🚀 [Product Case Study] Safety Strategies for CiCi AI: Ensuring Trust in GenAI

What is CICI AI?

Think of CICI AI as your AI chat companion for conversations, writing, translation, and emotional support.

https://www.ciciai.com/chat/

[Assumption] Business Goal:

Ensure a safe and trustworthy environment by complying with CiCi’s Community Guidelines and privacy policies.

Who are the Stakeholders of CICIAI?

How should we be segmenting these users?

Prioritizing Consumers: Building Trust Through Safety

Building trust is key. When users feel safe, they share data confidently, return for positive experiences, and recommend the app. Keeping users safe isn’t just good for them, it’s good for business. Nobody likes a product that sprouts harmful content.

I did a survey focusing on consumers to understand their key safety concerns when using CICI AI.

User Survey Findings

High level Summary

• Participants: 24 users aged 18–24, all from the Philippines.
• Usage: 54% use CiCi AI for academic purposes, 25% for personal assistance, and the remaining for emotional companionship or work.
• Safety Perception: Half of the participants felt safe using CiCi AI, while the other half had concerns.

Survey of 24 participants, mostly aged 18–24 and all comes from Philippines.

54% used CiCi AI for academic purposes, 25% for personal assistance, and the remaining for emotional companionship or work.

Understanding Safety Angles in Human-Cici Interactions

For users to feel safe, Cici AI needs to address various aspects of user well-being and platform integrity. These include:

How Safe Do Users Feel with Cici AI?

While half of the 24 participants who took the survey felt safe, which is a positive sign, the other half had some concerns 🤷.

Top 3 Key Safety Concerns

From Static to Smart: Why Building AI is Different

Building AI is different from building a traditional app. Traditional apps are like pre-programmed machines, offering the same features every time. CiCi AI, on the other hand, is constantly learning and improving. To do this, it needs high-quality data, just like a personal trainer needs clear instructions to create effective workouts.

This explains why transparency about data collection is so important. Users deserve to understand how their data helps CiCi AI learn and grow.

With this understanding, let’s explore the strategies for addressing the main user concerns: data privacy, AI hallucination, mathematical accuracy, and biased algorithms.

1. 📊 ADDRESSING DATA PRIVACY CONCERNS

User’s feedbacks on data privacy issues

Our recent survey revealed a key concern: some users felt Cici AI was asking for too much personal information. They wanted to know why and feel more in control.

While Cici AI already has a privacy policy listed, let’s further explore how we can address some of the existing gaps:

❌ Observed Gap (1): Inaccessible Privacy Policy

• Current Situation: The privacy policy requires multiple clicks to find, making users unaware of EXISTING data privacy guidelines.

Proposed Solution:

• Make it Findable: Provide a clear link to the privacy policy in the app’s main menu or where data is collected.
• Contextual Privacy Messages: Imagine you’re asking Cici AI, “What are some good restaurants nearby? Currently, there might not be an immediate explanation about location data collection. With contextual explanations, a small pop-up could appear:

“To find nearby restaurants, Cici AI will access your location with your permission. You can disable location sharing at any time in your privacy settings.”

❌ Observed Gap (2): Lack of perceived control over data

Current Situation: CICI AI app lacks an interface for users toeasily manage their data privacy preferences, leading to a feeling of powerlessness.

Proposed Solution

Data Privacy Dashboard: Develop a user-friendly dashboard where users can:

• Access and Review Data: View the data CiCi AI has collected (e.g., conversation history, saved preferences).
• Manage Data Usage: Control how their data is used (e.g., toggle on/off data collection for specific features).
• Data Deletion Controls: Delete data directly within the app (e.g., all data, specific data types, or data for a certain timeframe).

2. 🌀 ADDRESSING AI HALLUCINATIONS CONCERNS

Types of AI hallucination

Why does it happen?

Large Language Models (LLMs) sometimes “hallucinate,” or make things up, because of the data they’re trained on. They learn from huge amounts of text written by humans. If humans have biases or make mistakes, the AI can pick up those same biases and errors. When the AI tries to generalize this information, it might produce incorrect or biased responses.

From a fine-tuning model standpoint:

Product-Level Recommendation:

1. Empowering the user to select their data quality

To get accurate answers from AI, it’s important to specify the type of sources it should use. For example, if you’re looking for academic info, choose scholarly databases and peer-reviewed journals. This helps cut down on AI hallucinations.

Set a focus on which sources to improve the accuracy of the prompt by providing context

2. Research mode
If accuracy is super important, I would suggest a “RESEARCH MODE”. This mode can generate detailed reports with multiple citations for each sentence, linking back to verified sources. The trade-off here is speed versus accuracy, with research mode focusing on thoroughness over quick responses.

Experimenting on Cici’s AI

Trying Out Cici’s AI: When I tested Cici’s AI, it gave me exact sources the first time but none the second time. Showing sources is the first step to making the info more credible, but we can go deeper.

Get Detailed Citations: Tools like Perplexity AI provide citations at the paragraph level, but we can do better. By giving citations for each sentence, we can spot potential hallucinations when there are sentences without any citations.

Perplexity AI provides citation at a paragraph level but you can go a layer deeper and provide it for each sentences. Sentences without any citations = hallucination by LLMs.

3. ⚖️ ADDRESSING BIASED ALGORITHM CONCERNS

Biased information in CICIAI simply means the LLMs will try to answer in a specific way based on what it is trained on.

• AI systems rely on human-sourced information, and unfortunately, humans themselves can harbor unconscious biases
• Even the most well-intentioned data curation efforts might not eliminate this completely.
• Some content may lack the detail or complexity needed for a balanced perspective.

My proposed Approach to Making CiCi AI Fairer:

1. Finding Diverse Data

Actively seek out a wide range of data sources representing different backgrounds and experiences.

2. Improving Existing Data

Imagine CiCi AI recommends tech jobs based on skills and interests. If the training data associates “coding” or “engineering” mainly with male profiles, it could lead to biased recommendations for women in tech roles.

How We Can Improve Data for Gender Bias:

Imagine Cici AI is recommending tech jobs to users based on their skills and interests. If the training data primarily connects keywords like “coding” or “engineering” with male profiles, this could lead to biased recommendations for women interested in similar tech roles.

Here’s how we can improve the data for gender bias in the tech world:

• Include resumes and job descriptions from both men and women in various tech roles (e.g., software developers, data analysts, cybersecurity experts).
• Add open-source project contributions or coding competition entries from diverse teams and individuals.

* **Adding Variations:** We can add data variations by including:
    * Resumes and job descriptions from both men and women in various tech roles (e.g., software developers, data analysts, cybersecurity experts). 
    * Open-source project contributions or coding competition entries from diverse teams and individuals. 
* This helps the algorithm understand that these skills and achievements are not gender-specific.

3. Bias Evaluation and Testing

Regular evaluation and testing for biases are essential, utilizing both automated tools and human reviewers to assess LLM outputs across different scenarios and inputs.

Measuring Success: Key Metrics

Now that we have explored the solutions, let’s look at how we can measure their success. Here are some key metrics: