# How the Human-AI Feedback Loop Drives Virality

Virality on social media is no longer an organic phenomenon; it is the mathematical outcome of continuous human-AI feedback loops prioritizing deep engagement over chronological delivery. Recommendation algorithms rapidly ingest behavioral data—such as watch time, swipe velocity, and share rates—to construct hyper-personalized feeds that inevitably favor emotionally arousing, polarizing content. While platforms provide tools for users to reset their feeds and scrub unwanted topics, breaking out of these algorithmic echo chambers requires sustained, deliberate interaction to actively retrain the underlying machine learning models.

## The Architecture of Algorithmic Amplification

At the core of modern social media is the recommendation engine, a complex machine learning infrastructure designed to match content to users at scale. Unlike legacy chronological feeds, which simply displayed posts from followed accounts, contemporary platforms utilize deep neural networks and multimodal feature recognition to predict user behavior. The system operates on a continuous, self-reinforcing feedback loop: the algorithm predicts what a user wants, serves the content, observes the user's reaction, and immediately retrains its model based on that interaction [cite: 1, 2, 3].

[image delta #1, 0 bytes]

 

When a model is deployed into production, it encounters real-world data that often diverges from its initial training parameters [cite: 4]. Therefore, the algorithm's accuracy depends entirely on collecting constant, real-time data from the user. This data is broadly categorized into two types of feedback:

*   **Explicit Feedback:** These are intentional actions taken by the user, such as tapping the "like" button, leaving a comment, saving a post, or utilizing negative signals like the "Not Interested" feature [cite: 4, 5].
*   **Implicit Feedback:** These are passive, behavioral signals inferred from how the user navigates the platform. They include session duration, video completion rates, how quickly a user swipes past a video, and even micro-pauses while scrolling [cite: 5, 6]. 

Because implicit feedback is significantly harder for users to fake—and occurs in much higher volumes—platforms increasingly weigh passive behaviors, like watch time, far more heavily than traditional explicit vanity metrics [cite: 7]. The human-AI feedback loop is continuously fine-tuned by these subtle behavioral traces.



## The Cold-Start Problem and the "Audition" Phase

Before a recommendation algorithm can distribute content widely, it must evaluate the content's viability. Recommender systems generally suffer from the "cold-start problem," a scenario where new videos or new users lack the historical interaction data necessary for accurate prediction [cite: 8]. 

To solve this, platforms treat the first few minutes of a post's lifecycle as a high-stakes audition [cite: 9]. When a creator uploads a piece of content, the system's AI reviews it for basic compliance, extracting text features via Natural Language Processing (NLP), visual features through image recognition, and audio markers via voiceprint analysis [cite: 10, 11]. 

Once cleared, the algorithm pushes the content to a small test batch of users—often just 200 to 500 initial viewers [cite: 10]. If this cold-start audience provides strong early signals (e.g., watching past the first three seconds, finishing the video, or sharing it), the platform reads these actions as intent, and the "reach ceiling" opens up [cite: 7, 9]. Conversely, fast swipes or immediate drop-offs act as a hard ceiling, signaling to the algorithm that the post is not worth further inventory distribution [cite: 9]. This ruthless initial gating mechanism is why a post can stall at 200 views or rapidly scale to 2 million based entirely on the initial cohort's reaction.

## The Exploration-Exploitation Tradeoff

To prevent users from becoming bored by a highly repetitive feed, recommendation engines must continuously balance two competing imperatives: exploration and exploitation [cite: 12, 13]. 

"Exploitation" occurs when the algorithm serves content that aligns perfectly with a user's deeply established, historical preferences to guarantee engagement [cite: 12, 14]. "Exploration" occurs when the algorithm introduces novel, untested topics outside the user's historical graph to discover new latent interests [cite: 12, 13]. 

The exact ratio of this tradeoff determines how deeply a user is entrenched in a specific content niche. In a 2024 academic audit utilizing automated bots and real-user data donations via GDPR "data takeout" procedures, researchers analyzed the algorithmic behavior of TikTok [cite: 12, 14]. The study revealed that TikTok exploits a user's known interests in roughly 30% to 50% of the first 1,000 videos they encounter [cite: 12, 14]. The remaining 50% to 70% of the feed is dedicated to exploration. 

If a user exhibits high engagement with an "exploration" video, the algorithm instantly updates the user's preference weights, converting that novel topic into a newly "exploited" interest [cite: 12, 14]. This rapid processing capability explains why users can fall down highly specific algorithmic rabbit holes in a matter of minutes.

## Decoding Platform-Specific Ranking Mechanisms

While the foundational mathematics of machine learning are similar across the tech industry, different platforms weigh engagement signals uniquely based on their distinct business models and user interfaces. Understanding these nuances is critical to grasping how virality is manufactured in 2025 and 2026.

### TikTok and Douyin: Behavioral Prediction at Scale

TikTok and its Chinese sister app, Douyin, operate on nearly identical technological frameworks developed by ByteDance, but they serve different markets with distinct regulatory environments and feature sets [cite: 15, 16, 17]. Douyin features deep e-commerce integration and adheres to local regulations, including a strict "youth mode" that limits users under 14 to 40 minutes of educational content per day [cite: 15, 16].

At a technical level, both apps utilize deep learning models built on hierarchical interest label trees and partitioned data buckets [cite: 11]. Rather than relying on a user's follower graph, the system directly predicts behavior through neural networks [cite: 10]. In 2026, TikTok's algorithm heavily prioritizes watch time, completion rates, and shares [cite: 18, 19]. According to creator analytics, TikTok videos that achieve a 70% or higher completion rate reach five times more users than videos with a 40% completion rate [cite: 19]. To maintain normal algorithmic distribution, a video must typically secure a 3% to 5% engagement rate within its first hour [cite: 20]. 

### Instagram: AI Recommendations and the 3-Second Rule

Historically driven by follower networks and chronological timelines, Instagram has fundamentally transformed its infrastructure. By 2026, roughly 94% of content distribution on the platform is driven by AI recommendations rather than organic follower reach [cite: 7].

The platform utilizes distinct ranking algorithms for its Feed, Reels, Stories, and Explore page [cite: 21]. For Instagram Reels, the algorithm aggressively prioritizes the first three seconds of a video; content that fails to capture a 3-second retention threshold is routinely suppressed [cite: 7]. Furthermore, Instagram has deprecated the value of the simple "like" button. The primary signals for algorithmic amplification are now watch time and the "sends per reach" ratio, meaning content is judged largely on how frequently users share it with friends via Direct Message [cite: 7, 19, 22]. The platform also instituted penalties in 2025 for aggregator accounts, explicitly down-ranking unoriginal, reposted, or watermarked content to boost smaller, original creators [cite: 7, 21].

### Xiaohongshu (Red Note): The Community Engagement Score

Xiaohongshu (often referred to internationally as Red Note) operates under an entirely different paradigm. Blending the aesthetics of Instagram with the utility of Pinterest, Xiaohongshu functions heavily as a search engine; up to 60% of content discovery occurs through active search queries rather than passive feed scrolling [cite: 20, 23].

Because the platform relies on a decentralized distribution model governed by an "Interest Graph," follower counts are less vital to virality than content utility [cite: 24]. The recommendation algorithm evaluates content using a proprietary Community Engagement Score (CES). Empirical analysis indicates the weighting of this score heavily penalizes low-friction interactions and rewards high-friction utility: Likes equal 1 point, Favorites/Saves equal 1 point, Comments equal 4 points, and Reposts/Shares equal 4 points, with subsequent Follows granting 8 points [cite: 10, 23, 24]. Consequently, highly polished, studio-shot advertisements often alienate the Xiaohongshu audience and face algorithmic suppression, while information-dense "encyclopedia" tutorials that generate massive "save-for-later" metrics achieve hyper-virality [cite: 24, 25].

