How social workers use AI to help unhoused teens | Anamika Barman-Adhikari | TEDxMileHigh
A social work scholar proposes using AI and machine learning to improve interventions for vulnerable youth, arguing that custom grouping algorithms can reduce substance use by 40-70% compared to random assignments. She demonstrates this using a social network map of 160 homeless youth in Los Angeles, where an algorithm maximized positive friendships while minimizing substance users. The technology is presented as an empowering tool for social workers, not a replacement for their intuition or empathy.
## Speakers & Context
- Unnamed social work scholar.
- Has spent the last seven years developing algorithms with a computer scientist/technologist to help vulnerable youth, specifically those experiencing homelessness.
## Theses & Positions
- Artificial intelligence (AI) can provide solutions to major societal problems, including homelessness, poverty, and addiction.
- Pure peer support groups (for 6 to 8 weeks) are not effective for vulnerable populations because negative behaviors can be reinforced, and random group assignment is suboptimal.
- AI can mathematically optimize group formation by maximizing existing positive friendships while minimizing the presence of substance users.
- Technology should function as a tool to *empower* social workers to better utilize their limited time and resources, rather than replacing their intuition or empathy.
## Concepts & Definitions
- **Substance use:** The specific challenge area studied, noted as an "epidemic" among homeless youth.
- **Pure peer groups:** Groups where young people meet for 6 to 8 weeks to discuss problems and gain support from peers.
- **Developmentally appropriate:** Groups designed to match the inherent questioning and resisting nature of adolescents.
- **Social network map:** A visualization showing individuals (dots) and the connections between them (lines).
- **Systemic Racism:** A risk associated with unchecked AI deployment.
## Mechanisms & Processes
- **The problem with random grouping:** If highly connected youth use meth, random grouping allows negative behaviors to reinforce each other, with non-meth users following suit.
- **The AI solution:** The algorithm iteratively tests 30 to 50 group configurations to mathematically achieve two goals:
1. Maximize existing positive friendships within groups.
2. Minimize the number of substance users within groups.
- **Advanced prediction:** The algorithm also accounts for how relationships will *change* during the intervention, predicting the development of positive bonds and the subsidence of negative ones.
- **Predictive Analytics (General):** Used to screen for and triage serious child abuse cases in Allegheny County.
## Timeline & Sequence
- **Last seven years:** Duration of the speaker’s partnership developing the algorithms.
- **6 to 8 weeks:** Duration of the peer support group intervention.
- **Current/Recent:** Time when the speaker was able to obtain permission to use social network data from young people experiencing homelessness in Los Angeles, California.
## Named Entities
- **Allegheny County** — Location where social workers are using predictive analytics.
- **USC's Center for Artificial Intelligence and Society** — Institution partnered with for the research.
- **Los Angeles, California** — Location where the social network data was collected.
- **New York** — Location where researchers are studying tweets for violence prediction.
## Numbers & Data
- **7** years: Duration of work with the computer scientist/technologist.
- **50%:** Percentage of all young people experiencing homelessness who use some illicit substance (compared to < 5–10% for non-homeless youth).
- **6 to 8** weeks: Duration of peer support groups.
- **160:** Number of young people experiencing homelessness included in the social network map.
- **40 to 70%:** Reduction in substance use achieved with AI-generated groups compared to random allocation.
- **30%:** Accuracy rate of traditional substance use prediction methods.
- **80%:** Accuracy rate of the algorithm using Facebook posts to predict substance use.
## Examples & Cases
- **Homelessness illustration:** Depicted via a picture of a 19-year-old girl in Denver sleeping on a dirty stained mattress, with condoms and heroin needles on the floor.
- **The initial failure point:** The speaker's personal anecdote of smoking a banana peel because friends said it would get her high.
- **The social network map visualization:** Illustrates how highly popular nodes (individuals) draw others to them, potentially spreading negative behaviors.
- **Predictive modeling:** Developing an algorithm using young people’s Facebook posts to predict likelihood of substance use disorders.
## Tools, Tech & Products
- **Machine learning algorithms:** The core technology used to analyze complex data patterns.
- **Predictive analytics algorithm:** Specifically used in Allegheny County for child abuse screening.
- **Social network map:** The visualization tool used to represent relationships among the 160 youth.
- **AI/ML tools:** Used to run iterative testing on group configurations.
## References Cited
- *Schitt's Creek*, *The Office* — Examples of content governed by recommendation algorithms (Netflix).
## Trade-offs & Alternatives
- **Pure peer groups:** Low operational cost but high risk of negative reinforcement.
- **Random group assignment:** Simpler for social workers but ignores natural social influence.
- **Traditional prediction:** Predicts substance use accurately only about 30% of the time.
- **AI-optimized grouping:** Superior prediction accuracy (up to 80%) by accounting for network dynamics.
## Counterarguments & Caveats
- **Emotional/Intuitive vs. Machine:** Social services are understaffed and overworked, making manual gas calculations impossible for large groups.
- **Automation Risk:** Algorithms are not perfect and can exacerbate bias and systemic racism if misused.
## Methodology
- **Data input:** Social relationships, substance use behaviors, and other relevant behaviors of the youth.
- **Algorithm function:** Iterative testing of 30 to 50 group configurations to find the best balance between network stability and intervention goals.
- **Approach:** Comparison between AI-configured groups vs. randomly allocated groups.
## Conclusions & Recommendations
- **Goal:** To provide social workers with a powerful tool to improve service efficiency and reach more youth with limited resources.
- **Best outcome:** Utilizing AI to allow social workers to focus their time on the most critical cases, intervening where they will have the greatest impact.
- **Final appeal:** Hope that technology is used to enhance human care, not replace it.
## Implications & Consequences
- **Systemic Improvement:** AI can fundamentally improve intervention protocols for public health issues (homelessness, substance abuse).
- **Resource Allocation:** Increased prediction accuracy allows social services to prioritize caseloads, ensuring limited resources go to those who need help most urgently.
## Verbatim Moments
- *"Machine learning algorithms are the reason why a horror movie will never cross my Netflix screen."*
- *"Substance use is almost an epidemic among this group of young people."*
- *"They often use substances to cope with their trauma, to feel comfortable in an uncomfortable world."*
- *"Adolescents are almost biologically hardwired to question, resist and doubt what adults tell them to do."*
- *"The youth who are very popular and well-connected, find themselves at the centre of this network."*
- *"Meth use will probably go up instead of down."*
- *"The algorithm found a way of mathematically maximizing the number of existing friendships in each group while minimizing the number of substance users."*
- *"Compared to traditional methods that can predict substance use accurately, maybe 30% of the time, this algorithm can predict substance use almost 80% of the time."*
- *"This is not about machines replacing social workers."*