Relief-Seeking: Does Distress Make Emotional Eaters Believe Miracle Diet Claims?

About one in five U.S. adults frequently eats to manage their emotions rather than their hunger, and impatience (a weak capacity to wait for later rewards) is a leading predictor of who does (Emotional Eating in Adults…U.S. National Study). Separately, health misinformation spreads differently from political misinformation. Political falsehoods tend to run on anger, while false health claims tend to run on hope. In cancer tweets, false cure claims made up 55% of unique cure claims but drew 76.19% of engagement, and they were written in a joyful, trusting tone (Analysing the Extent of Misinformation in Cancer Related Tweets). This proposal connects the two findings with one testable hypothesis. Believing a "quick fix" diet claim and eating to feel better may be the same act of seeking fast mood repair. If so, a moment of distress should make people worse at telling true diet claims from false ones, but only when the false claims promise relief. The effect should be largest in people who already cope by eating.

We test this with a preregistered online experiment. We first induce or withhold distress through a realistic feed of upward social comparisons. Participants then rate true and false nutrition claims whose emotional tone is controlled: some offer hope or relief, others warn of dread. A secondary machine-learning question is whether persona-based LLM simulations of misinformation susceptibility can reproduce this interaction between a person's state and their traits. Such simulations are increasingly used as stand-ins for human participants. A third arm tests whether a brief perspective-taking prompt cancels the effect. Perspective-taking is known to buffer adolescents against comparison-driven emotional eating.

Core Claim

The claim is a specific three-way interaction. Distress induced by social comparison lowers a person's ability to separate true from false nutrition claims (we call this "discernment"). The drop is concentrated on false claims framed with positive, relief-promising affect. It is amplified by trait emotional eating and partly carried by impatience. Distress does not lower discernment for dread-framed diet claims or for non-diet health claims of matched tone. If the effect instead appears equally across all claim types, then the simpler story is right: distress just makes everyone sloppier. That result would itself be informative, because it would mean diet misinformation needs no special treatment.

This hypothesis resolves a tension the research file leaves open. Emotional eating is triggered by negative affect, through upward social comparison and lowered self-esteem (Social Networking Site Use and Emotional Eating…Chinese Adolescents). Yet people with binge eating disorder seek out food media for comfort and then get pulled into binges, a pattern called "impact ambivalence" (Understanding How People with Binge Eating Disorder and Bulimia Interact with Digital Food Content). Health misinformation, meanwhile, tends to be hopeful (Know it to Defeat it; Emotions in online rumor diffusion), whereas anger drives the deep cascades of false rumors generally (Emotions explain differences in the diffusion of true vs. false social media rumors). The relief-seeking hypothesis predicts which of these mechanisms applies when the content is about food: negative state paired with positive claim.

Positioning

The closest work falls into three groups, and none measures the person-by-state-by-claim interaction. The cancer-tweet study builds a pipeline that pulls dietary "anchors" such as sugar and apricot seeds out of tweets and documents the hope-valence pattern. It does not measure anything about the people who share the claims. We reuse its anchors to build stimuli and add the psychological measurements it lacks. The SMISTS work (Simulated Misinformation Susceptibility (SMISTS)) uses LLM personas driven by life stories to replicate human misinformation tests across COVID-19, politics, GMOs, and climate. It covers no nutrition and represents no affective states. We treat its method as the strongest predictive competitor and test whether it holds up under our design. The binge-eating qualitative study and the expert study of generative AI risks for eating disorders (From Symptoms to Systems) motivate the work. They show that distress, comfort-seeking, and "healthy" advice combine harmfully. However, they rest on 23 and 15 interviewees respectively and measure no belief outcomes.

Two papers constrain the design. The intervention framework in Ask or Answer models users only by health literacy and belief commitment. Our data would show whether an affective dimension belongs in such user models. The review of young Poles' dietary habits (Dietary Habits of Young Poles) reports that nutrition knowledge does not stop emotional eating under tension. This warns us against assuming knowledge is the active ingredient, so we measure knowledge rather than presume its role.

Research Program

We first build a stimulus set of about 160 short, social-media-style nutrition claims in a two-by-two design. Each claim is true or false, and each is framed with either relief and hope ("this one swap ends cravings") or dread ("this common food is quietly poisoning you"). Content anchors come from the cancer-tweet study's dietary anchors and from public health fact-check sources. Veracity is verified by two dietitians. Two matched control sets use the same tones: non-diet health claims drawn from PUBHEALTH and CounterHealth, and non-health consumer claims. We check that tone is balanced within each cell using human raters and an affective language model, following the implicit-emotion approach of ConspEmoLLM. That paper found that explicit emotion scores distract models, so we do not inject numeric emotion scores anywhere. Claims that promote extreme restriction are excluded on safety grounds.

