When AI Started Doing Science: Machines as Researchers

AI'S EXPANDING ROLE IN SCIENCE

PHASE 1: DATA ANALYSIS (1990s-2010s) ┌─────────────────────────────────────────┐ │ Machine learning for pattern recognition│ │ ↓ │ │ Examples: │ │ • Particle physics (find signals in │ │ detector noise) │ │ • Genomics (identify gene patterns) │ │ • Astronomy (classify galaxies) │ │ ↓ │ │ Role: Tool for humans │ └─────────────────────────────────────────┘

PHASE 2: PREDICTION (2010s) ┌─────────────────────────────────────────┐ │ AI predicts outcomes from data │ │ ↓ │ │ Examples: │ │ • Drug binding affinity │ │ • Material properties │ │ • Protein interactions │ │ ↓ │ │ Role: Hypothesis generator │ └─────────────────────────────────────────┘

PHASE 3: DISCOVERY (2020s) ┌─────────────────────────────────────────┐ │ AI discovers new knowledge │ │ ↓ │ │ Examples: │ │ • AlphaFold (protein structures) │ │ • AlphaTensor (matrix multiplication │ │ algorithms) │ │ • Drug discovery (new molecules) │ │ ↓ │ │ Role: Independent researcher │ └─────────────────────────────────────────┘

PHASE 4: AUTOMATION (Emerging) ┌─────────────────────────────────────────┐ │ AI runs entire experiments │ │ ↓ │ │ Examples: │ │ • Robot scientists (design, run, analyze│ │ experiments) │ │ • AI-designed chips │ │ • Automated labs │ │ ↓ │ │ Role: Replaces human scientists │ └─────────────────────────────────────────┘

ALPHAFOLD 2 (DeepMind, 2020)

THE PROBLEM: ┌─────────────────────────────────────────┐ │ Protein Folding Problem: │ │ ↓ │ │ Proteins = chains of amino acids │ │ ↓ │ │ Fold into 3D shapes │ │ ↓ │ │ Shape determines function │ │ ↓ │ │ Challenge: Predict shape from sequence │ └─────────────────────────────────────────┘

TRADITIONAL APPROACH: ┌─────────────────────────────────────────┐ │ 1. Crystallize protein (hard, often │ │ fails) │ │ ↓ │ │ 2. X-ray crystallography (expensive) │ │ ↓ │ │ 3. Solve structure (months, requires │ │ expertise) │ │ ↓ │ │ Cost: $100K-$1M per protein │ │ Time: Months to years │ │ Success rate: ~30-40% │ └─────────────────────────────────────────┘

ALPHAFOLD APPROACH: ┌─────────────────────────────────────────┐ │ 1. Train neural network on known │ │ structures (~170,000 proteins) │ │ ↓ │ │ 2. Learn patterns of folding │ │ ↓ │ │ 3. Predict structure from sequence │ │ ↓ │ │ Cost: Essentially free (after training) │ │ Time: Minutes │ │ Accuracy: ~90% │ └─────────────────────────────────────────┘

THE IMPACT: ┌─────────────────────────────────────────┐ │ 2022: AlphaFold predicts 200M+ protein │ │ structures │ │ ↓ │ │ Made freely available │ │ ↓ │ │ Entire fields transformed: │ │ • Drug discovery │ │ • Molecular biology │ │ • Biotechnology │ │ ↓ │ │ Work that would take 100+ years done in │ │ months │ └─────────────────────────────────────────┘

WHAT IT MEANS: ┌─────────────────────────────────────────┐ │ Structural biology no longer bottleneck │ │ ↓ │ │ Entire career paths obsolete │ │ ↓ │ │ New questions now askable │ │ ↓ │ │ Science accelerated by orders of │ │ magnitude │ └─────────────────────────────────────────┘

AI DRUG DISCOVERY

TRADITIONAL DRUG DISCOVERY: ┌─────────────────────────────────────────┐ │ 1. Identify target (disease protein) │ │ ↓ │ │ 2. Screen millions of molecules │ │ (5-10 years) │ │ ↓ │ │ 3. Test promising candidates │ │ (animal studies, clinical trials) │ │ ↓ │ │ 4. Most fail │ │ ↓ │ │ Timeline: 10-15 years │ │ Cost: $1-2 billion per approved drug │ │ Success rate: <10% │ └─────────────────────────────────────────┘

AI-ASSISTED DRUG DISCOVERY: ┌─────────────────────────────────────────┐ │ 1. AI predicts which molecules will bind│ │ to target │ │ ↓ │ │ 2. AI designs new molecules (never seen │ │ in nature) │ │ ↓ │ │ 3. AI predicts toxicity, side effects │ │ ↓ │ │ 4. Synthesize only most promising │ │ ↓ │ │ Timeline: 2-3 years (screening phase) │ │ Cost: Much lower │ └─────────────────────────────────────────┘

EXAMPLES: ┌─────────────────────────────────────────┐ │ INSILICO MEDICINE (2019-2021): │ │ • AI designs drug for fibrosis │ │ • 18 months from start to Phase 1 trial │ │ • Normal timeline: 5+ years │ │ ↓ │ │ EXSCIENTIA (2020): │ │ • AI-designed drug for OCD │ │ • 12 months to clinical candidate │ │ ↓ │ │ ATOMWISE (2020): │ │ • AI screens billions of molecules │ │ • Identifies COVID antivirals in weeks │ └─────────────────────────────────────────┘

