A 200-word lexicon already sees the US-vs-China AI sentiment gap.

A real reliability diagram on day one

US AI media coverage is shifting from boom-cycle enthusiasm toward skepticism. Chinese AI media coverage is moving the other way. A deliberately small bilingual word-list scorer — about a hundred positive terms and a hundred negative ones in English and Chinese — was run against the 60-article hand-coded calibration set built in the previous release. On a sentiment scale from −1 to +1, the scorer agrees with the human coders about 72% of the time on a three-way bull / neutral / bear classification, with a mean absolute error of 0.23 and a Pearson correlation of +0.76. Most importantly, the headline pattern — US coverage normalizing while Chinese coverage stays ascendant — reproduces cleanly even at this very simple level of modeling. The lexicon is now the baseline that the LLM scorer has to beat.

Why ship a deliberately weak scorer

The founding paper made a load-bearing claim: U.S. AI media is normalizing while Chinese AI media is ascendant, and the compression of the gap between the two is the indicator. The first release backed this with ten quarterly anchor values drawn from a careful narrative reading of the press of record. The seed is honest about what it is — ordinal anchors, not machine-scored data. But a narrative reading is exactly the kind of thing that pattern-matches to what its author already believes. The whole point of publishing a public, dated, scored index is to have a number that disagrees with you when you're wrong.

The fastest way to get a real number on the page was a word-list scorer. A bilingual lexicon and a simple hit-ratio is not a serious sentiment model. It can't handle negation, sarcasm, irony, or the case in which Chinese press of record runs U.S.-bear stories as Chinese bull. But it can be written in a day, runs on every machine without an API key, and produces a real number against a real calibration set on a real reliability diagram. That number then has to be beaten by the LLM scorer. If the lexicon already recovers the arc, the LLM run has a low bar to clear. If the lexicon misses the arc, the LLM has real work to do.

The lexicon recovers the arc. The cheapest possible model already sees what the dossier described. That is itself a substantive finding: the U.S.-normalizing, China-ascendant pattern is not a narrative artifact. It is robust enough that a two-hundred-word vocabulary picks it up.

Two hundred words, English and Chinese

The scorer is a bilingual word list: roughly a hundred positive tokens and a hundred negative ones, spanning English and Chinese. The positive list includes terms like breakthrough, frontier, capability, ascent, and their Chinese equivalents (突破, 领先, 跃升). The negative list includes slip, cancel, questioned, retreat, bubble, and their Chinese equivalents (放缓, 取消, 泡沫). Each article's sentiment score is a smoothed normalized hit ratio — positive hits minus negative hits, divided by the total — computed over the article's title, summary, and the human coder's one-sentence justification. The output sits on the same −1 to +1 scale the human coders used.

The calibration set is sixty articles — thirty U.S., thirty Chinese — spanning six blocks. On the U.S. side: peak bull (2023-Q1 through 2024-Q2), normalization (2024-Q3 through 2025-Q2), and anxiety (2025-Q3 through 2026-Q2). On the Chinese side: slow-build (2023 through 2024), confidence (2025), and ascendant (2026). Each article carries a hand-coded score on the same scale and a one-sentence justification. The hand-coding was done block-by-block, not article-by-article, to enforce internal consistency within each regime.

Within about a quarter of a unit, on a five-bin diagram

The headline numbers: the lexicon agrees with the human coders within an average of about a quarter of a sentiment-unit (mean absolute error of 0.227, root-mean-square error of 0.324, Pearson correlation of +0.76). On a three-way bull / neutral / bear classification the scorer agrees with the humans 71.7% of the time overall — 60.0% on U.S. articles and 83.3% on Chinese ones. The reliability diagram has four populated bins out of five, all sitting within 0.30 of the diagonal. The strongest agreement is at the bull end: in the most-bullish bin (twenty-three articles) the scorer and the humans agree almost exactly. In the mildly-bullish and neutral bins the scorer is within 0.10 to 0.12 of the human mean.

The honest failure mode is the bear side. On the mildly-bearish bin (fourteen articles) the scorer reads almost 0.30 more bearish than the humans. The lexicon over-fires on bear-register vocabulary — words like slip, cancel, and questioned — even when the article is a Chinese piece celebratingthe U.S. bear pivot. A People's Daily piece on the Stargate slip reads as bearish under the lexicon when the human coder marked it as cautiously bullish for the Chinese side. The most-bearish bin is empty: the lexicon never gets that bearish without an offsetting positive term firing.

The compression of the gap between US and Chinese AI sentiment — not its level — is the indicator.

The block-level ordering survives. On the U.S. side, the scorer nails peak-bull (+0.68 against the human +0.66) and anxiety (−0.33 against −0.33). On the Chinese side, it nails the slow-build (+0.30 against +0.31). It under-states the Chinese confidence and ascendant blocks — the bull register in Chinese AI coverage is muted, since the press of record talks about benchmark scores and capex commitments rather than “dominance” or “ascent.” It under-states the U.S. normalization block because half the bear pieces use technical capex language the lexicon doesn't catch. But the headline arc — U.S. normalizing while China stays ascendant— reproduces cleanly at the block level.

The thesis is robust to model choice

The substantive finding is that the U.S.-normalizing, China-ascendant pattern is not a function of model sophistication. The cheapest possible scorer recovers it. That makes it harder to dismiss the founding thesis as a confirmation-biased read of the press by an author who already believed it. A two-hundred-word list, run mechanically over a held-out calibration set assembled by a different process (block-stratified hand-coding rather than narrative reading), reproduces the same arc.

The implication for the next release is that the LLM scorer has to clear two bars. The first is calibration: the LLM should agree with the human coders more closely than the lexicon does, and should populate the empty most-bearish bin that the lexicon cannot reach. The second is the frame-aware handling of Chinese pieces celebrating U.S. bear stories — the lexicon scores those as bearish; the humans score them as bullish for the Chinese side. The LLM prompt is written to handle this case. The next release will tell us whether it actually does.

For policy and capital allocation, the under-stating of the Chinese confidence and ascendant blocks — by as much as 0.31 in places — is worth flagging. The bull register in Chinese AI coverage is muted: the press of record talks about benchmark scores and capex commitments rather than dominance or ascent. A sentiment model trained on Western AI bull-cycle vocabulary will systematically understate Chinese AI bull strength. The hand-coded scores caught this. The lexicon did not. The LLM should.

What this finding does not claim

Three limitations belong on this update rather than buried in an appendix. First, the calibration set is sixty articles. Sixty is small enough that a single mis-coded article can move headline metrics. A later release expands the set to five hundred or more. Second, the lexicon is hand-tuned. The word lists were assembled in an afternoon, not learned. A larger list, or a different smoothing parameter, would shift the numbers; we did not sweep hyperparameters because the point was to ship a baseline, not to polish the lexicon. Third, the reliability diagram has only four populated bins. The most-bearish bin is empty by construction of the scorer, not by absence of bear articles in the corpus. A real LLM scorer will populate that bin and the diagram will be richer.