200 AI Research Topics for High School Students

If you know you want to do AI research but haven’t found a question specific enough to actually start, you’re not alone. “I’m interested in machine learning” describes a field. “Does a transformer-based model fine-tuned on 10,000 labeled tweets detect sarcasm more accurately than a bag-of-words baseline, and where do the errors cluster?” describes a project. Most students stall in the gap between those two sentences.

This list gives you 200 specific AI research questions across ten subfields: computer vision, natural language processing, reinforcement learning, AI fairness and ethics, AI in healthcare, AI in climate and environmental science, generative AI, AI in education, robotics and embodied AI, and AI safety and interpretability. Each one is framed as a research question, not a topic area, and most name a specific dataset, method, or tool to get you started.

Strong AI research builds skills that carry directly into university coursework and research labs: framing a testable hypothesis, choosing appropriate model architectures and evaluation metrics, interpreting results honestly (including when they’re negative), and writing methodology that another researcher could replicate. Admissions benefits tend to follow from work like this, but they’re a byproduct of doing something real, not the reason to do it.

Key takeaways

• A workable AI research topic names a specific model, dataset, and evaluation metric, not just a general area like “AI in healthcare.”
• Most of these projects are achievable with a laptop, Python, and free datasets or APIs; lab access or specialized hardware isn’t required for the majority of them.
• Reading the limitations sections of recent papers on arXiv or in your subfield’s top venues (NeurIPS, ACL, CVPR) is the fastest way to find a real, unaddressed gap.
• Working with a mentor who has published in your specific AI subfield significantly shortens the time it takes to scope a project correctly.

Computer vision research topics for high school students

Computer vision is one of the more accessible AI subfields for students because labeled image datasets are widely available and tools like OpenCV, PyTorch, and pretrained models on the Hugging Face Hub lower the barrier to a working first model.

1. Can a convolutional neural network trained on the PlantVillage dataset distinguish between visually similar plant diseases with accuracy above 90%?
2. How does data augmentation strategy (rotation, flipping, color jitter) affect the robustness of an image classifier to lighting variation in real-world test images?
3. Can a YOLOv8 model trained on a custom-labeled dataset of recyclable materials accurately sort waste by category in real time?
4. How does model performance degrade when a facial recognition system trained primarily on one demographic group is tested on an out-of-distribution demographic group?
5. Can a Vision Transformer (ViT) outperform a traditional CNN architecture like ResNet-50 on a small, domain-specific image dataset with limited training examples?
6. How accurately can a CNN trained on satellite imagery from Sentinel-2 detect illegal deforestation compared to manually labeled ground-truth data?
7. Can semantic segmentation models distinguish between healthy and bleached coral in underwater photography datasets?
8. How does image resolution affect the accuracy-speed tradeoff of object detection models deployed on edge devices like a Raspberry Pi?
9. Can a model trained on the MNIST and EMNIST datasets generalize to handwriting samples collected from a different population (different age group, different language background)?
10. How does transfer learning from ImageNet-pretrained weights affect convergence speed and final accuracy compared to training a CNN from scratch on a small medical imaging dataset?
11. Can optical character recognition models accurately digitize handwritten historical documents, and where do error rates concentrate by handwriting style or document age?
12. How does adversarial perturbation (small, imperceptible pixel changes) affect the classification accuracy of a pretrained image classifier, and can simple defenses restore accuracy?
13. Can a CNN trained on x-ray images from the NIH Chest X-ray dataset detect pneumonia with sensitivity comparable to radiologist-level benchmarks reported in published literature?
14. How does the choice of loss function (cross-entropy vs. focal loss) affect model performance on imbalanced image classification datasets like rare wildlife species detection?
15. Can pose estimation models (such as MediaPipe or OpenPose) be used to quantitatively assess form in athletic movements like a squat or golf swing?
16. How accurately can a model trained on the FER-2013 facial expression dataset classify emotions compared to human inter-rater agreement on the same images?
17. Can image-based crowd counting models trained on benchmark datasets like ShanghaiTech accurately estimate attendance from drone photography of local events?
18. How does model accuracy on license plate recognition change under different weather conditions simulated through data augmentation?
19. Can a Siamese network architecture accurately perform one-shot facial verification with fewer training examples than a standard CNN classifier requires?
20. How do explainability methods like Grad-CAM reveal whether an image classifier is making decisions based on relevant features or spurious correlations in the background?

Natural language processing research topics for high school students

NLP research gives students access to some of the richest open datasets in AI, from the Hugging Face dataset hub to Twitter/X academic APIs, and most projects can be built using free-tier access to pretrained language models. If you’re not yet sure how to scope an NLP question into something testable, this is exactly the kind of decision a Horizon mentor working in machine learning or computational linguistics can help you make in your first session.

