In 2021 to 2026, projects outside the computing categories that used AI or machine learning won grand awards at 36 to 44%, against 22 to 26% for projects in the same fields without it. Behavioral and Social Sciences 43.7% vs 21.9%, Earth and Environment 40.9% vs 24.6%, Biology and Medicine 36.3% vs 24.5%. Inside Computing and Robotics the gap is 29.9% vs 24.4% (p=0.043). By subtype, reinforcement learning (46.7%), generative models (44.1%) and time-series forecasting (42.6%) lead, while projects built on an Arduino or microcontroller alone win at 23.1%, below the 27% base rate.
How the AI share grew
The share of projects whose title or abstract mentions machine learning, deep learning, neural networks, CNNs, transformers and similar terms: 6.8% in 2017, 17.9% in 2021, 22.5% in 2024, 27.1% in 2026. Outside the computing categories it went from 0.8% in 2014 to 21.0% in 2026. LLM and generative mentions were 1.4% in 2023 and 4.3% in 2026, still rare enough that a properly validated LLM project stands out.
| Year | Projects | AI/ML (all) | AI/ML (outside computing) | LLM / generative | Hardware / IoT | App / software |
|---|---|---|---|---|---|---|
| 2014 | 1394 | 2.8% | 0.8% | 0.2% | 14.2% | 26.6% |
| 2015 | 1347 | 3.8% | 1.9% | 0.3% | 17.3% | 30.5% |
| 2016 | 1371 | 4.1% | 2.4% | 0.2% | 19.3% | 29.7% |
| 2017 | 1403 | 6.8% | 4.0% | 0.3% | 19.5% | 31.4% |
| 2018 | 1390 | 9.5% | 5.3% | 0.1% | 19.5% | 32.5% |
| 2019 | 1426 | 11.6% | 6.6% | 0.1% | 22.3% | 35.6% |
| 2020 | 1027 | 14.4% | 9.6% | 0.4% | 22.2% | 34.9% |
| 2021 | 1452 | 17.9% | 12.5% | 0.3% | 21.2% | 37.1% |
| 2022 | 1400 | 18.1% | 13.4% | 0.6% | 21.7% | 35.2% |
| 2023 | 1316 | 17.8% | 13.0% | 1.4% | 22.3% | 34.7% |
| 2024 | 1345 | 22.5% | 15.8% | 3.0% | 23.3% | 42.5% |
| 2025 | 1328 | 27.2% | 20.8% | 3.1% | 25.4% | 39.2% |
| 2026 | 1366 | 27.1% | 21.0% | 4.3% | 28.3% | 38.1% |
Subtypes: share and odds
Not all AI projects are alike. Reinforcement learning (46.7% of 60), generative models (44.1% of 118), time-series forecasting (42.6% of 242), edge or embedded AI (40.7% of 54) and robots or drones (40.0% of 427) run well above the base rate. Computer vision (34.9% of 607) and NLP, LLM or chatbot projects (34.5% of 249) are above average but are the most common subtypes, so they rarely differentiate. Projects built on a microcontroller alone sit at 23.1%, below the 27% base; a microcontroller running a trained model (edge AI) sits at 40.7%.
Thin lines mark the 95% confidence interval; smaller samples give longer lines.
Show the full table (16 rows)
| Subtype | 2014 to 16 | 2017 to 19 | 2021 to 23 | 2024 to 26 | n 2021 to 26 | Grand rate (95% CI) | Special rate | Korean n | Korean grand |
|---|---|---|---|---|---|---|---|---|---|
| reinforcement learning | 0.2% | 0.3% | 0.5% | 1.0% | 60 | 46.7% (95% CI 35 to 59) | 28.3% | 3 | 33.3% |
| generative / diffusion / GAN | 0.0% | 0.4% | 1.1% | 1.8% | 118 | 44.1% (95% CI 35 to 53) | 29.7% | 2 | 50.0% |
| time-series / forecasting | 0.6% | 1.5% | 2.3% | 3.6% | 242 | 42.6% (95% CI 36 to 49) | 30.6% | 6 | 16.7% |
| edge / embedded AI | 0.1% | 0.2% | 0.4% | 0.9% | 54 | 40.7% (95% CI 29 to 54) | 38.9% | 2 | 0.0% |
| robot / drone | 3.1% | 3.8% | 4.5% | 6.0% | 427 | 40.0% (95% CI 36 to 45) | 28.3% | 9 | 22.2% |
| deep learning (generic) | 1.2% | 5.1% | 9.4% | 10.3% | 805 | 35.8% (95% CI 33 to 39) | 26.1% | 12 | 8.3% |
| classical ML (RF/SVM/XGB) | 0.8% | 1.6% | 3.0% | 4.2% | 294 | 35.4% (95% CI 30 to 41) | 24.5% | 2 | 50.0% |
| bioinformatics / omics ML | 0.9% | 1.4% | 3.0% | 5.1% | 333 | 35.1% (95% CI 30 to 40) | 17.4% | 2 | 0.0% |
| computer vision / image | 0.9% | 2.7% | 6.6% | 8.2% | 607 | 34.9% (95% CI 31 to 39) | 27.2% | 17 | 5.9% |
| NLP / LLM / chatbot | 0.7% | 0.9% | 1.6% | 4.6% | 249 | 34.5% (95% CI 29 to 41) | 31.7% | 11 | 0.0% |
| wearable / sensor device | 1.4% | 2.4% | 2.4% | 3.3% | 235 | 34.5% (95% CI 29 to 41) | 26.4% | 3 | 33.3% |
| app / web product | 4.5% | 6.2% | 7.7% | 9.8% | 718 | 33.8% (95% CI 30 to 37) | 25.9% | 21 | 14.3% |
| explainable / interpretable | 0.0% | 0.0% | 0.7% | 2.5% | 130 | 32.3% (95% CI 25 to 41) | 20.0% | 2 | 0.0% |
