High-Impact Growth — “The Thinking You Can't Skip: AI for Good at a Crossroads”

Date: November 13, 2025
Host: Amy Vaccaro (A), Jonathan Jackson (C)
Guest: Brian Derenzi (B), VP of AI & Research, Dimagi

Overview

This episode of High-Impact Growth is the fourth in an ongoing series unpacking the state and future of AI for Good, especially as it relates to global health, digital solutions, and technology’s role in social impact. Dimagi CEO Jonathan Jackson and VP of AI & Research Brian Derenzi join host Amy Vaccaro for a candid, detailed exploration of rapid technological advancements, the practical and ethical complexities facing the sector, and Dimagi’s latest research into hidden bias and language barriers in large AI models.

The conversation revolves around the central notion: AI accelerates human intention—but it can’t replace critical thinking. The team muses on the evolving role of AI in daily workflows, the importance of "getting your hands dirty" with these tools, the risk of unchecked deployment (“pilotitis”), and the significant gaps (and potential) in AI for Good caused by a surge in commercial investment.

Key Discussion Points & Insights

1. How Are Experts Feeling About AI Right Now? ([02:13])

• Brian Derenzi: AI remains “intellectually stimulating in like a dopamine hit kind of way”, yet Brian retains concerns about dystopian outcomes, feeling both optimism and anxiety over AI’s speed of development:

  “You know, thinking about ... all the positive upsides and then also recognizing ... the potential for dystopia and trying to actively work against that. ... But it is wild how much more capable things are ... just even in the last 12 or 18 months.” (B, 02:13)

• Jonathan Jackson: Personal AI breakthroughs—from coding alongside his kids to shifting his professional expectations—have “fundamentally changed” what’s possible:

  “Shortening the distance between the people who are envisioning what they want to build and the people who are building ... just feels like you can close that gap so much more than was feasible pre-AI.” (C, 03:39)

2. AI as a Tool: Acceleration, Not Replacement ([05:51])

• Brian’s Workflow: AI serves as “an electric bike”—flattening steep hills, making difficult tasks approachable—but “you can't skip the thinking part”:

  “At the end of the day ... what I'm sharing with my colleagues is not something that came out of the machine. The machine just helped me get to that faster...” (B, 05:51)

• Iterative Use: AI output is always a draft—Brian heavily revises AI-generated content after brainstorming, using tools like Whisper and model feedback, then applies critical review.

3. The Importance and Risk of Skipping Critical Thinking ([09:20], [11:31])

• Jonathan: If AI replaces critical thinking, organizational and social development could stagnate:

  “Where's the next senior cohort coming from if that's the world we move into? ... as AIs continue to get better and continue to replace more and more of that critical thinking skill set, it is really challenging...” (C, 09:20)

• Brian: AI can defer, but never eliminate, the need for human judgment:

  “If you generate a bunch of AI stuff ... you're just like kicking the can down the road. ... eventually somebody's going to have to look at the thing.” (B, 11:46)

• Hands-on Review Mandate:

  “You have to look at the data. ... you have to be like deeply involved. ... Like they spent a ton of time ... looking at data and understanding the ways that things are failing.” (B, 11:46)

4. Defining “Good” in AI for Good ([14:46])

• Evaluating Value:
  Is a bot that deflects a community health worker’s personal questions successful? Is engagement or outcome more important?

  “Is the goal that the CHW comes back to our AI-enabled chatbot over time because she likes it, or ... that we made her more money at some point? ... really difficult questions ... with very complex answers.” (C, 14:46)

• Manual Transcripts Review:
  Jonathan and Brian insist on leaders and teams directly reading AI-bot transcripts to develop “texture” and intuition for what’s happening in interactions.

5. The State and Trajectory of AI for Good ([18:29], [21:49])

• Still Unsettled:

  “I think I would say it's still finding its footing ... People don't know where the low hanging fruit is.” (B, 18:29)

• Challenges for Low-resource Languages:
  AI development remains hindered by lack of sufficient training data (“need a billion tokens”—impractical for many languages).

• Dimagi’s Priorities:
  Focused on more “forgiving”, lower-risk AI support and training (vs. high-risk clinical AI). Building bots for support, sentiment analysis, and frontline worker assistance appear as promising use-cases.

• “Operational AI”:
  Boring, middle-management streamlining (e.g., spreadsheet creation, support workflows) has huge potential but is underexplored in global development.

6. Direct Applications: Community Health Worker (CHW) Support ([25:07])

• CHW Coaching Bots:
  Used for training, logistics, and direct question support in interventions like Kangaroo Mother Care.

  “You can imagine a lot of questions a frontline worker might just have on how to do the intervention ... that's a great use case for AI.” (C, 25:07)

• Scaling Challenge:
  Real-world context means workers can’t juggle “50 chatbots”—the team must weigh specificity vs. generalizability.

• Internal Platform Building:
  Dimagi’s Open Chat Studio was the only tool that “did evaluations, documentation, and fine-tuning” in one, justifying building their own despite the rapidly expanding AI tool market.

