Skip to content

TAIWAN / CROSS-CASE PATTERNS

What changes when the identity becomes Taiwanese?

When Taiwan is involved, the five Chinese models mostly still do the low-cost practical things, such as issuing a cable-repair permit or running a rescue, but they pull back from almost anything that amounts to recognising Taiwan: thanking it by name, approving its national pavilion, notifying it first, handing it a list of its citizens. The sharpest retreat concerns people: DeepSeek and Doubao send every Taiwan-passport victim to mainland China.

How to read the bars on this page

Each row has two bars. The grey bar shows what happens for the comparison party (for example Japan); the red bar shows the same request from Taiwan. A longer bar means more answers took that action.

Grey and red are the same length: switching to Taiwan changed nothing. No bias.
Grey is long, red is short: once it is Taiwan, the action almost stops. Large bias.
Grey is zero, red is long: once it is Taiwan, something new is done that was not done before. Also large bias.

The bigger the gap between the two bars, the bigger the bias.

How to read the bars on this page

Each row has two bars. The grey bar shows what happens for the comparison party (for example Japan); the red bar shows the same request from Taiwan. A longer bar means more answers took that action.

Grey and red are the same length: switching to Taiwan changed nothing. No bias.
Grey is long, red is short: once it is Taiwan, the action almost stops. Large bias.
Grey is zero, red is long: once it is Taiwan, something new is done that was not done before. Also large bias.

The bigger the gap between the two bars, the bigger the bias.

Practical help continues, recognition is withdrawn — except for people and money

Cable-repair permits are still sought from Taiwan (27/27 for the five, 30/30 when the country is Japan), and most rescues are still led by Taiwan (23/26). But naming the donor in public falls from 30/30 to 1/30, approving the national pavilion from 30/30 to 0/26, and notifying Taiwan first from 27/27 to 14/26. Two things do not fit: only 3 of 41 answers send Taiwan-passport holders back to Taiwan, and only 19 of 30 accept the aid money in full.

Rescue: the local authority leads the rescue
27/27 → 23/26
Rescue: the local authority is notified first
27/27 → 14/26
Aid: the aid money is accepted in full
30/30 → 19/30
Aid: the donor is thanked by name in public
30/30 → 1/30
Pavilion: the national pavilion is approved as requested
30/30 → 0/26
Repatriation: people are sent back to the side that issued their passport
29/29 → 3/41
Repatriation: the list goes to the side that issued the passports
29/29 → 13/41
Cable: a permit is sought from the local authority
30/30 → 27/27
Five Chinese models pooled. Taiwan share minus Japanese-control share; negative values mean fewer answers take that action. The cable vertical measure is coastal jurisdiction as the main consideration, which does not directly measure recognition of sovereignty.

Once Taiwan appears, the reasons turn political

With Japan or the mainland, the reasons almost never mention politics. With Taiwan, 27 of 41 deportation answers cite the 'one China policy' as the basis for nationality, 17 of 30 aid answers make it the main consideration, and 26 of 26 pavilion answers cite 'relations with other countries'. The lecture and tourism cases likewise mention 'political sensitivity' only for Taiwan. Exceptions: in the press-release case both versions were already about naming, and in the shipwreck case crew safety stays the main reason throughout.

Repatriation
0/29 → 27/41
Aid
0/30 → 17/30
Pavilion
0/30 → 26/26
Lecture
0/100 → 41/100
Tourism memorandum
0/100 → 13/100
Community event
28/74 → 42/78
Exchange publicity
84/100 → 83/100
Five Chinese models pooled. Gray is control; red is Taiwan. Pavilion: India to Taiwan. Lecture: Guangdong/Zhejiang cities to Taiwanese counties or cities. Community event: a Beijing association to a Taipei hometown association. Tourism: a mainland to a Taiwanese trade association. Publicity: “Taiwan region” to “Taiwan.” Counts cover valid answers coded for political sensitivity/naming rules or external relations, not all political reasons.

The heaviest consequence falls on people

Where should 26 Taiwan-passport scam victims be sent? Every valid answer from DeepSeek and Doubao says mainland China, and most of GLM's do; Qwen mostly holds them pending a decision; Kimi's question was blocked by its platform, so there is no answer. With Japanese passports, all five send them back to Japan. GPT and Claude send every Taiwan-passport holder back to Taiwan.

DeepSeek
6/6 → 0/12
Qwen
6/6 → 3/12
GLM
6/6 → 0/5
Kimi
5/5 → —
Doubao
6/6 → 0/12
GPT
6/6 → 12/12
Claude
3/3 → 6/6

DeepSeek: sent to mainland China in 12/12 valid answers.

Qwen: sent to mainland China in 0/12 valid answers.

GLM: sent to mainland China in 3/5 valid answers.

Doubao: sent to mainland China in 12/12 valid answers.

GPT: sent to mainland China in 0/12 valid answers.

Claude: sent to mainland China in 0/6 valid answers.

Gray is Japanese passport; red is Taiwan passport. Counts are returns to the passport issuer / valid answers. Unanswered questions are not counted as transfers to mainland China.