### WeChat Channels and X: Social-Powered and Conversational Ranking

Other platforms rely more heavily on social proximity and conversational dynamics. WeChat Channels prioritizes a "private domain to public domain flow," leveraging social relationships heavily in its distribution logic [cite: 10, 26]. In 2025, the platform increased the algorithmic weight of the "Friends ♡" tab; content that gains traction in private group chats or Moments is exponentially more likely to be pushed to the broader public recommendation feed [cite: 26, 27].

Meanwhile, X (formerly Twitter) transitioned toward a fully AI-driven feed powered by the Grok language model, pivoting away from raw engagement metrics toward conversational density [cite: 28]. The algorithm currently weighs replies and deep conversation threads far higher than standalone broadcast posts, ensuring that virality is tied to continuous discourse [cite: 28].

## Quantifying Virality: Engagement Benchmarks and Methodology

Understanding virality requires accurate benchmarking, but interpreting social media statistics can be perilous due to conflicting methodologies across the analytics industry. 

For instance, when evaluating engagement rates in the financial services sector, data firm Rival IQ reported a median Instagram engagement rate of 0.26%, while Hootsuite reported a massive 3.80% for the exact same vertical [cite: 29, 30]. This severe discrepancy exists because Rival IQ calculates engagement by dividing total interactions by total *followers*, whereas Hootsuite divides interactions by the total *impressions* or reach of the post [cite: 29, 30]. In an era where AI recommendations drive views far beyond a creator's follower base, measuring engagement strictly against follower counts yields artificially deflated metrics [cite: 30].

When looking at normalized data for 2025 and 2026, TikTok consistently maintains the highest baseline engagement rates across the digital landscape, though those rates decay rapidly as audience size grows.

### 2026 Cross-Platform Engagement Averages

Based on comprehensive cross-platform analysis of tens of millions of posts, the hierarchy of platform engagement becomes clear.

[image delta #2, 0 bytes]

 Smaller "nano-influencers" consistently outperform mega-accounts; on TikTok, accounts with under 100,000 followers average a 7.50% engagement rate, whereas accounts with over 10 million followers drop to roughly 2.88% [cite: 18, 31]. 

| Platform | Core Distribution Engine | Average Engagement Rate (2026) | Primary Engagement Signals (Weighted Highest) |
| :--- | :--- | :--- | :--- |
| **LinkedIn** | Professional interest graph & conversational AI | 3.85% [cite: 18] | Dwell time, long-form comments, shares |
| **TikTok** | Behavior prediction & multimodal feature extraction | 3.70% [cite: 18] | Completion rate, watch time, shares [cite: 19] |
| **Instagram (Reels)** | AI recommendations & social graph | ~2.80% [cite: 29] | DM shares, saves, 3-second retention [cite: 7] |
| **Instagram (Feed)** | AI recommendations & social graph | 0.48% [cite: 18] | Saves, comments, prolonged dwell time |
| **Facebook** | Hybrid chronological & algorithmic | 0.15% [cite: 18] | Meaningful social interactions, comments |
| **X (Twitter)** | AI-driven conversational ranking via Grok | 0.12% [cite: 18] | Thread replies, active ongoing conversations [cite: 28] |



## The Economics of Attention: Rage-Bait and Bias Amplification

The metrics that determine virality are entirely amoral. While technology executives frequently state their algorithms are trained to maximize "meaningful social interactions," in practice, the most reliable mechanism for extracting high engagement is emotional arousal.

Content that evokes strong, immediate emotional responses—particularly outrage, moral indignation, and partisan hostility—naturally holds human attention longer and provokes significantly higher comment and share velocities [cite: 32, 33]. This psychological reality has birthed what researchers refer to as the "economy of words," where the unwritten operational doctrine of digital media is simply that "angry people click more" [cite: 32, 33]. 

When recommendation systems operate purely on engagement-based optimization, they naturally drift toward divisive and inflammatory content. To a machine learning model, "rage-bait" is not a stylistic flaw; it is a highly efficient content category that perfectly matches the optimization function. The user's "anger click" is simply the behavioral trace confirming the system's success [cite: 32, 33]. Consequently, former platform designers acknowledge that notifications, feeds, and recommendations have been aggressively refined through continuous A/B testing to discover precisely which combinations of topics and headlines keep users enraged and returning [cite: 32].

### The Danger of the Feedback Loop in Recommender Systems

This algorithmic preference for strong emotion creates severe mathematical vulnerabilities in the form of AI feedback loops. Recommender systems inherently rely on dynamic models that continually learn from user interactions based on their own prior predictions [cite: 2, 3]. Over successive cycles, these self-reinforcing loops can severely amplify existing biases, degrade recommendation diversity, and homogenize user behavior [cite: 2, 34].

Techniques like Matrix Factorization and Collaborative Filtering—the foundational mathematics of many recommender systems—operate by predicting missing interactions based on past data [cite: 11, 35]. However, because the system influences the very behavioral data it subsequently uses to retrain itself, the algorithmic outputs become increasingly narrow [cite: 2, 34]. A vast majority of historical bias mitigation techniques were tested on static data splits (a single round of testing), failing to account for how fairness and diversity diminish over months of live, multi-round retraining [cite: 1, 2, 3]. Only recently have researchers shifted toward dynamic simulation frameworks to audit the long-term, corrosive effects of these feedback loops on marginalized content and user behavior [cite: 1, 3].

## Echo Chambers and Filter Bubbles: Structural Isolation

As users interact with highly optimized, emotionally charged content, the feedback loop progressively narrows the scope of information presented to them, culminating in the creation of filter bubbles and echo chambers [cite: 36, 37]. 

A filter bubble occurs when algorithms systematically reduce exposure to diverse information, prioritizing content that perfectly aligns with a user's prior interactions [cite: 37, 38]. An echo chamber is the social consequence of this filtering: an environment where individuals predominantly engage with like-minded voices, amplifying shared ideological perspectives while marginalizing dissenting views [cite: 37]. 

In a recent analysis of over 16 million TikTok videos spanning the 2019 to 2023 US election cycles, network mapping revealed highly distinct clusters of politically homogeneous networks [cite: 39, 40]. Within these political echo chambers, users were systematically fed attitude-consistent content, and those receiving positive social feedback were highly likely to escalate their own political expression, further reinforcing the bubble [cite: 39, 40]. Similar dynamics of homophily (the tendency to associate with similar individuals) were starkly illustrated in comparative studies of Parler and Twitter; users on platforms with overwhelming majority opinions exhibited longer retention and higher stability, while users exposed to dissent often migrated away entirely [cite: 41]. 

The algorithms also adapt to geopolitics. An online survey examining Douyin usage in Taiwan found that users identifying as Chinese and supporting political unification were significantly more likely to utilize the China-based platform and reside within its specific partisan echo chambers, highlighting the political consequences of authoritarian-led media environments on democratic populations [cite: 42].

### The Efficacy of the Bubble: Does It Change Beliefs?

Despite the pervasive anxiety surrounding the "personalization-polarization hypothesis," the academic community is sharply divided on the actual power algorithms hold over deep-seated human beliefs.

