As artificial intelligence systems are integrated into critical infrastructure, from judicial sentencing to medical diagnostics, the necessity for rigorous ethical oversight has increased. For high school students, engaging in ethical AI research facilitates the development of a highly technical, multidisciplinary skill set. It teaches you to ask questions beyond basic programming and explore the intersection of data science, philosophy, and public policy. By conducting structured investigations into algorithmic fairness, transparency, and accountability, you can gain experience with advanced methodologies, including dataset auditing, sentiment analysis, and bias mitigation. This work prepares you for high-level academic research and professional roles in a field where technical proficiency must be balanced with ethical responsibility.
Why should students conduct research in ethical AI?
Engaging with the ethics of AI allows you to transition from being a passive consumer of technology to an active investigator of its societal impacts. In 2026, the demand for “AI Safety” and “Responsible AI” specialists has surged as regulatory frameworks such as the EU AI Act and NIST’s AI Risk Management Framework have become global standards. Exploring these topics lets you develop expertise in tools like Python-based fairness libraries (AIF360, Fairlearn) and explainability frameworks (SHAP, LIME), which are standard-use in industrial and academic R&D. Such research also hones your ability to perform empirical analysis on “black box” systems, a skill that is highly transferable across all STEM and social science disciplines.
Given these factors, we have outlined below a list of 30 ethical AI research questions to help you get started with some preliminary research.
30 Ethical AI Research Questions for High School Research
1. Quantifying demographic parity in automated resume screening
You can investigate how natural language processing (NLP) models used in recruitment might inadvertently penalize specific demographic groups based on proxy variables in resumes. The core research question focuses on whether removing explicit gender/race markers is sufficient to ensure neutral outcomes in high-volume hiring. It’s a good opportunity to utilize Python with the Scikit-learn and Fairlearn libraries to test for disparate impact and equalized odds. You can also refer to data sources such as the Kaggle “Resume Dataset” or to synthetic data generated to simulate diverse candidate pools. This project connects to Data Science, Sociology, and Human Resources Management.
2. Evaluating the accuracy of AI dermatology tools across Fitzpatrick skin types
The core question is whether computer vision models for skin cancer detection maintain consistent sensitivity and specificity across various skin tones. You can use PyTorch or TensorFlow to perform a subpopulation analysis on a pre-trained model like ResNet. The methodology involves auditing the ISIC Challenge Dataset to assess the underrepresentation of darker skin tones and then measuring false-negative rates for each category. Such research incorporates aspects of Biomedical Engineering, Computer Vision, and Public Health.
3. Detecting linguistic bias in AI plagiarism and content detectors
You can research whether AI-generated text detectors, such as GPTZero or Proofademic, exhibit higher false-positive rates for non-native English speakers. The core question is whether the perplexity and burstiness metrics used by these tools unfairly penalize structured, formulaic writing common in ESL (English as a Second Language) learners. It employs R or Python for statistical significance testing, using a corpus of essays from the TOEFL dataset vs. native-speaker datasets. This research sits at the intersection of Computational Linguistics, Education Policy, and Ethics.
4. Analyzing feedback loops in predictive policing algorithms
The core question is whether historical arrest data used by predictive policing software creates a runaway feedback loop that over-targets specific neighborhoods, regardless of actual crime rates. You will build a simplified simulation in Python (using Mesa or NetLogo) to model agent-based interactions. You will use open crime data from cities like Chicago or New York (via Open Data Portals) and apply GIS (Geographic Information Systems) to map predicted vs. actual crime hot spots. If you want to learn more about Urban Studies, Criminology, and Systems Modeling, this is a good topic to pick.
5. Measuring stereotype reinforcement in text-to-image generative models
You will investigate how Diffusion models (like Stable Diffusion or Midjourney) represent professional roles when prompted with neutral terms like “doctor” or “leader.” The research question focuses on the frequency of “savior” tropes or gendered occupational stereotypes in generated imagery. You will use CLIP (Contrastive Language-Image Pre-training) scores to quantify the alignment between images and stereotypical descriptors. Data sources involve generating a large sample size of images through API calls (Hugging Face) and using OpenCV for automated demographic tagging. This research sits at the crux of Media Studies, Psychology, and Visual Arts.
6. Transparency audits of Large Language Models using SHAP and LIME
The core question is how specific input tokens influence a Large Language Model’s (LLM) decision to categorize a text as “toxic.” You will use explainable AI (XAI) tools such as SHAP (SHapley Additive exPlanations) and LIME to visualize the attention weights of a transformer-based model. By feeding the model edge-case sentences from the Civil Comments Dataset, you can identify which words trigger false positives. Computer Science, Logic, and Communication Theory all play a vital part in this research.
7. Assessing the impact of echo chambers in AI-driven news feed algorithms
You will research the extent to which reinforcement learning (RL) in recommendation engines prioritizes high-engagement, polarized content over diverse viewpoints. The core question is whether click-through rate optimization inherently creates information silos. You can use SQL to query public social media API data (where available, such as the Reddit API) and NetworkX for social network analysis to visualize graph-based clustering. This work sits squarely at the intersection of Political Science, Network Science, and Behavioral Economics.
