Figure 1B summarizes the experimental approach. Participants were asked to choose a food item from pairs of food images of different healthiness and palatability in three previously used contexts (31): Health, typical, and unrestricted. In the healthy context, participants were instructed to select the food they should eat to be healthy, which aimed to reflect their ideal of a healthy diet. In the typical context, participants were instructed to select as they would do in their daily life, which aimed to reflect daily food choices (31). In the unrestricted context, participants were instructed to choose as if nothing was stopping them, which aimed to reflect situations where health concerns are minimized (i.e., dining out for pleasure). The healthy context was included to facilitate interpreting choice data, as this context drives healthier food choices compared to other contexts (26–31). Using different sets of food images in each context, we measured the probability of a healthy choice and conflict during food choice first without showing NWL for all pairs of food images and then showing NWL for the same pairs. This design created paths for each decision (Figure 1B) to test whether showing NWL would reverse unhealthy food choices made in the absence of NWL across the eating contexts. For all food choices, we estimated conflict by measuring the area under the curve (AUC, the difference in the mouse path from the direct line between the starting point and the chosen image) and reaction time (RT). A larger AUC and longer RT are interpreted as a larger conflict during choice (40–43). Finally, to gain insight into how NWL and context affect the decision process, we fitted a drift-diffusion model (DMM) to the choice data to examine whether changes in AUC could be related to changes in drift rate (how quickly a subject accumulates evidence toward a healthy choice) and decision bias (baseline probability of choosing a given option) toward healthier foods.

The Human Ethics Committee at Pontificia Universidad Católica de Chile approved this study (Protocol 201223001). Participants were recruited through social media (Instagram and Facebook) and email between June and November 2021 as a convenience sample (the only consideration was to recruit an equal number of males and females by biological sex) and offered a local retailer’s gift-card for completing the study (Cencosud, 20,000 Chilean pesos, ∼25 USD). Inclusion criteria were BMI between 20 and 35 kg/m², 18–45 years of age, absence of any disease, and stable body weight (less than 2 kg change over the last 3 months). Exclusion criteria were undergoing treatment for body weight loss, consuming any medication or nutritional supplement, tobacco or alcohol use, any eating disorder or restrictive dietary style, professional sports activity, pregnancy, or breastfeeding. Supplementary Figure 1 describes the workflow for participant selection. Briefly, participants that contacted investigators were sent a screening questionnaire on a rolling basis. Of the 716 individuals that received the screening questionnaire, 686 returned it. Of those, 496 were excluded, 181 were invited to the study, and 149 completed it. The main rejection criteria were the use of nutritional supplements (n = 230), weight change during the last 3 months (n = 138), and dietary restrictions (n = 128), with 66% of the individuals excluded based on any of these criteria. Among the 149 participants that completed the study, we eliminated 14 participants due to inconsistencies in demographic responses between the screening questionnaire and data collected during the study, one for completing the study twice, one for reporting not reading instructions, one for taking more than 3 hours in completing a single session, and two for having less than 2% of the mouse-tracking data recorded. Thus, data from 128 participants were considered for analysis.

The experiment was implemented in Spanish on the Gorilla Online Platform (44). Figure 1B summarizes the experimental design (Supplementary Table 1 details all the steps in the study). After screening, participants were instructed to complete the two sessions of the study at the same time of day within seven days. In the first session, participants completed a questionnaire for demographic information, exercise, the same questions from the screening questionnaire (body weight, height, age, variation of body weight in the last 2 months, dietary restrictions, alcohol and tobacco use, see Supplementary Table 2 for a Spanish and English version of the questionnaire), and the time of their last meal. Next, participants completed the Sleep Stanford Scale [SSS (45)], answered four questions about hunger and appetite using visual analog scales (VAS) from 0 to 100 (Questions 1−4 from Supplementary Table 3; 46), and the Adult Eating Behavior Questionnaire (AEBQ) (47). The SSS is a one-question form that measures the perceived level of sleepiness and was included to control for potential effects of sleepiness on the desire for high-calorie foods (48). The questions about hunger and appetite were included to control for appetite effects of choice as increased appetite measured by VAS is a predictor of calorie intake and can influence a subject toward choosing high-calorie foods (46, 49). The AEBQ measures different aspects of food approach and avoidance and has been validated in a sample of Chilean adults (47, 50) and was included to control for potential effects of eating behavior on food choice (51).

