Coherent texts, whether written or spoken, are built upon discourse relations linking utterances together through causal, temporal or contrastive connections, among many other types (Mann & Thompson 1988). Different types of relations are signalled by different types of markers, although there is no one-to-one mapping. These markers often belong to the functional category of discourse-relational devices, or “connectives”, such as however, because or in fact. Writers and speakers also have the option to use other signalling devices (e.g. lexical or syntactic patterns) or even to leave a discourse relation implicit (e.g. Taboada 2009). This study focuses on another strategy for discourse marking, namely the use of underspecified connectives (Spooren 1997). More particularly, we investigate the role of the additive conjunction and to signal relations of addition but also of consequence, contrast and concession. In these cases, the discourse relation is more specific than the information strictly provided by the connective: a consequence or contrast is more informative than a mere additive relation. Despite the low informative value of and, it is quite often found in authentic contexts where such enriched interpretations were assigned to the discourse relation (6% of and express a result in the Penn Discourse TreeBank 2.0, Prasad et al. 2008), which calls for more research on the conditions under which and can be used as an underspecified connective. Our research objective is thus to compare the linguistic and contextual features of utterances linked by and which either express addition, consequence, contrast or concession. To do so, we first need corpus-based data where such discourse relations are reliably identified. Discourse relation annotation is extremely costly in time and human resources, it requires heavy training and, even so, agreement scores are often rather low (Spooren & Degand 2010). As a result, researchers have recently started to turn to crowdsourcing as an alternative method to gather discourse relation disambiguations through a low-cost, non-expert workforce (Kawahara et al. 2014; Rohde et al. 2016). Scholman & Demberg (2017) report on the results of a connective insertion task which they used as an indirect method to annotate discourse relations: the sense of a relation can be retrieved through the selection of unambiguous connectives from a list to fill in a blank between utterances, provided this task is repeated by a large number of participants (around 20). The authors discuss the validity of this method and conclude that crowdsourcing connective insertions is reliable enough as an alternative to expert annotations. For connective insertion tasks, it's important to distinguish between originally implicit vs. explicit relations. For originally implicit relations, inserting a connective resembles the approach taken in PDTB annotation (except that the choice of connectives is more restricted in the crowdsourcing step in order to allow for disambiguation of relation type). In originally explicit relations, however, the meaning and interpretation may substantially change by removing the connective, see examples (1)-(3) below. In this study, where we investigate originally explicitly marked relations with the connective “and”, we will therefore compare a crowdsourced connective insertion task with a crowdsourced connective replacement task, in which the original connective “and” is not removed from the stimulus. (1) I am not going back to Germany. Therefore, I will not eat spätzle ever again. (consequence) (2) I am not going back to Germany. In fact, I will not eat spätzle ever again. (addition) (3) I am not going back to Germany. I will not eat spätzle ever again. (?cause) Our working hypothesis is that such differences in interpretations, triggered by the connective (or absence thereof), apply to utterances containing and as well. In this respect, we challenge previous experimental research on and which showed that and has a very small informative value and little or no facilitating effect on reading times (Murray 1994) or comprehension (Cain & Nash 2011). By contrast, we expect that connective insertion tasks will be affected by the presence of and, thus supporting the claim that and does trigger enriched pragmatic inferences (Blakemore & Carston 1999). We therefore propose to use crowdsourcing for the study of underspecified and by comparing stimuli with and without the original connective. In other words, we want to test whether connective elicitations will differ across stimuli which are identical except for the presence or absence of the conjunction and. To this end, we ran two crowdsourcing experiments on the Prolific Academic online platform. In the first one, we used 83 authentic pairs of utterances originally containing and. We collected them from the Loyola Corpus of Computer-Mediated Communication (Goldstein-Stewart et al. 2008), in order to avoid the high formality of existing corpora such as the Penn Discourse Treebank (economy newspaper articles). This corpus contains blogs and chat conversations between college students about topics such as gay marriage, gender discrimination or privacy rights. This data was pre-annotated by the first author as either expressing a relation of addition, contrast, concession or consequence. The stimuli are grouped in four lists of about 20 items each, which are balanced with respect to the pre-annotated relation type (about 10 addition, 6 consequence, 1 contrast, 3 concession in each list). The participants (paid 1€ per list) can choose from a list of eight connectives to fill in a blank between the two utterances (the original and has been removed). The connectives are in addition, plus, therefore, as a result, by contrast, whereas, nevertheless and yet. The second experiment uses exactly the same lists of items, except that the stimuli now show the original and connecting the utterances, and the participants are therefore instructed to substitute this and with one of the connectives from the same list of options. In the analysis, we first compare the connectives chosen by the participants with the relation type pre-identified by the expert annotator, in order to see whether they converge (e.g. therefore or as a result selected in case of a relation of consequence). This first step provides us with a dataset of utterance pairs with their disambiguated discourse relation, without resorting to costly (and partly subjective) expert annotations. The items thus classified into one of the four categories (addition, consequence, contrast, concession) will allow us to test the effect of additional variables (e.g. register) in further studies (Crible & Demberg 2017). We will replicate this analysis with the results of the second experiment. We will then compare whether the connectives selected by the participants are the same when and is present in the stimuli and when it is not. Preliminary results show that, when the relation was pre-annotated as additive, the participants tend to equally choose a consequence or an additive connective, with no significant difference, which suggests that consequence is often interpreted in the absence of a connective. For all other relation types (i.e. consequence, concessive and contrast), the great majority of participants’ choices match the pre-annotation, thus confirming that and can be used in contexts which express more than mere addition. The results from the second experiment are still pending, and should lead to interesting comparisons on the effect of and in connective elicitation (or connective substitution). This study has a number of implications on the informative value of and, which may be higher than what previous studies have suggested, and on the use of connective elicitation with or without the original connective included in the stimuli. We argue that, when dealing with authentic corpus-based stimuli, including the original connective in the experiment is a more accurate representation of the data and better reproduces the interpretation mechanisms as they would be processed in natural conditions. The experiments reported in this paper constitute the first step of a larger project on the contextual and cognitive constraints to the production and interpretation of underspecified connectives. They also relate to ongoing crosslinguistic projects on the meaning variation of and and its use across spoken registers (Crible, in press) and in translation (Abuzcki et al. 2017).