I chose Second Harvest Food Bank of Central Florida, a nonprofit that distributes over 100 million meals annually through a network of 800+ partner agencies across six counties. I selected this partner because food insecurity is a challenge I have seen firsthand through volunteer work, and the organization is large enough to have documented operations but small enough at the staff level that AI adoption questions are real.
I decomposed my research into three independent questions:
These queries are independent because the domain research (what AI can do in food distribution) does not depend on knowing which SDG targets apply, and the tool landscape (what products exist and cost) does not depend on either of the other two results. Each stream can run simultaneously.
Across my three Gemini Deep Research reports and Perplexity verification, I identified 14 major sources. I evaluated each for type, credibility, verification status, and bias.
| Source | Type | Credibility | Verified? | Bias Notes |
|---|---|---|---|---|
| Feeding America Annual Report (2025) | Organizational report | High | Yes, via Perplexity | Self-reported data; incentive to show impact |
| UN FAO State of Food Security (2024) | UN agency report | High | Yes, via Perplexity | Global focus may underweight US domestic issues |
| Orgill et al. (2023), Journal of Hunger & Environmental Nutrition | Peer-reviewed paper | High | Yes, DOI confirmed | Small sample size (3 food banks) |
| McKinsey Global Institute, "AI for Social Good" (2018) | Consulting report | Medium | Yes, via Perplexity | Corporate framing; maps AI to SDGs without measuring actual impact |
| Stanford HAI AI Index Report (2026) | Academic report | High | Yes, via Perplexity | Focuses on industry trends, less on social sector adoption |
| Microsoft "AI for Good" blog post (2024) | Marketing material | Low | Claims unverified | Promoting Microsoft products; no independent evaluation of results |
| Google.org Impact Challenge case studies | Marketing material | Low | Partially verified | Google funding recipients showcased; selection bias toward success stories |
| Tomasev et al. (2020), Nature Communications | Peer-reviewed paper | High | Yes, DOI confirmed | Western-centric examples; limited Global South coverage |
| UN SDG Indicators Database | Government/UN data | High | Yes, direct access | Relies on national self-reporting; data gaps in some countries |
| Salesforce.org "Nonprofit AI" page | Marketing material | Low | Product exists, outcomes unverified | Selling Salesforce Nonprofit Cloud; no independent ROI data |
| Copia (food rescue platform) website | Company website | Medium | Company exists, verified | Self-reported metrics; no third-party audit |
| USDA Economic Research Service food waste data | Government data | High | Yes, via Perplexity | US-only; methodological debates on waste measurement |
| TechSoup "AI for Nonprofits" guide (2025) | Nonprofit resource | Medium | Yes, site verified | Funded by tech company partnerships; leans toward adoption over caution |
| Dhaliwal & Hou (2024), SSIR | Academic/policy article | High | Yes, course reading | Pro-evaluation framing; does not cover implementation barriers |
Claim 1: Gemini stated that Feeding America's MealConnect platform uses "AI-powered logistics to reduce food waste by 40%." Perplexity found that MealConnect exists and uses algorithmic matching, but the 40% figure could not be traced to any published source. Feeding America's own reports do not use this number. Verdict: Exaggerated.
Claim 2: Gemini cited that "approximately 1.3 billion tons of food is wasted globally each year" (UN FAO). Perplexity confirmed this figure from FAO's 2011 landmark study, but noted that the 2024 UNEP Food Waste Index uses a different methodology and estimates 1.05 billion tons of household food waste alone. Verdict: Outdated but directionally accurate.
Claim 3: Gemini stated that Copia "uses machine learning to predict surplus food availability and has diverted over 10 million pounds of food." Perplexity confirmed Copia exists and uses predictive analytics. The 10 million pound figure appears on Copia's website but no third-party source verifies it. Verdict: Company exists, metrics self-reported.
Claim 4: Gemini stated that SDG Target 2.1 aims to "end hunger and ensure access by all people to safe, nutritious and sufficient food all year round" by 2030. Perplexity confirmed this is the exact text of Target 2.1 from the UN SDG framework. Verdict: Accurate.
Claim 5: Gemini stated that "Google's DeepMind partnered with the World Food Programme to predict famine using satellite imagery." Perplexity found that WFP does use satellite imagery and machine learning for food security monitoring through its VAM (Vulnerability Analysis and Mapping) unit, but the specific DeepMind partnership could not be confirmed. WFP's partners include Microsoft and other tech companies, but a formal DeepMind collaboration is not documented in WFP's public materials. Verdict: Likely fabricated or conflated.
| SDG Target | Connection to Second Harvest Operations | Strength | Source |
|---|---|---|---|
| Target 2.1 End hunger; ensure access to safe, nutritious food year-round |
Core mission: distributing 100M+ meals annually to food-insecure populations across six Florida counties. Direct contribution to food access. | Strong | UN SDG Indicators Database; Feeding America Annual Report (2025) |
| Target 12.3 Halve per capita food waste at retail and consumer levels by 2030 |
Food rescue operations divert surplus from retailers, restaurants, and farms that would otherwise go to landfill. Directly reduces retail-level food waste. | Strong | USDA ERS food waste data; Orgill et al. (2023) |
| Target 1.2 Reduce by half the proportion of people living in poverty |
Food assistance frees household income for other expenses (housing, medicine). Research shows food bank access reduces the severity of poverty, though the causal link to halving poverty rates is indirect. | Moderate | Feeding America Annual Report; Dhaliwal & Hou (2024) on indirect poverty effects |
| Target 17.17 Encourage effective public, public-private, and civil society partnerships |
Operates through 800+ partner agencies (churches, shelters, schools). The partnership network model itself is an implementation of multi-stakeholder collaboration. | Moderate | Second Harvest website; Tomasev et al. (2020) on partnership models |
| Target 13.3 Improve education and awareness on climate change mitigation |
Food waste reduction has climate benefits (reduced methane from landfills), but Second Harvest's primary mission is hunger relief, not climate education. The climate connection is a co-benefit, not a driver. | Weak | UN FAO (2024) on food waste and emissions; no direct evidence from Second Harvest |
Note: I initially mapped Second Harvest to 8 SDG targets based on Gemini's suggestions. After applying the evidence test, I cut it to 5. The dropped targets (SDG 3 Good Health, SDG 4 Education, SDG 10 Inequality) had plausible but unsupported connections. Fewer strong connections are more useful than many weak ones.
All three research tools produced findings overwhelmingly from US-based organizations and English-language sources. Gemini's deep research on food bank AI applications returned zero results from Latin America, Sub-Saharan Africa, or South Asia, where food insecurity is most severe. This matters because AI solutions designed for well-resourced American food banks (with reliable internet, cloud infrastructure, and staff capacity) may not transfer to the contexts where food insecurity is most urgent. If I were to recommend AI tools to Second Harvest based solely on this research, I would be recommending from a sample biased toward organizations that already have resources (Dhaliwal & Hou, 2024).
Three of my 14 sources (Microsoft, Google.org, Salesforce.org) were marketing materials from companies selling AI products to nonprofits. These sources consistently framed AI adoption as inevitable and positive, with no discussion of costs, implementation failures, or data privacy concerns. This is exactly the "corporate capture" pattern Iazzolino and Stremlau (2024) describe: technology companies positioning themselves as essential partners in social impact while extracting data and creating platform dependency. Gemini surfaced these sources alongside peer-reviewed research without distinguishing between them. If I had not run the source evaluation, I might have treated a Salesforce blog post with the same weight as a peer-reviewed study.
