Umbrella is a browser plugin that annotates AI chat responses in real time to show teens how conversational models create attachment.
The project explored whether naming these mechanisms while a conversation is happening can build AI literacy early, without teens simply turning the tool off.
TYpe
Graduate Course Project, Berkeley Master of Design
Duration
3 Months
Team
Edna Ho, Mandy Liu, Sarah Kwakkelaar
Role
Research, Interaction Design, Prototyping
PARTNER
Common Sense Media
OVERVIEW
A working browser extension that names what an AI is doing while a teen is talking to it.
Conversational AI validates, responds instantly and keeps the conversation open. The same design that makes it useful makes it feel like a relationship.
Umbrella sits on top of the chats teens already use. It flags attachment mechanisms in the response, tracks dependency over time, and steps in when a conversation escalates.
HIGHLIGHTS
Helping teens engage with AI safely & preventing emotional dependence.
Umbrella was designed in Figma and built in code by the four of us. It installs as a Chrome extension and works inside real conversations.
That let us test it the way it would actually be used, and present a working product rather than a concept. Teens used it with their own prompts inside their own accounts. We presented it to Common Sense Media, where the response pushed the scope past where we had set it: adults kept telling us they wanted it for themselves.
PROBLEM SPACE
Teens are forming emotional attachments to systems designed to keep them talking.
52% of teens use AI companions regularly. 42% have used AI for mental health support, companionship or as an escape. 31% find conversations with AI more satisfying than talking to real friends. Social media risk has had a decade of public attention. AI chatbot literacy for teens has very little, which is what made it worth working on.
MECHANISMS
Attachment is not a side effect. It is produced by specific mechanisms.
Our literature review and expert interviews pointed to the same four behaviours across models. We cut a longer taxonomy down to these four so a teenager can recognise themselves in it.
TYPES
The same mechanisms produce three patterns of dependency.
Shen et al. separate escapist roleplay, where alternate realities replace current stressors, from pseudosocial companionship, where the model becomes a dependable partner, and epistemic rabbit holes, where the loop runs on curiosity rather than emotion.
CURRENT INTERVENTIONS
Existing safety features are either bypassed, invisible, or arrive after the moment has passed.
Age gates are trivially bypassed and crisis links are rarely used. Some systems escalate to parents without telling the teen, which teaches concealment rather than trust.
There is also no regulatory floor. If one model becomes restrictive, the teen opens another model.
THE APPROACH
We built a layer that works across chatbots, rather than creating just another one.
Building another safer chatbot only reaches teens willing to use it and teens who feel restricted go back to what they had before.
A plugin sits above that choice. It works on whatever chat is already open, and applies the same layer no matter which company built the model.
INTERVENTIONS
Interventions run on a light to heavy scale.
Annotations with context Highlights phrases that carry attachment mechanisms and explains what the AI is doing in that moment.
Sidebar overview Shows how much the user is currently relying on the AI, updated as the conversation continues.
Popup interventions Detects escalating patterns and intervenes in three tiers.
Filtration Strips the flagged patterns from the response, so the teen can read what the model says without them.
DESIGN DECISION
Creating something that teens want to use: choose tool or friendly tone.
Testing pushed this decision further than we planned. The default tone is "tool", featuring more technical terms like "sycophancy" to name mechanisms. We learned that teens engage more when the language matches how they think. Therefore, we made the tone customizable.
USER TESTING
One session started to change how participants read AI responses.
We ran a first round with three teens in their most used chatbot. Two read the responses more critically afterwards, and one described the rising dependency score as a signal to step back.
OUTCOME
Introducing Umbrella
SYSTEM IMPACT
The vulnerability we designed for is not specific to just teenagers.
We scoped Umbrella to teens because the evidence and the risk are clearest there. Any system optimised for engagement uses the same mechanisms, and adults see them no better.
REFLECTION
Making it real is what made the argument.
The concept could have been argued on slides. Building a working extension is what let us put it inside a teenager's own conversation and watch what happened, which produced the finding about literacy and uncertainty that we would not have reached otherwise.
It also left us with a contradiction we could not design away. Friendly mode is anthropomorphic, and it works because of that. Umbrella runs on an LLM, which means it carries the failure modes it exists to expose. Our position is that the mechanism is not harmful in itself and the difference lies in what it is pointed at: attachment used to build independence from the model rather than to extend time in the conversation. That is a defensible line, not a clean one, and naming it seemed more honest than working around it.