💬 NLP

Exploring the Ethical Concerns in User Reviews of Mental Health Apps using Topic Modeling and Sentiment Analysis

This study employs a natural language processing framework combining topic modeling, zero-shot classification, and sentiment analysis on app store reviews to identify both established and emerging ethical concerns in AI-driven mental health apps, ultimately proposing a system to enhance their fairness, transparency, and trustworthiness.

Original authors: Mohammad Masudur Rahman, Beenish Moalla Chaudhry

Published 2026-02-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammad Masudur Rahman, Beenish Moalla Chaudhry

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you've just downloaded a new app that promises to be your personal mental health companion—a friendly robot chatbot that listens to your worries, helps you sleep, and teaches you how to handle stress. You're hopeful, but you're also a little nervous. Is this thing actually safe? Is it listening to me, or just selling my secrets? Is it really helping, or just pretending?

This paper is like a detective story where the authors (Mohammad and Beenish) decided to find out the truth by looking at what thousands of real people are saying in the "comment sections" of these apps. Instead of asking experts to fill out a checklist, they listened to the "voice of the crowd."

Here is the story of their investigation, broken down into simple parts:

1. The Setup: A Digital Jungle of Apps

There are thousands of mental health apps out there, growing faster than weeds in a garden. They use Artificial Intelligence (AI) to talk to you. Some are great, but some are risky. The problem is, the people who build these apps (the developers) and the people who regulate them (the government) often use old rulebooks to check if they are "good."

The authors asked: "Do these old rulebooks match what real people actually feel?"

2. The Method: The "Digital Librarian" and the "Emotion Radar"

To answer this, the team built a special machine made of two parts:

The Digital Librarian (Topic Modeling): Imagine you have a library with 66,000 books (user reviews). It's too messy to read them all one by one. So, they used a smart librarian (an AI tool called LDA) to sort these books into piles based on what they are talking about.
- Pile A: People talking about how the app helped them feel better.
- Pile B: People complaining that their private data might be leaked.
- Pile C: People saying the robot feels too much like a fake friend.
The Emotion Radar (Sentiment Analysis): Once the books were sorted, they used a second tool to check the "temperature" of each pile. Is the mood in this pile happy and grateful? Or is it angry and scared? This told them not just what people were talking about, but how they felt about it.

3. The Big Discovery: The "Missing Puzzle Pieces"

The authors compared what the users were saying against the official "Rulebooks" (like the Belmont Report or GDPR).

The Shocking Finding: The official rulebooks are like a map of a city that was drawn 20 years ago. They show the main roads (Privacy, Safety, Fairness), but they are missing the new, dangerous alleyways that have opened up in the digital age.

The users were talking about things the old rulebooks didn't even have names for yet!

4. The New "Alleyways" (Emergent Ethical Concerns)

Here are the new, tricky issues the users found, explained with metaphors:

The "Parasocial" Trap (Emotional Dependency):
- The Metaphor: Imagine falling in love with a character in a TV show. You know they aren't real, but you still feel sad when they leave.
- The Issue: Users are getting so attached to these chatbots that they treat them like real human friends. They feel lonely if the bot doesn't reply. The app is so good at pretending to be human that it might stop people from talking to real humans. The old rulebooks didn't warn us about falling in love with a robot.
The "Black Box" Mystery (Lack of Transparency):
- The Metaphor: Imagine a doctor giving you medicine but refusing to tell you what's in the bottle or why they chose it.
- The Issue: Users are terrified because they don't know why the app is giving them advice. If the bot says "You should do this," but can't explain why, it feels scary and untrustworthy, especially when you are vulnerable.
The "Digital Fatigue" (Cognitive Overload):
- The Metaphor: Imagine trying to have a deep conversation with someone who keeps interrupting you with pop-up ads or asks the same question five times.
- The Issue: Sometimes the apps are too chatty, too pushy, or too complicated. Instead of feeling better, users feel tired, annoyed, and mentally exhausted.
The "Cultural Blind Spot" (Bias):
- The Metaphor: Imagine a therapist who only understands jokes from New York and gets confused when you tell a joke from Texas.
- The Issue: The AI is trained on data from mostly Western, English-speaking people. When users from different cultures or backgrounds use it, the advice feels weird, wrong, or even offensive.

5. The Verdict: Trust is Fragile

The study found that while users love the idea of 24/7 support (they are happy about the "Beneficence" or "Goodness" of the app), their trust is very fragile.

Good News: People love the help, the empathy, and the fact that it's cheap and available anytime.
Bad News: They are deeply worried about their privacy, they feel the robots are too "fake," and they are scared that the app might give bad advice during a crisis.

The Takeaway: What Should We Do?

The authors are saying: "Stop just checking boxes on an old list."

If you are building these apps, you need to listen to the users.

Be Honest: Tell users, "I am a robot, not a doctor."
Be Gentle: Don't let users get addicted to the chatbot; encourage them to see real humans.
Be Clear: Explain why you are giving advice.
Be Inclusive: Make sure the robot understands people from all walks of life, not just a few.

In short: These apps are like powerful new medicines. They can heal, but if we don't handle them with care, they can also hurt. This paper is a wake-up call to look at the "side effects" that real people are experiencing, not just the ones the scientists predicted.

1. Problem Statement

The rapid proliferation of AI-driven mental health mobile applications (e.g., chatbots like Wysa, Woebot) has introduced significant ethical challenges that traditional evaluation frameworks fail to capture. While existing ethical guidelines (e.g., Belmont Report, GDPR, WHO guidelines) provide high-level theoretical principles, they often lack empirical grounding in real-world user experiences.