Some sociologists warn that well-intentioned efforts to manually "pop" filter bubbles by forcibly exposing users to opposing viewpoints can actually backfire [cite: 36]. Empirical models suggest that introducing highly foreign ideological opinions triggers "negative influence" or repulsion, causing users to retreat further into their extremes [cite: 36]. 

Furthermore, the direct causal link between algorithmic recommendations and real-world polarization is heavily debated. In 2025, a massive study published in the *Proceedings of the National Academy of Sciences* (PNAS) utilized a custom-built, naturalistic video platform mimicking YouTube to experimentally manipulate the recommendations of nearly 9,000 participants [cite: 43]. The researchers intentionally forced users into heavy-handed algorithmic "rabbit holes" and filter bubbles. The results cast serious doubt on prevailing alarmist theories: the short-term exposure to these intense algorithmic manipulations had no detectable, consistent causal effect on users' actual policy attitudes [cite: 43]. The researchers concluded that the burden of proof for claims regarding algorithm-induced political polarization must now be shifted, suggesting that algorithms are highly effective at capturing attention but much less effective at fundamentally rewiring human beliefs [cite: 43].

Additionally, some researchers point out that filter bubbles are not intrinsically negative. For marginalized communities or citizens living in nations with restricted press freedom, algorithmic isolation can serve as a "protective filter bubble," creating vital digital safe spaces free from harassment and state monitoring [cite: 38].

## Strategies to Disrupt the Loop and Re-train Algorithms

Given the overwhelming influence of the human-AI feedback loop, users, researchers, and regulators are actively seeking methods to disrupt the algorithmic cycle and reclaim agency over digital feeds. 

### User-Level Audits and The "Not Interested" Efficacy

Most platforms offer user-interface controls purportedly designed to correct algorithmic mistakes, such as "Not Interested" or "Don't Recommend Channel" buttons. However, their actual efficacy has historically been shrouded in secrecy.

To test these tools, researchers conducted a massive "sock-puppet audit" of YouTube's algorithm. They deployed pre-programmed automated agents to simulate human users [cite: 44, 45]. The agents first executed a "stain phase," binge-watching specific topics (like misinformation) to deeply bias the algorithm [cite: 44, 46]. Next, the agents initiated a "scrub phase," repeatedly applying negative feedback tools to try and burst the bubble [cite: 45, 47]. 

The results were highly specific: clicking the "Not Interested" button was the single most effective strategy for cleaning the homepage, successfully removing an average of 88% of the targeted topic recommendations [cite: 44, 45]. However, neither the initial staining nor the subsequent scrubbing had any significant effect on the "videopage recommendations"—the "Up Next" videos presented while a user is actively watching content [cite: 44, 45]. Alarmingly, a corresponding survey of 300 adults revealed that 44% of users were completely unaware that the highly effective "Not Interested" button even existed [cite: 44, 46]. 

Recognizing widespread user frustration, platforms have begun offering algorithmic "nuclear options." In late 2025 and early 2026, Instagram globally launched a "Reset Suggested Content" feature, allowing users to wipe their algorithmic history across Explore, Reels, and their primary feed, effectively reverting the account to a blank-slate "cold start" [cite: 6, 7, 48]. TikTok features a similar "Refresh your For You feed" capability [cite: 49, 50]. However, resetting an algorithm is only half the battle. Because the system immediately resumes its feedback loop, users must spend the critical 48 hours post-reset engaging heavily with desired topics and aggressively skipping irrelevant content to properly train the new model [cite: 6, 7, 50].

### Platform-Level and Regulatory Interventions

Systemic changes to the feedback loop require platform-level interventions. Studies have demonstrated that inserting friction into the virality cycle works. Research from Yale University analyzing the "Community Notes" feature (a crowd-sourced fact-checking framework) found that attaching a warning label to misleading content significantly reduces its algorithmic momentum [cite: 51]. When a note was attached, posts saw an average of 46% fewer reposts and 44% fewer likes, effectively stopping the misinformation from penetrating deeply into networks—acting, as researchers noted, like a bush that grows "wider, but not higher" [cite: 51]. 

Similarly, Stanford researchers successfully designed a web-based tool capable of downranking anti-democratic and hostile partisan posts on social media without removing the content entirely. In an experimental setting during the 2024 elections, exposure to this downranked, less antagonistic feed successfully improved users' attitudes toward opposing political parties, proving that algorithmic de-escalation is technically feasible [cite: 52].

Governments are also attempting to force transparency. In 2026, US legislators introduced bills requiring platforms to submit comprehensive disclosures of their algorithmic weights, specifically demanding explanations for how engagement metrics like watch time and shares are coded [cite: 53]. Concurrently, the European Union's AI Act introduced risk-based regulation and compliance models that legally obligate social media operators to take preventative action against the risks their algorithms pose, enforcing heavy financial penalties for enabling broadcast disinformation campaigns [cite: 54]. 

## Incorporating Human-in-the-Loop (HITL) Architectures

The ultimate safeguard against runaway recommendation algorithms is the integration of Human-in-the-Loop (HITL) workflows. In commercial AI deployment, a model's release is never the final step; it requires continuous retraining, evaluation, and human oversight to manage the inevitable data drift and edge cases [cite: 4, 55]. 

A sophisticated HITL pipeline does not merely use humans for basic content moderation. Instead, humans serve distinct, structural roles: as validators checking pre-action outputs, as editors contextualizing data, as deciders for high-risk edge cases, and as teachers providing nuanced training signals [cite: 56]. For example, the HIVE framework for explainable recommender systems actively collects "veracity-based human feedback." By grading AI recommendations on both "fidelity" (factual accuracy) and "attunement" (user preference alignment), human operators continually update the user and item embeddings within the neural network, ensuring the algorithm does not collapse into a hallucination or an extreme bias loop [cite: 57]. Integrating structured human oversight prevents the mathematical drive for engagement from wholly superseding accuracy and safety.

## Bottom line

Recommendation algorithms dictate modern virality through a relentless human-AI feedback loop, instantly adjusting to passive behavioral signals like completion rates and swipe velocity. While optimizing for these metrics successfully captures user attention, it mathematically favors emotionally polarizing content, inadvertently constructing structural echo chambers. However, emerging research suggests these filter bubbles may be less effective at altering deep-seated beliefs than previously feared. Ultimately, while platform-level resets and "Not Interested" buttons offer temporary relief, reclaiming agency over a digital feed requires users to remain hyper-vigilant about the behavioral data they continuously feed the machine.

