should be sent to Szilvia Biro, Department of Psychology, Centre for Child and Family Studies, Leiden University, Wassenaarseweg 52, 2333 AK Leiden, the Netherlands. E-mail: email@example.com
It has been shown that, when observing an action, infants can rely on either outcome selection information (i.e., actions that express a choice between potential outcomes) or means selection information (i.e., actions that are causally efficient toward the outcome) in their goal attribution. However, no research has investigated the relationship between these two types of information when they are present simultaneously. In an experiment that addressed this question directly, we found that when outcome selection information could disambiguate the goal of the action (e.g., the action is directed toward one of two potential targets), but means selection information could not (i.e., the action is not efficiently adjusted to the situational constraints), 7- and 9-month-old infants did not attribute a goal to an observed action. This finding suggests that means selection information takes primacy over outcome selection information. The early presence of this bias sheds light on the nature of the notion of goal in action understanding.
To infer that an observed action is goal directed (as opposed to random or accidental), that is, that it is performed in order to achieve a particular outcome, one can make use of two types of information that can be directly observed in a given situation. One type of information consists of the observation of the way in which the actor adjusts the action to the situational constraints in order to achieve the end state. Relying on this type of information implies the understanding of the relation between the means and the goal and the expectation that the actor will select the most efficient means action toward the goal. The other type of information consists of the observation that there are potential alternative outcomes and that the actor expresses a choice for a particular outcome via the action. We refer to these two types of information as “means selection” and “outcome selection” information, respectively.
Infants’ ability to interpret actions as goal directed has been investigated using two distinct experimental paradigms that have so far separately tested if and how infants can rely on one or the other type of information. Evidence for reliance on outcome selection information was found by showing that infants, after having watched an actor repeatedly acting upon (e.g., grasping) one of two objects, expect that the actor will continue to direct the action toward the same target after the spatial arrangement of the two objects has been altered (e.g., Biro & Leslie, 2007; Kiraly, Jovanovic, Prinz, Aschersleben, & Gergely, 2003; Woodward, 1998, 1999). Furthermore, when outcome selection information is not provided in this experimental paradigm, that is, when the actor’s action could not express a choice because there was only one object present, infants do not generate the expectation that the actor will act upon the same object after a novel object has been introduced (Biro, Verschoor, & Coenen, 2011; Hernik & Southgate, 2011; Luo & Baillargeon, 2005). Note that in these settings no means selection information is available because the information about situational adjustment of the action is insufficient (see Biro et al., 2011). Together, these findings suggest that outcome selection information is critical in infants’ goal attribution.
The second experimental paradigm demonstrates infants’ reliance on means selection information (e.g., Csibra, 2008; Csibra, Biro, Koós, & Gergely, 2003; Csibra, Gergely, Biro, Koós, & Brockbank, 1999; Gergely, Nádasdy, Csibra, & Biro, 1995; Sodian, Schoeppner, & Metz, 2004). Infants are typically presented with a scenario in which there is only one potential outcome, which the actor repeatedly achieves by adjusting the action to the situational constraints (e.g., approaching a target by getting around an obstacle). When the situational constraints change, infants are found to predict that the actor will perform a novel and efficiently adjusted means action to achieve the original outcome (as opposed to repeating the familiar but no longer efficient means). If, however, the actor’s action is not adjusted to the situation efficiently in the initial scenario, then infants do not generate a particular expectation about the new means action of the actor in a new situation. Thus, means selection information also has an important place in infants’ early goal-directed action interpretation.
In the present study, we combined the two kinds of paradigms to investigate how infants deal with situations which provide both types of information, that is, means selection and outcome selection, to disambiguate the goal. In the everyday lives of infants—as well as of adults—the two types of information are often present simultaneously. Actions do not take place in an empty space but in one with many potential targets and a constraining physical environment. We were most interested in infants’ goal attribution in a situation in which the two types of information lead to different interpretations about the goal directedness of an action. Specifically, when an action is directed toward one of multiple potential outcomes, but is clearly not adjusted efficiently to the situational constrains. Would infants interpret the action as goal directed by relying on the outcome selection information or would they be reluctant to make such interpretation due to the ambiguity of the means selection information.
