AI for Good

1. AI's Real Power Lies in Practical, Incremental Solutions.

Try to imagine the tech without the tech companies.

Beyond the hype. The true potential of AI isn't in the "awesome future" promised by accelerationists or the "human extinction" scenarios of doomers, but in its practical, often unglamorous, applications. General Gustave Perna's success with Operation Warp Speed demonstrated that AI could solve complex logistical problems by providing a "God view" of distributed data, enabling swift decision-making during a crisis. This real-world application, far from the jargon-filled industry, sparked the author's "AI awakening."

Iterative development. Khan Academy's journey with Khanmigo exemplified this incremental approach. Instead of aiming for a perfect, all-encompassing solution from day one, they focused on specific problems like helping students get "unstuck" during practice. This involved constant iteration, testing, and tweaking of prompts and features, demonstrating that effective AI solutions are built through continuous refinement rather than a single, grand design.

Quiet wins. Even in municipal government, small, focused AI applications can yield significant results. East Lansing, Michigan's recycling program used AI image recognition to identify contaminants and send targeted "emotional nudge" postcards to residents. This low-profile, practical application led to a 23% reduction in contamination and a 45% increase in participation, proving that impactful AI doesn't always require massive budgets or public fanfare.

2. Human-Centered AI Amplifies, Not Replaces, Expertise.

AI does not need to be perfect to be useful.

Empowering educators. AI tools like Khanmigo are not designed to replace teachers but to enhance their capabilities. Teachers like Alaina Richter in Hobart, Indiana, found that by treating Khanmigo as a way to affirm student progress rather than a primary teaching tool, they could differentiate instruction more effectively. Melissa Higgason used Khanmigo's "lesson hook" to transform a boring chemistry lecture into an engaging, hands-on lab, making her class "more human" by freeing her to focus on student interaction.

Supporting clinicians. In healthcare, AI acts as a powerful assistant, reducing cognitive burden and improving patient care. Cleveland Clinic's pilot of ambient scribe software, led by Eric Boose, showed that AI could automate note-taking, allowing doctors to increase face time with patients and close charts faster. This amplification of human effort led to a physician delaying retirement, calling the technology "life-changing."

Optimizing operations. Rita Pappas, Cleveland Clinic's medical director of hospital operations, leveraged Palantir's Hospital 360 to gain a real-time "God view" of admissions, discharges, and transfers. This AI-powered dashboard didn't replace her team but empowered them to streamline processes, reduce emergency room wait times by ninety minutes, and increase daily transfer volume by 10%, demonstrating AI's role in making complex systems more efficient for human oversight.

3. Bureaucracy, Not Technology, Is AI's Toughest Obstacle.

Every single thing I just told you is illegal in the federal government.

Inertia in crisis. Operation Warp Speed, despite its urgency, faced immense bureaucratic resistance. Deacon Maddox, General Perna's logistics expert, encountered federal agencies like the FDA refusing to share crucial data, prioritizing their prerogatives over the mission. This forced an agile, "embrace the chaos" approach, where Palantir's Aaron Jaffe had to build a functional system by "stringing together four singles" rather than waiting for a perfect, integrated solution.

"Frozen middle." Eric Schmidt, former Google CEO, observed that the federal government's acquisition rules and processes effectively make modern software development "illegal." The "Defense Innovation Board" identified a "frozen middle" of public servants, not lacking patriotism, but incentivized to comply with absurd rules rather than embrace innovation. This systemic inertia, rather than technological limitations, remains a profound barrier to change.

Political sabotage. The IRS, under Commissioner Danny Werfel, navigated a landscape where its budget was slashed, and its modernization efforts were often met with political disdain. The Department of Government Efficiency (DOGE), established by President Trump, exemplified this by attempting to "break things" through mass layoffs and freezing modernization programs, demonstrating how political agendas can actively undermine AI's potential for public good.

4. Data is AI's Fuel, But Human Empathy is Its Compass.

The model optimizes for fluency, not truth. It does not fact-check itself because it does not understand facts.

Fluency over truth. Large language models like ChatGPT are "giant prediction engines" that optimize for fluency, not truth. This means they can confidently invent facts or make mathematical errors, highlighting that raw data processing alone is insufficient. Khan Academy's "red teaming" revealed that without human guidance and ethical guardrails, AI could be easily misled or generate inappropriate content, underscoring the need for human oversight to ensure accuracy and safety.

