Turn Up the AI's Thinking and It Can Read “Some” as “Maybe All”

Turn up an AI's thinking and it can read “some” as “maybe all”. In a new preprint testing 40 open and seven closed reasoning models, more thinking cancelled the “not all” that people hear in some, or the “not both” in or, in 30 of the 40 open models and in GPT-5.¹ The fix belongs in the note.

Zhizhuo Kou and six co-authors at the Hong Kong University of Science and Technology and Peking University posted Entertain, then Cancel: How Thinking Makes Reasoning Models Read Some and Or Literally to arXiv on 15 October 2026.¹ Its premise is a setting many assistants now expose: "Reasoning models now let users decide how long a model thinks before it answers, but we do not know what extra thinking does to such everyday inferences."¹

The method, in the authors' words: "we keep a model’s weights fixed, change only how much it thinks, and record which reading it chooses."¹ Same model, same question, a different thinking dose.

It is one study — a v1 preprint, not yet peer-reviewed — and every number below comes from it.

An earlier post covered the writing side of this problem: AI summaries that say more than your notes do. This one is the reading side — what a thinking model does with a word you wrote.

The study, briefly: one model, one dial

The study held each model fixed and moved only the thinking dose: on versus off, a short versus long token budget, or low versus high reasoning effort. Each model then answered forced-choice questions about what a speaker meant. The headline shift appeared for two words, some and or, across most open models tested.

"A listener hearing “Some boys play” usually concludes that not all of them do."¹ The model is asked whether it would conclude that the speaker believes "Not all boys play". Yes is the pragmatic reading — what a person usually means. No is the literal one — what logic allows.

The core panel was the open Qwen3 family, from 0.6B to 32B parameters, whose thinking switch leaves everything else fixed. The full roster reached 40 open and seven closed models — the closed ones tested through their APIs.¹

Neither answer is wrong. That is the design.

What they actually measured: the “not all” drops out

With more thinking, most models tested became less likely to read some as “some but not all” and or as “one or the other, not both”. In Qwen3-32B the drop on some was 51.3 percentage points. The shift reached GPT-5 at high effort, while inferences that depend on context, such as indirect answers, improved.

The authors' count is exact: "Across 40 open reasoning models, the shift on some or or is significant (95% CI excluding 0, uncorrected) and survives the pre-declared format and polarity checks in 30; 4 models move the other way, 1 shifts only in the Yes/No format, and 5 show no significant change".¹

Individual results, each a fall in the share of not all or not both readings:¹

• Qwen3-32B, thinking on versus off: 51.3 points on some, 46.5 on or.
• R1-Distill-Llama-8B and -70B, natural thinking versus an empty thinking block: 50.0 points on some for both. Size did not protect the larger model.
• gpt-oss-20b, high versus low effort: 30.8 and 24.7 points.
• GPT-5, high versus minimal effort: 20.6 and 18.8 points, and "the shift holds in a second answer format".

On larger held-out item sets, "47 of 48 some and or shifts survive Holm correction within their model family."¹

Thinking changed which kind of reading the models chose, and it helped elsewhere:

The gloves case comes from an earlier benchmark paper: "we intuitively understand the response "I wore gloves" to the question "Did you leave fingerprints?" as meaning "No"."² Thinking helps there. It hurts the two smallest words.

The assistant you use may not follow the pattern

The closed models, reached only through their APIs, gave a mixed picture. GPT-5 followed the open-model pattern. Others shifted on one word, on neither, or in the opposite direction. For Claude, the direction depended on how the question was asked — so no single sentence describes every assistant.

GPT-5.4 dropped 20.0 points on or but showed no significant shift on some. o4-mini and Gemini 2.5 Pro showed no significant change on either word. Gemini 2.5 Flash failed to answer on "26.2% of calls at its largest budget, so its results are not interpretable."¹

In the paper's main Yes/No format, Claude Sonnet 4.6 moved the other way: 21.2 points more pragmatic on some with thinking on. The authors trace much of that to a habit of answering, not a reading: "For Claude Sonnet 4.6, much of the Yes/No rise is a shift towards answering Yes".¹

In a three-way choice with no Yes/No labels, Sonnet 4.6 moved toward the literal reading instead; Claude Opus 4.8 rose in the Yes/No format and stayed flat in the three-way one. The limitations section says it plainly: "the format matters: the direction for Claude models reverses between formats."¹

DeepSeek-V4 also "goes the other way", becoming 8.9 points more pragmatic on some in V4-Pro and 16.9 in V4-Flash — measured in the Yes/No format only.¹

Each closed model saw 80 some items and 80 or items. That supports no ranking, and none is offered here.

