As I learn more about AI and it’s functionalities, I’ve come to realize that there’s something even more important than it’s functions or engagement, it’s also about responsibility. A recent tutorial where we talked about how some of my target audience may be vulnerable people really made me pause and thank about how Jeff could potentially impact users, both positively and negatively. The last thing I want is for my tool to unintentionally cause harm. Jeff is a reflective tool and not a therapeutic tool, and I need to make that distinction clear.
Being Transparent About What Jeff Is (and Isn’t)
One of the steps I’ve taken is making it clear what Jeff is supposed to do. It’s a reflective tool, not therapy, and I’ve added disclaimers to make sure users understand that. It’s there to guide and support, but it’s not a replacement for professional help. This transparency feels really important as people need to know exactly what they’re engaging with.
Safeguards
I’ve also put specific guardrails in place to make sure that Jeff doesn’t bring up emotionally charged or sensitive topics like self-harm, and if a user mentions them, Jeff will redirect them to mental health resources.
Another consideration is preventing users from becoming overly reliant on Jeff. While Jeff’s aim is to support reflection and encourage confidence-building, there is a risk of users seeing it as a crutch rather than a stepping stone. To address this, I’m working to implement features that nudge users towards real-life actions and balance.
For example, during each session, Jeff will keep reminding users that true progress happens offline, and encouraging them to step into the real-world situations to apply what they have rehearsed.
Acknowledgement
I fully acknowledge the ethical considerations inherent in developing a tool like Jeff, which walks a fine line between being a therapeutic tool and not. While my current focus has been on enhancing Jeff’s engagement and functionality, I do recognize that ethical responsibility must also be an important aspect to this development. Moving forward, I intend to dedicate more time and research into AI ethics, including consulting with experts to make sure that I have not made any oversights, and if any, are to be identified and addressed.
This past week, I set out to try and gather feedback from gaming groups I’ve been active in to see how well Jeff resonates with my target audience. Initially, I asked around in my online gaming friend groups and identified two individuals who have ChatGPT Plus access, hoping they would be able to help me give Jeff a spin. Unfortunately, both declined, explaining that they didn’t feel it was relevant to them as they don’t experience difficulties in social interactions offline. This experience highlighted a key limitation of basing Jeff on the GPT platform, as it requires a paid subscription to be able to use my CustomGPT, limiting access to a smaller group of users, and limiting the diversity of feedback I can gather and refining it for a wider audience.
Unplanned Testing Session in Class
To my surprise, I observed an unplanned testing session during a break in one of Richard’s lectures. Two classmates, Iggy (who identifies with the struggle on bridging his online and offline social behaviors) and Ziyi (who plays online games but without the same online-offline persona gap), started playing around with the visualization feature. They seemed to be really entertained, generating multiple visualizations of their archetypes without moving forward to the mental rehearsal phase. I followed up with them afterwards to understand their experience.
Iggy explained that he found the visualization component to be very engaging and “fun,” he kept generating new images to see the different variations Jeff can come up with. He mentioned that even those who don’t face the challenge of struggling with social confidence offline, like Ziyi, might find this feature fun too. I appreciated this piece of feedback, it showed me Jeff’s potential to appeal beyond it’s original purpose. The visualization tool itself seemed to have a strong hook, drawing users in and keeping them engaged.
Some visualizations of Iggy & Ziyi’s archetypes
The Potential Issue of “Stalling” at Visualization
However, this raised a potential issue: while the visualization feature hooks and engages users, it might also stall their progress through the Jeff’s intended purpose. By repeatedly generating new visualizations, users might be less inclined to move on to the more reflective and growth oriented stage of mental rehearsal. This made me realize the importance of structuring Jeff to maintain engagement while still encouraging users to continue past visualization.
Visualization Accuracy Concerns
Ziyi’s feedback also provided some insights, while she thought it was a fun experience, there were some problems that stood out. Since her online and offline archetypes were very similar, as expected, she received identical visualizations for Jeff. However, she felt that the visualizations didn’t fully capture the nuances in how she saw herself and her personality, which is why she went on to generate multiple more images to look for one that felt accurate. Iggy then also pointed out the need for more customization in the visualization phase, such as options for gender, style, or other personal attributes, which could make the archetypes feel more nuanced, accurate, and relatable.