The main study is a preregistered online experiment with about 2,400 adults. Participants are screened out if they report a current eating-disorder diagnosis or are receiving treatment for one, and we will seek clinical advisory review. The sample size is chosen to detect a small three-way interaction with mixed-effects models and crossed random effects for participants and items. Participants are randomized to one of three conditions: a neutral feed; a feed of peers' idealized body and lifestyle posts, which mirrors the comparison mechanism in the adolescent social-media study; or the comparison feed followed by a short perspective-taking exercise. A manipulation check measures negative affect. Each participant then rates 40 claims for accuracy, intention to share, and intention to try. Trait measures include a standard emotional-eating questionnaire subscale, a delay-discounting task to measure impatience, nutrition knowledge, analytic thinking, and body dissatisfaction. The primary model is

$$\text{logit}\,P(\text{rated accurate}_{ij}) = \beta_0 + \beta_1 V_j + \beta_2 D_i + \beta_3 E_i + \beta_4 R_j + \beta_5\, V_j D_i E_i R_j + \dots + u_i + w_j,$$

where $V_j$ is the veracity of claim $j$, $D_i$ is participant $i$'s distress condition, $E_i$ is their standardized emotional-eating score, $R_j$ indicates relief framing, the omitted terms are all lower-order interactions, and $u_i, w_j$ are participant and item random intercepts. The hypothesis is that $\beta_5 < 0$: discernment shrinks most for distressed, high-emotional-eating participants when they judge relief-framed claims. A preregistered mediation model asks whether impatience carries part of this effect.

The ML component runs the identical protocol on LLM personas in the SMISTS style. We build one persona per human participant, conditioned on their demographics and traits, and compare two versions: one that has only traits, and one that is also told it just viewed the comparison feed. We then ask whether the personas reproduce the human item-level ratings and, more importantly, the sign and size of $\beta_5$.

As a field check of external validity, we use the Pro-ED Twitter corpus from ED-Filter. We classify diet claims within each user's timeline by veracity and tone, then test whether relief-framed false claims become more frequent in the days after posts in that corpus's Binge or BodyAndBodyImage topics, compared with the days after other posts. This analysis is observational and cannot establish cause. Extending it to a causal estimate with the counterfactual temporal model of CNTPP is follow-up work.

Baselines

The naive reference point is the neutral-feed condition: discernment with no induced distress. The first competing account is a general-distress model. It predicts that distress lowers discernment equally across all claims. We fit it on the same data as a restricted mixed model with no tone or domain terms. The relief-seeking account beats it only if adding the tone-by-domain-by-trait terms significantly improves held-out item-level prediction under participant-grouped cross-validation, and only if the effect for relief-framed diet claims exceeds the effect for matched non-diet health claims by a preregistered margin. The second competing account is an information-deficit model. Here nutrition knowledge and analytic thinking, rather than emotional eating, explain the pattern. We fit it with identical covariates and tuning. It loses if the emotional-eating interaction survives when knowledge is controlled.

The strongest existing predictive method is persona simulation in the SMISTS style. For human misinformation tests, that paper reports a Pearson correlation above 0.89 with human study results, along with a drop in LLM refusals from 20–50% to 0%. We run it with the same claims, the same instructions, and the same model family for every persona variant, and we allow equal prompt-tuning effort on a held-out pilot sample. We will say simulation fails on our question if it matches human item means but recovers the wrong sign or less than half the human magnitude of $\beta_5$. That outcome is plausible even if overall correlation stays high. The key ablation is a persona with traits but no mood state, which isolates whether representing the person's state matters. For human discernment on diet misinformation specifically, no reported result was found on Lacuna.

Why not just…?

Why not just assume distressed people are worse at judging all misinformation? This is the most serious threat, and the design is built to test it. The non-diet health controls and the dread-framed diet claims exist so the general-distress account can win. If it fits as well as the relief-seeking model, we report that, and we drop the claim that diet misinformation is special.

Why not just teach nutrition facts? The Polish review reports that knowledge fails to prevent emotional eating under tension. Experts also describe a "healthy advice paradox," in which standard calorie and BMI guidance worsens rigid, all-or-nothing thinking in vulnerable people. We do not rely on these reports alone: our knowledge covariate tests the information-deficit account directly. If knowledge absorbs the emotional-eating effect, that is a real finding against us.

Why not just simulate participants with LLM personas instead of recruiting humans? SMISTS's correlation above 0.89 makes this a credible competitor. However, its personas carry no momentary affective state, and the research file flags simulation fidelity under acute stress as an open problem. Running the simulation in parallel with the human study measures this gap directly rather than assuming it.

Why not just mine social media for correlations between eating-disorder posts and sharing diet misinformation? The CNTPP paper shows that non-causal models of social media influence reach a relative absolute error of 0.865 on its synthetic benchmark, against 0.234 with causal balancing. Selection effects swamp naive correlations. This is why randomized mood induction is the core of the design and the field analysis serves only as an external-validity check.

Why not just use an accuracy reminder instead of perspective-taking? It may well work as well. No reported result for accuracy reminders on diet claims was found on Lacuna. A fourth arm with a simple accuracy reminder costs little, and we add it, so perspective-taking must outperform it to count as a specific buffer.

Decisive Checks

The project succeeds if three things hold. The preregistered $\beta_5$ is negative and survives controls for knowledge and analytic thinking. The diet-relief effect exceeds the effect for matched non-diet health claims. The perspective-taking arm restores discernment beyond what the accuracy reminder does. The secondary result is whether adding mood state to personas brings the simulated $\beta_5$ closer to the human value. If it does not, persona simulations should not be used to evaluate interventions aimed at distressed users. That applies to dialogue-policy work like Ask or Answer and to empathetic-correction generators trained on VerMouth. A null three-way interaction alongside a strong main effect of distress would redirect the field toward state-general defenses against misinformation and away from defenses specific to diet content.