29 AI Across Disciplines
AI has escaped the computer science department. It is transforming medicine, physics, law, finance, education, art, and robotics, often in ways that researchers in those fields never anticipated. This chapter surveys the most important cross-domain applications, the unique challenges each domain presents, and the opportunities for anyone willing to bridge the gap between AI expertise and domain knowledge.
The most impactful applied AI researchers are T-shaped: deep in machine learning fundamentals (the vertical bar) and broad in at least one other domain (the horizontal bar). If you understand both transformers and protein biochemistry, or both attention mechanisms and contract law, you can identify problems that pure ML researchers overlook and build solutions that pure domain experts cannot.
29.1 AI in Healthcare and Medicine
Healthcare may be the domain where AI has the most potential to save lives, and also where the stakes of getting it wrong are highest.
29.1.1 Medical Imaging
Deep learning models now match or exceed radiologist performance in specific tasks: detecting diabetic retinopathy from retinal scans, identifying tumors in mammograms, and spotting pneumonia on chest X-rays. The key insight is that these models are not replacing radiologists but triaging their attention. A model that can reliably flag normal scans lets radiologists focus on the difficult cases.
29.1.2 Drug Discovery
The traditional drug development pipeline takes 10 to 15 years and costs over a billion dollars. AI is accelerating every stage:
- Target identification: NLP models mine biomedical literature to find promising drug targets.
- Molecular generation: Diffusion models and VAEs can propose novel molecules with desired properties.
- Protein structure prediction: AlphaFold (Jumper et al. 2021) revolutionized structural biology by predicting 3D protein structures from amino acid sequences, earning Demis Hassabis and John Jumper a Nobel Prize in Chemistry. AlphaFold3 extends this to protein complexes and drug-protein interactions.
- Clinical trial optimization: AI helps identify suitable patients, predict adverse effects, and design more efficient trials.
29.1.3 Clinical NLP
LLMs are being applied to clinical notes for diagnoses coding, patient summarization, and information extraction. But medical AI has a uniquely challenging failure mode: hallucinations. A chatbot that invents a plausible-sounding but incorrect drug interaction can cause real harm. This drives the need for retrieval-augmented generation and careful human oversight.
Healthcare AI must navigate HIPAA (US), GDPR (EU), and other regulations that restrict how patient data can be used. Federated learning, where models are trained across hospitals without sharing raw data, is one promising approach. Synthetic data generation is another. The technical challenge is ensuring that privacy-preserving methods do not degrade model quality.
29.2 AI in Scientific Research
29.2.1 Physics and Chemistry
AI is becoming an essential tool in the physical sciences:
- Molecular dynamics: Neural network potentials (like MACE and NequIP) learn energy surfaces from quantum chemistry calculations and run simulations 1000x faster than traditional methods.
- Materials discovery: GNoME (Merchant et al. 2023) (Google DeepMind) predicted 2.2 million new stable crystal structures, vastly expanding the catalog of known materials.
- Particle physics: Deep learning classifies collision events at the Large Hadron Collider, separating signal from enormous backgrounds.
- Weather forecasting: GraphCast (Lam et al. 2023), Pangu-Weather, and GenCast now rival or outperform traditional numerical weather prediction models that took decades to develop, and they run in minutes instead of hours.
Geometric deep learning (Bronstein et al. 2021) (Chapter 22) is central to many of these applications because physical systems have inherent symmetries that models must respect.
29.2.2 Biology and Genomics
Beyond AlphaFold, AI is transforming biology:
- Foundation models for biology: ESM (Meta) for protein sequences, Evo (Arc Institute) for DNA, and scGPT for single-cell transcriptomics are the biological equivalents of GPT: large models pre-trained on massive biological datasets that can be fine-tuned for specific tasks.
- CRISPR design: Models predict the efficiency and off-target effects of guide RNAs, making gene editing safer and more precise.
- Ecological modeling: Computer vision models identify species from camera trap images and satellite photos, enabling large-scale biodiversity monitoring.
29.3 AI in Law
The legal profession generates enormous amounts of structured text (statutes, case law, contracts, regulatory filings), making it a natural target for NLP:
- Legal research: LLM-powered tools search case law and statutes far more efficiently than keyword search.
- Contract review: Models identify risky clauses, missing provisions, and inconsistencies across hundreds of pages in minutes.