8. Evaluating privacy risk in differential privacy for health data
The core question focuses on the trade-off between data utility and individual privacy when adding noise to medical datasets. You will utilize the Google Differential Privacy Library to apply Laplacian noise to a sample health dataset (e.g., CDC National Health and Nutrition Examination Survey). You can measure the accuracy loss of a predictive model vs. the level of epsilon privacy protection achieved. This connects to Cybersecurity, Statistics, and Healthcare Policy.
9. Investigating cultural exclusion in low-resource multilingual AI
You will research the performance gap of LLMs when processing “low-resource” languages compared to English. The core question is whether the lack of representative data leads to hallucinations or cultural inaccuracies in translation. You will use the FLORES-200 dataset and BLEU/METEOR scores to compare translation quality across diverse language families. It’s a good opportunity to analyze racial discrimination at the crux of Anthropology, Linguistics, and Global Studies.
10. The ethics of Agentic AI: defining autonomy thresholds
The core question is at what point a semi-autonomous AI agent (like those in AutoGPT) should be required to seek human intervention to prevent cascading errors. It starts by designing a stress-test protocol using Python-based simulations, in which an agent is given a multi-step task with conflicting constraints. You then analyze its failure modes and propose a “Human-in-the-Loop” (HITL) framework. This connects to Philosophy (Deontology), Law, and Robotics.
- Copyright infringement in training data: Tracing originality
Arguably one of the hottest debated questions of the decade. You will investigate whether generative models can be prompted to reproduce copyrighted artworks or text verbatim. The core question is the degree of similarity required to constitute a derivative work. You will use Cosine Similarity measures and PyTorch to compare model outputs with a database of known copyrighted works (e.g., The Pile dataset). This is a highly relevant topic at the intersection of Intellectual Property Law, Digital Humanities, and Fine Arts.
12. Carbon footprint analysis of transformer-based model training
Another hotly contested topic in AI is the core question of how the scale of a model’s parameters correlates with its total CO2 emissions during training. You will use tools such as CodeCarbon or the ML CO2 Impact Calculator to estimate the environmental cost of fine-tuning different model versions (e.g., Llama-3 vs. Mistral). This research employs Environmental Science data on regional grid carbon intensity, incorporating aspects of Energy Policy and Computational Efficiency as well.
13. Algorithmic bias in credit scoring for young adults
The key question is whether alternative data, such as rent payments and utility bills used by AI-driven fintech apps, creates new forms of financial exclusion. You will use the UCI Machine Learning Repository’s Credit Approval Dataset and apply AIF360 to check for equal opportunity metrics. You will conduct a Sensitivity Analysis on variables such as zip codes to determine whether they serve as proxies for race. This connects to Finance, Economics, and Social Justice.
14. Detecting shadowbanning and bias in automated content moderation
You will research whether specific political or social keywords lead to lower visibility in social media algorithms. Are safety filters being applied unevenly across the political spectrum? You will use Web Scraping (BeautifulSoup/Selenium) to track the reach of “paired” accounts that share identical content but use different keywords. If you are interested in Political Science, Data Ethics, or Software Engineering, this is a good topic to pursue.
15. The impact of AI tutors on educational equity
Another important and contested topic is whether adaptive learning platforms provide equivalent hints and support levels to students from different socioeconomic backgrounds. You will analyze metadata from open-source platforms such as Khan Academy (where available for research) or Open edX. You will use T-tests and ANOVA to compare learning gain scores across demographics. This requires you to have an interest in Pedagogy, Educational Psychology, and Statistics.
16. Analyzing gender bias in voice-activated personal assistants
You will research whether voice recognition software has higher error rates for female voices or specific regional accents. You will analyze the impact of training data imbalances in LibriSpeech or Common Voice datasets. You will use Praat (Phonetic Software) to analyze pitch and frequency distributions and correlate them with word error rates (WER). This connects to Acoustic Physics, Gender Studies, and Signal Processing.
17. Ethical implications of deepfake detection in political campaigns
An increasingly resonant concern worldwide, this research examines whether the availability of deepfake detection tools (such as DeepWare) increases or decreases public trust in authentic video footage. You will conduct a Survey Design using Qualtrics or Google Forms to measure trust levels in a controlled group exposed to both real and manipulated videos. This connects to Cognitive Psychology, Digital Forensics, and Journalism.
18. AI in legal sentencing: Auditing the COMPAS algorithm
The core question is whether the COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) tool’s recidivism predictions are racially biased. You will use the ProPublica COMPAS dataset and apply Confusion Matrix analysis to compare false positive rates for Black vs. White defendants. This project uses R for statistical modeling and logistic regression and sits at the intersection of Law, Criminology, and Ethics.