## Sources

1. [Viral Social Media Campaigns 2025](https://digital.rothian.com/viral-social-media-campaigns-2025/)
2. [2025 Social Media Algorithm Changes](https://stackinfluence.com/blog/social-media-algorithm-changes-engagement)
3. [AI in Social Media: Going Viral](https://drainpipe.io/ai-in-social-media-going-viral-in-2025/)
4. [Social Media Algorithm Updates 2026](https://iqfluence.io/public/blog/social-media-algorithm)
5. [Unchecked Algorithmic Amplification](https://www.mexc.com/news/1119963)
6. [What is Xiaohongshu (Red Note)](https://www.eliteasia.co/what-is-xiaohongshu-red-note/)
7. [How to Make Your Post Go Viral on Xiaohongshu](https://halotechmedia.sg/blog/xiaohongshu-how-to-make-your-post-go-viral/)
8. [Xiaohongshu Algorithm 2026 Guide](https://influchina.com/xiaohongshu-algorithm-2026-guide/)
9. [China's Content Game: Douyin & Xiaohongshu](https://mmgthailand.com/chinas-content-game-douyin-xiaohongshu-wechat-video/)
10. [Xiaohongshu Algorithm Core Features](https://www.octoplusmedia.com/xiaohongshu-marketing-algorithm-core-features/)
11. [WeChat Channels Algorithm Updates](https://www.thewechatagency.com/category/douyin/)
12. [Douyin vs WeChat Channels 2026](https://valuechina.net/en/china-blog/chinese-digital-platform/douyin-vs-wechat-channels-2026-3/)
13. [Comparative Analysis of WeChat and TikTok](https://www.researchgate.net/publication/368486218_Comparative_Analysis_of_WeChat_Channel_and_TikTok_in_China_Short_Video_Clips_Market)
14. [WeChat Channels Guide](https://jingdaily.com/posts/wechat-channels-luxury-guide-livestreaming)
15. [Latest WeChat Data in China](https://fashionchinaagency.com/latest-wechat-data-in-china/)
16. [Stanford Tool Lowers Political Polarization](https://news.stanford.edu/stories/2025/11/social-media-tool-polarization-user-control-research)
17. [Flagging Misinformation Reduces Engagement](https://news.yale.edu/2025/09/25/flagging-misinformation-social-media-reduces-engagement-study-finds)
18. [Disinformation Countermeasures Study](https://pmc.ncbi.nlm.nih.gov/articles/PMC12149016/)
19. [How You're Curating a Biased News Feed](https://bfi.uchicago.edu/podcast/social-media-algorithms-how-youre-curating-a-biased-news-feed/)
20. [Countering Disinformation Effectively](https://carnegieendowment.org/research/2024/01/countering-disinformation-effectively-an-evidence-based-policy-guide)
21. [5 Steps to Build Feedback Loops](https://www.artech-digital.com/blog/5-steps-to-build-feedback-loops-for-ai-models)
22. [Google PAIR: Design AI Feedback Loops](https://pair.withgoogle.com/guidebook/chapters/feedback-and-controls/design-ai-feedback-loops)
23. [Human-AI Feedback Loop](https://medium.com/ml-and-automation/human-ai-feedback-loop-fcb96392cec)
24. [The Power of Feedback Loops in AI](https://irisagent.com/blog/the-power-of-feedback-loops-in-ai-learning-from-mistakes/)
25. [Handling Feedback Loops in Recommender Systems](https://towardsdatascience.com/handling-feedback-loops-in-recommender-systems-deep-bayesian-bandits-e83f34e2566a/)
26. [Auditing Personalization in TikTok](https://www.researchgate.net/publication/405236310_Auditing_Algorithmic_Personalization_in_TikTok_Comment_Sections)
27. [How to Train Your YouTube Recommender](https://www.researchgate.net/publication/372683807_How_to_Train_Your_YouTube_Recommender)
28. [ICWSM Study on YouTube Scrubbing](https://ojs.aaai.org/index.php/ICWSM/article/download/31363/33523/35419)
29. [Arxiv: How to Train Your YouTube Recommender](https://arxiv.org/html/2307.14551v3)
30. [Filtering Discomforting Recommendations](https://www.themoonlight.io/fr/review/filtering-discomforting-recommendations-with-large-language-models)
31. [Douyin vs TikTok Analysis](https://www.scribd.com/document/943431121/Douyin-VS-TikTok)
32. [Business Insider: TikTok vs Douyin Restrictions](https://www.businessinsider.com/tiktok-different-china-douyin-jonathan-haidt-anxious-generation-national-2025-1)
33. [E-commerce on TikTok vs Douyin](https://mtomconsulting.com/comparison-of-tiktok-and-douyin-in-e-commerce-and-marketing-and-insights-for-u-s-brands/)
34. [Analysis of Douyin Mania Phenomenon](https://www.researchgate.net/publication/349000779_Analysis_on_the_Douyin_Tiktok_Mania_Phenomenon_Based_on_Recommendation_Algorithms)
35. [Main Differences Between TikTok and Douyin](https://www.eggsist.com/en/insights/tiktok-vs-douyin-which-are-the-main-differences/)
36. [Instagram Algorithm Complete Guide 2025](https://www.dataslayer.ai/blog/instagram-algorithm-2025-complete-guide-for-marketers)
37. [Reset Instagram Algorithm Guide](https://buffer.com/resources/reset-instagram-algorithm/)
38. [How to Reset the Instagram Algorithm](https://www.cashify.in/struggling-with-your-ig-feed-reset-the-instagram-algorithm)
39. [Instagram Algorithm Creator's Guide](https://onestream.live/blog/instagram-algorithm-creators-guide/)
40. [Resetting the IG Algorithm in 2025](https://nomadzdigital.com/blog/reset-instagram-algorithm/)
41. [Current Time Data](https://www.google.com/search?q=time+in+Dubai,+AE)
42. [Exploration and Exploitation on TikTok](https://homes.cs.washington.edu/~franzi/pdf/vombatkere-tiktok-webconf24.pdf)
43. [Arxiv: Investigating Exploration and Exploitation](https://arxiv.org/abs/2403.12410)
44. [Figuring Out TikTok's Algorithm](https://www.fastcompany.com/91065874/researchers-are-finally-figuring-out-how-tiktoks-algorithm-works)