According to the teleological stance theory (Gergely & Csibra, 2003), infants are equipped with an abstract interpretational system, which produces a teleological (goal directed) representation of an observed action if, and only if, its central component, the principle of efficient action, is satisfied. This theory thus predicts that infants only attribute a goal to an action if unambiguous means selection information is available, that is, the action can be evaluated as an efficient action toward the end-state in the given situation.
However, the available evidence that has so far been cited to support the crucial role of means selection information in disambiguating the goal is indirect. First, the cited studies (see above) explored infants’ predictions about the means, and not the goal, of the action. Thus, the findings that infants do not generate a specific expectation about the new means if the initial action is non-efficient only suggests, but does not prove, that infants do not in the meantime attribute a goal to the initial action. Infants may have given up on predicting the means, but could in principle have relied on the end state (the outcome) of the action to infer that it is the goal. Second, since in these experiments only one potential target was available, it is not yet known how infants interpret a non-efficient action when clear outcome selection information (multiple potential targets) is present to disambiguate the goal.
In our current study we therefore tested the hypothesis that ambiguous means selection information would prevent infants from disambiguating the goal even if clear outcome selection information is available. This hypothesis predicts that when an action is performed in the presence of multiple potential outcomes, infants would only attribute a goal to this action if it is adjusted efficiently to the situational constraints. On the other hand, if it is the outcome selection information that governs infants’ developing ability to attribute goals, then infants would interpret the action as directed to a particular goal regardless of whether the means is non-efficient toward this outcome.
Infants participated in three conditions. The Efficient Action condition involved familiarizing infants with a hand that first opens one of two transparent boxes that each contain a toy and then grasps the toy. In the non-efficient Gratuitous Action condition, the hand performs the same action, but the toys are in front of the boxes. Opening the box is therefore unnecessary and does not relate efficiently to the grasping.1 In the non-efficient Superfluous Action condition, the boxes are not present, the hand grasps one of the toys after making some unnecessary movements in the air above and behind the toy (which actually mimic the opening of a box). The test events were identical for all conditions: two toys are present in swapped positions (no boxes) and the hand simply grasps either the same toy as before or the other toy. If infants rely on outcome selection information, then they should expect the hand to grasp the same toy as before in all conditions. This expectation would be indicated by looking longer in the New toy test event than in the Old toy test event in all three conditions. If means selection is the crucial source of information, then infants should attribute a particular goal and thus look longer in the New toy test event than in the Old toy test event in the Efficient Action condition, but not in the non-efficient conditions. Furthermore, the two types of non-efficient approaches allow us to investigate whether the inefficiency of the action can only prevent goal attribution if there is a single action involved (as in the Superfluous Action condition) or when the means can be considered as a part of a sequence, a subaction leading to an overarching goal (as in the Gratuitous Action condition). One could argue that inferring the goal by considering the inefficiency of a subaction might be too demanding for young infants, and thus they might only rely in their interpretation on the second reaching action containing only outcome information. In the case of the presence of this transitional period, we would expect longer look in the New toy than in the Old toy test event in the Efficient and the Superfluous Action conditions but not in the Gratuitous Action condition.
Sensitivity to each type of information separately has been shown to be present from as early as 6 months (Csibra, 2008; Woodward, 1998). The combined presence of both types of information creates a visually complex stimulus and could thus be cognitively demanding; therefore we included 7- and 9-month-olds in the study to check for possible age effects.
Seventy-two 7-month-olds (mean age = 31.10 weeks; SD = 1.7 weeks; 36 girls and 36 boys) and seventy-one 9-month-olds (mean age = 39.60 weeks; SD = 1.08 weeks; 39 girls and 32 boys) were randomly assigned to one of three conditions: Efficient Action (23, 24), Gratuitous Action (24, 24) and Superfluous Action (24, 24), respectively. An additional eight 7-month-olds and fourteen 9-month-olds were excluded due to experimental error (3, 7) and fussiness (5, 7), respectively.