- Case summarization: LLMs summarize lengthy court opinions into key holdings.
- Compliance checking: Models map regulations to specific business processes and flag potential violations.
In 2023, a lawyer submitted a brief citing six cases that GPT had hallucinated. None existed. This case became a cautionary tale about using LLMs in domains where factual accuracy is non-negotiable. RAG systems that ground LLM responses in verified legal databases are essential. Legal AI must cite its sources, and humans must verify them.
29.4 AI in Finance
Finance has been an early adopter of machine learning:
- Algorithmic trading: RL agents and time-series models execute trades based on market signals.
- Credit scoring: ML models assess creditworthiness from alternative data (transaction patterns, behavioral signals).
- Fraud detection: Anomaly detection models identify unusual transaction patterns in real time.
- Document analysis: LLMs extract information from earnings calls, SEC filings, and financial reports.
Key concerns include model interpretability (regulators require explanations for credit decisions), adversarial robustness (trading models must handle market manipulation), and the risk of feedback loops where AI trading systems amplify volatility.
29.5 AI in Education
AI-powered education has the potential to provide every student with a personal tutor:
- Intelligent tutoring: LLMs can explain concepts, answer follow-up questions, and adapt explanations to the student's level. Khan Academy's Khanmigo is an early example.
- Automated assessment: Models grade essays, provide feedback on writing quality, and generate practice problems at appropriate difficulty levels.
- Curriculum adaptation: Systems adjust content presentation based on student performance, spending more time on weak areas.
In 1984, Benjamin Bloom (Bloom 1984) showed that one-on-one tutoring improved student performance by two standard deviations compared to conventional classrooms. The challenge has always been cost: human tutors are expensive. LLMs could finally solve Bloom's two-sigma problem by providing personalized tutoring at scale. Whether current AI tutors actually achieve two-sigma improvements remains an open question and an exciting research direction.
29.6 AI in Creative Arts
Generative AI has ignited both excitement and controversy in the arts:
- Visual art: Stable Diffusion, Midjourney, and DALL-E create images from text descriptions, enabling anyone to visualize ideas. Professional artists use these tools for concept art, storyboarding, and rapid iteration.
- Music: MusicGen (Meta), Suno, and Udio generate music from text prompts or continue existing melodies. The quality has improved rapidly from “interesting noise” to “indistinguishable from human production” in some genres.
- Writing: LLMs assist with brainstorming, drafting, editing, and translation. They are not replacing authors but changing the writing process.
- Video: Sora (OpenAI), Kling, and Veo generate video clips from text or images, though long-form coherent video remains challenging.
These applications raise unresolved questions about copyright (can AI-generated content be copyrighted? does training on copyrighted works constitute fair use?), attribution, economic displacement of creative workers, and what we value about human creativity.
29.7 AI in Robotics
Robotics is experiencing a transformer-driven revolution:
- Vision-Language-Action models: RT-2 (Brohan et al. 2023) and OpenVLA (Kim et al. 2024) connect visual perception with language understanding and physical action. A robot can follow instructions like “pick up the blue cup and put it on the second shelf” by grounding language in perception and translating both into motor commands.
- World models: Rather than learning reactive policies, world models (Chapter 19) let robots plan by simulating the consequences of actions before executing them.
- Sim-to-real transfer: Training in simulation and transferring to physical robots avoids the expense and danger of real-world training. Domain randomization and physics-accurate simulators are key enablers.
If you have expertise in any of these fields and are learning AI, you have a significant advantage over pure ML researchers. The bottleneck in applied AI is rarely the model architecture; it is understanding what problem to solve, what data to use, and how to evaluate the result. Domain expertise makes you invaluable.
29.8 Exercises
- Choose a domain you are passionate about (medicine, physics, law, music, or another field). Find three recent papers applying AI in that domain. Summarize the most important open challenge that AI has not yet solved.
- Build a simple domain-specific RAG chatbot: ingest a set of domain-specific documents (e.g., legal statutes, medical guidelines, or academic papers) and evaluate whether the LLM's answers are factually correct. How often does it hallucinate domain-specific facts?
- Discuss the ethical implications of deploying AI in one high-stakes domain of your choice. What safeguards should be in place before deploying an AI system that makes consequential decisions?
- Explore AlphaFold's predictions on the EMBL-EBI database. Pick a protein relevant to a disease you find interesting and examine the predicted structure and confidence scores.