Cutting Through the Noise: Reimagining Tech for Good

The term “Tech for Good” has been in use for over a decade. At its core, the movement rests on the idea that governments and nonprofits—institutions explicitly charged with advancing social outcomes—can benefit from the tools and talent of technology companies. In practice, this has taken the form of technology transfer programs where firms provide concessional software and engineering support to public institutions.

There have been notable successes: in 2013, when the Affordable Care Act’s marketplace website, HealthCare.gov, famously failed at launch, the Obama White House responded with a “tech surge,” recruiting elite technologists from the private sector to stabilize the product. My own work has benefited from this Tech for Good model. While leading humanitarian response at GiveDirectly, we partnered with Google.org, which provided mapping and machine learning tools that enabled faster aid deployment after natural disasters.

The problem is a mismatch in scale

Image

Using WhatsApp’s AI chatbot to complete math homework takes seconds

While technology transfer can help public institutions build beneficial AI tools, this model of social impact is coming under increasing strain. Without significant policy intervention, the reach of public sector tools may be eclipsed by the scale of harmful private-sector products.

Take education—an AI tutor in Ghana delivered sizable learning gains equivalent to a year of schooling at a fraction of traditional costs. The specialized product scaffolds learning—guiding students through progressively harder material without giving away answers. But as public institutions try to scale these specialized tools, students are already using general-purpose AI chatbots on Meta—reaching 3.5 billion daily users across WhatsApp, Instagram, and Facebook—to generate homework answers.

Evidence is showing that the use of generic AI platforms for schoolwork can harm learning. A study in Turkey found that students using a generic AI chatbot performed better when using it but worse once it was removed. Without educational guardrails, the chatbot gave answers away and students’ agency and self-directed learning degenerated. Neuroimaging research on “cognitive debt” also suggests that outsourcing critical thinking to AI weakens neural connectivity linked to long-term memory and recall. Additionally, such direct-to-consumer applications often lack wraparound structures such as teacher guidance and support, which has been shown to dramatically improve student learning when using AI tools.

The debate is no longer whether AI brings benefits or harms—it clearly can do both. The question is whether harmful uses outscale beneficial ones. The current outlook is not encouraging. In India, a Central Square Foundation survey found that only 6 percent of students using edtech rely on specialized learning tools, while 67 percent use WhatsApp. In rural Colombia, teachers report a surge in “erudite” AI-generated essays on Meta’s platforms alongside rising exam failures. In the US, a Pew study found that 54 percent of students ages 13–17 have used chatbots like ChatGPT or Copilot for schoolwork, with nearly 60 percent of respondents telling Pew that students at their school used chatbots to cheat “very often” or “somewhat often.” The Center for Universal Education at Brookings isn’t optimistic either, concluding after reviewing the evidence and interviewing more than 500 people in the education sector that while AI can certainly be used beneficially, its current use is delivering more harm than benefit.

It’s not solely the fault of generic AI chatbots. These tools are layered onto social media, messaging, and productivity platforms with massive reach and strong network effects: the more people use them, the harder they are to leave. Optimized for attention and advertising revenue, these platforms make switching unlikely. If students rely on an AI bot on WhatsApp for homework, the education sector cannot expect them to move to pedagogically sound alternatives without major intervention. Classrooms and self-study once anchored learning; now it is increasingly shaped by chatbots in students’ pockets.

Public institutions have long struggled to keep pace with new technologies. Health agencies, accustomed to being the primary source of health information, were slow to respond to the surge of misinformation on social media during the COVID-19 pandemic. More than two decades after Facebook’s launch, we are starting to see lawsuits and bans over social media’s potential harms to mental well-being. In the age of AI, the pathways to harm may be more immediate and far-reaching. Generative AI chatbots can be sycophantic and flattering, persuading users to believe falsehoods in ways far more personalized than social media. Additionally, policymakers who once focused on issues like health misinformation must now contend with risks in domains such as education.

Where we go from here

To be clear, while tech giants may be gaining influence over domains traditionally led by public institutions, that influence is potentially uneven across sectors. This blog has focused on education, where generative AI may pose distinct risks to the development of core cognitive processes. In other sectors, the risks may differ. Farming advice delivered through a consumer AI product, for instance, may not erode capacity in the same way; effective farming may depend less on pedagogy and more on access to high-quality information—something general-purpose AI can often provide. Even within education, effects may vary. Primary school children often have caregiver-controlled phones, and younger students may not adopt social media organically, creating greater opportunity for use of approved, purpose-specific tools.

The point is not to overgeneralize harm—if well developed and scaled, AI can deliver substantial positive impact. It’s to understand under what conditions harmful uses scale so dramatically that they create “overmatch”: a situation where positive applications are dwarfed by the reach of harmful ones. Education may be a realm where we are starting to see overmatch. There are a number of policy and programmatic questions worth exploring to address it:

• “Reverse flow”: Traditional Tech for Good models transfer technology from private firms to public institutions. But what if we reversed the flow—channeling public-sector data, contextual insight, and institutional knowledge to tech firms to better align widely adopted products with the public interest? While reaching millions with a public-sector app would be a success, improving already widely used private-sector products could shape outcomes for billions of users and the many applications built on top of them. Rather than just viewing tech firms as distributing less beneficial products than the public sector, we could more actively explore how they can serve as channels for developing and delivering publicly beneficial tools at scale.