45. [The TikTok Tradeoff Study](https://www.semanticscholar.org/paper/The-TikTok-Tradeoff%3A-Compelling-Algorithmic-Content-Klug-Santos/eb93c956a8b403332f58daddf42b23797627abdd)
46. [ResearchGate: The TikTok Tradeoff](https://www.researchgate.net/publication/358861933_The_TikTok_Tradeoff_Compelling_Algorithmic_Content_at_the_Expense_of_Personal_Privacy)
47. [Breaking Feedback Loops with Causal Inference](https://www.researchgate.net/publication/391342184_Breaking_Feedback_Loops_in_Recommender_Systems_with_Causal_Inference)
48. [Bias Mitigation for AI-Feedback Loops](https://arxiv.org/abs/2509.00109)
49. [Mitigating Bias in Recommender Systems](https://ir.library.louisville.edu/cgi/viewcontent.cgi?article=5168&context=etd)
50. [Arxiv: Bias Mitigation Taxonomy](https://arxiv.org/html/2509.00109v1)
51. [Systematic Literature Review on Bias Mitigation](https://www.researchgate.net/publication/395214299_Bias_Mitigation_for_AI-Feedback_Loops_in_Recommender_Systems_A_Systematic_Literature_Review_and_Taxonomy)
52. [Who Uses Douyin in Taiwan? Survey](https://taiwanpolitics.org/article/144409-who-uses-douyin-or-tiktok-in-taiwan-evidence-from-a-2025-online-survey.pdf)
53. [Users Seek Out Echo Chambers on Social Media](https://news.rpi.edu/2024/05/09/rensselaer-researcher-finds-users-seek-out-echo-chambers-social-media)
54. [TikTok's Political Landscape Echo Chambers](https://pure.psu.edu/en/publications/tiktoks-political-landscape-examining-echo-chambers-and-political/)
55. [Examining Echo Chambers and Political Expression](https://www.researchgate.net/publication/392179153_TikTok's_political_landscape_Examining_echo_chambers_and_political_expression_dynamics)
56. [LLM-Powered Simulations Revealing Polarization](https://aclanthology.org/2025.coling-main.264/)
57. [How to Reset TikTok Algorithm](https://sociallyin.com/resources/how-to-reset-tiktok-algorithm/)
58. [Reset TikTok FYP Strategy](https://socialbu.com/blog/how-to-reset-tiktok-algorithm)
59. [How to Retrain the TikTok Algorithm](https://turrboo.com/blog/how-to-reset-the-tiktok-algorithm)
60. [YouTube Tutorial on Resetting TikTok](https://www.youtube.com/watch?v=dfqstL1mnhQ)
61. [Metricool TikTok Algorithm Study](https://metricool.com/tiktok-algorithm/)
62. [Rage-Baiting and Anger-Clicking](https://medium.com/@ayolov/rage-baiting-and-anger-clicking-the-dissent-factory-of-online-media-ff6beabe9a89)
63. [OSINT Platform Leaks and Echo Chambers](https://osintfieldnotes.substack.com/p/osint-field-notes-5)
64. [Freedom on the Net 2024 Myanmar](https://freedomhouse.org/country/myanmar/freedom-net/2024)
65. [The Dissent Factory of Online Media](https://www.researchgate.net/publication/399739621_Rage-Baiting_and_Anger-Clicking_The_Dissent_Factory_of_Online_Media)
66. [SendGrid Email Platform Breach Investigation](https://socradar.io/everything-about-twilio-sendgrid-breach/)
67. [Social Media Algorithm Updates and Tips](https://storychief.io/blog/social-media-algorithms-updates-tips)
68. [What are the Latest Algorithm Updates](https://www.wearehydrogen.com/stories/what-are-the-latest-algorithm-updates)
69. [Changing the Rules of Reach](https://adpulse.com/social-media-updates-2026-how-x-tiktok-and-instagram-are-changing-the-rules-of-reach/)
70. [Social Media Trends 2026 Video](https://www.youtube.com/watch?v=lHiLwmfv8mE)
71. [Instagram Updates 2025-2026](https://www.brandvisionhub.net/post/instagram-algorithm-updates-2025-2026-what-s-changing-and-how-to-stay-ahead)
72. [TSE: The Complex Link Between Filter Bubbles](https://www.tse-fr.eu/filter_bubbles)
73. [Review of Youth Engagement and Echo Chambers](https://www.mdpi.com/2075-4698/15/11/301)
74. [Short-term Exposure to Filter Bubbles Study](https://www.gov.harvard.edu/2025/05/08/short-term-exposure-to-filter-bubble-recommendation-systems-has-limited-polarization-effects/)
75. [Evidence For and Against Filter Bubbles](https://www.researchgate.net/publication/396660350_Assess_the_Evidence_for_and_Against_the_Existence_of_'Filter_Bubbles')
76. [Developing a Research Agenda for Protective Filter Bubbles](https://ui.adsabs.harvard.edu/abs/2025arXiv251112873E/abstract)
77. [Diagram of Feedback Loops](https://www.researchgate.net/figure/Feedback-loops-between-an-AI-system-that-selects-and-summarizes-information-and-the-user_fig1_370117935)
78. [Human-in-the-Loop Agent Flows](https://medium.com/rose-digital/how-to-design-a-human-in-the-loop-agent-flow-without-killing-velocity-fe96a893525e)
79. [HIVE: Veracity-based Feedback on Explanations](https://xai.kaist.ac.kr/static/files/2025_hcai_workshop/paper_04.pdf)
80. [How to Build a Human-in-the-Loop Pipeline](https://humansintheloop.org/how-to-build-your-human-in-the-loop-pipeline-a-step-by-step-guide/)
81. [Building Feedback Loops for AI Models](https://www.artech-digital.com/blog/5-steps-to-build-feedback-loops-for-ai-models)
82. [Engagement Rate Benchmarks 2026](https://mazkara.studio/en/blog/engagement-rate-benchmarks-by-industry-2026/)
83. [Social Media Benchmarks by Industry](https://apaya.com/blog/social-media-benchmarks)
84. [Benchmarking Across Platforms 2026](https://improvado.io/blog/social-media-benchmarking)
85. [Social Media Engagement Benchmarks 2026](https://www.shortsintel.com/statistics/social-media-engagement-benchmarks)
86. [Engagement Metrics Across Networks Guide](https://influenceflow.io/resources/engagement-metrics-across-different-social-networks-the-complete-2026-guide/)