To complete the first session, the participants answered whether they recognized 94 food images and then were asked to rate those images for healthiness, liking, and wanting using a VAS scale from 0 to 100 (52) (Question 5–7 from Supplementary Table 3).

In the second session, participants recorded the time of their last meal and answered the SSS and the same questions about hunger and appetite as in the first session. Next, participants completed a training task instructing them to choose the desk item between images of a desk and a kitchen item (50 pairs of images). Next, participants were randomized to one of six possible sequences of contexts (typical, healthy, unrestricted) for the food choice task. In each context, participants had to choose one food image from 90 pairs without displaying the NWL. After a 5 min break, the same procedure was repeated, but now the NWL were show for each food of the same 90 pairs. Participants also had a 5-min break between contexts. Finally, participants completed a survey where they selected the option that best described how they used NWL in the task (counted the number of NWL, read the information displayed by NWL, or did not use NWL).

Food images (FoodPics database) (53) were classified into the following categories: bread (n = 3), breakfast food (n = 4), cake (n = 9), cheese (n = 3), cookie (n = 5), dessert treats (n = 22), fruit (n = 25), pasta (n = 4), pastries (n = 7), pie cakes (n = 1), pizza (n = 3), prepared meals (n = 28), salads (n = 3), seafood (n = 1), seeds nuts (n = 6), snacks (n = 14), soups stews (n = 5), and vegetables (n = 24). Online nutritional information for each food image was used to assign NWL following the Chilean Food Labeling Law (14) (See Figure 1A for the NWL used in the study). A pilot test (N = 8 adult subjects) indicated that 18 images were not recognized and thus were removed from the dataset. We also selected ten images of typical Chilean preparations assigned to the categories of prepared meals and dessert treats that were validated for recognition in a separate pilot test (N = 6). All food images (N = 94, see Supplementary Information for detailed information including food pictures) were assigned with 0 to 3 NWL as foods with 4 NWL (high in calories, sugar, saturated fat, and sodium) were infrequent. Images were assigned to each context following a stratified sampling strategy from different food categories. The health context included 32 images (0 NWL: 8, 1 NWL: 6, 2 NWL: 8, 3 NWL: 10), the typical context 29 (0 NWL: 7, 1 NWL: 6, 2 NWL: 7, 3 NWL: 9), and the unrestricted context 33 images (0 NWL: 8, 1 NWL: 6, 2 NWL: 8, 3 NWL: 11). The calories, saturated fat, sugar, and sodium were not different among images used in the different contexts (Supplementary Figure 2 and Supplementary Table 4).

The eating contexts were defined by the instructions given to participants (31). In the healthy eating context, participants were instructed to “select as fast as possible the food that you would eat to stay healthy,” in the typical eating context to “select as fast as possible the food that you would usually select in your daily life,” and in the unrestricted eating context to “select as fast as possible the food that you would eat if nothing stopped you.”

Each trial started with a white screen that displayed a button labeled “START” at the bottom center of the screen. Once participants pressed the start button, they had up to four seconds to select between food items that appeared in the upper left and right corners with or without NWL displayed. After clicking on the selected food image, a white screen with a fixation cross displayed for a randomly selected period between 100 and 500 ms before the next pair of images was presented. For each participant, the image pairs used in each context were generated pseudo-randomly as follows: (1) Each participant saw 15 food image pairs for each of the six possible combinations of numbers of NWL per food image (0 vs. 1, 0 vs. 2, 0 vs. 3, 1 vs. 2, 2 vs. 3 NWL); (2) Any food image had to skip two pairs before being shown again; (3) Food images with the same number of NWL were not shown on the same side for more than 3 consecutive pairs. Food images were not repeated between contexts to prevent participants from recalling NWL associated with each image.