21. Can a fine-tuned BERT model detect sarcasm in social media text more accurately than a bag-of-words logistic regression baseline, and where do its errors cluster?
22. How does few-shot prompting compare to fine-tuning when adapting a large language model to classify legal documents by case type?
23. Can a transformer-based summarization model produce summaries of scientific abstracts that domain experts rate as more accurate than extractive baseline summaries?
24. How does tokenization strategy (byte-pair encoding vs. WordPiece vs. SentencePiece) affect downstream performance on low-resource language translation tasks?
25. Can NLP methods reliably distinguish AI-generated text from human-written text across different LLMs (GPT-4, Claude, Gemini), and how does detection accuracy change as models improve?
26. How does the inclusion of named entity recognition as a preprocessing step affect the accuracy of a fake news classification model?
27. Can sentiment analysis models trained on English-language data be effectively adapted to code-switched text (e.g., Spanglish, Hinglish) using transfer learning?
28. How accurately can topic modeling methods like LDA versus BERTopic identify emerging themes in a corpus of recent news articles on a single topic?
29. Can a question-answering model fine-tuned on a domain-specific corpus (e.g., medical FAQs) outperform a general-purpose model like GPT-3.5 on domain-specific factual accuracy?
30. How does prompt engineering (zero-shot, few-shot, chain-of-thought) affect a large language model’s accuracy on multi-step arithmetic reasoning problems?
31. Can stylometric features combined with a classifier reliably attribute authorship of anonymous or disputed texts, and how does accuracy change with text length?
32. How well do toxicity detection models trained on one platform’s data (e.g., Reddit) generalize to comments scraped from a different platform (e.g., YouTube)?
33. Can named entity recognition models accurately extract structured data (dates, locations, organizations) from unstructured historical newspaper archives?
34. How does the presence of negation (“not good” vs. “good”) affect the accuracy of sentiment classifiers, and which model architectures handle it best?
35. Can a retrieval-augmented generation (RAG) pipeline reduce hallucination rates in an LLM’s answers to factual questions compared to the same model without retrieval?
36. How accurately can NLP models classify the political leaning of news articles based on lexical and syntactic features alone, without metadata?
37. Can paraphrase detection models distinguish AI-paraphrased plagiarism from original student writing with accuracy useful for academic integrity tools?
38. How does fine-tuning a language model on a specific author’s body of work affect its ability to generate text in that author’s style, evaluated using stylometric similarity metrics?
39. Can multilingual BERT models perform cross-lingual sentiment transfer, training on English data and evaluating on a low-resource language with no labeled sentiment data?
40. How do different word embedding methods (Word2Vec, GloVe, FastText, contextual embeddings from BERT) compare in capturing semantic similarity for domain-specific vocabulary like medical or legal terms?

Reinforcement learning research topics for high school students

Reinforcement learning research at the high school level typically uses simulated environments like OpenAI Gym (now Gymnasium) or simple custom environments, which makes experimentation accessible without needing real-world deployment. RL is also one of the subfields where a mentor’s guidance pays off quickly, since choosing the wrong algorithm for your environment’s action space (discrete vs. continuous) can cost weeks of debugging that an experienced Horizon mentor would flag immediately.