19. Mental Health and AI: Dependency patterns in Gen Z
Another topic that has been steadily gathering public attention is AI’s psychological effects on younger generations. You will investigate whether prolonged interaction with AI chatbots leads to a decline in real-world social self-efficacy, and investigate the correlation between AI companionship hours and scores on the UCLA Loneliness Scale. You will use Quantitative Survey Methods and Correlation Coefficients. This connects to Clinical Psychology, Sociology, and Human-Computer Interaction (HCI).
20. Smart Home Surveillance: Consent in multi-user environments
Another point of view in the privacy debate is that the core question focuses on how IoT AI systems handle data privacy for non-consenting third parties, such as guests or children. You will utilize Threat Modeling frameworks to map data flows from a smart camera or speaker and propose a Privacy-by-Design technical solution. This connects to Cybersecurity, User Experience (UX) Design, and Ethics.
21. Algorithmic accountability in autonomous vehicle collisions
A highly future-relevant topic, you will research the “Moral Machine” problem: how should an AI prioritize lives in an unavoidable accident? Analyze how varied cultural values influence the “target” of the algorithm’s decision, using data from the MIT Moral Machine project, and perform Multivariate Analysis on global preference patterns. This connects to Ethics, Automotive Engineering, and Public Policy.
22. Philosophy of AI personhood and digital rights
Your research will examine the legal and philosophical precedents for “corporate personhood” and whether they provide a viable framework for future AI digital rights. You will debate whether an AI that simulates empathy or creativity warrants moral status. This involves a comparative analysis of legal texts and philosophical arguments using the Stanford Encyclopedia of Philosophy. This work is relevant to Jurisprudence, Philosophy, and Political Science.
23. Bias in AI-Driven Disaster Relief Prioritization
This investigation examines whether AI models used for resource allocation after natural disasters tend to overlook rural or impoverished areas due to lower data density. You will compare Sentinel-2 satellite imagery with FEMA ground-truth damage reportsusing GIS. Your goal is to determine whether prioritization correlates with property values rather than with human needs. This connects to International Development, Geoscience, and Humanitarian Law.
24. Identifying Ageism in AI Hiring Visuals
You will research whether AI video-interview analysis tools show bias against older candidates’ facial expressions or speech cadences. Using Facial Action Coding Systems (FACS) and computing tools like OpenFace, you can analyze video datasets for age-related features. You will then measure the correlation between these features and “hireability” scores. This study links Gerontology, Industrial-Organizational Psychology, and Computer Science.
25. The “Filter Bubble” in AI Search Engine Result Pages (SERPs)
This project explores whether AI-integrated search engines (such as Perplexity or Google SGE) provide more biased information than traditional search by summarizing complex topics into a single viewpoint. You will test if LLM-based summarization creates a narrower information landscape. Using content analysis and VADER sentiment analysis, you can compare summaries of controversial social issues. This research concerns Information Science, Logic, and Communications.
26. Algorithmic paternalism and dark patterns in recommendation engines
You will research whether recommendation algorithms “nudge” users toward behaviors that infringe upon human autonomy. The focus is on “dark patterns” in social media interfaces and their impact on freedom of thought. Utilizing a technical audit of recommendation logic and user behavior data, you can identify paternalistic nudges. This connects to Behavioral Economics, HCI, and Ethics.
27. Accessibility in voice-to-text for speech impairments
The study evaluates the effectiveness of Automatic Speech Recognition (ASR) for users with dysarthria or other speech differences. You will utilize the Torgo Dataset and calculate the Word Error Rate (WER) using Python’s SpeechRecognition library. You can then propose a fine-tuning methodology to improve model inclusivity. This project bridges Speech-Language Pathology, Accessibility Engineering, and Computer Science.
28. Adversarial robustness and safety-critical AI deployment
Your research focuses on how adversarial attacks can cause AI systems to make catastrophic errors in safety-critical environments. You will test whether current robustness techniques are sufficient for medical or industrial AI applications. Using the Foolbox library in Python, you can evaluate model vulnerability to “evasion attacks.” This connects to Cybersecurity, Mathematics, and Systems Engineering.
29. AI-Driven urban planning: Bias in public transit optimization
This inquiry looks at whether AI models that optimize transit routes lengthen commute times for marginalized populations while prioritizing high-traffic commercial zones. Using GTFS (General Transit Feed Specification) data and Pandas, you can analyze route efficiency across different socioeconomic zip codes. This aligns with Civil Engineering, Urban Planning, and Social Policy.
30. The “Black Box” of AI in college admissions
The research investigates whether AI models used for “holistic” admissions use data points strongly correlated with family wealth to filter applicants. You will build a synthetic model using Random Forests to determine the “Feature Importance” of application variables. This study is relevant to Education Policy, Data Science, and Public Administration.
One more option – Horizon Academic Research Program
If you’re looking for a competitive mentored research program, consider applying to Horizon’s Research Seminars and Labs! This is a selective virtual research program that lets you engage in advanced research and develop a research paper on a subject of your choosing. Horizon has worked with 1000+ high school students so far and offers 600+ research specializations for you to choose from. You can find the application link here!
Image source – Horizon Inspires