WebPerceptor: Evaluating Browsing an Ai Mediated Web
Supervisor: Dr Joseph O'Hagan
School: Computing Science
Description:
WebPerceptor ( https://github.com/theartofhci/WebPerceptor ) is a browser extension which, for any web page, identifies text content, relays this to a local or cloud-based LLM with a user-defined prompt, then automatically replaces the identified text with the LLM response. As the user, browsing the web, all you’ve done is open a web page as you normally would but in reality its content has been completely rewritten by an LLM.
Browsing such an AI mediated web has many potential benefits (e.g. adapting content to specific reading levels, altering tone, providing in-line fact checking, censoring triggers, etc) and harms (e.g. introducing bias, censoring information, amplifying information disorder / extremism, etc).
This research project will (1) contribute to the on-going development of the WebPerceptor tool, and (2) map how AI could be used to alter web content in this way, for what reasons, what benefits/vulnerabilities are exposed, and the attitudes of individuals when confronted by an AI mediated web.
Using an online survey (e.g. n=100) considering different website types – news, shopping, social media, search – we will ask respondents to:
(a) Identify potential stakeholder(s) and envision a reason why they might use AI to alter the perceived website
(b) Describe the alterations they envisage occurring, and classify the significance and extent of the alterations proposed
(c) Rate whether the proposed reason would be of benefit, neutral, or detriment to each of the parties noted
(d) Rate the likelihood of use, overall concern, and likelihood of misuse/abuse
We will then confront respondents with example alterations and usage scenarios and capture their attitudes towards these (e.g. concern, acceptability, etc). Given said survey results, we will thematically cluster the types of alteration and reasons for alteration to develop a taxonomy of use cases and examine the degree of benefit, concern, prospect of mis-use/abuse to identify the most prevalent, desired, or risky types of AI-web alteration.
Information For Applicants:
This project will suit someone with prior experience in web development (JavaScript, CSS, HTML) and Chromium plugin development.
Applicants should also have an interest in and ideally prior experience with human-computer interaction and user experience research.
Applicants are encouraged to watch the WebPerceptor concept video ( https://youtu.be/MPSisruuTY0?si=xpgC4RCpUgIIAEIb ) and look at the project repository ( https://github.com/theartofhci/WebPerceptor ) when preparing their application.