41. How does the choice of reward shaping affect the speed and stability of convergence for a Q-learning agent trained on a custom grid-world navigation task?
42. Can a Deep Q-Network (DQN) trained on Atari environments from Gymnasium generalize to slightly modified versions of the same environment (different colors, different obstacle positions)?
43. How does the exploration-exploitation tradeoff parameter (epsilon in epsilon-greedy policies) affect total reward accumulation across different environment complexities?
44. Can a reinforcement learning agent trained to play a simplified version of a real-world resource allocation problem (e.g., classroom scheduling) outperform a greedy heuristic baseline?
45. How does the choice between policy-based methods (REINFORCE, PPO) and value-based methods (DQN) affect sample efficiency in environments with continuous action spaces?
46. Can multi-agent reinforcement learning produce emergent cooperative behavior in a simulated predator-prey environment without explicit cooperation incentives in the reward function?
47. How does reward sparsity affect training stability, and can reward shaping or curriculum learning meaningfully accelerate convergence in sparse-reward environments?
48. Can a reinforcement learning agent trained on simulated traffic light control reduce average wait time compared to a fixed-timing baseline across varying traffic density scenarios?
49. How does the choice of discount factor (gamma) affect an agent’s behavior in environments requiring long-term planning versus short-term reward maximization?
50. Can model-based reinforcement learning methods achieve comparable performance to model-free methods with significantly fewer training episodes in a sample-limited simulated environment?
51. How does adding noise to an agent’s observations (simulating sensor uncertainty) affect the robustness of a trained policy, and can adversarial training improve resilience?
52. Can a reinforcement learning agent learn an effective bidding strategy in a simulated auction environment, and how does its strategy compare to known game-theoretic equilibria?
53. How does curriculum learning, training an agent on progressively harder versions of a task, affect final performance compared to training directly on the hardest version?
54. Can reinforcement learning be used to optimize a simple inventory management policy, and how does the learned policy compare to standard operations research heuristics like (s, S) policies?
55. How does the size and architecture of the neural network function approximator affect a DQN agent’s performance on environments of varying state-space complexity?
56. Can an agent trained via reinforcement learning in a simulated environment transfer its learned policy to a physical robot with minimal additional fine-tuning (sim-to-real transfer)?
57. How does the choice of action space discretization (fine-grained vs. coarse) affect learning speed and final policy quality in continuous control tasks?
58. Can self-play reinforcement learning produce strategies in a simple two-player game that outperform strategies designed by human players?
59. How does experience replay buffer size affect the stability of DQN training, and is there a point of diminishing returns?
60. Can hierarchical reinforcement learning, decomposing a complex task into subtasks, accelerate learning compared to flat reinforcement learning on a multi-step navigation task?

AI fairness and ethics research topics for high school students

Fairness research is one of the most active and high-impact areas in AI, and it’s well suited to high school research because it combines technical analysis with substantive ethical reasoning, both of which are gradeable on rigor rather than access to specialized equipment.

61. How does training data demographic composition affect the false positive rate of a facial recognition model across different racial groups, measured against benchmarks like the NIST FRVT?
62. Can fairness-aware machine learning techniques (reweighting, adversarial debiasing) reduce demographic disparities in a loan approval prediction model without significantly reducing overall accuracy?
63. How does the choice of fairness metric (equalized odds, demographic parity, predictive parity) affect which model is judged “fairest,” and can a model satisfy all three simultaneously?
64. Can large language models exhibit measurable bias in resume screening tasks when candidate names are swapped between demographically associated name sets, holding qualifications constant?
65. How does algorithmic bias in predictive policing models, trained on historical arrest data, compound over time when the model’s outputs influence future data collection?
66. Can explainability methods like SHAP values reveal whether a credit scoring model is implicitly using protected characteristics through proxy variables like zip code?
67. How do recommendation algorithms affect content diversity exposure across different demographic user groups, measured using simulated user profiles?
68. Can synthetic data generation methods reduce bias in training datasets for underrepresented groups without introducing new forms of distributional distortion?
69. How does model compression (pruning, quantization) for deployment on edge devices disproportionately affect accuracy for underrepresented subgroups in the training data?
70. Can crowdsourced content moderation labels show measurable inter-rater disagreement correlated with the labelers’ own demographic background?
71. How does the use of AI in college admissions screening tools affect outcomes for first-generation applicants compared to traditional review processes, based on published case studies?
72. Can differential privacy techniques applied to a machine learning model meaningfully protect individual data points while preserving acceptable model accuracy?
73. How does algorithmic decision-making in hiring tools handle candidates with employment gaps differently depending on the stated reason for the gap?
74. Can a sentiment analysis model trained on general social media data show measurably different accuracy when applied to text written in African American Vernacular English versus Standard American English?
75. How transparent are publicly available AI model cards (Hugging Face, Google Model Cards) in disclosing known limitations and bias evaluations, assessed across a sample of widely used models?
76. Can counterfactual fairness techniques, testing whether a model’s prediction changes when a protected attribute is hypothetically altered, reveal bias not caught by standard fairness metrics?
77. How does dataset imbalance in medical AI training data (underrepresentation of certain demographics in clinical trial data) propagate into diagnostic model performance gaps?
78. Can students design a bias audit framework for a publicly available AI tool (a resume screener, a chatbot) and document measurable disparities using a structured testing protocol?
79. How does the framing of AI risk in media coverage (existential risk vs. present-day harms) correlate with public survey data on AI policy preferences?
80. Can interpretable machine learning models (decision trees, rule-based systems) achieve comparable fairness outcomes to black-box models while offering greater transparency to affected individuals?

AI in healthcare research topics for high school students

Healthcare AI research is highly active in published literature and offers strong access to public datasets like MIMIC-III/IV, making it one of the more rigorous subfields for ambitious student researchers. It’s also one of the areas where Horizon’s mentors most often help students navigate data access requirements (MIMIC requires a short credentialing course before you can download it) before the research itself even begins.

