Human advisors increasingly have access to AI recommendations and use them to shape the advice they give clients, often without disclosing their AI use to their clients. Disclosure of AI use involves informing individuals when AI is being used on their behalf by another person, such as disclosing to clients that advisors are using AI to formulate their expert advice. Our study aims to investigate whether and how disclosing advisors' use of AI assistance to their clients influences the advisors' perceived responsibility for decisions and degree of AI reliance. Recruiting financial advisors to perform a personal finance advising simulation that manipulated whether clients would know that advisors used an AI system (disclosed vs. not disclosed), we found that financial advisors felt less responsible for their recommendations when they believed their use of AI assistance would be disclosed rather than undisclosed to clients. We also found that advisors' perceived personal responsibility was higher when their reliance on AI was lower. Advisors' perceived self-competence increased their perceived personal responsibility for investment decisions relative to AI, and their trust in AI decreased their perceived responsibility. We conclude by discussing how our findings can inform disclosure schemes for improved human-AI collaboration in advising.

Social media platforms are widely seen as a valuable medium to spread a wide range of information including charitable causes and health awareness. But given the flexibility provided by the social media platforms, it is important to ensure that the right kind of information is delivered to the right audience when needed. The pilot study presented in this paper considered a sample of Zika related tweets that were classified into different prevention techniques. The classification categories were drawn from the guidelines by CDC. Training a logistic regression model on the annotated data we found the accuracy to be 72%. The findings are significant in studying the effectiveness of social media platforms in spreading the right kind of information in time. This in turn can be useful in informing health care officials to take necessary steps with the help of real-time communication for such unfortunate events in future.

Generalization performance of classifiers in deep learning has recently become a subject of intense study. Deep models, which are typically heavily over-parametrized, tend to fit the training data exactly. Despite this “overfitting", they perform well on test data, a phenomenon not yet fully understood. The first point of our paper is that strong performance of overfitted classifiers is not a unique feature of deep learning. Using six real-world and two synthetic datasets, we establish experimentally that kernel machines trained to have zero classification error or near zero regression error (interpolation) perform very well on test data. We proceed to give a lower bound on the norm of zero loss solutions for smooth kernels, showing that they increase nearly exponentially with data size. None of the existing bounds produce non-trivial results for interpolating solutions. We also show experimentally that (non-smooth) Laplacian kernels easily fit random labels, a finding that parallels results recently reported for ReLU neural networks. In contrast, fitting noisy data requires many more epochs for smooth Gaussian kernels. Similar performance of overfitted Laplacian and Gaussian classifiers on test, suggests that generalization is tied to the properties of the kernel function rather than the optimization process. Some key phenomena of deep learning are manifested similarly in kernel methods in the modern “overfitted" regime. The combination of the experimental and theoretical results presented in this paper indicates a need for new theoretical ideas for understanding properties of classical kernel methods. We argue that progress on understanding deep learning will be difficult until more tractable “shallow” kernel methods are better understood.