|
Law in the Internet Society
|
|
Predictive Text-- Convenience or Catastrophe?-- By GabrielaFloresRomo - 21 Oct 2021How did Google Know?In a minute, my final draft was gone. After calling every technology support available, receiving an extension, and re-doing an entire issue for my brief, I began solely working on Google Docs (“GD”). GD’s autosave and history browsing were lifesavers, until a new feature piqued my interest, both in Gmail and GD. My first thought when I noticed the auto-complete feature on Google’s Gmail and GD was, “how did they know what I was going to say?” Google’s “Smart Compose” is a neural-network model—machine learning models in which a computer learns to perform a task by analyzing training examples—that classifies and predicts sequences of natural language (Hardesty, "Explained: Neural Networks"). Something invisible to untrained eyes was now put at the centerstage for the sake of convenience. Whether I was taking class notes or drafting a personal email, Google apparently “read my thoughts.”Smart Compose-- Is it Inherently Smart?GD’s Smart Compose “assists with composing new messages from scratch and provides much richer and more diverse suggestions along the way,” and it is trained by email data, which includes machine-collecting data from email chains, email subjects, date and time of emails, and location of the user (Chen et. al., 3). While Google claims individuals cannot individually read these emails in order to protect privacy, this data compilation must be stored to determine individuals’ often repeated written sequences of words. Thus, while humans may not be actively reading personal information, a database tracking from whom and when and attaching strings of words to certain individuals is likely readily available. Regardless of these assurances, however, “even if sensitive or private training-data text is very rare [given the omission of some text and focus on commonly stringed words], one should assume that well-trained models have paid attention to its precise details” (Carlini, et. al., 1). Since these systems store information based on repeated phrases, individuals may mine for private details by using sequencing of auto-fill to determine private information. Since Google’s Smart Compose uses location, although tedious, it is not unfathomable that two competing individuals in the same proximity may attempt to steal information from each other in this manner. Additionally, given that some companies’ sole purpose is to buy and sell individuals’ information, they may also mine for private information, given their interest in attaining and selling as much information as possible. Thus, private information can be made public not only to one’s surprise but to the benefit of others as well. Since humans had to develop this Smart Compose system based on unattainable data, biases likely promulgate the system. Suggestions of what comes next, thus, are more likely to adhere to developers’ vernaculars. One way Google has combated this is by removing gender-based pronouns so that it reduces the risk of incorrectly suggesting one’s gender identity and suggesting biased autofill phrases (Dave, Reuters). Although this reduces a variable, it is not enough. While Google claims that its language model adapts to a user’s personal mail data, thus personalizing text, biases may persist based on gathered information as well. For example, cultural group associations, geographical location, and socioeconomic information may suggest preferences based on string of words associated with and linking individuals from different areas. As use becomes more prevalent, groupthink on what is correct may become more pervasive between groups, especially amongst those who adhere to groups that have historically had power. This could further ignite ongoing debates about English, and language’s social and political implications (Fasold, 150). The implications of predictive words surpass convenience; they include biases and promote further divisions in language.Problems and ResponseThese Smart Compose concerns are not limited to Google, however; Apple and Samsung, among others, use predictive text software. While Google does allow its users to turn off the Smart Compose feature so that they can no longer receive predictive text suggestions, it is uncertain whether this simply limits individuals from seeing the predictive text or whether individuals have fully opted out of the system. This begs the question of whether Smart Compose can still access individual’s information regardless of the opt-out feature. While the first step is to turn off Smart Compose software on devices, the goal should be to work on systems that either fully allow you to disengage from these monitoring and scanning technologies or switch to software that does not actively use any of these systems.SourcesCarlini et. al.: https://arxiv.org/pdf/1802.08232.pdf Chen, et. al.: https://arxiv.org/pdf/1906.00080.pdf Dave: https://www.reuters.com/article/us-alphabet-google-ai-gender/fearful-of-bias-google-blocks-gender-based-pronouns-from-new-ai-tool-idUSKCN1NW0EF Fasold: https://www.jstor.org/stable/pdf/43025634.pdf Hardesty: https://news.mit.edu/2017/explained-neural-networks-deep-learning-0414You are entitled to restrict access to your paper if you want to. But we all derive immense benefit from reading one another's work, and I hope you won't feel the need unless the subject matter is personal and its disclosure would be harmful or undesirable. To restrict access to your paper simply delete the "#" character on the next two lines:
|
|