Contextual understanding. In sepsis prediction, Bayesian Health's AI model, while powerful, struggled with the Cleveland Clinic's complex patient mix. Allie Tallman, a nurse practitioner, provided crucial human feedback, identifying that the model needed to differentiate between various patient subpopulations and learn from subtle clinical cues that the data alone couldn't capture. This iterative process of human correction and model refinement was essential for improving accuracy.

Meaningful signals. Kristy Johnson's research into nonverbal communication for her son, Felix, demonstrated the challenge of translating lived experience into quantifiable data. Her Comalla and ROSCO projects relied on caregivers to label vocalizations and gestures, providing the empathetic context that AI needed to learn. This human-labeled data, though smaller in scale, was vital for teaching AI to recognize the rich, individualized communication patterns that sensors alone couldn't interpret.

5. AI Can Bridge Profound Communication Gaps.

If we ask, ‘How are they already communicating?’ then I think we can get somewhere.

Beyond translation. The evolution of machine translation, from rule-based systems to neural machine translation (NMT), shows AI's increasing ability to bridge linguistic divides. NMT models process multiple languages simultaneously, mapping them into a shared numerical space to achieve "zero-shot translation" between language pairs they've never explicitly seen. This leap transformed translation from word-matching to something resembling genuine comprehension.

Real-time transformation. Google's AudioLM takes this further by bypassing text entirely, transforming audio in real-time. Awaneesh Verma's team developed a model that listens to a few seconds of speech and predicts its continuation while mimicking the speaker's voice and affect. This allows for natural, real-time conversations between speakers of different languages, preserving crucial emotional nuances that traditional translation often misses.

Unlocking nonverbal worlds. Kristy Johnson's work aims to bridge communication for nonverbal individuals. By shifting the question from "Why don't they speak?" to "How are they already communicating?", she developed systems like Comalla and ROSCO. These platforms collect caregiver-labeled audio and video data, allowing AI to learn and eventually translate the unique vocalizations and gestures of nonverbal people, offering a path to connect them more deeply with the world.

6. Ethical AI Requires a "Human-in-the-Loop" Approach.

If an IRS employee is playing chess... the idea is it’s still an IRS employee in the chess match. Now, we might set up the computer next to the employee to advise them on what the move is to make... But the computer advises. The human makes the final decision on what move to make.

Guarding against harm. The rapid proliferation of AI companions, from Character.ai to Meta's bots and Grok, highlights the ethical imperative for human oversight. Instances like the Chai bot-driven suicide and Grok's antisemitic responses underscore the dangers when AI is left unchecked or designed with problematic incentives. Rosalind Picard's work at MIT emphasizes that humans are "ontologically superior" to AI, necessitating guardrails to prevent harm and ensure well-being.

Preserving human judgment. The IRS, under Commissioner Danny Werfel, adopted a "human-in-the-loop" approach to AI, ensuring that employees retain final decision-making authority, especially in "inherently governmental" functions like audits. This framework protects against AI replicating biases, as seen in past IRS algorithms that disproportionately audited Black taxpayers. AI advises, but humans decide, maintaining accountability and trust.

Continuous feedback. Human-in-the-loop means active participation in training, testing, and refining AI models. In the Cleveland Clinic's sepsis pilot, clinicians regularly reviewed Bayesian Health's alerts, overriding or confirming them. This continuous feedback loop allowed the AI to learn from human judgment, refining its predictions and ensuring that the technology served as a trusted assistant rather than an autonomous decision-maker.

7. Even Imperfect AI Delivers Life-Saving and System-Improving Benefits.

If you tell someone that even imperfect AI has contributed—even in a small way—to reducing sepsis mortality by 40 percent, you would have to call that success.

Warp Speed's triumph. Operation Warp Speed, despite its chaotic beginnings and imperfect data, successfully distributed COVID vaccines to all fifty states simultaneously. The Tiberius platform, built by Palantir, provided General Perna with a real-time "God view" of the supply chain, enabling him to track every vial and respond to governors' concerns with precise data. This demonstrated that even a rapidly deployed, evolving AI system could achieve monumental logistical feats.

