# Machine Learning Practices Outside Big Tech: How Resource Constraints Challenge Responsible Development

**Authors:** Aspen Hopkins, Serena Booth  
**arXiv:** [2110.02932](https://arxiv.org/abs/2110.02932)  
**Date:** 2021-10-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2110.02932v1/

## Key points

- Machine learning (ML) practitioners from various backgrounds increasingly use ML methods
- Studies often focus on Big Tech and academia, neglecting startups, non-tech companies, and the public sector
- Selection bias excludes broader, lesser-resourced ML community facing challenges in deploying ML with limited resources and increased existential risk
- Tensions identified in qualitative analysis of 17 interviews: privacy vs. ubiquity, resource management vs. performance optimization, access vs. monopolization
- Challenges for responsible ML development in resource-constrained organizations include company expectations, bias, explainability, data literacy, model lifecycles, and privacy concerns
- Mixed opinions on value of implementing ML technology among underrepresented organizations highlight need for further study to address challenges
- Explainable AI community seeks meaningful mechanisms for understanding ML models; some advocate transparency through example testing
- Desires for explanations vary among practitioners with concerns about uncertainty communication and overconfidence
- Onboarding processes for ML products/tools should focus on continual evolution based on user interactions with the model to ensure effective human-AI teaming
- Holistic understanding of limitations in responsible ML development is essential to guide future research agendas benefiting all stakeholders in the machine learning community

## Summary

In the realm of machine learning (ML), practitioners from various backgrounds are increasingly utilizing ML methods. However, studies often focus on populations from Big Tech and academia, neglecting those in startups, non-tech companies, and the public sector. This selection bias excludes the broader, lesser-resourced ML community facing challenges in deploying ML with limited resources and increased existential risk. A qualitative analysis of 17 interviews with stakeholders from underrepresented organizations reveals tensions between privacy and ubiquity, resource management and performance optimization, and access and monopolization. The challenges of responsible ML development for resource-constrained organizations include company expectations, bias, explainability, data literacy, model lifecycles, and privacy concerns. Despite mixed opinions on the value of implementing ML technology among these organizations, there is a need for further study to address these challenges. The explainable AI community seeks meaningful mechanisms for understanding ML models while some practitioners advocate for transparency through example testing. However, desires for explanations vary among practitioners with concerns about uncertainty communication and overconfidence. Onboarding processes for ML products and tools remain underexplored but should focus on continual evolution based on user interactions with the model to ensure effective human-AI teaming. Overall,a more holistic understanding of limitations in responsible ML development is essential to guide future research agendas that benefit all stakeholders in the machine learning community.