81. Can a gradient-boosted classifier trained on MIMIC-III vital sign features predict ICU readmission within 30 days more accurately than a logistic regression baseline?
82. How does class imbalance correction (SMOTE, class weighting) affect sensitivity and specificity tradeoffs in a sepsis early-warning model trained on EHR time-series data?
83. Can a CNN trained on dermatology image datasets (such as HAM10000) classify skin lesions with accuracy approaching published dermatologist-level benchmarks?
84. How does feature selection method affect the interpretability and accuracy tradeoff in a diabetes risk prediction model trained on the Pima Indians Diabetes dataset?
85. Can natural language processing applied to de-identified clinical notes detect undiagnosed depression at rates exceeding standard screening questionnaires?
86. How accurately can wearable sensor data (heart rate variability, accelerometer data) be used to predict the onset of atrial fibrillation episodes using a recurrent neural network?
87. Can a machine learning model trained on publicly available cancer genomics data (TCGA) predict treatment response with accuracy comparable to existing clinical staging systems?
88. How does model calibration (the alignment between predicted probability and actual outcome frequency) differ across demographic subgroups in a published clinical risk prediction model?
89. Can graph neural networks trained on molecular structure data accurately predict drug-target binding affinity compared to traditional QSAR modeling approaches?
90. How does the use of synthetic patient data generated via generative adversarial networks affect the performance of downstream diagnostic models trained on it, compared to real data?
91. Can a model trained on speech recordings detect early signs of Parkinson’s disease with accuracy comparable to clinical assessment, using publicly available voice datasets?
92. How does temporal data resolution (hourly vs. daily vital sign measurements) affect the predictive accuracy of an early deterioration warning system in ICU data?
93. Can federated learning approaches train an effective diagnostic model across multiple simulated hospital datasets without sharing raw patient data between sites?
94. How accurately can a CNN trained on retinal fundus images detect diabetic retinopathy compared to ophthalmologist grading in publicly available datasets like Kaggle’s APTOS dataset?
95. Can survival analysis methods combined with machine learning (random survival forests) improve prediction of time-to-event outcomes compared to standard Cox proportional hazards models?
96. How does data drift, changes in patient population characteristics over time, affect the long-term accuracy of a deployed clinical prediction model evaluated on more recent data?
97. Can a model trained on electronic health record data predict medication non-adherence with sufficient accuracy to be clinically actionable?
98. How does combining structured EHR data with unstructured clinical notes (via NLP) improve prediction accuracy compared to using structured data alone?
99. Can wearable-derived sleep data combined with self-reported mood logs predict depressive episode onset using time-series machine learning models?
100. How does algorithmic triage prioritization in emergency department settings compare to nurse-led triage in terms of both accuracy and resource allocation efficiency, based on published comparative studies?

AI in climate and environmental science research topics for high school students

Environmental AI research is well suited to students because satellite imagery, climate model outputs, and ecological monitoring data are largely open access through NASA Earthdata, NOAA, and Copernicus.

101. Can a CNN trained on Sentinel-2 satellite imagery accurately detect illegal deforestation events compared to ground-truth verified reports from monitoring organizations like Global Forest Watch?
102. How does a recurrent neural network trained on historical weather station data compare to traditional numerical weather prediction methods for short-range (24 to 48 hour) precipitation forecasting?
103. Can machine learning models trained on ocean buoy data predict harmful algal bloom onset earlier than current threshold-based monitoring methods?
104. How accurately can a model trained on camera trap images automatically identify and count wildlife species compared to manual review by ecologists?
105. Can graph neural networks model the spread of wildfire risk across a landscape more accurately than traditional fire behavior models like FARSITE, using historical fire and weather data?
106. How does a machine learning model trained on urban sensor network data predict air quality index changes compared to standard dispersion modeling approaches?
107. Can satellite-derived vegetation indices (NDVI) combined with machine learning predict crop yield more accurately than historical average-based forecasting?
108. How does a model trained on acoustic monitoring data classify bird species calls compared to expert ornithologist identification, using datasets like the Cornell Lab’s BirdCLEF?
109. Can unsupervised clustering of satellite nighttime light data identify illegal fishing zones when cross-referenced with vessel tracking data?
110. How accurately can machine learning models predict coral bleaching events using sea surface temperature data combined with other oceanographic variables?
111. Can a model trained on historical glacier imagery predict future ice mass loss trajectories with uncertainty bounds comparable to physically-based glaciological models?
112. How does incorporating remote sensing soil moisture data improve drought prediction model accuracy compared to precipitation data alone?
113. Can reinforcement learning optimize renewable energy grid dispatch decisions to reduce curtailment compared to standard rule-based dispatch algorithms?
114. How accurately can computer vision models applied to drone imagery estimate methane leak locations at industrial sites compared to specialized gas detection equipment?
115. Can machine learning models trained on historical flood data combined with real-time precipitation forecasts improve flood warning lead time compared to threshold-based systems?
116. How does a model trained on microplastic concentration data from coastal monitoring stations predict accumulation hotspots, and how do predictions compare to subsequent field validation?
117. Can time-series forecasting models predict urban heat island intensity changes based on land use and tree canopy coverage trends?
118. How accurately can a CNN classify invasive plant species from drone-collected imagery compared to manual field survey identification?
119. Can machine learning models combining satellite and ground sensor data improve estimates of carbon sequestration in reforested areas compared to standard allometric equations?
120. How does an ensemble of climate model outputs (CMIP6) combined with machine learning bias correction improve regional temperature projection accuracy compared to a single model?

