📄 Free choice for all: a response to Emmanuel Chemla

Summary

This brief reply responds to Chemla's (2009) experiments on "embedded implicatures." Chemla claimed his results trouble both globalist and localist accounts of quantity inferences, and globalism most of all in the case of free choice. Geurts and Pouscoulous first challenge how he frames the debate. Treating it as a question of where an exhaustivity operator O may be inserted already takes the localist view. The real issue is whether these inferences are Gricean pragmatics or products of grammar.

They then reassess his data. For scalars embedded under universal or negative quantifiers, Chemla's ratings match their own earlier findings and count against localism. For free choice, they argue three things:

• The negative items needed marked contrastive stress on "and," so they cannot fairly test either theory.
• The strong universal result may not generalise to other modals or to non-modal free choice.
• A Gricean account is available. Addressing a group distributively amounts to separate speech acts to each member, which reduces the universal case to ordinary singular free choice permission.

Key points

• In Geurts & Pouscoulous (2009) Experiment 1, a plain scalar inference was accepted 93% of the time, while its embedded universal counterpart was accepted only 27%. That study used "some" with yes/no answers; Chemla used "most" with scale ratings, yet the pattern was the same.
• In their verification tasks (Experiments 3 and 4), only one participant, on a single trial, answered in line with localism. Together with Chemla's data, four distinct paradigms point against localist predictions.
• On a standard reading of O, deriving free choice requires recursive exhaustification (OO, following Fox 2007). Chemla's single-operator "globalist" version effectively defines a new iterated operator O*. The authors see this as a technical device, not a pragmatic explanation.
• The exhaustivity operator builds in the speaker-competence assumption. It therefore yields only the strong implicature (speaker believes not-all) and cannot express the weak one (speaker does not believe all).
• Inference-endorsement tasks are biased toward "yes," so absolute rates mean little. Only contrasts between conditions are informative.

Interesting ideas

• I1. Implicature can be modelled as two steps: a weak epistemic inference first, then strengthening only if the context supports speaker competence; this keeps the competence assumption explicit and defeasible.
• I2. A distributive utterance to several addressees can be treated as several parallel speech acts, one per person, which turns a group-level puzzle into a solved individual case.
• I3. When a task inflates acceptance overall, trust differences between matched conditions rather than raw scores.
• I4. Readings that depend on marked prosody, such as a contrastive or "together with" sense of "and," fall outside the general theory and should be set aside rather than forced into it.
• I5. A claimed generalisation should be tested across embedding environments: deontic, ability and counterfactual modals, and non-modal existentials.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): crate prag could store both the weak and the strengthened implicature, linked by an explicit competence-assumption flag that the world layer can switch off when the speaker seems poorly informed.
• I2 (frames): the speech-act layer could expand plural "you" permissions or orders into per-addressee acts, so each entity in the world of the text receives its own permission frame.
• I3 (learn): evaluations of inference or implicature modules should pair each target item with a minimally different control and report the gap, not the bare accuracy.
• I4 (frames): coordination frames could carry a "possibly marked reading" flag, so downstream answers on such sentences are given lower confidence instead of a confident wrong inference.
• I5 (answer): QA test sets for disjunction inferences could vary the embedding systematically, showing whether a reader overgeneralises free choice beyond the cases where it is robust.