These unintended negative consequences of the introduction of intelligence often have no direct relationship with the intended benefits, just as the adverse effects of a medication may bear no obvious relationship to the intended benefits of taking that medicine. Therefore, these negative consequences can be seen as side effects. The purpose of this article is to give designers, developers, and users of interactive intelligent systems a detailed awareness of the potential side effects of AI. As with medications, awareness of the side effects can have different implications: We may be relieved to see that a given side effect is unlikely to occur in our particular case. We may become convinced that it will inevitably occur and therefore decide not to "take the medicine" (that is, decide to stick with mainstream systems). Or most likely and most constructively, by looking carefully at the causes of the side effects and the conditions under which they can occur, we can figure out how to exploit the benefits of AI in interactive systems while avoiding the side effects.

The new social media sites - blogs, wikis, del.icio.us and Flickr, among others - underscore the transformation of the Web to a participatory medium in which users are actively creating, evaluating and distributing information. The photo-sharing site Flickr, for example, allows users to upload photographs, view photos created by others, comment on those photos, etc. As is common to other social media sites, Flickr allows users to designate others as ``contacts'' and to track their activities in real time. The contacts (or friends) lists form the social network backbone of social media sites. We claim that these social networks facilitate new ways of interacting with information, e.g., through what we call social browsing. The contacts interface on Flickr enables users to see latest images submitted by their friends. Through an extensive analysis of Flickr data, we show that social browsing through the contacts' photo streams is one of the primary methods by which users find new images on Flickr. This finding has implications for creating personalized recommendation systems based on the user's declared contacts lists.

The new social media sites -- blogs, wikis, Flickr and Digg, among others -- underscore the transformation of the Web to a participatory medium in which users are actively creating, evaluating and distributing information. Digg is a social news aggregator which allows users to submit links to, vote on and discuss news stories. Each day Digg selects a handful of stories to feature on its front page. Rather than rely on the opinion of a few editors, Digg aggregates opinions of thousands of its users to decide which stories to promote to the front page. Digg users can designate other users as ``friends'' and easily track friends' activities: what new stories they submitted, commented on or read. The friends interface acts as a \emph{social filtering} system, recommending to user stories his or her friends liked or found interesting. By tracking the votes received by newly submitted stories over time, we showed that social filtering is an effective information filtering approach. Specifically, we showed that (a) users tend to like stories submitted by friends and (b) users tend to like stories their friends read and liked. As a byproduct of social filtering, social networks also play a role in promoting stories to Digg's front page, potentially leading to ``tyranny of the minority'' situation where a disproportionate number of front page stories comes from the same small group of interconnected users. Despite this, social filtering is a promising new technology that can be used to personalize and tailor information to individual users: for example, through personal front pages.

We develop a new collaborative filtering (CF) method that combines both previously known users' preferences, i.e. standard CF, as well as product/user attributes, i.e. classical function approximation, to predict a given user's interest in a particular product. Our method is a generalized low rank matrix completion problem, where we learn a function whose inputs are pairs of vectors -- the standard low rank matrix completion problem being a special case where the inputs to the function are the row and column indices of the matrix. We solve this generalized matrix completion problem using tensor product kernels for which we also formally generalize standard kernel properties. Benchmark experiments on movie ratings show the advantages of our generalized matrix completion method over the standard matrix completion one with no information about movies or people, as well as over standard multi-task or single task learning methods.

Awerbuch et al.'s approach to distributed recommender systems (DRSs) is to have agents sample products at random while randomly querying one another for the best item they have found; we improve upon this by adding a communication network. Agents can only communicate with their immediate neighbors in the network, but neighboring agents may or may not represent users with common interests. We define two network structures: in the ``mailing-list model,'' agents representing similar users form cliques, while in the ``word-of-mouth model'' the agents are distributed randomly in a scale-free network (SFN). In both models, agents tell their neighbors about satisfactory products as they are found. In the word-of-mouth model, knowledge of items propagates only through interested agents, and the SFN parameters affect the system's performance. We include a summary of our new results on the character and parameters of random subgraphs of SFNs, in particular SFNs with power-law degree distributions down to minimum degree 1. These networks are not as resilient as Cohen et al. originally suggested. In the case of the widely-cited ``Internet resilience'' result, high failure rates actually lead to the orphaning of half of the surviving nodes after 60% of the network has failed and the complete disintegration of the network at 90%. We show that given an appropriate network, the communication network reduces the number of sampled items, the number of messages sent, and the amount of ``spam.'' We conclude that in many cases DRSs will be useful for sharing information in a multi-agent learning system.