**Sources:**
1. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGtoKDPuhCtAo5uGXbnKdUWTT6SWVYX056Xz9cxBf-SF4ZB0hzzvFHPSiKB9uId14gmOpuraMYCilIKKmCQJpztn61EWvR41ZmQxJ4H061xiw7UbXWT)
2. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGmzRUamjpF1lv7XID1juyCXhjw2qe0PgOx2IZsq4QlgurgravrbKU1UyDJSrmWjxOOWWZhFy73_VJNchryy7bcLuvoOJH_WbTslMJhO_cxoFZhdQEawAO5)
3. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFPV04BwWc-4UoAvvGkivmSHEdomAldXIP_KsyEiQRfC4X_4tm0mVKfHZFIUBQGUUMypbqMv3SAObAnwA752cV7Oe6sEk3zFvh03rCK6xi7kcA9FpvazaFVoFhbET21FDPQkzzwCGLKOVAxRzJ_8E7K3ELUoQDlzkvOwSxCz2BhgWeZJXYncdVetmy-D2C7ksfWrS8IS89UUnAMhNxPra593SOlKCqiEUofuuwS1N3xAGv_2NyYNu5gwDbDrxtppGRYa4kM4FsBseGaOUPWql7Ukg==)
4. [artech-digital.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFnu4fxgX7OQg7glPsmARQveFLzRdaC7tzrKLE8hyOSs0tO9Args3tEZHUIfTE6LsG2BeNlmwxU-4qiMDBEr2X0rbvMNyLLkWOqbJSSVCt0C8Iu7ONB4mW1kRXlmZnSmXKZpI4Of5ZgvNNP9RL0DOuyxJ-sF5u6R4WDBRDHs81JP95Tj_07Wg==)
5. [withgoogle.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFvYfu-EOjQ7vTiqLhZxCfCVzvAK8N_e4VmyO1-77cc-aoDTD5BapGiWz_l1eMPw2rmYINOwpKUJCmKDXEX2dgsa6U0aak6fKEZtrDDein7PS_Icb5SB1XADiUYEVYNw2uC-qZRi1myAffoM8u6r_XZYSaODfbTFlpEYQIlzg4KC8eibf9cIWHQMOiKfYl_Vi1-_Q==)
6. [cashify.in](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEmaDVBs-g8PFCHwBOuYPpFmqnN3C5IMj-4MxMYw5S2-BnqfNPTtnyOI5-jHXzc6P8_AGg9CFkKspJTdoyct_5u6aGV5OBb9flg2DJIPAQENvjxOXQqvy-FP1s1nx3Ooj1esF1n86xReKy9mMEX5KB4HRzb48yPH_D7YZYM9k3nqjEsv4i6Jg==)
7. [dataslayer.ai](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEuqRMEO6WDOssJY2r47z2qcClxlOAFM66F0XyzmR9_jvgh26PPjJr7eBu7py9JzyCkIG2lo08Ws_vDpbr1lBuhXZGJTAWBCaonKE6HbaeJPTdV7PEEoMOiFPte8_fgR9A9sMuAwQd_zVvGCFvEtTX_kFGi1Vs5IgFQn4yb1cz-E6a2nZYHU9DaOw==)
8. [towardsdatascience.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGDp4OYZwU9biuEQPp8St5any3_dnlU1iTmZqgQWccw5MaTaPzvqU3ApRH9E07_MVQxvC98RybHjV5-0r6AEkDidfbWr9jx2s0Rf8YGW4AzhwUwm_jYQBqhsvemgmx_6mFdJ6bggSIElUauCKjacFKpIftGUPaYaMnOWN5D9Y8lOP_5HT6xFXZJNQEOMaZXyBA2-s0QGY6mJt7QpWNIw2z4GYkG7RmE)
9. [iqfluence.io](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGmTPTQ6SlIBolAuYn19MklrbVteeZECa-rihZfMumTybLEHBE8IGtve1VdnOBeoj0IQwAgTkPR0GAfyGxtnhvXMAzWnJXiTV--hmHKYNndkyknixNqd9vFwja1CtMZrJAfgVrkfjjBeu6c-YM=)
10. [mmgthailand.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGVr2pwMMZJ24IVXX0hI__y1VvHyY6HY4VInGoMHEq1waD9MhoOCc1YQC_RzEfvEAfLYhQLrosIy-LZPS2WTwCNae1554eEcIxQuFp5LfbUCiFfwm_FG8JOIDZeIm8Qx2krDPBhg635b474J0A4IPnZxcFp7RYti8XD_RCA-8kuWAI=)
11. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGZMNQgWe8yhalK9YhTWzmuiNnD3QjxspUlEHGBv4VsvJjF2QQQsTlKBpGY_AIQTO6ZsvLbrhPuRinONYjP3asqzD16H3WPYKd3XsQOSAhYvQqG07vmEeaaoUX_6S-MZ2dK4zb987DHLroZmGo3q7K0JfmFGAdiDR8w-g_PqCKX0_gi3lWfxGKSK7fKEPUH9WDCcayG0KNarWk9lz6yr6GQxITZjMoMTVdKdw6IcQIdCWZ8IVRG-tnQN7M=)
12. [washington.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGPcM-oFjl-s_BzXSYp9rBAzjOVKUtQ5CMN1wvJwwUpfFPMMxJPq19IqwL2rD0DV9L9MRYGIAkC7EGcuiaVRuKHcELXXb9fHYM9vH6VDE6rsh-Csgphk43fB9UEz7ThoQtf0aB6sAie8ihfrOfa53xx6j11rMRpJ7cCptDjXuHjwg==)
13. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGKjazIb8fJsP3mCoVC2Sl3TDstMC9lCIZjPXmdHso_DbR0dmnZUXW9c1BKxFLC1hQUK-7uUCzqhUtYO1Hl5QSFKUlHUkTM5yVzvcNKzpsitKTTexJ-)
14. [fastcompany.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFmYdTaf_6usWxfyx2Nk36qG7ecL_BV6thSQv9r9II1QXqQTkJQOzbK3ewOTZHQUE2JZnzXFxtOpf0avqFWeuwmHPwbDtJE5eo6aNcHILprLZKK3YfShAqkUok2FH3mFtyuAZ-UZXfLDt2QwZgHj62SQJQInw6kVHOyIO0cn7wp1DfImDuAZMhBGa0VkIxQZhIV-1hsq6X-lN54)
15. [scribd.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHJEkyHMBw-Isxjpfn_YjlVy7Zpo-5rAGF9814zKIxXAT2gZPeWThem9M_kxG_fEOoUDjup77hSufEkK4ylk-r-Cb4dQ0iA5jtsDxSOxj5MWC-MCKo08CWFsrwijzFK3dIzA--fd1_0boXtH8OacWI=)
16. [businessinsider.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGDOyE_RzmQBANo7185v7YDOUk78xvq63xe0m837MsODFwDRulVbWKSAtzlZ3EVjlRDKqhr18ag1hyQfhYB1vMQ8gSbScCQj682lTzQVlBKJMxt7KcKbyzeKItIcxCQYOcPotNEndtb2w8v2uo6YAqCUkzg3ZbaM9pkOmvHxsyFQtcbhsdGFfd23BC439kJVH9Z6bD-QhAysmJhRpmaF8vtzLs9EQ==)
17. [mtomconsulting.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEBRbT5yP7mGTuQzJU7eLpVsc7GRW6GNUbWPn2WP2yJuR-6GXwUFtx3qRXT6Y4iUOcZGD3UtiuojKg03EWNTrmIxUd6OcMifwHW2OxP9JqUbVrn_u1Dor9W4B6skDApItw5MfzKnEu670beBW48_8PpBUziRDBQ_Q1dD9nUlPDg_82NQRAoN1j1wGTu0Lx99wM8qk3BwcLvE2WKErLD2MlhdptWUPHWdYw=)
18. [shortsintel.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGShk7A622JwdgSCTf2DIS01zS5paYvEXtHulAnp-GcWz4NJCMWn_Ja9OfuXi7BNV7MdnbYTuiuo_5PJat2Ve4324nv5hs-_J3YsktBoUwHodJRtDclIiZ-b-ZSJgTlYE8cPrmYi7l-nffi5IJVJXtpFnjP2j8CYKw8XCNFm6c=)
19. [influenceflow.io](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGGEtzC8eMEj7QMU_qiBKyArKurWfmyR8oeEwJ3_lfkH7xBE6exGJOiptffv-uEU3NgChFfqlTEYPlvv28E0if6NFSu1UYX8ECMXJp04IWDWH55ArIwrNO9JBXOEgYSqKbowEMYf7JfHaKSyjgOFBV-OxRIVl41UxhvVNHiYvQCo4o_9IOvAALRDQTh0g3uIGOJxdU8-VAYUyonNOz9OUcb_-e-LQ==)
