanism or question it speaks to, not just its topic |
| `facets` | List of facet IDs from the table above |
| `reliance` | `documented`, `plausible` or `unclear` (see below) |
| `reliance_evidence` | For `documented`: who relies on it and where, with a link. Otherwise one short line or blank. |
| `ea_network` | `true` or `false` |
| `contrasts_with` | Titles of items in this list that reach conflicting conclusions, if any |

**Reliance.** *Documented* means you found a named funder, policy body, forecaster, lab or advocacy group citing or using it (look at grant pages, parliamentary or congressional testimony, government reports, think-tank footnotes, forecasting rationales). *Plausible* means it circulates in relevant discussion but you found no specific use. *Unclear* means you don't know. Never upgrade plausible to documented. An honest "unclear" is more useful than a confident guess.

## How many

As many as genuinely fit. The topic is narrow enough that something close to a full list of hosted work is the goal; 40 to 100 items is a reasonable expectation. Don't pad to reach a number, and don't stop at the first 20 because the leads are covered. Breadth requirements come before covering every lead.

## What not to do

- Don't rate, score, rank or order by importance. Scoring happens separately.
- Don't assess quality, methods or credibility, not even briefly.
- Don't write a synthesis, taxonomy or "what the literature shows". Facet tags are the only classification.
- Don't generate research questions or agendas. That is the working group's job.
- Don't polish. Partial and clearly provisional is the aim, so people feel free to add to it and contradict it.
- Don't reproduce substantial text from sources.
- Don't name or speculate about who is in the working group.

## Output

Return, in this order:

1. **The list**, as a single JSON array in one code block, one object per item with the fields above, in date order (oldest first).
2. **Unverified leads**: items you half-recognise but couldn't confirm exist or couldn't open. Title, what you think it is, and why you couldn't confirm it.
3. **Known but not linkable**: items you're confident exist but that have no stable public link. Full citation and why it matters.
4. **Facet gaps**: one line per facet with fewer than three items, saying whether you think the gap is real (little work exists) or a search failure (work likely exists but you couldn't reach it).
5. **Sourcing note**: the searches you ran (including the non-English ones), databases and sites you checked, what you couldn't access, and the date you compiled the list.