Infants were shown video recordings (Fig. 1). In all conditions infants saw six familiarization events. In the first familiarization event of the Efficient condition, two toys (a bear and a cube) were sitting on a stage which had a curtained opening on the right side. In the second familiarization event, infants were shown that the two toys were sitting inside two transparent boxes with red lids. In the 3–6 familiarization events a hand reached through the opening, opened one of the boxes by removing the lid, and grasped the toy inside. The hand remained still and kept grasping the toy until the end of the event. The Gratuitous Action condition was identical to the Efficient Action condition except that the toys were not inside but in front of the boxes. Thus, after the hand removed the lid, it grasped the toy in front of the box. In the Superfluous Action condition the boxes were absent. In the first two familiarization events only the two toys were sitting on the stage. In the 3–6 familiarization events the hand mimicked the action of removing a lid of a box and then grasped the toy in front of the location where the action had taken place.
Following familiarization, the same pre-test event was presented in all conditions. The boxes were absent and the two toys were shown in switched positions. This pre-test event was followed by two test events. In the Old toy test event the hand reached through the opening and simply grasped the same toy as in the familiarization events. In the New toy test event the new toy was grasped instead.
2.3. Apparatus and procedure
During the experiment infants sat on their caregivers’ lap in a curtained experimental booth. The stimuli were presented on a 31-inch widescreen monitor that was placed approximately 70 cm in front of infants’ eye level. Two hidden cameras recorded the infants’ faces and the stimuli. The on-line experimenter, who was blind to the condition, used specially built software to control stimulus presentation and to register infants’ looking. At the start of each trial the experimenter used an attention getter sound. When the infant was looking at the screen, the experimenter started the presentation of the event. If the infant watched the test event for less than 5 s, then the trial was ignored and started again. (This was the minimum time required to insure that the infants saw which toy was grasped in the test events.) If the infant looked away for more than 2 s or if 120 s had elapsed, the test event ended. Each infant saw both types of test events. The order of the presentation of the test events (New goal first/second), the type of object touched during familiarization (bear/cube), and the object’s position (left/right side of the stage) were counterbalanced in all conditions.
The looking times of the test phase and familiarization phase were re-coded off-line by a secondary coder. The intercoder reliability for the test events was 96.86%.
Looking times during familiarization were analyzed first. A multivariate anova revealed no significant difference between the three conditions and between the two age groups in the looking times for the familiarization and the pre-test events (Fig. 2).
Second, the effects of order, gender, and toy position on the looking times in the test events were investigated. None of these factors had any main or interaction effect. These variables were therefore omitted from further analyses. A repeated measures anova was conducted on the looking times in the test events with test event type (new toy vs. old toy) as a within-subject variable, and with condition, age group and toy type (cube vs. bear) as between-subject variables. A significant main effect of condition was found (F(2, 131) = 4.258, p =.016, effect size: 0.06). Pairwise comparisons indicated that infants overall looked longer in the Gratuitous Action than in the Superfluous Action condition (p =.005). There was, however, no difference between the two non-efficient conditions and the Efficient condition. No main effects or interactions with age group were found (Table 1).
Table 1. Mean looking times (s) and standard errors for the test events in the three conditions by the two age groups
A strong interaction between test event type and condition was also found (F(2, 131) = 7.139, p =.001, effect size = 0.098). Paired t tests revealed that infants in the Efficient condition looked longer in the New toy test event than in the Old toy test event (t(46) = −2.583, p = .013, two-tailed). In the Gratuitous Action condition, the opposite pattern was found: infants looked longer in the Old toy test event than in the New toy test event (t(47) = 2.310, p =.025, two-tailed). No difference was found between the two types of test events in the Superfluous Action condition (Fig. 3). A non-parametric Wilcoxon signed rank test confirmed these results (Efficient condition: Z = −2.423, p =.015, Gratuitous Action condition: Z = −2.595, p =.009).
