What the Rise of Data-Driven Social Science Centers Means for You

Why Data-Driven Social Science Centers Became Research Infrastructure

Princeton named Jonathan Mummolo interim director of its Data-Driven Social Science initiative in August 2026, placing an associate professor of politics and public affairs atop a unit now folded into Data and Intelligent Systems. The move formalizes something already true at a handful of research universities: the center is no longer a seminar series with a mailing list.

These centers give a researcher three kinds of help. People: consultants and staff scientists who write production code. Tooling: locally hosted AI models and pipelines that turn millions of documents into machine-readable data. Money: seed grants that fund a pilot before a federal agency will.

The binding constraint is no longer a method. It is access.

What a Data-Driven Social Science Center Actually Does

Princeton's Data-Driven Social Science initiative now operates inside the Statistics and Data Science pillar of Data and Intelligent Systems, the university's newly formed academic unit, with associate professor of politics and public affairs Jonathan Mummolo serving as interim director since August 2026.

A Unit, Not a Department

The category is easy to blur. A data-driven social science center is a standing research unit that supplies consulting, training, computing, tooling, and internal funding to social scientists. It grants no degrees, so it is not a department. It is also not central IT: the staff read working papers, understand identification strategies, and know what a reviewer will ask about a table.

What the Menu Looks Like

At Princeton the offerings run from one-on-one consultations and workshops to symposia, research grants, and computational tool development. Two tools show the ambition. Blackfish lets researchers run open-source AI models at no cost, without programming and without shipping data to an outside vendor. TigerFlow automates large multi-step jobs such as converting millions of documents or thousands of hours of video into machine-readable data.

The Staffing Model

The work depends on embedded technical staff and staff scientists who build instruments other researchers reuse, rather than one-off scripts for a single paper. Lori Bougher, director of research and strategy at Data and Intelligent Systems, frames that as a deliberate team strategy rather than an accident of hiring.

Why the Model Is Spreading

Three pressures push individual labs toward a center. Data arrives at a scale one principal investigator cannot process alone; restricted federal data must be analyzed inside certified environments; and AI tooling is hard for a lone scholar to vet. Mummolo's warning sets the tone: off-the-shelf AI tools, he argues, are often a poor fit for research at the technical frontier.

From Consultations to Code: The Core Service Menu, Compared

The tradeoff is rarely quality. It is depth against access. Two centers can both advertise "consulting," yet one replies to a help request within two business days while the other assigns a research engineer to your project for a semester. Knowing which kind of shop you are walking into matters more than the brochure.

The Same Five Lines, Very Different Depths

Nearly every center organizes its menu around the same five services. What varies is how far each one goes.

Service

What it typically covers

Cost signal

One-on-one consulting

Study design, code, models, results

Free up to a cap; Harvard IQSS covers three hours per publication, Harvard CGA the first hour, then $100 to $200 per hour³

Workshops and short courses

Methods and tool training

Usually free to affiliates

Symposia and speaker series

Outside researchers and industry guests

Free

Seed grants

Pilot funding and internal awards

Competitive

Tool development

Production software such as Blackfish and TigerFlow

Staffed capacity, not drop-in help

Some centers stop at office hours. Princeton's Data-Driven Social Science initiative builds and maintains its own software, including an interface for running open-source AI models locally and a pipeline that converts millions of documents into machine-readable data.¹

Ask Three Questions Before You Book

• Is consulting free and unlimited, or capped in hours? A three-hour cap per publication is generous for a regression question and thin for a multi-year project.²
• Do graduate students get the same access as faculty? Princeton lists faculty, graduate students, and research staff; Harvard's services extend to postdocs, visiting scholars, and thesis-writing undergraduates.
• What is the turnaround? Two business days is a common response target for a first reply, not for finished data processing.²

The Cheapest Service Is Usually the Highest-Leverage One

A two-hour consultation before data collection can save months of rework. Bring your research question, data source, and timeline to that first meeting, not a finished dataset that needs rescuing.

Inside the Secure Enclave: Title 13, HIPAA, and Restricted Cloud Environments

What actually happens to an analysis when you run it inside a secure enclave?

The Room Without a Door

An enclave is an isolated computing environment: restricted data stays on the enclave's own infrastructure, you submit code rather than download files, and nothing leaves until a reviewer has inspected each table, figure, or coefficient you want to take out. In Census Research Data Centers, that means Special Sworn Status, a security clearance, an approved project scope, and a disclosure form for every output request². Title 13 bars releasing anything that identifies a person or business, and even sample sizes and estimated effects can need review before you discuss them outside the room¹.

What the Compliance Layers Look Like

Health data adds another layer. NCHS restricted-use files cannot be exported to a network or carried out on removable media, access is limited to NCHS staff and Designated Agents under CIPSEA, and both materials entering and output leaving are reviewed³. In practice: no USB drives, no screen capture, no moving files out except through review.

