Taiwan Strait, AI and conflict: what research exists?
A provisional longlist of publicly hosted research and analysis on how advanced AI development changes strategic incentives between the US, China, Taiwan and their partners. It is shown in date order, with no ratings or rankings displayed, so readers can form their own view of what matters. This is an optional resource for the working group.
Short on time? Start with the discussion guide: the main questions, contested claims and a handful of key readings in each area, readable in 15–20 minutes. Then come back here to check what is missing: which part of the problem does this list leave unrepresented?
Scope. In scope is AI as the stakes: states expecting AI to shift relative power, competing to control chips and compute, using dependence as leverage, and negotiating over AI, including beliefs about AI's future military value. Out of scope is AI as an instrument: how AI performs in weapons, operations, command and control, intelligence or information operations. The list shows in-scope items by default; background and out-of-scope items are one click away.
Read with care. Only work with a public link is listed. Books, paywalled work without a landing page, classified or internal analysis and much non-English work are under-represented or missing. The list was assembled by AI-assisted web search with no human screening yet; it is not a systematic review. Summaries may be wrong; links and dates were checked by the search agents but not yet by a person.
How this list was made: sources, filters, model and prompt
| Who and when | Compiled on for The Unjournal's working group on AI geopolitics and the Taiwan Strait, as a first pass to be corrected and extended. |
|---|---|
| Model | Claude Opus 5.5 (Anthropic), run as research agents inside Claude Code at the model's default reasoning effort, with web search and page access. First sweep (23 Oct): five agents, sharing a cap of 200 web searches. Scope follow-up (28 Oct): three more agents searched inside the narrowed scope, and Claude Sonnet 5.5 re-screened every earlier item against it from the item's summary. |
| Scope follow-up | On 26 Oct the working group narrowed its focus to AI as the stakes rather than AI as an instrument (see the scope note at the top). The follow-up started from the convenor's annotated bibliography, verified and linked each item in it, then searched for more work in four core areas, in English, Chinese and Japanese. It also admitted expert analytical commentary in policy outlets and theses, which the first sweep excluded, because the convenor's bibliography relies on them. Read the follow-up prompt. |
| How the work was split | Every agent got the same instructions (below) plus one slice of the problem: (1) chips, upstream inputs, costs of a conflict and economic interdependence; (2) escalation, military AI, AI-enabled hybrid operations, scenarios and wargames; (3) US–China AI competition, AI governance and third-party decision-making; (4) PRC and Taiwanese sources, searched mainly in Traditional Chinese, Simplified Chinese and Japanese; (5) forecasting, critiques of prediction and work from the effective-altruism / global-priorities network. Each aimed for 20–30 items. |
| What counted | Academic articles and working papers, think-tank and policy reports, datasets and trackers, forecasting questions and elicitations, government, parliamentary and AI-lab publications, scenario and wargame reports, and EA / global-priorities analyses that visibly inform funding or advocacy. Excluded: op-eds and commentary with no underlying research, news articles, and general AI-capability work with no stated link to relations between states. |
| Filters | Only work with a stable public link to the work or its official landing page. Each agent had to open every link and confirm title, authors and date before listing an item; items it could not open went to "unverified leads", and work it knew of but could not link went to "known but not linkable". The agents were told not to rate, rank, assess quality or synthesise, and to include both sides where pieces disagree. |
| Breadth rules | Search in Chinese and Japanese as well as English; deliberately look for Taiwanese, PRC, European, Japanese and other non-US work; include at least three critiques of conflict forecasting; flag items from the EA / global-priorities network so readers can see how much of the list comes from one source. |
| Merging | The five outputs (175 items) were merged and deduplicated by DOI, link and title into items. Facet tags, reliance levels and "contrasts with" links are the agents' own. |
| Non-English items | Where the source publishes an official English version, the item links to it. Other web pages get a Google Translate link. For PDFs and long items, working-group members can ask The Unjournal for a private machine translation (Claude Sonnet 5.5); these are not published, since the originals are copyrighted. |
| Known gaps |
Abbreviated instructions given to every agent
- Build as complete a list as possible 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 AI governance.
- This is an intake, not an assessment: record what each item is, what it claims, why it bears on the problem and who visibly relies on it. No ratings, rankings, quality judgements, synthesis or research agendas.
- Tag each item with the facets it substantively addresses (the 15 facets shown below), so gaps are visible.
- Reliance: "documented" only with a named organisation citing or using the work; otherwise "plausible" or "unclear". Never upgrade a guess.
- Breadth requirements (non-English searches, Taiwanese and PRC institutions, non-US sources, other disciplines, sceptics of prediction, both sides of disagreements, an EA-network flag).
- Leads to verify and expand from: RAND, CSIS China Power, CSET, CNAS, Mercatus PCI-Tensions, Metaculus and other forecasting questions, EA Forum and Founders Pledge work, Forecasting Research Institute, Martin, Mayer & Thoenig (2008), SIPRI.
- Output: one record per item, plus unverified leads, known-but-unlinkable work, facet gaps and a log of searches run.
Read the full shared prompt (each agent also got a short note naming its slice and the 20–30 item target).
Searches the agents reported running
Coverage by area
Finer facets (from the first sweep)
Click a facet to filter. Facets with fewer than three items are shaded. An item can carry several facets. Counts follow the scope setting above.
Suggest something missing
Research, data or analysis that belongs here, or a part of the problem the list misses. One suggestion per submission; it goes into the next intake pass after a quick check, not straight onto the page.
How ratings and comments are used
Each item has a "Rate or comment" button. Ratings use The Unjournal's usual priority scale for deciding what to evaluate (-- = 10th percentile, - = 30, ~ = 50, + = 70, ++ = 90, relative to all research we might evaluate). They feed the same aggregates as the prioritization dashboard, but this page never shows other people's ratings, so the list stays a neutral starting point. Your own rating is remembered in this browser only.
Comments can be public (anonymous or named), visible only to The Unjournal team, or confidential to the people who administer the form. Your email is never published. Working-group members who would rather not be publicly associated with views on Taiwan should choose "Unjournal team only" for comments and leave ratings anonymous. General comments on the page are also welcome through Hypothes.is (the sidebar on the right).
Known but not linkable (0)
Work the search agents were confident exists but could not link to a stable public copy. Excluded from the main list for that reason only.
Unverified leads (0)
Items the agents half-recognised but could not confirm or open. They may not exist in the form described.