Qualitative UX research exploring how users respond to different prototypes of a browser-based tool for identifying online misinformation.
Qualitative UX research exploring how users respond to different prototypes of a browser-based tool for identifying online misinformation.
Cindy Gilbertson and Matthew Costa, co-founders of Euclid, an early stage startup tackling online misinformation
Matt and I have been working in cybersecurity-adjacent fields for over a decade, primarily focusing on mission-critical products in highly regulated environments. At Callsign, we worked on securing one of the world’s first behavioural biometrics tools in global payments, and at Optimizely my product line provided critical backend systems for most European governments. We are used to working in adversarial environments, where threats have been evolving in complexity and scale, and failure would have had catastrophic impact.
We are deeply concerned about the impact AI has on the scale and spread of misinformation online. We've seen the impact directly, our friends and acquaintances have become increasingly divided and polarised on politics. The information they receive is delivered through algorithms intended to maximise engagement at the cost of the truth. Having built multiple commercially successful products end to end, we have turned our attention full time to restoring trust and transparency on the internet. In our first iteration, we are looking at creating a browser extension that alerts users to online misinformation and AI-generated content.
Our solution is at an early prototype stage, and we have completed two rounds of user testing to learn how Internet users currently navigate and verify online information, and their willingness to use a browser-based tool to help them better evaluate information.
Our user testing revealed that people are
Generally, we received positive feedback on our proposed solution, but our individual users differed in how they would like the UI to behave in various alert scenarios. For example, some wanted more layers of information present, whilst others wanted to minimise cognitive overload. Since we want to ensure that even non-technical users are comfortable using the product, and can understand its warnings, we need to invest more time and effort into improving the user experience.
We are asking for funding to help optimise the user experience by investing time into redesigning the user flows of our browser extension, expanding the types of scenarios where one may encounter warnings, and testing the new prototype on a larger, more diverse group of users. Our previous cohort included 7 users, we would like to increase this to 30 users, from a wide range of backgrounds and level of familiarity with technical tools, and AI.We will conduct moderated, qualitative user interviews to understand what works well, and what might cause issues to users, in particular vulnerable users. In these interviews, we will model realistic scenarios that the users could encounter on the Internet, and observe how they interact with our prototype in those settings.
Example questions we would like to get answered include:
We will produce a research report that will contain our key findings from these interviews. This will inform our product development, but we would also be happy to share non-sensitive findings with the rest of the industry. It is not within the scope of this proposal to build a production-ready solution or evaluate any particular backend tools’ ability to identify misinformation, we are only focused on understanding how to create an optimal user experience.
AI can be used by malicious actors to generate large volumes of content, which can easily be tuned to manipulate public opinion at scale. Deepfakes have been used in the context of elections in a number of countries, including the US and the UK, and are often deployed by hostile states, for example Russia’s ‘Operation Overload’ which sought to reduce international support for Ukraine. Similar campaigns are operated by Iran, with other examples having been identified in countries such as Nigeria and India.
Trust erosion is a key problem we need to be able to address. Since November 2024, the majority of online content has been generated by AI rather than created by humans. In our user research, participants could not quantify how much AI generated content they are interacting with, and as the technology improves, it is increasingly difficult to identify. However, signs of mistrust are increasing - 31% of Europeans believe AI has directly influenced their voting.
There are worrying signs of societal tensions being increased by the use of AI-generated content. For example in the aftermath of the Southport terror attack in the UK, AI-generated social media content was used to spread Islamophobic sentiments. Deepfake tools also increase gender-based violence, with 98% of all deepfake videos being non-consensual deepfake pornography targeting women.
Team Member
No comments yet. Be the first to share your thoughts.
Some technical tools are being developed by generative AI providers to check if a particular piece of content has been generated by them, but these tools are not embedded into the workflows of ordinary users, require additional time and effort to use, and are not even particularly accurate. There are broader solutions, such as ZeroGPT, that are trying to look at identifying AI-generated content from different sources, but these are also not in the natural workflow of internet users, and are unable to identify or alert on misinformation.The people most at risk from misinformation campaigns do not go to external tools to verify individual pieces of content - we need a solution that comes to the user and provides a solution where they naturally interact with the content, rather than putting additional burden on them.
Some legislative changes, for example via the EU AI Act, are coming to help ordinary users, for example by forcing platforms to declare when a piece of content is generated by AI. Unfortunately, these have limited regional impact, take a long time to roll out, and do not cover misinformation as a broader category. They are also ineffective against malicious actors, such as hostile-state-driven propaganda campaigns.
Tools that ordinary citizens can use to identify and tackle misinformation will be key to fight back against the erosion of trust, but poor user experience that puts additional burden on the user will likely make the technology go unused. This is why we want to focus our research on finding out what good user experience would look like, even for the least technical users. If we can create a tool that is understandable by any Internet user, they will be able to fight back even where legislative changes are not rolled out, or are insufficient to ensure safety.