  Efforts along these lines are emerging in health and education. In education, Anthropic, Google, and OpenAI have launched “learn modes” that embed pedagogical frameworks co-developed with public and private sector education experts into their foundational models. These tools go beyond generic content delivery to incorporate instructional best practices—such as managing cognitive load and withholding answers from students. Beyond education, organizations like Karya are building pipelines of validated datasets of Indian dialects that companies like Google and Microsoft are incorporating.

  Technology companies have long delivered publicly beneficial products. However, how those products are chosen and built could be more systematic and done in collaboration with the public sector. For example, when Google Maps first launched, it lacked public transit navigation. The feature was built because staff who frequented public transport used a Google policy which allowed employees to spend 20 percent of their time on self-directed projects. Public-interest features should not emerge in major platforms solely because of the interests of tech employees. What would it look like if tech companies built a pipeline that systematically incorporated insights from thousands of nonprofits—covering early childhood education in rural India to business practices in Kenya—to consistently shape the features and performance of foundational AI models? While we often discuss the harms of a new technology, perhaps the greater loss to society would be from the beneficial products that do not get created.

• Benchmarking: Once firms incorporate public-sector know-how and content, they can test model performance. OpenAI has released HealthBench, which provides a standardized measure of how AI models deliver health information. In education, the independent organization AI-for-Education.org has created a standardized pedagogical benchmark against which foundational models can be compared. Similar benchmarks should be systematically established at greater scale and cover topics ranging from agriculture to minority language performance. 

  Image

  

  Benchmarking against domains in the public interest can help tech companies steer their generative AI products towards beneficial outcomes.

• Heavier safeguards: ChatGPT, Gemini, and Meta’s AI products have parental controls and age-estimation tools that aim to safeguard the user experience for minors. They vary in the level of constraints placed on the user experience—ChatGPT and Gemini's teen-appropriate setting aims to reduce exposure to topics like graphic content and extreme beauty ideals, while Meta restricts access to the feature Character AI, which received fierce criticism for allowing minors to create flirty chatbots. Should these parental controls allow parents to only allow their child access to “Learn Mode”? Instead of merely filtering harmful content, should companies build more options like a “learn mode” that actively steer users toward healthy use? For example, should there be a “teen health mode” constrained to promoting healthy behaviors like exercise while discouraging harmful ones such as vaping? Or should such restrictions be applied to all users?
• Regulation: At what point should governments require such controls? OpenAI introduced parental safeguards without regulation, and Meta removed teen access to Character AI—but only after highly visible harms. Firms may have limited incentives to adopt safeguards voluntarily, especially when risks like public health or learning erosion are less visceral. They may also lack incentives to support “reverse flow” or adhere to public-interest benchmarks when these do not align with their business interests. That said, regulating firms to align their products with a large number of public outcomes should be carefully weighed. It assumes a technocratic government acting in the public interest, yet many states have strong political incentives to manipulate the technology for political ends under that same public-interest guise.
• Significantly larger financing and tech transfer for public-sector products: While this blog argues for expanding the Tech for Good toolkit, we should continue doubling down on technology transfer to the public sector. But if the public sector is to compete with harmful products at scale, it will need an order-of-magnitude increase in capital. Today, nonprofits building AI applications are lucky to receive $2–5 million for product development and go-to-market, while startups raise multiples more each funding round and major tech firms invest vastly more in their own products. It remains to be seen how Microsoft’s $50 billion pledge to the Global South will be spent, but regional investments at that scale bring us closer to what is needed.
• Sector-specific training: While we can try to steer tech companies to meet socially optimal benchmarks or introduce more safeguards, we have to develop strategies to deal with the scale of harmful products that will exist. Schools and caregivers need to be trained to guide students on how to learn with the help of AI instead of devolving their agency. Governments should guide the public on how to interpret AI-generated information on topics from public health to elections. Governments could regulate and control access, but citizens also need to be empowered to engage beneficially.

Part of the motivation for this blog is to expand the Tech for Good sector’s approach to impact and ask how it can achieve scale in the age of generative AI. It also underscores the urgency: without broadening our toolkit, we risk being outmatched by products that do not serve the public interest. Those products are already here today. This may be a difficult message for governments and technology companies—it implies greater public-sector involvement in the tech sector’s products and public institutions expanding their capabilities and thinking differently. But if even a fraction of the tech industry’s claims about AI dramatically transforming society are true, we must be prepared to radically transform how we approach Tech for Good.

Many thanks to Dave Evans, Dan Bjokegren, Markus Goldstein, Charles Kenny, and Tim Ohlenburg for their comments and ideas. Thanks to Thet Htar for research support.