The Newer Restricted Cloud Model

Instead of a physical room, some university centers now run restricted cloud environments stocked with research-ready voter files and commercial data at scale. Princeton's Data-Driven Social Science initiative says the environment saves researchers weeks to months of coding time.

Budget for Review, Not Just Analysis

Output review is not instant. The Census Bureau advises allowing about six weeks after you produce output, and reviews can send work back for revision⁴. Design analyses that yield a handful of clean, pre-specified tables rather than thousands of diagnostic plots. If a reviewer has to read it, it has to earn the trip.

Data-Sovereign AI: Why Blackfish and TigerFlow Beat Off-the-Shelf Chatbots

Data-sovereign AI means the model runs on your institution's own infrastructure, or under a contract tight enough that your data never becomes a vendor's asset. Princeton's Data-Driven Social Science initiative built two tools on that premise. Blackfish runs open-source models at no cost, with no programming required, and without sending data to an outside company. TigerFlow automates long multi-step jobs, such as converting millions of documents or thousands of hours of video into machine-readable data.

The Risk, in Research Terms

Interim director Jonathan Mummolo, an associate professor of politics and public affairs, puts it bluntly: off-the-shelf AI tools are often inappropriate for serious research, and AI brings "great opportunities, but also great risks." For researchers those risks are concrete. A model that updates without notice breaks reproducibility. A chat window that logs prompts can breach confidentiality. Undocumented model drift silently changes the variables in your dataset.

What Institutional Guidance Now Says

Pasting restricted or human-subjects text into a commercial interface can violate a data use agreement or IRB protocol on its own⁴. Guidance issued through 2025 and 2026 has moved toward requiring local or contractually constrained models. NIH's 2025 notice prohibits sending controlled-access genomic data through public generative AI tools and treats resulting model parameters as data derivatives¹. Campus policies now commonly require IRB approval before AI touches human-subjects data, name the specific tools in consent forms, and bar identifiable participant data from non-secure services². Wharton prohibits confidential data in personal or free tools entirely³.

Prompt-and-Paste Is Not a Method

The pedagogical point follows. Students who learn machine coding by pasting text into a chatbot learn a workflow, not a method. They need validation against hand-coded samples, reported error rates, and documentation of the exact model, version, and prompt behind every machine-coded variable.

Related Articles

Interdisciplinary Research Funding for Social ScientistsLLMs for Social Science ResearchThe Data Boom and Social Science Methods

Who Pays? Funding Models, Fees, and Faculty Seed Grants

Center support gets paid for in one of three ways: institutional subsidy, where services are free at the point of use; recharge or fee-for-service billing, where a project pays for compute or staff time by the hour; and grant funding, where center costs are written into a proposal as a service or personnel line. The seed programs below show what the grant-funded route looks like in practice, including how far eligibility reaches into graduate students and postdocs. Watch the cycles closely: several internal calls run on short windows and surface in center newsletters before they reach a campus-wide audience.

How to Pitch a Project to a Center Director

The temptation when you first approach a center director is to ask for everything at once: compute, a seed grant, staff time, a letter of support. The stronger move is the reverse. Directors triage by what is actually being requested, and a narrow, concrete ask with a visible return gets read before a broad one does. Ask early in the term, before a queue forms, and expect to be routed onward to a postdoc or staff scientist rather than handled personally.

Five Lines, No More

Write the pitch so it fits in an email that needs no scrolling. Cover five things in order:

• The question: one sentence, framed as a question with a stake for a real audience.
• The data: source, approximate size, and sensitivity. Say plainly whether it can sit on a laptop or needs a restricted environment.
• The bottleneck: the specific step you cannot clear alone, whether that is entity resolution across millions of records, video transcription, or secure access.
• The timeline: months, not aspirations.
• The deliverable: a paper, a proposal, a dataset, a tool.

Bring Evidence, Not a Wish List

A sample of the data, a draft codebook, or one pilot result separates a serious request from a fantasy. Directors field many ideas; few arrive with a file attached and a plausible path to a result.

State the Ask, and the Return

Hours of consulting, compute, a seed grant, or a staff scientist as co-investigator: name one as the primary ask. Then name what the center gets. Proposals are weighed by whether they generate a grant, a publication, a workshop, or a reusable tool, because that portfolio is how the budget survives the next cycle. Say which of those you expect to produce, and roughly when.

The Institutional Landscape: Who Is Investing and Why

Centers cluster in three kinds of homes: interdisciplinary institutes, professional-school research arms, and library-based data services. That placement tells you who the center actually serves and whose problems it is built to solve. The funding logic is similar everywhere, with provost and dean seed money meant to buy external grants, so centers are judged on the grant dollars they help attract and on the researchers they keep. In 2026 several centers are adding AI and research-engineering staff rather than more statistical consultants, and vendor partnerships plus guest lectures from labs such as Google DeepMind, Anthropic, and OpenAI now double as recruiting tools for methods training. When comparing institutions, ask three questions: is the center permanent or grant-funded, does it have dedicated technical staff, and does it support restricted-data work.