Sepsis mortality reduction. At Cleveland Clinic, the implementation of AI for sepsis prediction, alongside other human-led initiatives, contributed to a 40% reduction in sepsis mortality from 2021 to 2024. While the AI model didn't achieve 90% accuracy on the most critical alerts, its ability to flag cases earlier and standardize response protocols across a massive system proved invaluable. This validated the principle that AI doesn't need to be flawless to make a significant, life-saving impact.

Municipal efficiency. East Lansing's AI-assisted recycling program, though a "soft" application, yielded impressive results. The system's 0.5% error rate in identifying non-recyclables and its ability to generate targeted feedback postcards led to substantial improvements in contamination rates and participation. This small-scale success showcased how AI, even in its early stages, can bring transparency and efficiency to everyday public services.

8. Transformative AI Needs Stubborn, Empathetic Leaders.

I look to see: Do I have a partner who will live and die with this problem, or is this entertainment?

Relentless advocates. Leaders like Debbie Kwon at Cleveland Clinic exemplify the stubbornness required. Despite knowing "nothing" about AI, she gambled her career leverage to establish an AI-fueled cardiac imaging lab, battling bureaucratic resistance and internal sabotage to recruit Chris Nguyen and secure resources. Her persistence, driven by a desire to improve patient care, ultimately led to groundbreaking advancements in cardiac MRI.

Unwavering dedication. Kristy Johnson's decade-long journey to understand and translate her nonverbal son's communication is a testament to empathetic leadership. Facing academic skepticism and financial hardship, she built her research around her personal mission, creating Comalla and ROSCO. Her unwavering commitment attracted brilliant students like Siddhant Shah, who were inspired to apply their AI skills to a deeply meaningful, underfunded problem.

Community champions. Peggy Buffington, superintendent of Hobart, Indiana schools, championed Khanmigo despite initial teacher skepticism and technical glitches. Her trust in Sal Khan and her teachers, combined with her "Be Excellent on Purpose!" ethos, fostered an environment where AI could be experimented with and refined. Similarly, Cliff Walls in East Lansing, Michigan, quietly navigated bureaucratic hurdles to implement an AI recycling program, driven by a desire to improve his community.

9. The "AI Counterculture" Focuses on Public Good Over Profit.

Companies have their AI priorities, but we’re here for improving human lives.

Mission-driven innovation. The book highlights a crucial distinction between AI developed primarily for corporate profit and that focused on public good. While companies like Meta and xAI prioritize engagement and market dominance, organizations like Khan Academy are driven by a mission to provide "a free, world-class education for anyone, anywhere." This difference in intent shapes everything from ethical guardrails to business models.

Nonprofit integrity. Sal Khan's decision to partner with OpenAI was a "wager" on his reputation, but it was rooted in Khan Academy's commitment to educational integrity. They meticulously red-teamed GPT-4, built robust guardrails against cheating, and focused on making Khanmigo student-safe and teacher-friendly. This contrasts sharply with the "move fast and break things" ethos often seen in the private sector, where ethical considerations can be secondary to market speed.

Leveraging big tech for niche problems. Rosalind Picard's Affective Computing Research Group and Kristy Johnson's lab embody this counterculture. They repurpose powerful, often expensive, AI models developed by large corporations (like OpenAI's CLIP or Alibaba's Video-LLaMA3) to solve niche, underfunded problems that directly improve human lives, such as translating nonverbal communication. This demonstrates a functional arrangement where public good can benefit from the infrastructure built by profit-driven ventures.

10. User Behavior Shapes AI's Future: Choose Wisely.

The question is who it’s empowering.

Conscious engagement. The epilogue serves as a call to action, urging readers to actively shape AI's future. It emphasizes that user behavior is the most powerful feedback mechanism for tech companies. By consciously choosing which AI features to use or ignore, individuals can influence what companies prioritize, ensuring that AI development aligns with societal values rather than just profitability.

Informed choices. Understanding the "swap" inherent in every AI service—convenience for data or attention—is crucial. The author encourages users to read privacy policies, or even use AI to summarize them, to make informed decisions about what they are giving up. This vigilance helps ensure that the terms of engagement with AI are clear and that the benefits outweigh the costs.

Supporting human amplification. Ultimately, the future of AI depends on empowering people. This means actively supporting AI that amplifies human judgment and capabilities—like teachers customizing lessons or doctors spotting patterns in scans—rather than AI that replaces human roles or denies services. By advocating for funding and directing attention towards AI solutions for critical societal problems, individuals can ensure that AI serves humanity's best interests.