On social platforms, cheap and fast automated fact‑checking tools are becoming widespread, yet their accuracy is imperfect. This paper studies when such tools backfire within a sequential social learning model. Each agent first observes a reliability signal of the article $s\in\{0,1\}$, then sees the actions of predecessors—likes indicate belief in reliability, dislikes the opposite. The agent may pay a cost $c>0$ for manual fact‑checking, which is perfectly accurate, or use a free automated tool that errs with probability $\epsilon$. Welfare is defined as the probability of a correct cascade $P_{\text{correct}}$.
Two opposing effects arise from introducing the tool. Wider fact‑checking can help later agents escape wrong cascades because more agents obtain reliable signals. Conversely, errors from the automated tool can mislead agents who would otherwise have verified the article perfectly, reducing $P_{\text{correct}}$. We prove that when the article’s intermediate reliability $\theta$ lies in a certain range, truthful articles are not too informative, and manual checking is sufficiently costly, adding the tool strictly lowers welfare. Simulations show the tool may improve outcomes outside the intermediate range and outperforms manual‑only checking on average only when its accuracy exceeds a threshold.
$$ P_{\text{correct}} = \Pr{\text{final cascade matches the true state}} $$
Review: The study highlights that in environments with uncertain information quality and high verification costs, indiscriminate deployment of imperfect automated fact‑checking can be detrimental, urging platforms to balance tool accuracy against incentives for manual verification.