Generative AI research topics for high school students

Generative AI is the fastest-moving subfield in AI research, and student projects here benefit from being narrowly scoped, since the field changes quickly enough that broad surveys go stale within months. This is precisely the kind of subfield where working with a Horizon mentor who tracks the literature actively makes the biggest difference: knowing which open-source models and benchmarks are current saves you from building a project around a tool that’s already outdated by the time you publish.

121. How does fine-tuning a small open-source language model (such as Llama or Mistral) on a domain-specific corpus compare to prompting a larger general-purpose model for the same task, in terms of accuracy and cost?
122. Can a diffusion model fine-tuned on a narrow image domain (a specific art style, a specific species) produce outputs that domain experts rate as comparably authentic to other examples in that domain?
123. How does the presence of watermarking or provenance metadata in AI-generated images affect viewers’ trust ratings compared to unmarked images, based on a controlled survey?
124. Can retrieval-augmented generation meaningfully reduce factual hallucination rates in a chatbot answering questions about a specific, narrow knowledge domain (a textbook, a local government’s public records)?
125. How does prompt structure affect the consistency of outputs from a text-to-image model across repeated generations with the same seed?
126. Can a fine-tuned language model trained on a specific genre of writing (legal contracts, scientific abstracts) be reliably distinguished from human-written text in that genre by both humans and automated classifiers?
127. How does model size (parameter count) correlate with hallucination rate on a fixed benchmark of factual questions across a family of open-source models?
128. Can synthetic data generated by a language model improve the performance of a smaller downstream classifier trained on a low-resource NLP task, compared to training on real data alone?
129. How does temperature setting in text generation affect the tradeoff between output diversity and factual accuracy in a constrained question-answering task?
130. Can voice cloning models be reliably detected using audio forensics techniques, and how does detection accuracy change as the cloned sample length increases?
131. How does instruction fine-tuning change a base language model’s behavior on tasks it wasn’t explicitly trained for, measured through a held-out generalization benchmark?
132. Can a generative model be used to produce synthetic training images for a rare medical condition, and does a classifier trained on the augmented dataset outperform one trained on the limited real data alone?
133. How does chain-of-thought prompting affect the accuracy of a language model on multi-step logical reasoning problems compared to direct prompting?
134. Can text-to-speech models trained on a specific regional accent produce outputs that listeners from that region rate as more natural than a generic accent model?
135. How does model output consistency change when the same prompt is submitted multiple times to a language model at non-zero temperature, and what does that imply about reliability for high-stakes use cases?
136. Can a fine-tuned generative model produce synthetic tabular data that preserves statistical properties of a real dataset closely enough to be useful for downstream machine learning without exposing original sensitive records?
137. How does prompt injection vulnerability differ across language models with different safety training approaches, tested through a structured red-teaming protocol?
138. Can a language model accurately summarize its own confidence or uncertainty about a generated answer, validated against actual correctness rates?
139. How does combining a generative model with a symbolic verification step (checking generated code against unit tests) improve reliability compared to the generative model alone?
140. Can style transfer techniques applied to generated music be reliably evaluated by listeners as belonging to a specific genre or composer’s style, using a blind listening test?

AI in education research topics for high school students

AI in education research is well suited to students because it can combine technical model-building with original data collection through surveys or classroom-based experiments, and it has direct relevance to the researcher’s own daily experience. It’s also a field where Horizon mentors with backgrounds in learning science can help you design the controlled comparison correctly, since these projects often hinge on getting the experimental design right before any modeling begins.