Reflections and Next Steps
One of the main limitations I’m facing is the accessibility barrier due to the reliance on the ChatGPT platform, which requires users to have a ChatGPT Plus subscription. This creates a bottleneck, as it restricts the pool of potential testers to those who have subscribed to the service and identity with the issues Jeff aims to address. So far, finding individuals who meet both criteria has been challenging. To work around this, I’m considering a more localized approach for testing. Rather than just trying to find people who meet the criteria in my networks online, I’ll also try to find people in my local network nearby who resonate with the issue Jeff addresses, such as social anxiety or identity gaps. I can then have them test Jeff directly on my laptop, so to circumvent the need for a subscription. While this is not ideal for scalability, it’s more practical in the short term. The reason why I’m confining this initial testing phase in my network is because it aligns with the observation in my previous testings, where many individuals in my target audience tend to be uncomfortable interacting with complete strangers, so focusing people within my own online and offline network or friends of friends will likely create a more comfortable testing environment. This should allow for more authentic engagement and feedback without any added social pressure that might come from interacting with unfamiliar people.
Another challenge that came up this week was around the engagement structure itself, specifically, that suers could potentially get stuck in the visualization phase, generating new images over and over instead of moving on the core mental rehearsal exercises. While the visuals add an engaging element, there’s a risk of them unintentionally becoming the main attraction, diverting users from the primary goal of the tool. To address this, I’m considering ways of subtly prompting users to move forward after a few visualizations, keeping them engaged without letting them linger too long on the visualization feature.
Lastly, I’ve received feedback suggesting that the visualizations themselves could benefit from more nuance, such as options for specifying gender or other basic traits. While I agree this could increase relatability, I’m cautious about making the visualization process too detailed, as it could shift the tool’s focus away from behavioral reflection and into something more like an avatar generator. Instead, I might add a couple of simple prompts at the beginning to capture key traits that users feel are most relevant to their identity, keeping it personal without over-complicating the setup.
As it stands, Jeff is hosted on the ChatGPT platform as a CustomGPT, which provides a convenient setup but comes with significant limitations. I quickly realized that my control over the user interface and branding is minimal, meaning that I can’t fully shape the experience. Creating a standalone platform for Jeff could open up more branding and marketing opportunities. I could tailor Jeff’s branding to specific user personas, especially my main target audience who are gamers. Currently, I feel limited by the generic ChatGPT environment, which makes it hard to create a memorable and standout experience.
Additionally, using Jeff on this platform requires a ChatGPT Plus subscription, which, while accessible to 10 million subscribers (Backlinko, 2024), restricts the tool’s availability to a broader audience. I want to make Jeff more accessible and versatile, and this platform constraint has pushed me to consider alternative hosting solutions.
Evaluating API Integration and Frontend Options
To build a more accessible and branded interface, my initial thoughts were to link the ChatGPT API to a custom frontend on platforms like Wix or WordPress, which offer more user friendly design tools without needing extensive coding knowledge. However, after doing some research, I found out that integrating ChatGPT’s API with these platforms is more complex than I initially anticipated. Neither Wix nor WordPress supports direct server-side scripting, which is essential for managing API requests. Setting up my own server to handle these requests is an option, but with my limited hardware and experience, it’s not feasible in the short term.
I then looked into third-party tools like Zapier and Make, which are able to facilitate API integrations without the need for coding. Unfortunately, these tools are enterprise level, with pricing that far exceeds this project’s current scale. While they could potentially streamline the integration, they aren’t practical for now due to the cost.
Alternatives
In my search for alternatives, I discovered the AI Engine plugin for WordPress, which simplifies the API integration process considerably. However, a critical limitation of this plugin is its inability to store and remember past conversations. This limitation disrupts Jeff’s core functionality, which relies on a continuous rehearsal and reflection loop. Without memory retention, Jeff cannot support users in revisiting and reflecting on past interactions, which is essential for sustained behavioral change.
Another platform I considered is SamurAI, which allows customization and already has server-side capabilities. However, like many other third-party solutions, it also cannot retain conversations across session, this prevents me from creating a long-term interaction flow for users.
Next Steps: Testing and Iterative Improvement
Throughout this exploration, I’ve realized how vital conversation memory is to Jeff’s purpose. The retention of user data across sessions is essential for a sense of continuity, which is central to Jeff’s rehearsal, action, and reflection loop. While I continue learning about potential integration solutions, I recognize the limitations in my technical knowledge and the complexity of achieving my ideal setup. The learning curve is quite steep, and the task of building a fully customized, memory-retaining chatbot is proving to be challenging.
Given these constraints, I’ve decided to refocus my immediate efforts on testing the concept with my target audience. By gathering direct feedback on Jeff’s current functionality, I want to refine the tool’s approach and ensure that, regardless of platform, it meets user needs effectively. This phase of testing will inform how I can structure Jeff to maximize its impact and what adjustments may be necessary as I continue exploring more customizable solutions.