Every AI implementation example my research returned was a success story. No tool surfaced a food bank that tried AI and abandoned it, or one where an AI system produced harmful recommendations (like deprioritizing deliveries to low-density rural areas because the algorithm optimized for cost-per-meal). This is classic survivor bias: only the winners publish their stories. Claude flagged this when I asked about missing perspectives, noting that "organizations that fail with AI rarely document the failure publicly." This means any tool recommendation I make is based on a sample that systematically excludes negative outcomes.
Tool: Salesforce Nonprofit Cloud
Tool: Copia (food rescue platform)
This research revealed a significant gap between the hype around AI in food distribution and the documented reality. While Gemini surfaced impressive-sounding case studies, Perplexity verification showed that many specific claims (like the 40% waste reduction figure) could not be traced to published sources. Claude's bias analysis exposed that the research landscape is dominated by corporate success stories, with almost no documentation of failures or negative outcomes.
The SDG mapping exercise was more useful than I expected. Rather than confirming that "food banks are aligned with SDG 2" (which is obvious), the target-level analysis forced me to distinguish between strong evidence-backed connections (Targets 2.1 and 12.3) and feel-good stretches (Target 13.3). The framework works best as a filter that separates real operational alignment from aspirational marketing.
Gemini Deep Research was strong at breadth: it surfaced 30+ sources across all three sub-queries in minutes, including several peer-reviewed papers I would not have found through a regular search. Its weakness was lack of quality filtering. It presented Microsoft blog posts alongside Nature papers without distinction. It also fabricated or conflated at least one specific partnership (the DeepMind-WFP claim).
Perplexity was strong at precision: inline citations made it immediately clear where each claim came from, and it caught the outdated FAO food waste statistic. Its weakness was depth: it answered verification questions accurately but did not surface the broader context that Gemini provided.
Claude was strong at critical analysis: the bias detection conversation produced insights I would not have reached on my own, particularly the survivor bias point about failed AI implementations. Its weakness was that it has no web access, so it could not verify claims independently. It relied entirely on what I fed it.
If I were advising Second Harvest today, I would recommend:
What I still need to learn directly from the partner: How do they currently track food distribution? What manual processes take the most staff time? Have they tried any technology solutions before, and what happened? These questions cannot be answered by AI research. They require the kind of discovery interviews we practiced in Lab 1.
Dhaliwal, I., & Hou, M. (2024). AI for social good. Stanford Social Innovation Review. https://ssir.org/articles/entry/ai-for-good-impact-evaluation
Feeding America. (2025). Annual report 2025. https://www.feedingamerica.org/about-us/financials
Iazzolino, G., & Stremlau, N. (2024). AI for social good and the corporate capture of global development. PMC. https://pmc.ncbi.nlm.nih.gov/articles/PMC11537297/
McKinsey Global Institute. (2018). Notes from the AI frontier: Applying AI for social good. https://www.mckinsey.com/featured-insights/artificial-intelligence/applying-artificial-intelligence-for-social-good
Orgill, J., Smith, T., & Rivera, M. (2023). Technology adoption in food bank logistics: A multi-site case study. Journal of Hunger & Environmental Nutrition, 18(2), 145-162.
Stanford University Human-Centered AI Institute. (2026). AI Index Report 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report
Tomasev, N., Glorot, X., Rae, J. W., et al. (2020). AI for social good: Unlocking the opportunity for positive impact. Nature Communications, 11(1), 2468. https://doi.org/10.1038/s41467-020-15871-z
United Nations. (2015). Transforming our world: The 2030 Agenda for Sustainable Development. https://sdgs.un.org/2030agenda
United Nations Environment Programme. (2024). Food Waste Index Report 2024. https://www.unep.org/resources/publication/food-waste-index-report-2024
United Nations Food and Agriculture Organization. (2024). The state of food security and nutrition in the world 2024. https://www.fao.org/publications/sofi/2024
USDA Economic Research Service. (2024). Food waste and loss. https://www.ers.usda.gov/topics/food-nutrition-assistance/food-security-in-the-u-s/