141. Can an adaptive learning algorithm that adjusts question difficulty based on real-time performance improve learning gains compared to a fixed-difficulty practice set, measured in a controlled study?
142. How does AI-generated feedback on student writing compare to teacher feedback in terms of specificity, actionability, and student-reported usefulness?
143. Can a machine learning model trained on learning management system clickstream data predict which students are at risk of falling behind, with enough lead time to be useful for intervention?
144. How does the use of an AI tutoring chatbot affect self-reported student motivation compared to a static FAQ-style help resource, in a controlled comparison?
145. Can natural language processing applied to student forum posts in an online course detect confusion or frustration early enough to trigger useful instructor intervention?
146. How accurately can a model trained on past exam performance and study habit data predict standardized test score improvement, and which features carry the most predictive weight?
147. Can automated essay scoring models achieve inter-rater agreement with human graders comparable to the agreement between two human graders on the same set of essays?
148. How does personalized AI-generated practice problem difficulty affect knowledge retention measured by delayed post-tests compared to immediate post-tests?
149. Can a recommendation system for supplementary learning resources, built on a small classroom dataset, measurably improve engagement compared to a non-personalized resource list?
150. How does student trust in AI-generated grading feedback vary based on whether the AI’s reasoning is shown alongside the score, tested through a controlled survey?
151. Can speech recognition technology accurately assess pronunciation in a second-language learning context compared to a human language instructor’s assessment?
152. How does an AI-assisted Socratic questioning chatbot affect depth of student reasoning in written responses compared to direct AI-generated answers?
153. Can a model trained on classroom participation data (anonymized) identify patterns associated with disengagement that aren’t visible from grades alone?
154. How does the framing of AI tutoring tools as “helper” versus “tutor” versus “study buddy” affect student willingness to use the tool, based on a controlled survey design?
155. Can a knowledge-tracing model accurately predict which specific concept a student is struggling with, based on their pattern of correct and incorrect responses across a problem set?
156. How does over-reliance on AI writing assistance correlate with measured changes in students’ unassisted writing quality over a semester-long period?
157. Can AI-generated study guides tailored to an individual’s quiz performance outperform generic study guides in measured exam preparedness?
158. How does an AI plagiarism detection tool’s false positive rate vary across writing samples from English language learners compared to native English speakers?
159. Can sentiment analysis of anonymous end-of-course student feedback reliably identify specific pedagogical issues that numeric course ratings miss?
160. How does access to an AI homework helper affect long-term problem-solving skill development compared to traditional homework without AI assistance, measured through delayed skill transfer tests?

Robotics and embodied AI research topics for high school students

Robotics research can be approached computationally through simulation (using tools like PyBullet or Gazebo) even without physical robot hardware, which makes several of these projects accessible without a lab.

161. How does sensor fusion (combining camera and ultrasonic data) improve obstacle avoidance reliability in a simulated mobile robot compared to using a single sensor type?
162. Can a reinforcement learning policy trained entirely in simulation transfer successfully to a low-cost physical robot platform (such as a Raspberry Pi-based rover) without significant performance loss?
163. How does the choice of control method (PID control vs. a learned neural network controller) affect stability and energy efficiency in a simulated quadruped robot’s gait?
164. Can a swarm of simple simulated robots following local interaction rules produce emergent collective behavior (flocking, foraging) comparable to biological swarm systems?
165. How does end-to-end learning (camera pixels directly to motor commands) compare to a modular pipeline (perception, planning, control) in terms of robustness for a simulated line-following robot?
166. Can a robot arm trained via reinforcement learning in simulation learn to grasp irregularly shaped objects with a success rate comparable to scripted grasping approaches?
167. How does adding proprioceptive feedback (joint angle sensing) improve a simulated legged robot’s ability to recover from being pushed off balance?
168. Can SLAM (simultaneous localization and mapping) algorithms implemented on a low-cost robot platform produce maps accurate enough for reliable indoor navigation?
169. How does communication bandwidth between robots in a multi-robot system affect task completion time in a simulated cooperative search-and-rescue scenario?
170. Can imitation learning from a small number of human-demonstrated trajectories produce a viable control policy for a simple manipulation task, compared to reinforcement learning from scratch?
171. How does domain randomization during simulation training (varying lighting, friction, object properties) affect a learned policy’s real-world transfer performance?
172. Can a low-cost robot equipped with a simple camera and a pretrained object detection model reliably sort objects by category in a constrained environment?
173. How does battery and computational power constraint affect the tradeoff between model complexity and real-time inference speed for an embedded vision system on a small robot?
174. Can a simulated robot learn an energy-efficient gait through reinforcement learning that uses measurably less simulated power consumption than a hand-tuned baseline gait?
175. How does haptic feedback (force sensing) improve a robot’s ability to handle delicate objects without damage, compared to vision-only feedback?
176. Can a robot navigation policy trained using reinforcement learning generalize to environment layouts it wasn’t trained on, and how does generalization performance scale with training environment diversity?
177. How does the presence of dynamic obstacles (moving people, other robots) affect path-planning algorithm performance compared to static-obstacle environments?
178. Can a low-cost drone equipped with a vision model autonomously navigate a simple obstacle course with a success rate comparable to manually piloted runs?
179. How does sim-to-real transfer performance change when training in a more visually realistic simulator versus a simplified one?
180. Can a robot trained through reinforcement learning develop a stacking strategy for irregular objects that performs comparably to a human-designed heuristic stacking algorithm?

