Blog

The Daily Pennsylvanian investigated the impact of federal NIH grant terminations on Penn researchers, including Sharath Guntuku whose grant studying how online information ecosystems impact health behaviors was cut three years in with no prior notification. The piece highlights the broader consequences for public health research infrastructure at Penn. Featured: Sharath Guntuku.

Penn researchers developed “evaluator agents” — specialized AI systems that test large language models for potential cyberbullying behavior by generating nuanced, demographically diverse prompts. The study uncovered reasoning blind spots in some leading language models that could lead to harmful cyberbullying-adjacent outputs, with implications for responsible AI deployment. Featured: Shreya Havaldar, Eric Wong, Lyle Ungar.

Analysis of 1.5 million Yelp reviews found that positive patient sentiment dropped sharply after COVID-19, with disparities emerging along racial and geographic lines — areas with higher Black populations reported more billing complaints, while predominantly white areas cited wait times. The findings highlight how digital review data can surface systemic inequities in healthcare quality. Featured: Neil Sehgal, Sharath Guntuku, Anish Agarwal, Raina Merchant.

AWS awarded $700,000 to Penn Engineering’s ASSET Center to support 10 PhD students researching safe and trustworthy AI, with Young Min (Jeffrey) Cho named as one of the recipients for his work on LLM-based conversational agents for mental health support. The funding supports research on fairness, explainability, privacy, and human-AI interaction. Featured: Young Min Cho, Lyle Ungar.

Penn Today covered the PNAS study on racial disparities in AI depression detection, explaining how perceived stigma may cause Black individuals to express depression differently on social media — and how standard AI models fail to account for this. The findings underscore the need for more inclusive training data and culturally aware model development. Featured: Sharath Guntuku, Sunny Rai, Lyle Ungar.

Reuters covered a PNAS study finding that AI depression-detection models are more than three times less predictive for Black Americans than for white Americans on Facebook, exposing a critical racial disparity in AI mental health tools. The research showed that linguistic markers previously associated with depression — such as first-person pronouns and self-deprecating language — do not generalize across racial groups, calling for culturally tailored AI models. Featured: Sharath Guntuku, Sunny Rai, Lyle Ungar.

Machine learning analysis of emergency physicians’ Twitter posts from 2018–2022 found marked increases in loneliness, anxiety, anger, and depression during COVID-19. The study, published in JAMA Network Open, suggested that real-time social media monitoring could serve as an early warning system for physician burnout. Featured: Sharath Guntuku, Anish Agarwal.

An in-depth interview exploring AI and positive technology in which Ungar defines positive technology as tools fostering human connection and praises AI’s potential to democratize education. He raises three key concerns: misinformation eroding societal trust, synthetic companions replacing real relationships, and large-scale job displacement — urging mindful use of technology and greater awareness of language’s emotional impact. Featured: Lyle Ungar.

The Computational Social Listening Lab offers cell phone data collection services for digital phenotyping and personal sensing research projects using the open-source AWARE framework.

Penn researchers discussed their NIH-funded research on how social media accelerated the spread of COVID-19 misinformation, outlining plans to use machine learning to analyze misinformation patterns and develop community-targeted counter-messaging strategies. The $3.8M NIH grant supports a multi-year effort to understand and disrupt health misinformation ecosystems. Featured: Sharath Guntuku.

Researchers analyzed 3.4 million Facebook posts and found that depression and loneliness have distinct linguistic signatures — depression centers on emotional expression while loneliness focuses on cognitive processes — suggesting loneliness may stem from maladaptive social perception rather than isolation alone. The findings could help clinicians and AI tools distinguish between the two conditions. Featured: Sharath Guntuku.

A recap of Ungar’s AAAS Leshner Fellowship activities — including the NYT mask op-ed, free machine learning ethics courses through NeuroMatch Academy, informal advisory conversations with major tech platforms on content moderation, and Capitol Hill visits advocating for AI research funding. Ungar called for deeper investment in understanding the socio-political contexts of technology, especially amid U.S.–China AI competition. Featured: Lyle Ungar.

A PNAS study found that George Floyd’s killing caused unprecedented emotional distress across the US, with depression screening rates jumping nearly 30% among Black respondents — approximately 900,000 additional Black individuals screening positive — compared to 1.2% among white respondents. The research used Gallup surveys and Census data to quantify the disproportionate psychological toll. Featured: Sharath Guntuku, Lyle Ungar, Johannes Eichstaedt.

Analysis of 78 million vaccine-related tweets revealed that distinct US community types — from African American South communities to evangelical hubs to urban suburbs — had strikingly different vaccine concerns correlated with actual vaccination rates. The findings offer a roadmap for designing community-specific public health messaging around vaccine hesitancy. Featured: Sharath Guntuku.

Penn researchers found that county-level unhappiness was a stronger predictor of the 2016 Trump vote than income, employment, race, religion, or traditional demographics — based on Gallup poll data from over 2 million respondents. The research suggests that dissatisfied communities mobilize more strongly behind anti-establishment candidates, with implications for how public health and economic policy intersect with political behavior. Featured: Lyle Ungar, Johannes Eichstaedt.

An in-depth AAAS member spotlight on Ungar’s career — from heart disease prediction via Twitter to real-time COVID symptom tracking that surfaced signals “before it showed up in the New York Times.” He discusses goals of helping public health officials craft better interventions and using language analysis to predict addiction relapse risk. Featured: Lyle Ungar.