Furthermore, a three-way interaction between test-type, condition and toy type was found (F(1, 131) = 4.056, p =.020, effect size = 0.058). Follow-up tests revealed that only in the Gratuitous Action condition and only when the cube was the target, infants looked longer in the Old toy test event than in the New toy test event (t(23) = 3.130, p =.005, two-tailed), in the other conditions infants produced the same looking patterns in the two test events for both toys.
We have investigated the relationship between outcome selection and means selection information in young infants’ goal attribution. Infants were presented with situations in which both types of information were available to interpret and generate an expectation about the goal of an observed action. We found that 7- and 9-month-old infants only expected the actor to continue to direct the action toward the same target if both types of information were converging on the same goal-directed interpretation (Efficient Action condition). In the other two situations, in which the outcome information could be used to disambiguate the goal but the means selection information did not justify this interpretation, infants did not use the outcome selection information to generate expectations about the actor’s goal in a new situation (Gratuitous and Superfluous action conditions). These findings together support the hypothesis that ambiguous means selection information prevents infants from relying on outcome selection information in their action interpretation.
In neither of the two non-efficient action conditions did the infants expect the actor to pursue the original outcome. This suggests that inefficiency of means action prevents goal attribution not only when there is a single action (with unnecessary detours, Superfluous condition), but also when the action can be considered to be part of an action sequence (Gratuitous condition). Our finding that infants as young as 7 months can evaluate the causal efficiency of a subaction (opening the box) leading to an overarching goal (getting the toy) is quite remarkable as this ability had previously only been demonstrated at 10–12 months (Sommerville & Woodward, 2005; Woodward & Sommerville, 2000). The early presence of this ability is also supported by previous evidence showing that by 5½ months infants can understand the physical constraints of containment events in goal attainment (Baillargeon, Graber, DeVos, & Black, 1990), which is certainly necessary for the evaluation of the efficiency of the means in our containment event.
Recall that infants looked equally long at the two test events in the Superfluous Action condition. This suggests that they had indeed no specific expectation about which toy would be grasped. In the Gratuitous Action condition, however, infants looked significantly longer at the Old toy than the New toy test event. While this looking pattern also shows that the infants did not expect the same goal to be achieved, it might reflect the fact that they encoded the action in terms of its spatial properties (see also Woodward, 1998, Experiment 2). Infants might thus have expected to see the action to be performed on the same side of the stage as in the new toy event. One can speculate that the presence and the manipulation of the box in the Gratuitous Action condition highlighted the spatial position of the action, while the mimicked action in Superfluous Action condition did not tie the action as clearly to one side of the stage because no box was present to emphasize a particular location.
A further caveat that needs to be considered is whether infants in the Gratuitous Action condition might have evaluated the action of opening the empty box as being efficient action—despite our intention for it to look inefficient—by inferring that the goal of the action was to put a toy into the box.2 Since the boxes were no longer present in the test events, this interpretation was not tested in the test events and this possibility can thus neither be verified nor rejected. However, we argue that this interpretation is quite unlikely, because it requires the inference of a hypothetical state of affairs since this goal has not been seen achieved. Inferences about hypothetical scenarios are above the level of the representational ability of infants at the age tested in the current study (Bellagamba & Tomasello, 1999; Csibra et al., 2003). Furthermore, in the familiarization events, after opening the box the hand kept grasping the toy for 5 s until the end of the event, which would be unnaturally long if the goal had been putting the toy into the box. Note, however, that such an interpretation of the familiarization phase of the Gratuitous Action condition would have involved relying on the means selection information rather than on the outcome selection information. Hence, it would not change our main argument about the relation between the two types of information.
