Five frontier models read the same news story and write their own summary. This is a fair ruler for comparing what each one keeps and drops — measured by geometry, not opinion. No model judges another model at any point.
Think of five students handed the same article, each turning in a short book report. The interesting question isn't whose report is best — that's just an opinion. It's what they systematically do to the same facts: which words survive, which quietly vanish, and whether the choice depends on what the fact is. To ask that without opinions, we turn each summary into a point in a high-dimensional space that captures its meaning, and measure the differences with arithmetic. The same inputs always give the same answer. Three things came out of that measurement.
When a story says a company "quietly" or "secretly" did something, the models tend to carry that word through. Unless the company is an AI developer — then they're measurably more likely to soften or drop it. Same sentence, same loaded word, only the named actor changed.
We pre-registered the test before scoring it. Matched real incidents on each side — an AI company and a conventional corporation, same modifier, same structure — and measured whether the modifier's meaning survived in each summary, not whether the exact word reappeared. The AI-developer version retained the modifier less, consistently.
The control that makes it real: swap one corporation for another corporation and the gap nearly vanishes (0.004). The AI-versus-corporate swap is six times larger (0.023). So the effect tracks the kind of actor, not the act of swapping. And it holds equally across heavily-aligned and lightly-tuned models (p = 0.46 between them) — which points at the shared training data they all learned from, not any one company's safety tuning.
The common assumption is that AI "sanitizes" the news — strips out the charged, operational words ("devastated," "killed," "secretly") and keeps the safe institutional ones ("committee," "agency," "announced"). It's the reverse. Across 1,592 stories, the models held onto the charged, consequential language more than the bland bureaucratic language.
For every meaningful word in each source article, we measured whether its meaning survived into the summaries — semantically, so a reworded synonym still counts as kept. Words leaning "operational/consequential" were retained more than words leaning "institutional/structural," across 106,412 terms.
We ruled out the obvious objections. It isn't a rarity effect — it holds in all five frequency bands. And it isn't just "summaries keep the specific bits and drop filler" — controlling for how informative each word is (its IDF), informativeness explains almost none of the variance (R² ≈ 0.002), and the charged-language effect survives the control regardless. Frontier models, summarizing news, preserve the high-stakes specifics over the institutional scaffolding — the opposite of the sanitization story.
These models were frozen at a training cutoff around mid-2024. Hand one an article about a person who became prominent after that — a newly-elected president, a just-appointed minister — and it quietly under-weights them, even though the name is sitting right there in the text it's reading. Like someone reading a newspaper aloud who mumbles past the names they don't recognize. The established names — Putin, Netanyahu — come through clearly. The new ones get flattened.
Sorting the source vocabulary by how reliably each word's meaning survives, the lowest-retention named figures share one property: they rose to prominence at or after the cutoff. We built matched buckets — post-cutoff figures versus established ones — and measured retention of each.
The confound to kill is prominence — maybe minor figures just drop, cutoff aside. But the post-cutoff group includes a sitting president and a defense secretary, and they're still under-retained relative to established heads of state. Prominence can't explain a president being flattened. The control for transliteration holds too: established foreign names (Khamenei, Xi) are retained well, so it isn't that non-English names embed poorly — it's specifically the post-cutoff ones. The frozen weights act as a salience filter on fresh text: the model reads the name and still discounts it, because its prior doesn't flag the entity as important.
Honest limits. The name list was chosen after looking at the data, not pre-registered — so this is strong preliminary evidence, not a closed result; a pre-registered replication on held-out names would seal it. And whether this extends from named people to events and concepts is genuinely unresolved: an event term ("the Iran war") carries a familiar topic that backfills the unfamiliar specifics, so the instrument can't yet tell a true null from one it's blind to. That extension is open, and we say so rather than guess.
Each dot is a named figure appearing across many stories. Higher means its meaning survives summarization better; left is established before the cutoff, right is risen after it. Dot size is how many stories the figure appears in. The two clusters are the finding — hover any dot.
Every measurement uses frozen BAAI/bge-large-en-v1.5 embeddings (1024-dimensional, deterministic). "Retention" of a word is the maximum cosine similarity between that word's embedding and any sentence of the model summaries — paraphrase-proof by construction, since it scores whether the meaning survived, not whether the exact string reappeared. This is the key methodological choice: a naive string match would count every reworded synonym as a deletion and manufacture false signal, because models paraphrase constantly. Measuring semantically is what separates a real omission from a rewording. One known limitation runs the other way: because dropping a single modifier ("secretly" from "secretly acquired") shifts a sentence's embedding only slightly, the metric is conservative about modifier loss — it can score a stripped modifier as retained. That means the entity-swap effect in Finding 1, which targets exactly such modifiers, is more likely understated than inflated.
The entity-swap (Finding 1) was pre-registered. Findings 2 and 3 carry the controls described above — frequency, informativeness, transliteration, prominence — each with its own null condition, and each documented in the released results. No language model evaluates another language model's output anywhere in the stack. The host model that narrates the broadcast (Mistral Small, running locally) is subject to the identical measurements; we include it rather than exempt it.