The rapid diffusion of generative AI raises two main questions: who adopts it, and how adoption reshapes online behavior. Using high-frequency mobile and desktop data from a large U.S. consumer panel, we address both. Adoption is strongly predicted by pre-existing digital activity. Users oriented toward productivity and information-seeking are more likely to adopt, while demographic characteristics such as gender and age add little once activity is observed. To study behavioral change, we combine measures of task-level complementarity, captured by which services are used alongside AI, with post-adoption reallocation. This approach reveals three main patterns: productivity and information-seeking services are complementary to AI and increase in usage, consistent with Jevons-type mechanism; education records the highest co-use, yet loses share for adopters under 30, where educational tasks are concentrated, consistent with efficiency gains; and leisure services exhibit low complementarity and declining usage, suggesting substitution.

Assessing competition among digital services requires evidence on substitution patterns that is difficult to obtain in the absence of observable prices. We exploit the unexpected outage of all Meta Platforms, Inc. (then Facebook, Inc.) services on October 4, 2021, which displaced its global user population for around six hours. Using high-frequency device tracking data from 14,000 individuals in the United States and Spain, we estimate effective substitution rates, the reallocation of usage time per minute of Meta usage lost. Non-Meta social media and messaging absorb the largest share, but substitution crosses conventional category boundaries and a substantial share of displaced time leaves the device entirely. Multi-homers substitute at rates an order of magnitude larger than single-homers, younger users drive social media substitution, and the composition of within-Meta usage shapes substitution destinations differently across countries. Users with the heaviest pre-outage Meta reliance exhibit persistent reductions in Meta's usage share over the four post-outage weeks. Because the outage displaced entire social graphs, our estimates capture substitution when users must coordinate their reallocation, drawn from a broad cross-section across two countries with different platform ecosystems.

Red teaming, where testers attempt to elicit harmful outputs from a large language model (LLM), has become central to AI safety practice. Yet without coordination, testers duplicate effort on familiar attack vectors and leave others underexplored. We test whether real-time novelty incentives can address this coordination problem. In two preregistered experiments (N=1,075), participants attempt to elicit harmful outputs from an LLM. Treatment participants earn bonuses based on both the harmfulness and the novelty of their outputs, while control participants earn bonuses based on harmfulness alone. Contrary to our hypothesis, treatment groups achieve lower joint scores than control in both experiments. This ``backfiring'' effect is driven by a decrease in harmfulness with no compensating gain in novelty. The effect is concentrated in an excess mass of near-zero-harmfulness outputs. Once a minimum harassment threshold is imposed, treatment achieves higher novelty, and at some thresholds higher joint scores. These results reveal a limit of multi-dimensional incentive design: objectives that are theoretically complementary can crowd each other out when participants must optimize both in real time, even in a setting explicitly designed for coordination gains.