Another related point that needs to be addressed is how our design and research questions differ from those of Woodward and Sommerville (2000). In that study 12-month-olds were shown the same familiarization events as in the Efficient and the Gratuitous Action condition of the present study. However, in the test events, only the two means actions were shown and not the goal attainment: The boxes were still present, the toys were either inside or in front of the boxes, and the hand only touched one of the boxes without grasping the toy. Therefore, the finding that infants did not have specific expectations about the means in their non-efficient condition did not necessarily prove that they did not interpret the grasping of one of the toys in the familiarization phase as goal-directed. While both the current and Woodward and Sommerville’s experiments provided simultaneous presence of outcome selection information and means selection information, only the present study clarified that ambiguous means selection information prevents infants from relying on outcome selection information in specifying the goal of the observed action.
Two more recent studies are relevant concerning the relationship between means selection and outcome selection information. These studies have tested whether 9- and 12-month-old infants can transfer their goal attribution from a situation that contained only means selection information to another situation in which only outcome selection information was available (Biro et al., 2011; Hernik & Southgate, 2011). In these studies infants were shown a situation in which initially only one target object was present (no outcome selection information) while the efficiency of the action (means selection information) varied across conditions. In the test events a new target was introduced (outcome selection information was provided) in all conditions. It was found that infants only expected the action to be directed at the original target if the initial action was clearly efficient. Importantly, infants did not attribute a goal if the action was clearly non-efficient or if no situational constraints were present to evaluate the efficiency of the action. These findings, together with the current study, show that while unambiguous means selection information disambiguates ambiguous outcome selection information (Biro et al., 2011; Hernik & Southgate, 2011), unambiguous outcome selection information does not disambiguate ambiguous means selection information (current study). This suggests that means selection information takes primacy over outcome selection information in infants’ interpretation of goal-directed actions.
It is an interesting question whether there exists any period during the development of action interpretation in which infants would preferentially rely on outcome selection information if means selection information is also available. On the one hand, processing means selection information is more demanding in terms of computational resources and background knowledge, which would suggest that at the very beginning infants might primarily rely on encoding the end state of an action in goal attribution. On the other hand, from birth infants perceive a flow of action sequences in an ever-changing rich environment. To interpret every perceptually salient end state as a goal, even if that interpretation is not justified by the means in the given context, would be counterproductive since it could lead to misinterpretation and false action predictions, and would certainly not allow infants to start to parse the flow of actions into meaningful and hierarchically related goal-directed action units (Baldwin, Baird, Saylor, & Clark, 2001). This view is consistent with studies demonstrating that not only means selection information but also other types of observable information, such as emotional and facial expressions of the actor, are utilized by young infants to make differential interpretations of, or reaction to, actions with identical outcomes (e.g., Behne, Carpenter, & Tomasello, 2005).
The demonstration of the primacy of means selection information does not imply that outcome selection information is unimportant in action understanding. On the contrary, it is very useful information when it is the only observable information one can rely on and it is very likely crucial for understanding and attributing preferences. In fact, there is an on-going debate about whether the experimental paradigm that investigates sensitivity to outcome selection information alone (Woodward, 1998) is a test for preference attribution, rather than for goal attribution, or both (Hernik & Southgate, 2011; Jacob, in press; Luo & Baillargeon, 2005).
Overall, our current findings provide strong evidence, which has hitherto been missing, for the main claim of the teleological stance theory concerning the crucial role of the efficiency of the action in goal attribution. By providing both means selection and outcome selection information, our current setting allowed us to directly test the expectations about the goal of an observed action. The early presence of the primacy of means selection information suggests that this bias might be a core property of the cognitive mechanisms of goal attribution and thus sheds light on the nature of infants’ notion of goal.
These two conditions were based on a previous study in which it was shown that 12-month-olds understand the causal relationship between opening a box and getting the object inside, and that they understand that opening the box is unnecessary if the toy is not inside but in front of the box (Woodward & Sommerville, 2000).
Adult viewers sometimes came up with this interpretation of the familiarization events.
This research was supported by a grant to the second author from the Netherlands Organization for Scientific Research. The study was carried out by the first author in partial fulfillment of the requirements of the PhD degree at Leiden University, and under the supervision of the second author. We thank Gergely Csibra and two anonymous reviewers for their valuable comments on the manuscript. We are also grateful to the parents and infants for participating in the study.