All analyses were adjusted for covariates. The covariates were trial order during the task (numeric), age (numeric), sex (male/female), body mass index (BMI, numerical), SSS score (numerical), AEBQ score (numerical) (47), physical exercise (yes/no), appetite score (the mean of the four questions about hunger and appetite, see Supplementary Table 1), hour of test start (numerical) and time since last meal (hour, numerical). All continuous covariates were centered and scaled. For all linear mixed models, participant was included as random effect, and each term’s statistical significance was calculated using Wald Type III ChiSquare test. Adjusted estimates and posthoc tests for dependent variables were done on estimated marginal means (emmeans R package). The mixed effects logistic regression (glme function) and linear mixed models (lme function) functions were from the lme4 R package. All results are presented as mean and standard error. Post-hoc comparisons were corrected by Tukey’s Honest Significant Difference.

Participants were balanced by sex with an average BMI of 23.3 ± 0.2 kg/m² (all participants that completed the study had a BMI < 30) and 24.8 ± 0.5 years old (Table 1). Men and women did not differ in age (P = 0.79), BMI (P = 0.33), frequency of physical exercise (P = 0.89), percent of overweight participants (P = 0.28), and education level (P = 0.59).

Health ratings decreased as the actual number of NWL per food increased (X_3 = 12052.16, P < 0.01; Figure 2A). Like and want ratings had a U-shape as decreased from foods without NWL to foods with 2 NWL and increased to foods with 3 NWL (Like: X_3 = 184.55, P < 0.01; Want: X_3 = 85.81, P < 0.01; Figures 2B, C). Among covariates, increasing BMI reduced health (β = −0.73 ± 0.27, P = 0.01) and like ratings (β = −1.05 ± 0.37, P < 0.01) but had no effect on want ratings (P = 0.84). Supplementary Figure 3 shows the estimates for health, like, and want ratings separated by demographic variables. Overall, participants recognized objectively healthier foods (based on the number of NWL per food image) and showed higher like and want ratings for foods with 0 and 3 NWL.

When participants had to choose between pairs of food images in the absence of NWL, the probability of a healthy choice (choosing the food item with the actual lowest number of NWL between the two options) was 82.4, 73.9, and 68% in the health, typical, and unrestricted contexts respectively (Figure 3A; main effect of context: X_2 = 286.76, P < 0.01; P < 0.05 between all contexts). Among covariates, an increase in one unit of BMI increased the probability of a healthy choice by 8.3 ± 3.9% (P = 0.01), while appetite had no significant effect (Supplementary Tables 5, 6 show the complete output of the logistic regression). In all contexts, the probability of a healthy choice increased as the difference in health ratings between images increased (βΔHealth ratings = 3.86 ± 0.05%, P < 0.01; Figure 3B) and decreased as the difference in like ratings between images increased (βΔLike ratings = −0.06 ± 0.005%, P < 0.01; Figure 3C). Overall, in the absence of NWL, the probability of a healthy food choice was higher in the healthy context and lower in the unrestricted context.

There were significant interactions between food choice type (healthy vs. unhealthy) and context for AUC and RT (AUC: X_2 = 10.52, P < 0.01; RT: X_2 = 242.47, P < 0.01; Figures 3D, E). In a healthy choice, the AUC was largest in the unrestricted context and larger compared to the health context (9.7 ± 3.8%, P = 0.01, Figure 3D). For an unhealthy choice, the AUC in all contexts was larger compared to the healthy choice (P < 0.05, Figure 3D), and AUC was now largest in the typical context and different from the unrestricted context (larger by 5.9 ± 2.5%, P = 0.01, Figure 3D). For RT, healthy choices were slower in the unrestricted context (P < 0.05 compared to all contexts, Figure 3E); but unhealthy choices were slower compared to healthy choices only in the health and typical contexts (P < 0.05, Figure 3E). Overall, healthy choices in the unrestricted context showed higher AUC and were slower than in other contexts. Compared to healthy choices, unhealthy choices had higher AUC in all contexts and were slower only in the health and typical contexts.