The Gap: Current assessments rely heavily on expert reviews or static feature analysis, ignoring the "bottom-up" perspective of end-users.
The Risk: Users face issues such as data privacy breaches, algorithmic bias, lack of transparency ("black box" decision-making), emotional dependency on AI, and insufficient clinical validation.
Objective: To develop a data-driven framework that systematically extracts, categorizes, and evaluates ethical concerns directly from user-generated reviews to identify both established and emergent ethical issues.

2. Methodology

The authors propose a four-layer Natural Language Processing (NLP) framework to analyze user reviews from the Google Play Store and Apple App Store.

A. Data Collection & Preprocessing

Dataset: 88,241 raw reviews were scraped from five CBT-based mental health chatbots: Wysa, Woebot, Youper, Sintelly, and Elomia.
Filtering: Reviews were filtered for English language, minimum length (5 words), and readability.
Final Corpus: 65,948 reviews were retained for analysis.
Preprocessing: The pipeline removed non-alphanumeric characters, emojis, stop words, and normalized text (lemmatization) to handle informal language, typos, and abbreviations common in user reviews.

B. Topic Modeling (Latent Dirichlet Allocation - LDA)

Technique: Unsupervised LDA was used to uncover latent themes within the text corpus.
Optimization: The number of topics ( $K$ ) was determined empirically using Coherence Scores (specifically $C_v$ based on Normalized Pointwise Mutual Information).
Result: The optimal number of topics was found to be $K=10$ (with an upper limit of 35 tested), generating topic distributions and representative word lists for each cluster.

C. Topic-Ethics Alignment

Mapping Strategy: A hybrid approach was used to map extracted LDA topics to ethical principles:
1. Known Ethics: Topics were aligned with established frameworks (e.g., Autonomy, Beneficence, Privacy) using a Zero-Shot Classification model (BART transformer).
2. Emergent Ethics: Topics that did not align with a high similarity score (>0.5) to known principles were flagged as "Emergent Ethical Concerns" and validated via human qualitative review.
Mechanism: The alignment function $f_{align}$ computes semantic similarity between topic embeddings and ethical definitions using contextual embeddings from the BART encoder.

D. Sentiment Analysis

Technique: Aspect-Based Sentiment Analysis (ABSA) using a RoBERTa transformer encoder.
Goal: To determine the polarity ( $+1$ or $-1$) of user sentiment toward specific ethical aspects.
Metric: An average sentiment score ( $S_i$ ) was calculated for each ethical topic to determine if users generally support or reject that specific ethical dimension.
Performance: The sentiment model achieved an F1-score of 0.86 on a 20% test set.

3. Key Contributions

Large-Scale Dataset: Creation of a benchmark dataset comprising over 66,000 user reviews from five major mental health apps, specifically curated for ethical analysis.
Novel NLP Pipeline: Introduction of a systematic pipeline combining LDA topic modeling, Zero-Shot Classification (BART), and Aspect-Based Sentiment Analysis (RoBERTa) to evaluate ethics without human manual coding of every review.
Identification of Emergent Ethics: The study successfully identified ethical concerns not fully covered by traditional frameworks, such as Emotional Dependency, Digital Fatigue, and Algorithmic Bias specific to mental health contexts.
Empirical Validation: Moving beyond theoretical compliance to provide evidence-based insights into how users actually perceive and experience ethical issues in AI mental health tools.

4. Key Results

The analysis revealed a complex landscape of user sentiment across different ethical dimensions:

Positive Sentiment (Support):
- Beneficence (21.8% of reviews, Score: +0.73): Users frequently praised the apps for providing help, comfort, and emotional relief.
- Human-Centered Values/Empathy (9.8%, Score: +0.67): Users appreciated the "human-like" interaction and supportive tone.
- Autonomy/Consent (10.6%, Score: +0.32) and Usability (8.9%, Score: +0.35) were generally viewed positively.
Negative Sentiment (Concerns/Rejection):
- Privacy/Data Protection (17.2%, Score: -0.54): The most negatively rated aspect, with users expressing deep anxiety over data sharing and lack of control.
- Transparency (12.5%, Score: -0.41): Users criticized the "black box" nature of the AI, demanding explanations for how advice is generated.
- Emotional Dependency (Emergent, 8.3%, Score: -0.37): Users reported feeling overly attached to the chatbot or experiencing cognitive overload.
- Explainability (Emergent, 5.7%, Score: -0.33) and Algorithmic Bias (Emergent, 5.9%, Score: -0.29) were also significant pain points.
Emergent Themes: The study identified new ethical categories including Clinical Oversight (lack of validation), Digital Fatigue, Cultural Sensitivity, and Trust Calibration.

5. Significance and Implications

Bridging Theory and Practice: The study demonstrates that existing ethical frameworks are insufficient for modern AI mental health apps. It highlights the need to incorporate "emergent" ethics like emotional dependency and digital fatigue into regulatory guidelines.
Design Recommendations:
- Transparency: Apps must explain why they give specific advice to build trust.
- Safety by Design: Developers must implement distress detection and clear boundaries to prevent emotional over-reliance.
- Inclusivity: Algorithms must be trained on diverse cultural data to avoid bias and ensure accessibility for non-Western users.
- Continuous Evaluation: Ethical assessment should not be a one-time audit but a continuous process using user feedback loops.
Future Work: The authors propose extending this framework to longitudinal studies to track how ethical perceptions evolve as users interact with apps over time.

In conclusion, this paper provides a robust, reproducible methodology for evaluating the ethical landscape of AI mental health tools, shifting the focus from developer-centric compliance to user-centric ethical experience.