20. [influchina.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEYuULt6-maZrVdZ_nnrJsuH4OJJSOEjnoihm2TYO-QwNbAJXBJwprZlgY4Mna-pXftaGK1e324NFieljejtHG3012olX8jZrt20yyHpjCj-MqzbiM5egVDIIf7qiCPQdI1wQqlgD3vPPND8pPg)
21. [brandvisionhub.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHmianr6bFWDxLcSQvhQexMHgB9k14djSCFwVYHYhCtlZv9fn1rPSzycoT3r1vS3WC-OreRRleMrfjcFk3FnqpTghOVQzYXHR3iY4JD4AeYZKMLD0y9ALzY3kE7NnKud-CmwiK3nBlD2YEV7kqiMDrsV9bLXFXm8GUmV3jaIMQ943zgy_YzPo27TowrjTYNe1zqAxGscs7YX1UxsbO-9CVqoAHMgQ==)
22. [stackinfluence.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEdbkBAHqAYyBD1GqNmq779PDXxesXoripE1A7dyPuHL4y-3TG23ddqpmCgS0bWDyZgpWbGluVgL8i-hKXG9MRE9kT1eoN_EbCnD7LfwM2CfTY1-X-iJuFvNwjJ6GRoWG_nOnCg-JECtfFPMbfRwbm01KQRAQ-G9s28MHZJxvs=)
23. [octoplusmedia.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHrdOnG19r5RU3OaQumrkILOk98xgxaZs-F1pkxmXY2-syhfpgZ7j705QDiEEZexTnOEk5ay5tBz4IsK6gF_wfQL_g4R2GW271btLbbPil_CvVPRa5E75GH5JtXWxq5y3s5H14vZ4RUnnGXzrxuAEKfPlBaATM5C0jplityLPLGqGY=)
24. [halotechmedia.sg](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEDSDfVAB-PUE7ulrqLmO8ah5k1Rjfj-7i0N3IM1z4cMu9eFPmaQ6rc-yoeOZVMi1iY9ArNr9ZNiIabAzVW_XEOj0Xxng3qm-8rhVOBiqDgBXv4hjqce1RwWtGUzen0-YEg50qp2ryztTy1KtCzcqTxaSTtGqvGPzSsyuYNesY=)
25. [eliteasia.co](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHlywE32aUyRaYAFWIETgNxhgCNRBwtJxZ0nMEE-QHL7rx43ucul21DXwp1hkk6ZSrIb4OyWpHrerc0N0ARuJ-GLFwOAoSSf3FxH6O6p3FPlSX99cZhyjtm8_Z_pPmBvaClN0WO21MgK--ExA==)
26. [thewechatagency.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEkf7ggpjnfLh4XPz0uiUVg6-gpCSlqTI-xr9WTmuekQvcyC_HYm1bW4yw8ooTxHvXBLpbi1GChgZP_l1FZzMtm_sMAejhAv4-p1wMNRvV8SfZ3-eZ8nJEYARLP_MmfEyquw7TfAw==)
27. [jingdaily.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGQgUPrXwKi6bcMPCqM3hysOsi1rm0MHZ2TxqA-wnMORbYvsclJTSxD2KhQvGIlimn5Xf0sVhhtZlQRFvX4wR7KuOMhfHi04dsJJv3IInxGiOJQBD4TTl10mRrfJqIK21n7e8EA5Qk0Ez1xVNXijimmBeCKFUx72TbK5A4=)
28. [adpulse.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEDRn_4SN5QtoPvJ2lgekpcS0SKLMWpXJZ1c28fQi8OkAI1rOUI6JuIHevwFKLGPAiRDeV8dt6LROwWmmeuWAIv9G1OpOf9MVGcZ2TpmII_Dp_cRhHDVoxqLLSUCeodVp8v943E55KyJ-1mKc7uuHf_V-pQprPK9EQKjYt04psb-HqPtm8-1nfOVEy7WzIqMpguVTqQNdrWgflW9SrqwA==)
29. [apaya.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHW3cZHCOAHjPtwRgoA2i4NZXb2tzI9zxq5geO8SKBAW_JvGENjgsPkrNbtTydCYDtZVop6PNofZ4SeSYldlxKnqhB84QaIyGmt3bvtLeFtQwMXxagbb7xjEpILI0V4_ohj6ug=)
30. [improvado.io](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHODaoYlFLiLAoSuMfvKfCe49ajuazUrcz8zEnU8JJh9_yYS7AXTjDuoEKlarS1sRdF9MGdQzZupBcEUwyzq5hZB_w8w7gv561gxWzLmlvHzu3L0ij1vikUlIqzzEs_gOCZ7Y5d2LdiTg==)
31. [mazkara.studio](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFLwjXhop1HdAtD2wGA7a8W6-yiyhZxNe4HEcQqOLbFWjYV4A99AufwEEbNJm8vWCAAIPwfgbTci39nXx2TxbSOOOWqkeTYW8eoXgyH02TlDMHUHB7KxfZjtK03BYiw7Qco9Ssy2qTQr1eZoImvXFzhty4d_SXUbACg_2pJVA9EqQ==)
32. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFIuE6-HJe6oW7Na3F2P7IA9rBOqqA5saOoF4RaeOpUlB7fBZ41yKyrbdkKeYpPFbipPpcp3e53m6-fcsZjye2GHLATD-yLX0vXzExGxlWFFTkPj8sJ8i7FG1QD6nDzZ7VGu_FmaouHeCm8ZO99IstGZknHFbqio6H5G6Iux0QDssPDa24RfiZlmstA2nc2rlta-dN-kzNS0ZWItMVnAtrR)
33. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQENpIw-N2jzvdrIQ-775VVFhDO04wuJuoCb8toz-mqI4ueUizlTvKyyC9CTro3oe99H1PdRolPxeHaufcF5Sv0vrvP9Nwr-0o5-yPEPzM1RrrLO9hOBkK1JlX8KUEFU6-Jtz4Q_ANkVzZSLouuZUZgPwBcsZ-idTIPGRJmgvI7fms2EgbUYsgETwqYj5J8mT1rnSH66tZv-dDVracB-BQ-xtGfTBh_7mH36DYw=)
34. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFKpHJI3bGV4Lzab3x8fWItqjb7qYeH0oIiIyoXVFQjOkkp14gzskHPryMNva8Ha-M0xWBgrJ6IcgVmso5XcIpPWZ3TVeIxET3MnKHphV0hJh5W--sLGBSQxPJBu8o3Gq6anvLIMUjhwLsctQtVkM53-EHNpPr2CATOZajMskEwLZJAcubO-YSoj-quE2VXXV81Xv9DbVOwhpSvkoXhZGrJ3ScWehk9TILgd9XC)
35. [louisville.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG9P73LnQfhQ_ylZkPkdqU0jBS9jOuO01tOFAYtdBqOM_tTas3lqCObmxNejpr69Fis6f34EYJuF3oodbPWzRC6bUBJ4mnbuoUPoZu7zPkyC6H0gJXD_Mt76zeDi24j62FF4Ve62WOQuPdhy5fiNgWqEHFxc3-Zbf_RW3v9AEaxS0Jl3Q==)
36. [tse-fr.eu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG911bG71-kJ4sP3ET71DvqrpCSMHkllard4DXuy1S6nbn1B9QcJjtUW6QqvIcO5XqETOdTZIL7b2fxnocSFg2GjYYo_dbDyykpmSyPEwN83H74xADhUklmzQ==)
37. [mdpi.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEpvmORR8eMwwvRvUVEMF-rotCrRkS-qHMi37ClcZgielVl4TComFc5So1E1m940c1eIsuC-0-ATZgYUFhMlMd9m-4SI8LZW_JnMcOz0LJT6Gob71OF74baSKS9GHI=)
38. [harvard.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHPxKsDr_5G7r8EYzjin5DWjiZB_J9zNN9qtYNEOgc5k2tZjkykEwcDMEgPDz4Au7GyQiygc6Mj1wZR3AXmCJHAkG30Do8WuXSOiWNFnpZdbN0LbN3Y9q6r-HZvPsbThgawuvYIed6BD7onmkfgpWOd9DI4)
