Harvard's Martha Minow Tells MIT: Build AI That Asks Questions, Not AI That Answers Them

Harvard legal scholar Martha Minow used this fall's Karl Taylor Compton Lecture at MIT to make an argument that sits uncomfortably with most AI product roadmaps: the dominant consumer AI interfaces are optimized for exactly the behavior democracies can least afford. Her talk, "Constituting Democracy in the Era of AI," described AI as both a corrosive force on civic life and an underused instrument for repairing it. The pitch to the room was direct — build the better tools yourselves.

The framing: four "ingredients," all under strain

Minow, the 300th Anniversary University Professor at Harvard and a Harvard Law School faculty member since 1981, opened with a blunt assessment: this is not a good time for democracy or constitutionalism. According to MIT's account of the lecture, she catalogued violations of laws and norms, constraints on free speech, reduced congressional authority, disregard of legal rulings, and routine vilification of others in public life. She cited recent Gallup polling in which 77 percent of respondents said the country's founders would be disappointed in its current condition.

She then organized her remarks around four components she treats as prerequisites for a functioning democracy: education, reliable and accessible information, security from violence, and fairness. The structure matters for a technical audience because it reframes AI's societal effects as inputs into specific civic systems rather than as a generalized ethics debate. Each of those four areas has measurable outputs, and each is being reshaped by products that were not designed with those outputs in mind.

The education argument is the sharpest one

The bulk of the lecture concerned schooling. Minow noted that the United States already carries large local and regional disparities in education funding, and that schools have become significantly more segregated along economic and ethnic lines over the past three decades. AI, she argued, lands on top of that baseline and threatens to further weaken independent-thinking skills at a moment when civic understanding is most needed.

Her mechanism claim is worth quoting precisely because it names design choices rather than the technology in the abstract. Dominant AI platforms, she said, supercharge strategies already deployed by social media companies and induce passivity; leading models, interfaces and sycophantic elements endanger young people's developing capacity to think, remember and reason critically.

That is a strong causal claim, and the lecture as reported did not present underlying studies to support it. Sycophancy in large language models is a documented and widely discussed failure mode among model developers, and the research literature on AI's effects on learning outcomes is young and contested. Readers should treat the specific claim about cognitive development as an argument from a legal scholar surveying the field, not as a settled empirical finding. The broader point — that engagement-optimized interfaces shape behavior — rests on more familiar ground.

Minow's proposed alternative is where product people should pay attention. Existing LLMs, training methods and interface design, she said, can offer nudges, questions and hints instead of replacing student work, and much more could be built that is trustworthy and deepens curiosity rather than substituting for the learner. She also called for better assessments — more meaningful and more resistant to cheating — and told the audience they could build better tools than what currently exists.

This is a familiar tension inside edtech. Socratic tutoring modes exist across major assistants, but they are usually optional layers on top of general-purpose products whose default behavior is to produce a finished answer as quickly as possible. A model that withholds the answer is, by most consumer metrics, a worse product. Nothing in the lecture addressed who pays for the slower, more pedagogically useful version, or how schools with the funding disparities Minow described would procure it. That gap between the civic case and the business case went unexamined.

Information: an "information commons" without a funding model

On the second ingredient, Minow observed that modern media technology has coincided with a 75 percent reduction in journalists in recent decades, with a corresponding rise in corruption in places that lost local coverage. Her proposal was that civic-minded technologists work toward a widely used information commons, or several of them, to give people access to more reliable information.

The idea is not new, and the lecture did not specify governance, funding, editorial standards or how such a commons would compete for attention against existing distribution channels. Those are the questions that have sunk previous attempts. It is also unclear from the account whether Minow envisions the commons as human-curated, model-generated, or some hybrid — a distinction that determines nearly everything about trustworthiness. The absence of detail is not a flaw in a lecture, but it does mean the proposal is a direction rather than a design.

Security, fairness and the Enlightenment worry

On security from violence, Minow said the country is on the brink, noting that every judge she knows faces a credible threat of violence. She also addressed misinformation and multiple dimensions of societal fairness, though MIT's account gave less detail on those portions.

Asked during a Q&A moderated by MIT President Sally Kornbluth to name her central concern, Minow pointed to what she called the loss of the Enlightenment: the erosion of belief in reason, hypothesis testing, rejecting myths when evidence contradicts them, and the toleration and expanded participation that accompanied that period. She tied disinformation and the decline of critical thinking in schools back to that single worry, and connected it to her long study of mass violence and what happens when people neither understand nor trust each other.

She did not end on resignation. A turnaround is genuinely possible, she said, if people at MIT and elsewhere direct time and talent toward real solutions, adding that the world people want has to be invented.

Why the venue matters

The Compton Lecture Series, established in 1957, is MIT's highest-profile event of its kind, named for the Institute's ninth president, who served from 1930 to 1948 and chaired the MIT Corporation from 1948 to 1954. Kornbluth, introducing Minow, described Compton as having guided MIT through the Great Depression and World War II and transformed it into a global research university, and called Minow a legal scholar of extraordinary intellectual scope.

The placement is the message. Minow was not addressing regulators; she was addressing builders, at an institution that supplies a meaningful share of them, and arguing that the civic repair work is an engineering problem as much as a policy one. Whether that argument produces anything beyond applause is the open question. The lecture offered no commitments, funding, partnerships or programs — none were announced — and the incentives that shape consumer AI interfaces were not addressed in the remarks as reported.

What to watch

• Whether any model developer ships a default-on pedagogical mode — hints and questions rather than completed answers — instead of burying it as an optional setting.
• Credible peer-reviewed evidence on AI use and critical-thinking outcomes in K-12 and higher education, which would either substantiate or complicate Minow's central education claim.
• Movement on assessment design: cheating-resistant, meaningful evaluation is the bottleneck for schools trying to adopt AI without abandoning measurement.
• Any concrete attempt at an information commons with a disclosed governance and funding structure, as opposed to a prototype without an operating model.
• Whether sycophancy continues to be treated by labs as an alignment bug to be measured and reduced, or quietly tolerated because users prefer agreeable assistants.