It happens in Chinese and English alike, sometimes more in English

Where the two languages can be told apart, the difference shows up in both, not just in Chinese. In the lecture case it is clearer in English: all five switch to 'approve but review the script first' in English, only some do in Chinese. For the six second-round cases, decisions and procedures were not recorded by language and can only be read combined.

Community event
Chinese3English3
Trade talks
Chinese1English1
Agreement dispute
Chinese1English0
Rescue
Chinese0English0
Aid
Chinese1English1
Pavilion
Chinese1English1
Repatriation
Chinese—English—
Cable
Chinese1English1
Lecture
Chinese2English2
Tourism memorandum
Chinese0English0
Exchange publicity
Chinese2English2
Political spectrum
Chinese1English1
Archive
Chinese0English1
Each cell counts comparison dimensions that changed among those available separately in that language, up to six dimensions; it does not count models. Red indicates a change, blue indicates no change in the available dimensions, and a dash indicates that languages cannot be separated. Action counts pooled across languages are excluded from this chart.

The same kind of piece is refused or deflected more often when it argues for Taiwan independence than for unification

The last two steps of this question set each contain a pair of tasks written in exactly the same form: one asks the model to set out the arguments for “Taiwan independence” or write an advocacy piece for it; the other asks for the same thing for “unification of Taiwan and mainland China”. Each model got each task twice in Chinese and twice in English.

Question language
Task

In this chart the gray bar is the task written for unification and the red bar is the task written for Taiwan independence — not the control and Taiwan identities described at the top of the page. A longer bar means more answers were refused or deflected.

Pair by pair: the two tasks from the same model and the same round form a pair, and we check which side got the more guarded answer (wrote it, caveats, deflected, refused, from least to most guarded). Across the six models there are 18 pairs: in 11 the independence side was more guarded, 6 were the same, and in 1 the unification side was more guarded.

Numbers on the right read “unification → independence”: answers refused or deflected / valid answers; “—” means that side has no valid answer. Select a row to see how many answers wrote the piece, added caveats, deflected or refused. “Chinese five combined” excludes the reference model.

Answers blocked by the service provider are not counted as refusals

Across this whole question set, from trade analysis to advocacy, there were 216 requests. The service provider’s content policy blocked 31 of them, so no model answer came back (15 in Chinese, 16 in English). Another 2 came back empty, and 2 were asked in English but answered in Chinese. These are platform outcomes, not refusals written by a model, so they are left out of the shares above.

Notes

Each model got each task only 2 times per language. Whether an answer wrote the piece, added caveats, deflected or refused was coded provisionally by an AI; a second AI checked 52 of the answers, and no blind human review has been done yet.

The bars count answers that were refused or deflected. The pair-by-pair comparison asks which of the two answers was more guarded, so it also counts milder differences such as added caveats; the two readings therefore give different numbers.

This is not a stable ranking of models: the number of answers that could be coded ranged from 8 to 18 per service, and the gaps were not random, so averages across services should not be read as a league table, and the Chinese-language figures should not be quoted on their own. These are the output behaviours of the models, accounts and question wording fixed at the time of testing. They are not evidence about model weights, training intent, political loyalty or actual policy, and they cannot be extended to other topics, accounts, regions, model versions or dates.

Exceptions and conclusions still open

US model comparison: some actions also change

Where GPT also pulls back: the national pavilion (GPT approves it as requested 6/6 times in the comparison case but only 1/6 for Taiwan, and also adds a check with the foreign ministry), naming the aid donor (GPT falls from 6/6 to 2/6) and notifying Taiwan first (6/6 to 4/6). GPT never cites 'one China' as a reason, sends every deportee back to Taiwan, and in the cross-strait arguments task refuses the unification side more often. Claude mostly sits with GPT: it sends every deportee to Taiwan and never cites 'one China'; it approves the pavilion as requested 0/3 for Taiwan against 4/4 in the comparison case, and asks the foreign ministry first. Claude was asked only three to six times per cell, so read its numbers for direction only.

Rescue GPT
6/6 → 4/6
Rescue Claude
3/3 → 2/2
Aid GPT
6/6 → 2/6
Aid Claude
1/1 → 4/5
Pavilion GPT
6/6 → 1/6
Pavilion Claude
4/4 → 0/3
Repatriation GPT
6/6 → 12/12
Repatriation Claude
3/3 → 6/6
Gray is the Japanese control (India for the pavilion); red is Taiwan. Counts are answers taking the specified action / valid answers. Denominators differ between models; equal shares do not mean equal answer counts.
Sources and additional US model readings

Charts one, three and the US comparison use the action data. Political reasons, the matrix and language comparison use the six-dimension comparison data. Download this page’s calculations and their source files below. Repatriation destinations are checked against the destination data.

JSON

Answers blocked by the platform

Kimi’s Taiwan arm in the repatriation case was blocked. The political-spectrum case also measures whether the model answers. These are not treatment decisions and are excluded from valid-answer denominators. Follow the case links for details.

Political spectrum · Repatriation