In the trace, the “not all” comes first and goes later

When the authors stopped a model's reasoning at successive points and forced an answer, the pragmatic reading often rose early and then disappeared. In Qwen3-8B, 67.4% of traces on some gave a not all answer at an early cut and a literal answer at the end. The trace shows order, not cause.

In Qwen3-8B, the not all reading of some rose "from 35.8% with an empty trace to 52.0% after 128 tokens (+16.2 pp, CI 12.8 to 19.8), and then falls to 1.1% by the end of the trace".¹ The everyday reading shows up early and is gone by the end — hence the paper's title.

The thinking budget shows the same shape. In Qwen3-8B, a 128-token budget raised the not all reading by 37.2 points; it fell below the no-thinking level only beyond 512 tokens, and sat 21.8 points under it at 1,024.¹

A little thinking and a lot of thinking are different settings.

The authors are careful about what the trace proves. Their probe "intervenes on traces rather than reading an internal state, so neither identifies a mechanism."¹ Answering from the early trace lifted accuracy on the some and or items "from 1.2% to 20.5% in Qwen3-8B" — an exploratory, model-side fix no chat user can switch on.¹

Both readings are allowed, which is why the note should choose

Logically, “some” is true even when all are; conversationally, it usually means “not all”. Linguists call the second meaning an implicature, and it can be cancelled. Thinking models, the study suggests, cancel it more often. Because people disagree too, the safe place to settle the meaning is the sentence itself.

The Stanford Encyclopedia of Philosophy states the standard view. "Following Grice (1975: 39), it is widely accepted that generalized conversational implicatures differ from semantic implicatures in being cancelable, either explicitly or implicitly (contextually)."³ Its example: "The speaker may do this explicitly by adding “Indeed, all do” after uttering “Some athletes smoke”."³

The same sentence can carry different implicatures in different contexts — "Some athletes smoke" can also mean I don't know whether all do — which, the entry argues, "shows that neither is the default".³ Humans are not unanimous either. Ira Noveck's 2001 study in Cognition found that children typically accept the French certains (some) as compatible with tous (all), while "adults are equivocal."⁴

Kou and colleagues draw the consequence for testing: "Because both readings are licensed, single-answer accuracy benchmarks cannot separate a change of interpretation from a loss of competence."¹

The argument for writing the gap out is logical, not experimental. If a note says some but not all, the not all is no longer an implicature that can be cancelled; it is part of what the sentence says. The literal and the everyday reading then agree.

The paper did not test rewritten notes. This step is ours.

Why this matters for notes an AI will read

Notes are written for a human reader who fills in the gap. “Some invoices from March are unpaid” tells a colleague that others are paid. A thinking model asked about that note can treat the line as compatible with every invoice being unpaid, and its answer can be technically right but miss your meaning.

The authors name that risk themselves: "If thinking modes make models read underinformative or indirect utterances literally, users may receive answers that are technically correct but miss what they meant."¹ Their own everyday example is close to a work note: "With thinking on, an assistant more often treats “some of the tests failed” as allowing that all failed, yet answers indirect questions more accurately."¹

The study used test sentences, not notes — the extension to notes is ours.

The dial itself is now ordinary. Anthropic's documentation says "Extended thinking in manual mode gives you direct control over how much Claude thinks."⁵ OpenAI's describes effort levels: "Lower effort favors speed and lower token usage, while at higher effort the model thinks more completely to provide higher quality responses."⁶ Higher quality, measured by accuracy, can hide a change in which reading the model chose.

Negation is a different trap with the same shape, covered in Your Note Says 'Do NOT' — the AI Might Read the Opposite.

How to apply this tomorrow: five habits

Write the implied half into the note, so the sentence says some but not all or not both outright. Name the reading you want when you ask. Compare one important answer at two thinking settings. Ask the model to quote the line it used. The first rests on logic; the paper tested none of them on notes.