AI safety and interpretability research topics for high school students

AI safety and interpretability is a rapidly growing research area, well suited to students with strong technical skills who want to engage with the most actively discussed open problems in the field. Anthropic’s interpretability research and OpenAI’s published safety work are good starting points for understanding what current open questions look like at the frontier. Students working with Horizon on AI safety topics are typically matched with mentors who have direct research experience in interpretability or alignment, which matters given how fast the open problems in this subfield shift.

181. Can mechanistic interpretability techniques (probing classifiers, activation patching) identify which layers of a small transformer model are responsible for a specific learned behavior, such as arithmetic computation?
182. How does adversarial robustness training (training a model on adversarial examples) affect a model’s accuracy on clean, unperturbed data, and is there a measurable tradeoff?
183. Can attention visualization in a transformer model reliably indicate which input tokens the model is actually relying on for its prediction, validated through targeted ablation experiments?
184. How does model size affect susceptibility to “jailbreaking” prompts designed to bypass safety training, tested across a family of open-source models of varying sizes?
185. Can a smaller model be trained to predict the failure modes of a larger model, functioning as an early-warning system for unreliable outputs?
186. How consistent are explainability method outputs (SHAP, LIME, integrated gradients) when applied multiple times to the same model and input, and what does inconsistency imply about their reliability?
187. Can red-teaming a publicly available chatbot using a structured testing protocol reveal systematic categories of failure that aren’t disclosed in the model’s published safety documentation?
188. How does fine-tuning a language model on a narrow task affect its behavior on unrelated tasks, and can this reveal unintended capability or safety regressions?
189. Can sparse autoencoders trained on a language model’s internal activations identify human-interpretable features corresponding to specific concepts the model has learned?
190. How does a model’s stated confidence in its own answer correlate with actual accuracy across different question types, and where does miscalibration concentrate?
191. Can watermarking techniques embedded in AI-generated text survive common paraphrasing and editing operations while remaining detectable?
192. How does training data deduplication affect a language model’s tendency to memorize and reproduce specific training examples verbatim?
193. Can a classifier be trained to detect when a language model’s output appears to be “hallucinating” by analyzing internal model signals rather than just the output text?
194. How does the order in which examples are presented during in-context learning affect a language model’s output accuracy and consistency?
195. Can interpretability tools reveal whether a model trained to refuse harmful requests has learned a genuine concept of harm or a shallow pattern-matching shortcut?
196. How does scaling model size affect the emergence of specific capabilities (such as multi-step reasoning), and is the relationship smooth or threshold-like across a benchmark suite?
197. Can a model’s behavior under distribution shift (test data that differs systematically from training data) be predicted in advance using metrics computed only on the training distribution?
198. How does combining multiple independent safety classifiers (an ensemble approach) reduce the false negative rate for harmful content detection compared to a single classifier?
199. Can techniques from mechanistic interpretability identify circuits responsible for specific factual recall in a language model, and can those circuits be selectively edited to update a fact?
200. How does a model’s susceptibility to prompt injection attacks change when the injected instruction is presented in different formats (plain text, code comments, foreign language)?

How to turn one of these topics into a real research project?

A topic from this list is a starting point, not a finished proposal. Before you commit, run it through a quick feasibility check: identify the exact dataset or simulation environment you’ll use, confirm it’s actually accessible (many academic datasets require institutional access you may not have), and read three to five recent papers close to your question to understand how researchers in that specific area design their methods and report their results. A tool like Connected Papers can help you map a field quickly from a single seed paper, showing you where the literature is dense and where it’s sparse.

Pay particular attention to the “Limitations” and “Future Work” sections of those papers. Researchers regularly flag exactly the kind of follow-up questions that make strong, well-scoped student projects, and starting from an explicitly identified gap is far more efficient than guessing at one.

