Dissertation Title: From Data to Intelligence: Shaping, Scaling, and Scrutinizing AI Systems

Abstract:
A defining endeavor of our time is the transmutation of humanity's informational morass into artificial intelligence. This thesis contributes to the lifecycle of that transformation—from raw data to deployed intelligence, and back—with data understanding and optimization at the heart of this research program.

First, we establish that training data is the single most consequential and most neglected lever in AI development. Through empirical studies of data quality, toxicity, composition, multilingual scaling, and instruction tuning, we demonstrate that deliberate data curation often rivals or exceeds the gains from scaling model size. Second, we conduct the largest audits of AI training data to date, spanning nearly 4,000 datasets across text, speech, and video, and uncover pervasive failures in licensing, consent, and representation—alongside the rapid erosion of the web data commons on which open AI development depends. Third, we develop frameworks for evaluating AI at the ecosystem level, measuring developer transparency, concentration of power, and benchmarking integrity, and find that existing evaluation infrastructure is ill-equipped for the societal stakes it adjudicates. Fourth, we propose concrete institutional interventions—legal safe harbors for AI safety research, coordinated flaw disclosure protocols, and evaluation standards—that have already influenced the EU AI Act, prompted industry policy changes, and supported new legal exemptions for independent AI research.

This thesis argues that how we curate, govern, and hold accountable the data from which intelligence is forged will shape AI's trajectory more than any architectural or algorithmic advance. The alchemy runs both ways: the same data-driven methods that build these systems can be turned toward auditing, evaluating, and improving them. We will only get the future we deliberately design.

Committee members:
Alex 'Sandy' Pentland
Professor of Media Arts and Sciences
MIT Media Lab

Sara Hooker
CEO
Adaption Labs

Peter Henderson
Assistant Professor
Princeton University