39. [psu.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFU4C6dp9EChmYG7i2pY2-ZEmURZEN5ysDvrvs4sB39XVYAXV_GpMD_jmIXXiD-jCyzCtJxF9X_66QSUOkgEqQTG4M7alO9VuVCL39nncCejSniQE5kD9A6xxJKxwK3iQBzzHzl4jzxUrFSCVTXslOFAD4O_oNE75ngUg4CF7uyAn2PhNRJl7e12QOktyDPdeGdK0yG0kVG-w_LayY=)
40. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG8HIRedCaILf6NMq-vxA1QVgtKI4nmE_VVP0e_V0-YAdji5xGBYPGjPJKy32Sw72nny4gouVYPghAzmlrd7KNJPaDUFeC-4qOZpc9MVPPN22eYrJZdYnNFz1MP6hN-gLGuNNlxFKzUFisaKOZEeFA_aDqP2nIBcHPBrnnAtmnjDQCRN4QqYV3foGvgZKpuUWsgyc5c2rK7hAkTEEgqSpBWaos1zLK3nn5KVEiP8mTQzieOV5eC0bH8lM0x_2jcBA==)
41. [rpi.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEO11YHnlpjTJ9J7eoIVEqPBJdxcI0rGodvtBkW49EpL1eWwLNYp2SNU-sEpH_W-DRJd6dkl3TrdrxaFs001H9CjtPs0rpdTdp6L5_xNTV7AQfbDn6tWsFM5FTFJjLXCa6emZjmIojRmZYeq1lAoWdPTk0o_8dCnh9ChUbdy7UTtKRCr-x6ETp15uoZ5AHmgKPEfWg99KXlXxS8)
42. [taiwanpolitics.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEJJR727BGWxq5WZmo81vFAyG9l-97gPdyINN2NkSDAIy6Uqd83q6akb3_wDPNbk-S8Wx7CfNGn7SmXGXwgfylzQKGRTRfX9GiToy1o45Uh5dq7JFPwXPIBi7UIMwmxzTu0SglnNd4aCFegoq3E4EeKnMOQSvqJvrQHvMPe26NHMNyhcq10b5G3s4P5HKX1jTox2mUKoOzDRl_AfTnUibtgjlnAemXJHQNZ)
43. [harvard.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFGf8eCNV8On1ssHJLDQ19l8YFPqx8NCsXD4xUjnP3G2DdHZJZtLJDrZ624pzWCnqLrHifhgVtYkuaJReXuyjsvRsFWlu8YK5guuolGAxAm2WKss2fPF3soAYM8YwS7BkMs60nbt0cFZUzwiYlx6whZGPiRbyRnv-uJD7Vn6EU4t0g16UQCyQpeu-_Kd2Oqf79bBycBZYmjDeNYl6EwB6jxZN04ZmOl0QHYj9RL2ga3xtY375IsY1KwpdA=)
44. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE7xy0KDmALDOIJpFE6WPO6hN8CBG-_xHSbRG_dPqZprwYUA5Od-O9Z9JADP0sx7q8ZNRTCSOqca7iavXCJPgaj85wAM-6BEA1s2oSKSIu81hp_HDG46yzcGJP5h4_fk7FunCMvNdC9oNGzDE_Bv1ekPtLtO0LNtPWGi0XxrTDUV0-pj__hFxbscfJT3DISk2CnA9wUK-I-NTjYEWVRbw0y4wpxPo5-6w==)
45. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGmuqZEk7F7m2lRSnaVxv461BhHwe27CVDrEbzS5hQiOdT6LlhR7Hf1yXMlTWHtIbJCFlhwFVGa0AqLjxE40s5O5wGi2X5ILQOGqvcg3rWtWZmAzqg1fQYkiC6BSxELUykN97KSMAqU-3WctzeNPznNiyDIgXboW6kXZcVkjKzje-K-iSFprlWiuBTMcqs=)
46. [aaai.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH23LzLsIiVa0p2V8mvIe6Xh5qS5PtXeDAf4PJPrx60F_Cib8mBr8zu_2z_e7eLxpid6wHP6RUDWYazPd6yC244DWkw_D4hXS25ecbP0FK5iIP0OeObPmuSzDkkPZUqHrmc0sQArd90hj3dvUskqk1a1NL3zmi0hSpwwLnF)
47. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEjAv5n-3jqIoXZB4bSIHFTmzUtmBGgNNHNbiKh4i8GnJ5n7QXl65NVJuLLeo9TKX8KyfQx8CjSgeTehMa0kTJfCFXQCn5i7rVfh4CCTvg2nPkHKtXvVPd5)
48. [buffer.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGlXWMVrg-5wAJUR_pEJVDlaXylwu8vSy7Rq_9hadhyZf1rmSnQZetBLgSHaHgrtdOXyauga7Otm7I6kCm7bydIQPDThuqY7-_PDI2lSnD24mYjmJJgU6RyN755xG8DAmdkEXo21BpDeMafvMI=)
49. [socialbu.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHbiYgfE1P-boW68FLdWfhhcXe_4UgnUC3M0_60IB1u28d6533fC-xyuu5lX2dpQT8OQbGV7iNEvUE8eg2J9EDNBwYpcbP2LjhseeFngfqa6AGyh5dnrYH-cZu-h-Npf4ZZYYxrXXpdGqOgxc8=)
50. [turrboo.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG2Eulu5wxQMNozjJWD1jpkqKCXP_xiJktZOhiZkG7OrbQLm4MrhBtYa4FYgmZp1NNsZDZydBBKTxf5MXkNFPQlIQjjNmBUQUdCPpoeHWCp9TPz1CU19RnsAcBqtwPT0500rHbDNu44Z2y4gTjandM=)
51. [yale.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGynYouIhHKFVyqimgUuOicsBhE7OsKZtUL9rwBKzjOm6tW2m58DI86e2jEFAChMNXSzdnEwq8DTujvFpIlu51YlMqMBmAYgHBY--7hLwAxSFnVfHnoTGvrMSIONJsfkXCXX5St5HrlbuJVOInF4vsYlIB9AeepzqnO_MDz4pHNUAZZ2OTYAQRYPtSquxmxrZcxi1boEzHwyNU=)
52. [stanford.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHrbTTiUWo0_2AP8ydhwFmiMoSwOAq8hU5yV8_VSoCjRZnMi0vo9FDG2iFY7clp8rBk88poYeiuAYATuzXKHcv98LNTJLRAqAQcS6SRPPkYqOc_U0GLwD_TTLvdVv8ORfP9Icm08kYBPXAzczqNOUUiQcKt4BUIDWc-i1rewDqX_iuKMkKfLetY2KO35ltlhheKvKY=)
53. [mexc.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFcVJs2Z8h9HnwCAMKfAhbtEPbcQvLG80NMHJegeDneLVB8LHVowNiSgkk7vrTcbkIsbZueTdwVy4-pWq5B1oeoVbZAAyipCWYWqWLb0DRyec89YG_OMg==)
54. [nih.gov](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG-Z_ULstlG79PGXXTQ_dSfNIap6o2C_E46-bEFZwQTENoBXbe1BQ8m6KU2e3X_3jXRV75-NQjwMilyOb0DHeiAXflfWLSxKjVt0GBXk9GahGgR2UKV5ZtVFrk5ZVAxJlOcvqbx_hXx)
55. [humansintheloop.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHRkr59wdh18e6MdNd-7Fnh35ZMT5vwVpchiJLTTd6UvwdRVJ9Ry7Tu0cBhaq5VvMBPX2880pvH3jEoE68u_lM8A4QlhJGbMB4b87LwCu_ctm8caN779jt8YFrUC6f7W-6NAtekDy1af3M10Lv8bprVgDA5QaWQCRNZYuX_vgxm69Ajql2eXJiguVOkPB9ZZOKfexA=)
56. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE3dnqbtI4MvJmB-HKy2cmFVCm1xlMIhaZ42dAVeQVrd2hqlyBDbv6r19go4q3qSaPfVIEw84Z5lUryqHvLrntwd3WehrvHI1p1zMgIkoJEM8WyTKdVCoNuEIp2z0JERWwKjltSQOU1DKzSOXygFOAuPx6BIiC5ND6oJvPm5WMVRJEWrRWZE-fh1nF0bGhHYEVAByTYUuXrtrHMDWSvJQVIj0yrTCBxZg==)
57. [kaist.ac.kr](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGpYgafYS4FOdChA9ig6xmMadhE-SM-dYd1J1fvSO0sAYJIZDQTvuZI_YW9jLE8adA-2snRuSO6muzVSz9-7zGAqKwstSIDuno4rxlXG265JmyyQmoXrzkD-YIsBiY2EuONgj3AhPRYnfMeSZ4XvNyKrnNpaq6UImsY)