| simulation / numerical | 8.7% | 9.0% | 10.1% | 15.1% | 1029 | 31.4% (95% CI 29 to 34) | 23.0% | 15 | 20.0% |
| blockchain / crypto | 0.8% | 0.9% | 0.5% | 0.7% | 52 | 30.8% (95% CI 20 to 44) | 42.3% | 3 | 0.0% |
| Arduino / microcontroller | 3.1% | 3.8% | 3.8% | 4.1% | 324 | 23.1% (95% CI 19 to 28) | 21.6% | 5 | 0.0% |
Instrument inside a field, or the field itself
Grand-award rates for AI and non-AI projects by macro group, 2021 to 2026: Biology and Medicine 36.3% vs 24.5% (n=706 vs 2,759), Earth and Environment 40.9% vs 24.6%, Behavioral and Social 43.7% vs 21.9%, Physical Sciences and Math 39.6% vs 26.1%, Engineering and Energy 40.4% vs 25.9%, Computing and Robotics 29.9% vs 24.4%. By category the AI lift is largest in Animal Sciences (+34 points), Earth and Environmental Sciences (+27) and Behavioral and Social Sciences (+22), and near zero in Robotics (+2), Embedded Systems (+1) and Environmental Engineering (+1).
The premium is shrinking. Outside computing, AI projects won 43% vs 24% in 2021 and 30% vs 26% in 2026, while the AI share rose from 12.5% to 21%. Inside computing the gap disappeared in 2022 to 2024 and reopened in 2025 to 2026 (32% vs 21%) as AI reached 65% of entries, which means the non-AI computing projects are now the ones struggling. The durable advantage is method, not the presence of a model; see winning abstracts.
AI = machine-learning or neural-network mention; device = sensor, microcontroller or hardware mention in the abstract.
Korea's AI paradox
Korean entries mention AI at the world rate (23% vs 22.7% in 2021 to 2026), yet Korean AI projects won at 10.9% (46 projects) and Korean non-AI projects at 16.8% (226), the reverse of the global pattern. Korean computer-vision entries: 17, one grand award. Korean NLP or LLM entries: 11, none. Korean app or web products: 21, three. The same subtypes worldwide win at 34 to 35%. The difference reads as design, validation and framing rather than topic.
The AI projects that took First Awards in 2024 to 2026 share a shape: a named system (SmartBlink, MalariaX, HandTalk, HARMONI), an opening sentence that sizes the human problem (11 million Americans with severe hearing loss; a $35,000 microscope), concrete components (YOLOv11, MediaPipe, INT8 quantization, Geant4), numbers in the result sentence (95% detection accuracy for wheelchair users; lift-to-drag ratio three times the original), and usually hardware or a field deployment attached to the model.
Worldwide the same types win grand awards 34 to 35% of the time.
Frequently asked questions
Data and method
Two public sources. From the Society for Science abstract database (abstracts.societyforscience.org) we compiled every ISEF finalist project from 2014 to 2026: 17,565 projects with title, category, fair country, awards and full abstract. From the isef.net project library we took the award-tier flags for the 8,134 ISEF projects of 2021 to 2026 and joined them to the abstracts by year and booth id. Compiled on 2026-10-11.
A grand award means a First to Fourth Award or Best of Category, roughly the top 27% of each category. A special award is given by a sponsoring organization with its own judges. ISEF 2020 was held without judging and is excluded from every rate. The abstract database carries no 2025 awards and misses most Materials Science grand awards after 2017, so award flags for 2021 to 2026 combine both sources. An AI mention means the title or abstract contains a machine-learning term; it does not mean the project is about AI. All numbers are correlations, and Korean per-category figures rest on 2 to 24 projects, so read them as direction rather than precision.
Sources
- https://abstracts.societyforscience.org/
- https://isef.net/search-results/projects
- https://www.societyforscience.org/isef/awards/
The figures on this page are category-level base rates compiled from public ISEF results. They do not predict or promise any student's award, selection or admission. CIT does not guarantee awards, selection or admission.
ISEF data series
Want to pressure-test a project idea against this data?
Tell us the student's grade, field of interest and target fair. We will map a realistic preparation order against these numbers.