Lyle Ungar co-authored a New York Times opinion piece with Angela Duckworth and Ezekiel Emanuel arguing that effective mask adoption requires making mask-wearing “easy,” “understood,” and “expected” — citing Hong Kong’s near-universal adoption versus New York City’s early struggles. The op-ed drew on behavioral science principles and Penn’s concurrent COVID Twitter tracking data. Featured: Lyle Ungar.

Penn’s World Well-Being Project analyzed 1.5 billion tweets to predict county-level well-being, launching a real-time COVID-19 wellness map showing dramatically higher stress, anxiety, and loneliness compared to prior years. A counterintuitive finding emerged: counties tweeting more words like “love” were often less satisfied, as people frequently use positive language to describe what they lack. Featured: Lyle Ungar.

Lyle Ungar was named a 2020–21 AAAS Alan I. Leshner Leadership Institute Public Engagement Fellow in the AI cohort, recognizing his work applying machine learning and text mining to social media for physical and mental health research. As a fellow, Ungar planned to foster national conversations about how AI transforms work, human identity, and well-being. Featured: Lyle Ungar.

The Philadelphia Inquirer reported on Penn research analyzing ~400 million tweets from Pennsylvania users to identify linguistic markers of loneliness. People expressing loneliness also frequently posted about substance use and relationship difficulty, suggesting that social platforms could enable earlier mental health interventions. Featured: Sharath Guntuku.

The Atlantic examined how social media language — in tweets and Facebook posts — can reveal emotional states and predict mental health conditions. Penn research on computational approaches to detecting mood disorders through digital footprints was featured as part of a broader investigation into the future of AI-based mental health tools. Featured: Sharath Guntuku.

Penn researchers used machine learning on nearly one billion social media posts from the US, UK, Canada, Japan, and China to study how emoji usage differs across cultures. Western users favored negative facial expression emojis when discussing health, while Eastern users chose healthcare object emojis — findings with implications for cross-cultural health communication. Featured: Sharath Guntuku, Lyle Ungar.

NBC News profiled emerging technologies — including wearables and apps — that track emotional biomarkers to predict mental health crises before they happen. Penn research on predicting mood through social media language and sensor data was featured as part of this broader look at computational psychiatry. Featured: Sharath Guntuku.

WIRED covered Penn research using Facebook post data and machine learning to identify signals of mental illness, exploring both the promise of AI-based mental health screening and the ethical questions around using social media data for clinical purposes. The piece highlighted how language patterns in posts can serve as early indicators of depression and other conditions. Featured: Sharath Guntuku.

Penn and Stony Brook researchers showed that Facebook language — including hostile content, expressions of loneliness, and elevated first-person pronouns — can predict depression up to three months before a formal medical diagnosis, matching the accuracy of clinical screening tools. The study analyzed 524,000+ Facebook posts from ~700 emergency department patients who also shared their medical records. Featured: Lyle Ungar.

Sharath Guntuku appeared on the APA’s flagship Speaking of Psychology podcast to discuss Penn research analyzing over a million tweets from individuals with ADHD. He described how ADHD users post more at night, use more anxious language, and show distinct mood fluctuations — and how these patterns could inform clinically-aligned AI tools. Featured: Sharath Guntuku.

US News & World Report covered Penn research analyzing 1.3 million tweets from 1,400 users with self-reported ADHD diagnoses. The study found distinctive patterns including late-night posting, negative self-talk, and discussions of substance use — insights that could help personalize mental health interventions. Featured: Sharath Guntuku, Lyle Ungar.

Healio’s psychiatry publication covered the Penn study on ADHD behavioral patterns in Twitter data, reporting that users with ADHD show distinct emotional and behavioral signals including nighttime posting and negative emotion language. Guntuku discussed how social media data could supplement clinical diagnosis and enable real-world mental health monitoring. Featured: Sharath Guntuku.

A profile of the World Well-Being Project describing how Penn researchers analyze Facebook and Twitter data to assess mental wellness at community and individual levels, with applications for government partnerships and healthcare. Researchers describe the well-being map and emphasize that social media analysis provides “huge samples” at low cost without intrusive data collection. Featured: Lyle Ungar.

Penn researchers unveiled EmoNet at the ACL conference — a deep learning system trained on 1.6 million tweets and labeled with 24 emotion categories, achieving 87% accuracy matching human performance. Potential applications range from emotionally-aware healthcare tools that understand patient sentiment to more sophisticated customer service chatbots. Featured: Lyle Ungar.

Penn’s World Well-Being Project launched a freely available interactive map showing county-level well-being, personality traits, and health indicators drawn from 37 billion geo-tagged tweets across the US. The five-year project involved ~20 Penn scientists and opens a new way to compare traits across all US counties — with applications in public health, policy, and community planning. Featured: Lyle Ungar.

An in-depth interview covering Penn research analyzing social media language to understand personality, well-being, and health — from predicting heart disease via angry tweets to potential early warnings for addiction relapse. Ungar discusses the practical applications of the “quantified self” and how community-level social media insights could inform public policy. Featured: Lyle Ungar.

A Penn study published in Psychological Science found that Twitter language patterns can predict county-level coronary heart disease mortality rates more accurately than traditional risk factors like smoking, income, or diabetes. Expressions of anger, hostility, and stress in tweets were linked to higher heart disease risk, while optimism was protective — demonstrating that community-level language reflects population health. Featured: Lyle Ungar.

Penn researchers enrolled 75,000 Facebook users in a personality study, using the open vocabulary of their status updates to reveal personality traits — predicting gender 92% of the time and age within three years. The “open vocabulary” approach let social media data itself surface psychological patterns rather than relying on pre-selected word lists, laying groundwork for large-scale computational psychology. Featured: Lyle Ungar.