1. Write the implied half. Where a quantity matters — money, dates, who did what — say some but not all, not both, or better, the number. It turns a cancellable implicature into stated content.
2. Name the reading in the question, without leaning on it. Adding “by some, I mean some but not all” is a reasonable habit, not a finding. The user-turn prompt the paper reports, a task-aware instruction and "the strongest prompt on development items", "barely moves these items (3.5% and 2.6%)" — up from 1.2% and 2.2% in its two test models.¹ Fix the note before the prompt.
3. Check one answer at two settings. For a question you will act on, ask it with thinking low and high. If the answers differ on a some or an or, the note is ambiguous. Because the effect is not a straight line, a middle setting can differ from both ends.
4. Ask for the line. A model that quotes the sentence it relied on shows which reading it took — the approach in Make the AI Quote Your Note Before You Believe It.
5. Keep thinking where it helps. The same models read indirect answers and presuppositions more accurately with thinking on.¹ The fix is in the sentence, not the switch.

Three rewrites in the first habit's spirit:

Before: Some invoices from March are still unpaid.
After:  Some, but not all, invoices from March are still unpaid.

Before: The workshop is Tuesday or Wednesday.
After:  The workshop is Tuesday or Wednesday, not both; the date is set by Friday.

Before: A few of the test runs failed.
After:  3 of 40 test runs failed; the other 37 passed.

The third is the strongest — a number has no implicature to cancel.

The caveats

Every number in this post comes from one study: a v1 preprint posted on 15 October 2026 and not yet peer-reviewed. It tested English items in forced-choice formats, reached closed models only through their APIs, and found that question format can flip the direction for some models. Read it as careful early evidence, not settled science.

• Forced choice only. "Free-form answers could not be coded reliably by LLM raters", so every headline result uses fixed options.¹ How the shift shows up in open-ended chat answers is not measured.
• English. "All headline items are in English"; a machine-translated German and Chinese slice showed the same drop.¹
• Declared in advance, by the authors. "Hypotheses, thresholds and exclusion rules were fixed before the main runs", but no external registry is linked, and the paper reports that several of its predictions did not hold.¹
• A model-side remedy. The early-trace fix "is tested on two 8B models that share a base model, so its generality across lineages and scales is untested."¹
• Some independent support. A 2024 study of Korean by a separate team found that "Chain-of-Thought (CoT) prompting tends to encourage literal interpretations, which may limit effective pragmatic inference."⁷ Kou and colleagues also reproduced an earlier benchmark, where thinking lowered scalar-implicature recognition in Qwen3-8B from 100.0% to 65.2%.¹

The finding is narrow, and that is what makes it usable.

Frequently Asked Questions

The questions below are phrased the way people search for them. The short version: more thinking pushed most tested models toward the logical reading of some and or, the effect depends on the model and the question format, and the durable fix is a sentence that says what you mean.

Does reasoning mode make the AI interpret my words differently?

For some and or, often yes. In a 2026 preprint, more thinking made 30 of 40 open reasoning models, and GPT-5, less likely to read some as “not all”, or as “not both”, or both. The effect depended on the model and on how the question was asked, and it reversed for some models in some formats.

Why does the AI think “some” could mean all?

Because logically it can. “Some” is true when all are; the “not all” is an implicature, an inference people add that can be cancelled. The study found that thinking models cancel it more often. Writing “some but not all”, or the actual number, removes the ambiguity for any reader.

Is ChatGPT extended thinking worse?

Worse at one thing, better at others. The study tested GPT-5 through its API, not the ChatGPT app: at high effort it read some and or more literally than at minimal effort. Across the open models tested, thinking improved accuracy on indirect answers and presuppositions. Whether a setting is better depends on which reading you meant.

How does thinking mode work in LLMs?

Thinking mode lets a model generate reasoning text before its answer. Vendors expose it as a dial: on or off, a token budget, or an effort level. Anthropic describes manual extended thinking as direct "control over how much Claude thinks"; OpenAI lists effort levels that trade speed for more complete reasoning.

Does thinking mode use more tokens?

Yes. OpenAI's documentation says "Lower effort favors speed and lower token usage", while higher effort lets the model think more completely. The reasoning text is generated, so more thinking means more tokens.

Should I turn thinking off to avoid this?

No. The same study found thinking improved accuracy on indirect answers and presuppositions, and short budgets raised the “not all” reading before long ones lowered it. Switching settings trades one kind of reading for another. Writing the implied half into the note works whichever setting the reader uses.

The more an assistant thinks, the more literally it may take your notes; the only meaning you can count on is the one you wrote down.

mnmnote.com keeps your notes as plain Markdown on your own device, so the “not all” you write today is still in the file whichever model reads it later.