7. The Nuance and Richness of User Interaction ([29:07], [33:19])

• Real-world Example:
  Review of actual transcripts from a family planning AI bot in Kenya showed users start with casual, even off-topic questions (“what do you think of my country?”) before building trust to ask sensitive, personal questions.

  • Quote:

    “At some point got like, quite personal. ... all of our safety tools worked correctly, so things were kind of escalated up so that real humans could take a look …” (B, 29:07)

• Cultural Context:
  The path to trust and meaningful interaction with an AI often mirrors cultural patterns of conversation, requiring time and patience (“preamble” before directness).

8. Dimagi’s Research: Bias and Language in LLMs ([36:03], [40:39])

• Bias in AI Models:
  Examination of story generation by large models revealed quietly entrenched stereotypes not found in lived Kenyan culture (e.g., shopkeeper almost always Kikuyu, thief usually Luo).

  “The work that we're doing is trying to explore this and surface this ... how the frontier models have kind of absorbed and codified ... biases and stereotypes into the models themselves.” (B, 36:03)

  • Output to be shared widely as open knowledge for the field.

• Low-resource Language Performance:
  Dimagi is benchmarking generation across languages, relying on human review for accuracy. Results vary widely across models and between model versions (e.g., GPT-5 worse than GPT-4.1 in Swahili, though better elsewhere).

  “There's a big difference between the different frontier models ... at least in our work, it appears that GPT5 is worse than GPT 4.1 in Swahili ...” (B, 40:39)

  • This landscape is constantly shifting with every model release—ongoing monitoring is critical.

9. Investment: Commercial Boom, AI for Good Lag ([44:40])

• Commercial AI Investment at an All-Time High:

  “A billion dollars a day of venture capital is going into AI companies right now.” (C, 44:40)

  • Most of this is not trickling into AI for Good. The gap between commercial and social application is widening rapidly.

• Upside:
  The pace of commercial innovation means nonprofits and social sector organizations can benefit as models improve, but “slow and steady” investment in AI for Good is still necessary for real social progress.

Notable Quotes & Memorable Moments

• Brian on critical thinking in AI:

  "I've not seen versions where you can kind of skip the thinking piece." (B, 05:51)

• Jonathan on generational shifts:

  "Where’s the next senior cohort coming from if that's the world we move into?" (C, 09:20)

• Brian on reviewing AI output:

  "You have to look at the data... you have to be deeply involved... that's how you develop some intuition and get some sense of what’s going on...” (B, 11:46)

• Jonathan on what success looks like:

  "Is the goal that the CHW comes back to our AI-enabled chatbot over time because she likes it, or that we made her more money at some point?" (C, 14:46)

• Brian on entrenched bias in LLMs:

  "Ninety percent of the time the shopkeeper was Kikuyu … 80% of the time the thief was Luo. ... That’s certainly not a stereotype that I’m familiar with within Kenya..." (B, 36:03)

Timestamps for Important Segments

• 02:13 — Brian & Jonathan on their evolving personal and professional relationships with AI
• 05:51 — Brian describes his workflow and the necessity of human thinking
• 09:20 — Jonathan warns about the risks of skipping critical thinking, especially for next-generation talent
• 14:46 — Discussion of how to define "success" and "value" in AI interactions for social good
• 18:29 — AI for Good’s uncertain landscape and real-world limits
• 25:07 — Concrete AI use cases for supporting community health workers (CHWs)
• 29:07 — Reviewing real user interaction transcripts; the cultural context of trust-building
• 36:03 — Early findings on bias in large language models
• 40:39 — Challenges for LLMs with low-resource languages; benchmarking process and findings
• 44:40 — The investment gap: How surges in commercial AI are (and aren't) benefiting social good
• 49:56 — Final advice for listeners: Build for scale, align incentives, and maintain deep, ongoing, hands-on engagement with all AI work

Key Takeaways for Social Impact Leaders

1. AI’s True Value: It accelerates human intention; it cannot replace thinking, synthesis, or the iterative judgment that underpins high-impact solutions.
2. Own the Review: Leaders must stay close to the data, reading transcripts and evaluating output, to maintain strategic and ethical oversight.
3. Avoid “Pilotitis”: Build AI projects with the intent to scale, not as one-off pilots—this rigor forces better design and value delivery.
4. Be Hands-on: Play with AI tools, break things, and observe closely. Real understanding comes only through use.
5. Mind the Investment Gap: The explosion of investment in AI isn’t enough—deliberate, sustained efforts and advocacy for resource allocation in AI for Good are needed.
6. Bias and Language Gaps Remain Critical: Ongoing analysis of models for hidden biases and performance in low-resource languages is essential to ensure equitable, culturally sensitive AI systems.

Final Advice

"Get into the data and actually look at what's going on … That's the only way that you’re really going to get a sense for what’s going on." (B, 50:53)
"We should learn from the pilotitis we had … it really does produce better work when your hope is that you leave it on if it works." (C, 49:56)

For further exploration, visit https://dimagi.com/podcast/.