Also be realistic about your current technical level. If a topic requires fine-tuning a large language model and you haven’t worked with the Hugging Face Transformers library before, budget several weeks for that learning curve before you start your actual analysis, or choose a topic on this list that uses more accessible tools and smaller models. This is also the stage where a quick conversation with a mentor pays off most: a Horizon mentor can typically tell you within one session whether a topic is scoped appropriately for your current skill level, before you’ve invested weeks into the wrong direction.

Build your skills before you start

If you want to build foundational skills before committing to one of these topics, our 15 free AI courses for high school students and 15 free online AI courses for high school students cover everything from introductory machine learning concepts to hands-on work with libraries like scikit-learn and TensorFlow. If you’re looking for a shorter, more curated starting point with fully worked example questions, our 25 AI research ideas for high school students covers a smaller set of projects in more depth, each with specific datasets and tools named.

If you’re interested in structured research opportunities beyond self-directed projects, 10 AI research programs for high school students and 10 online machine learning research opportunities for high school students cover a range of programs, including mentored, application-based options.

Work on your AI research with expert mentorship

Scoping an AI research question correctly, and knowing whether your chosen architecture and dataset are actually appropriate for your question, is exactly the kind of judgment that takes researchers years to build. Getting that judgment from a mentor who has published in your specific subfield can save you weeks of misdirected effort.

Horizon Academic Research Program pairs high school students 1:1 with PhD scholars and professors from institutions including Oxford, Stanford, and Columbia for trimester-long research projects. Within AI specifically, Horizon offers tracks across machine learning, data science, AI and social media, and applied machine learning, alongside its 600+ specializations spanning the rest of STEM and the social sciences. The program is structured around producing a 20-page university-level research paper, and Horizon’s in-house publication support team helps students target appropriate journals and navigate peer review once the research is complete.

If you’ve found a topic on this list that genuinely interests you and want guidance on scoping it correctly, choosing the right tools, and executing it at a publishable standard, Horizon is built around exactly that. Application details and the full list of research tracks are at horizoninspires.com.

Frequently asked questions

What is a good AI research topic for a high school student? A good AI research topic names a specific model or method, a specific dataset, and a specific, measurable comparison or outcome. “AI in healthcare” is a field. “Can a gradient-boosted classifier trained on MIMIC-III data predict ICU readmission more accurately than logistic regression” is a topic. The narrower and more specific your question, the more executable your project will be.

Do I need to know how to code to do AI research in high school? Basic Python proficiency is necessary for most computational AI projects, including familiarity with libraries like NumPy, pandas, and scikit-learn, and ideally some exposure to PyTorch or TensorFlow for deep learning projects. If you’re starting from zero, budget four to six weeks to build foundational skills before beginning your actual research, using free resources like our list of free AI courses for high school students.

What datasets can high school students use for AI research? Widely accessible options include Hugging Face’s dataset hub for NLP tasks, Kaggle for a broad range of structured and image datasets, the UCI Machine Learning Repository, OpenAI Gym/Gymnasium for reinforcement learning environments, MIMIC-III and MIMIC-IV for healthcare AI (these require a free credentialing process), and NASA Earthdata or Sentinel-2 imagery for environmental AI projects. Most of these are free and don’t require institutional affiliation.

What AI subfield is best for a beginner doing research for the first time? Computer vision and basic supervised learning projects on structured data (tabular datasets from Kaggle or UCI) tend to be the most approachable for first-time researchers, because the tools are mature, well-documented, and have large communities producing tutorials. NLP with large language models and reinforcement learning both have steeper initial learning curves but are very achievable with several weeks of preparation.

How long does an AI research project take a high school student to complete? Most well-scoped AI research projects take 10 to 15 weeks from finalized topic to a complete draft, assuming the student already has basic Python proficiency. Projects requiring substantial new skill development (learning a new framework, fine-tuning a large model for the first time) often need an additional 4 to 6 weeks of preparation time built into the schedule.

Can high school students publish AI research? Yes. Journals like the Journal of Emerging Investigators, the International Journal of High School Research, and the Journal of Student Research all accept original AI and computer science research from high school students, provided the work involves a genuine original contribution, whether that’s a novel application, a rigorous comparison of methods, or a systematic evaluation of an existing model’s limitations. Students who complete their research through Horizon also have access to in-house publication support and the Horizon Scholar Journal as an additional pathway.

Is it better to build a novel model or evaluate an existing one for a student AI research project? Evaluating and rigorously testing existing models, rather than building entirely novel architectures, is usually the stronger choice for high school research. Questions like “how does an existing model perform on an underexplored dataset” or “where does this model’s accuracy break down” are well within reach technically and can produce genuinely original, publishable findings without requiring you to out-engineer professional AI labs.

Image source: Horizon Academic Research Program