TwinLink AI

Inspiration

TwinLink AI grew out of our discussions about how people currently use AI and messaging apps in everyday life.

We identified several common problems.

First, most messaging platforms have not yet deeply integrated AI into communication between users. AI is often used for moderation or customer support, while the core experience of connecting and communicating with other people remains largely unchanged.

For example, when someone is temporarily unavailable, a basic automatic reply cannot understand relationship differences or privacy boundaries. A family member, coworker, and business contact should not receive the same information, even though some conversations still require a quick response.

Second, valuable conversations with GPT are usually separated from real-world group discussions. A user may ask GPT to create a travel plan, copy the result into a messaging app, collect feedback from family members, manually organize everyone’s opinions, and then send everything back to GPT. This process is fragmented and inefficient.

Third, communication between people often involves reminders and coordination among family members, coworkers, or organizations. Most existing reminder tools focus only on reminding the user, rather than supporting reminders within real social relationships.

Fourth, conversations between people contain valuable memories and contextual information. They reflect relationships, preferences, plans, opinions, and shared experiences, but most messaging platforms do not make meaningful use of this information.

These problems inspired us to bring AI conversations and human communication together in one system.

How We Built It

We focused TwinLink AI on two connected features.

The first is a permission-aware digital twin.

Before the AI responds on behalf of a user, the system checks who is asking, what relationship they have with the user, what information they are allowed to access, and whether the request should be answered, refused, or handed back to the real user.

The second feature is the distribution and collaborative discussion of GPT-generated content.

Users can chat with GPT directly inside TwinLink, turn a GPT response into a shareable card, and send it to selected contacts. Participants can discuss the same card together, and the complete discussion can then be returned to the original GPT conversation. GPT can use everyone’s feedback to create an updated summary, plan, or recommendation.

We built TwinLink as a working application for the web, Windows, and Android. The OpenAI Responses API powers AI conversations, content summaries, memory processing, and digital-twin replies. Codex helped accelerate requirement analysis, development, testing, deployment, and bug fixing.

Challenges We Faced

Our biggest challenge was not generating AI responses. It was deciding when the AI should be allowed to respond.

We learned that simply telling a model not to reveal private information is not reliable enough. Permissions must be checked by the application before the model is called. Only information that the current contact is authorized to access should ever be sent to the model.

We also needed to prevent the digital twin from pretending to be the real person. Every digital-twin response is clearly labeled as AI-generated, and the twin can operate only within a limited scope. Questions involving sensitive information, important decisions, commitments, or personal judgment are handed back to the real user.

Another challenge was preventing features from being abused. For example, social reminder tools could potentially be used to send repeated or unwanted messages, so they require recipient permissions, usage limits, and expiration controls.

We also had to preserve context when moving information between a private GPT conversation and a group discussion. The system must correctly connect the original GPT response, the shared card, participant feedback, and the updated result generated by GPT.

What We Learned

Through this project, we learned that trustworthy social AI requires more than a powerful model.

It also requires clear identities, relationship categories, permission controls, memory boundaries, visible AI labeling, and a reliable way for humans to take control.

We also learned that AI does not need to replace communication between people. It can help people continue conversations, organize information, collect different opinions, and turn group discussions into better shared outcomes.

The goal of TwinLink AI is to make AI a transparent, controlled, and permission-aware participant in human communication—not a replacement for the people involved.

What's next for TwinLink AI

Our next step is to improve the user interface and expand TwinLink AI into a more complete AI-powered social platform.

Planned features include group chats, richer social and media-sharing tools, personal social feeds, business communities, and professional networks.

We also plan to explore digital memory legacy and inheritance, allowing users to securely preserve selected memories, experiences, and personal knowledge, and authorize trusted family members to access them in the future.

These additions will extend TwinLink AI beyond messaging and turn it into a platform where personal AI, digital twins, memories, relationships, and communities can develop together.

Contributors

TwinLink AI was developed as a family collaboration.

Yixuan Li contributed to the development and testing of the application. She also appeared in and recorded the project introduction video, helping present TwinLink AI’s core features and user experience.

Built With

• api
• gpt-5.6
• openai
• responses