Long-term health monitoring is becoming increasingly important with the rising prevalence of chronic disease in the U.S. While many researchers are investigating the use of remote biological monitoring and telemedicine technologies, the use of frequent self-report in long-term health monitoring remains a relatively unstudied area. We discuss some of the many cognitive, affective and contextual issues that must be addressed in maintaining a long-term stream of quality data from patients at home or in the field, and how many of these issues can be addressed through the use of conversational agents.

Virtual patients are viewed as a cost-effective alternative to standardized patients for role-play training of clinical interviewing skills. However, training studies produce mixed results. Students give high ratings to practice with virtual patients and feel more self-confident, but they show little improvement in objective skills. This confidence-competence gap matches a common cognitive illusion, in which students overestimate the effectiveness of training that is too easy. We hypothesize that cost-effective training requires virtual patients that emphasize functional and psychological fidelity over physical fidelity. We discuss 12 design decisions aimed at cost-effective training and their application in virtual patients for practicing brief intervention in alcohol abuse. Our STAR Workshop includes 3 such patients and a virtual coach. A controlled experiment evaluated STAR and compared it to an easier E-Book and no-training Control. E-Book subjects displayed the illusion, giving high ratings to their training and self-confidence, but performing no better than Control subjects on skills. STAR subjects gave high ratings to their training and self-confidence and scored better higher than E-Book or Control subjects on skills. We invite other researchers to use the underlying Imp technology to build virtual patients for their own work.

We describe CARDIAC, a prototype for an intelligent conversational assistant that provides health monitoring for chronic heart failure patients. CARDIAC supports user initiative through its ability to understand natural language and connect it to intention recognition. The natural language interface allows patients to interact with CARDIAC without special training. The system is designed to understand information that arises spontaneously in the course of the interview. If the patient gives more detail than necessary for answering a question, the system updates the user model accordingly. CARDIAC is a first step towards developing cost-effective, customizable, automated in-home conversational assistants that help patients manage their care and monitor their health using natural language.

Collaborative recommendation is an information-filtering technique that attempts to present information items (movies, music, books, news, images, Web pages, etc.) that are likely of interest to the Internet user. Traditionally, collaborative systems deal with situations with two types of variables, users and items. In its most common form, the problem is framed as trying to estimate ratings for items that have not yet been consumed by a user. Despite wide-ranging literature, little is known about the statistical properties of recommendation systems. In fact, no clear probabilistic model even exists allowing us to precisely describe the mathematical forces driving collaborative filtering. To provide an initial contribution to this, we propose to set out a general sequential stochastic model for collaborative recommendation and analyze its asymptotic performance as the number of users grows. We offer an in-depth analysis of the so-called cosine-type nearest neighbor collaborative method, which is one of the most widely used algorithms in collaborative filtering. We establish consistency of the procedure under mild assumptions on the model. Rates of convergence and examples are also provided.

Traditional recommendation systems make recommendations based solely on the customer's past purchases, product ratings and demographic data without considering the profitability the items being recommended. In this work we study the question of how a vendor can directly incorporate the profitability of items into its recommender so as to maximize its expected profit while still providing accurate recommendations. Our approach uses the output of any traditional recommender system and adjust them according to item profitabilities. Our approach is parameterized so the vendor can control how much the recommendation incorporating profits can deviate from the traditional recommendation. We study our approach under two settings and show that it achieves approximately 22% more profit than traditional recommendations.