The probability of a healthy choice when NWL were shown was higher than 90% in all contexts when participants first made a healthy choice in the absence of NWL (Figure 4A) and was lower in all contexts when their first choice was unhealthy (P < 0.05 for pairwise comparisons within contexts across decision paths, Figure 4B). Thus, the probability that showing NWL would reverse an unhealthy choice was 79.6 ± 1.8, 36.6 ± 2.1, and 18.6 ± 1.4% in the health, typical, and unrestricted contexts, respectively (P < 0.05 between contexts), and regardless of how participants declared to use NWL during food choices (Among participants, 51% declared to count NWL, 37.5% to read them, and 10.9% to not use them during choice, Supplementary Figure 4). Increasing ΔNWL (the difference in the number of NWL between foods in each pair) increased the probability of a healthy choice only in the typical context (up to 13.5 ± 3.3% for ΔNWL = 3, P < 0.05, Figure 4C), regardless of how participants declared to use NWL during food choices (Supplementary Figure 5). Overall, showing NWL had a probability to reverse an unhealthy choice of 80% in the healthy context and lower than 50% in the typical and unrestricted contexts.

The effect of showing NWL on the change in AUC (ΔAUC) and RT (ΔRT) compared to the choice made without showing NWL depended on the decision path (interaction between decision path and context: ΔAUC, X_6 = 14.36, P < 0.01; ΔRT, X_6 = 140.58, P < 0.01). There were no significant effects of BMI, appetite score, or time from the last meal on ΔAUC or ΔRT (Supplementary Table 7). The confirmation of a healthy choice made when NWL were shown was associated with reduced AUC and faster decisions (negative ΔAUC and ΔRT) in all context, and these effects were the largest in the healthy context. Making an unhealthy choice in the presence of NWL, either by reversing a healthy choice or confirming an unhealthy choice made without NWL, was associated with increased AUC (positive ΔAUC) in the typical and unrestricted contexts. Still, these decisions were slower (positive ΔRT) in the typical context and faster (negative ΔRT) in the unrestricted context. Finally, in the presence of NWL, the reversal of an unhealthy food choice made without NWL was associated with reduced AUC (negative ΔAUC) in the typical and unrestricted contexts but faster decisions (negative ΔRT) only in the health and typical contexts (Figures 4C, D).

Overall, compared to decisions made in the absence of NWL, showing NWL reduced the AUC during healthy choices and increased the AUC during unhealthy choices across eating contexts. The reversal of unhealthy food choices by NWL in the healthy context (80% probability) is associated only with faster choices, in the typical context (36% probability) is associated with reduced AUC and faster choices, and in the unrestricted context (14% probability) is associated only with reduced AUC.

There was an interaction between context and presence of NWL for decision bias (X₂ = 6.97, P = 0.01), drift rate (X_2 = 76.19, P < 0.01) and non-decision time for the healthy choice (X_2 = 13.08, P < 0.01). Regardless of the presence of NWL, the healthy context showed the largest decision bias, drift rate, and shorter non-decision time toward a healthy choice (Figure 5). The unrestricted context showed the opposite pattern (i.e., the smallest bias, drift rate, and longer non-decision time toward a healthy choice, Figure 5). Compared to the choice made without NWL, the presence of NWL increased decision bias for a healthy choice in the health (7.6 ± 1.8%, P < 0.01) and typical (6.9 ± 1.9%, P < 0.01), but not the unrestricted context (P = 0.45; Figure 5A). Also, the presence of NWL reduced non-decision time in all contexts with the largest reduction in the health context (−16.8 ± 1.6%, P < 0.01) compared to the typical (−9.6 ± 1.5%, P < 0.01) and unrestricted context (−8.8 ± 1.5%, P < 0.01; Figure 5B). Finally, the presence of NWL almost doubled the drift rate in the healthy (88.3 ± 6.2%, P < 0.01) and typical contexts (89.1 ± 21.8%, P < 0.01) and increased it by threefold in the unrestricted context (367.5 ± 117.9%, P < 0.01), but the drift rate remained highest in the health context and lowest in the unrestricted context (Figure 5C). Overall, compared to a typical and unrestricted context, a healthy context reduces non-decision time, increases decision bias and drift rate toward healthier foods. NWL reduce non-decision time and increase drift rate toward healthier foods in all contexts but increase bias toward healthier foods only in the health and typical contexts.