cs.CLOct 19, 2026
SaveGRID: Modeling Health Influencer Discourse as Argument Networks for Credibility Assessment
Organizations: Purdue University
Abstract
Health-related discussions on social media are increasingly shaped by influencers who offer diverse, and often competing perspectives on the same issue. Yet, such content is typically evaluated in isolation, without considering the broader discourse in which it appears. We introduce GRID (Grounded Reasoning over Interconnected Discourse), a framework for evaluating influencer content through collective inference over the surrounding discourse. GRID represents each video as an argumentation graph connecting claims to its central message, or takeaway, links related claims and takeaways across videos, grounds claims in scientific evidence, and jointly infers credibility over the resulting network. We evaluate GRID on an end-to-end health-related content validation task using 569 influencer videos grounded in expert assessments. Our results show that accounting for the broader discourse improves influencer content validation. Finally, we use GRID-inferred credibility to demonstrate how argumentative strategies and moral framing differ between low- and high-credibility content.
Figures & tables
Figure 1: Treating influencer content in isolation may be insufficient: a claim can be well-justified within a video yet still be incorrect. Evaluating it in isolation also ignores the broader distributed information environment , where related influencers may provide supporting or competing perspectives. We contend that incorporating this surrounding discourse enables a more holistic evaluation of the target content.
Figure 2: GRID collects related influencer videos, connects their claims and takeaways in an networked argumentation graph, incorporates scientific evidence, and jointly infers per-video credibility. Note that GRID operates on influencer video transcripts.
| Class-wise F1 | ||||
|---|---|---|---|---|
| Method | Macro F1 | Inc. | Part. | Corr. |
| Direct prompting | 45.2 | 86.9 | 12.3 | 36.4 |
| Per-video pragmatic | 46.0 | 82.1 | 18.4 | 37.6 |
| Per-video aggregation | 53.1 | 89.8 | 0 6.2 | 63.5 |
| LOKI Li et al. (2025) | 30.8 | 52.8 | 14.5 | 25.0 |
| GRID (ours) | 67.6 | 92.1 | 50.8 | 60.0 |
Table 1: Main results for content validation: GRID outperforms all baselines in Macro-F1, showing the benefit of the context of broader influencer discourse. Inc./Part./Corr.: incorrect / partially correct / correct .
Figure 3: Larger networks protect GRID from very poor predictions. In networks of up to videos, prediction quality spans nearly the full range from to ; in larger networks, every prediction stays above . Points show individual target videos, plotted by number of videos in their network.
Figure 4: Does GRID conform to the majority view when flooded with information? While majority voting’s prediction degrades with information flooding, GRID maintains and even improves in content validation.
| Class-wise F1 | ||||
|---|---|---|---|---|
| Removed connection | Macro F1 | Inc. | Part. | Corr. |
| Takeaway–takeaway | 51.4 | 84.5 | 27.5 | 42.2 |
| Cross-video claim–claim | 60.9 | 90.3 | 34.9 | 57.4 |
| PubMed–claim | 37.8 | 85.1 | 0 4.7 | 23.5 |
Table 2: Every connection type contributes, and scientific evidence contributes most. F1 (%) after removing each connection type; full GRID reaches Macro F1 (Table 1 ). Without PubMed–claim connections, Macro F1 drops by points and partially correct F1 nearly vanishes. Inc./Part./Corr.: incorrect / partially correct / correct .
Figure 5: Which framings are most credible? Claims citing empirical data or causal mechanisms, and claims framed around care/harm, receive above-average trust, while claims framed around fairness, loyalty, or liberty receive the least. Points show mean claim trust relative to the mean over all claims with bootstrap intervals.
| Argumentative strategy | Moral foundation | ||
|---|---|---|---|
| Label | Definition | Label | Definition |
| Causal/mechanistic | Explains a causal process connecting the claim to the takeaway. | Care/harm | Emphasizes preventing suffering, injury, or illness. |
| Associational/empirical | Uses an observed association, measurement, or empirical result. | Fairness/cheating | Invokes equitable treatment, honesty, or accountability. |
| Consequence/trade-off | Reasons from expected outcomes or balances costs and benefits. | Loyalty/betrayal | Invokes obligations to, or betrayal of, a group. |
| Classification/definitional | Defines or categorizes an entity, event, or intervention. | Authority/subversion | Defers to or challenges institutions, expertise, or hierarchy. |
| Sign/diagnostic | Treats an observation as a sign of an underlying condition. | Purity/sanctity | Invokes contamination, naturalness, or bodily integrity. |
Table 3: Argumentative-strategy and moral-foundation labels used in the discourse analysis.
Figure 6: Do low-credibility videos invoke authority more often? On average, no (all topics ). But in COVID-19 vaccine and mask topics they invoke authority more, whereas in treatment and diet , high-credibility videos do.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
| Removal (%) | F1 (Macro) | Class-wise F1 | ||
|---|---|---|---|---|
| Incorrect | Partially Correct | Correct | ||
| 25 | ||||
| 50 | ||||
| 75 | ||||
Table 4: Content-validation performance after randomly removing related videos. Results aggregate random trials. Removing all related videos corresponds to the Per-video Pragmatic Inference baseline in Table 1 .
Figure 7: Median prediction quality across information-flooding levels for small and larger networks; shaded regions show the interquartile range. The dashed line marks the point at which opposing perspectives become the majority.
Figure 8: Does a strategy receive similar trust regardless of the moral foundation it invokes? No: the same strategy receives much higher trust when paired with care/harm than with fairness, loyalty, or liberty. Cells show mean claim trust for each strategy–foundation pair relative to the mean over all claims; N/A marks pairs with fewer than claims. Sign/diagnostic is omitted because most of its pairs fall below this threshold.
| Validated output | Result | Agreement |
|---|---|---|
| Takeaway–video match | ||
| Takeaway–expert alignment | ||
| Source invoked | ||
| Specificity | ||
| Argumentative strategy | ||
| Moral foundations |
Table 5: Human validation results and raw inter-annotator agreement. The result column reports the percentage judged correct, aligned, or yes, as applicable.
| Influencer Popularity | # Labeled | # Unlabeled |
|---|---|---|
| 0K–50K | 466 | 442 |
| 50K–100K | 43 | 51 |
| 100K | 60 | 183 |
| Total Videos | 569 | 676 |
| Overall Dataset | 1,245 videos | |
Table 6: Dataset distribution for labeled (expert-annotated) and unlabeled videos by influencer popularity.