THE REVOLUTION: ┌─────────────────────────────────────────┐ │ AI doesn't just speed up existing │ │ process │ │ ↓ │ │ It designs molecules humans wouldn't │ │ think of │ │ ↓ │ │ Explores chemical space humans can't │ │ (10^60 possible drug-like molecules) │ │ ↓ │ │ True discovery, not just assistance │ └─────────────────────────────────────────┘

AI IN MATHEMATICS

DEEPMIND'S ALPHATENSOR (2022) ┌─────────────────────────────────────────┐ │ Problem: Matrix multiplication │ │ ↓ │ │ Find most efficient algorithm │ │ ↓ │ │ AlphaTensor discovers new algorithms │ │ better than human-found (50+ years) │ │ ↓ │ │ Finds thousands of efficient algorithms │ │ for different matrix sizes │ └─────────────────────────────────────────┘

LEAN THEOREM PROVER + AI: ┌─────────────────────────────────────────┐ │ AI assists mathematicians in proving │ │ theorems │ │ ↓ │ │ Suggests proof strategies │ │ ↓ │ │ Checks logical validity automatically │ │ ↓ │ │ Speeds up proof verification │ └─────────────────────────────────────────┘

AI-GENERATED CONJECTURES: ┌─────────────────────────────────────────┐ │ Ramanujan Machine (Technion, 2021): │ │ • AI generates mathematical conjectures │ │ • Some novel, some known │ │ • Mathematicians verify/prove │ │ ↓ │ │ AI as hypothesis generator in pure math │ └─────────────────────────────────────────┘

THE CONTROVERSY: ┌─────────────────────────────────────────┐ │ If AI finds proof humans can't follow: │ │ ↓ │ │ Do we "understand" the theorem? │ │ ↓ │ │ Is computer-verified proof sufficient? │ │ ↓ │ │ What is mathematics if not human │ │ understanding? │ └─────────────────────────────────────────┘

AI'S SUPERHUMAN CAPABILITIES

CAPABILITY 1: PATTERN RECOGNITION IN HIGH DIMENSIONS ┌─────────────────────────────────────────┐ │ Humans: Can visualize 3 dimensions │ │ ↓ │ │ AI: Operates in thousands of dimensions │ │ ↓ │ │ Finds patterns invisible to humans │ │ ↓ │ │ Example: Gene interactions (20,000 │ │ genes = 20,000 dimensions) │ └─────────────────────────────────────────┘

CAPABILITY 2: EXHAUSTIVE SEARCH ┌─────────────────────────────────────────┐ │ Humans: Intuition guides search │ │ ↓ │ │ AI: Can search billions of possibilities│ │ ↓ │ │ Example: Drug molecules, material │ │ combinations │ │ ↓ │ │ Doesn't get tired, bored, biased │ └─────────────────────────────────────────┘

CAPABILITY 3: INTEGRATION OF VAST DATA ┌─────────────────────────────────────────┐ │ Humans: Read ~1000 papers max │ │ ↓ │ │ AI: Process millions of papers │ │ ↓ │ │ Synthesize connections across all │ │ scientific literature │ │ ↓ │ │ Example: Meta-analysis, cross-field │ │ discovery │ └─────────────────────────────────────────┘

CAPABILITY 4: SPEED ┌─────────────────────────────────────────┐ │ Protein structure prediction: │ │ • Human: Months │ │ • AI: Minutes │ │ ↓ │ │ Enables iterative design impossible for │ │ humans │ └─────────────────────────────────────────┘

CAPABILITY 5: HYPOTHESIS GENERATION ┌─────────────────────────────────────────┐ │ AI generates thousands of hypotheses │ │ ↓ │ │ Humans test most promising │ │ ↓ │ │ Expands possibility space │ └─────────────────────────────────────────┘

CURRENT AI LIMITATIONS

LIMITATION 1: CAUSAL UNDERSTANDING ┌─────────────────────────────────────────┐ │ AI finds correlations, not causation │ │ ↓ │ │ Can predict "X correlates with Y" │ │ ↓ │ │ Can't explain "X causes Y because..." │ │ ↓ │ │ Problem: Science needs causal mechanisms│ └─────────────────────────────────────────┘

LIMITATION 2: EXPERIMENTAL DESIGN ┌─────────────────────────────────────────┐ │ AI good at analyzing results │ │ ↓ │ │ Still weak at designing novel │ │ experiments │ │ ↓ │ │ Requires creativity, domain knowledge, │ │ "what would be interesting to test?" │ │ ↓ │ │ Improving, but not yet human-level │ └─────────────────────────────────────────┘

LIMITATION 3: ASKING NEW QUESTIONS ┌─────────────────────────────────────────┐ │ AI optimizes within defined problem │ │ ↓ │ │ Humans: "What if we asked a completely │ │ different question?" │ │ ↓ │ │ Paradigm shifts still require human │ │ insight │ └─────────────────────────────────────────┘

LIMITATION 4: INTERPRETABILITYThe extent to which a system can explain its own outputs. Interpretability is about tracing causes, not just achieving performance. ┌─────────────────────────────────────────┐ │ Neural networks = "black boxes" │ │ ↓ │ │ Can predict, but can't explain HOW │ │ ↓ │ │ Science wants understanding, not just │ │ predictions │ │ ↓ │ │ Problem: "It works but we don't know │ │ why" │ └─────────────────────────────────────────┘

LIMITATION 5: COMMON SENSE ┌─────────────────────────────────────────┐ │ AI lacks intuition about physical world │ │ ↓ │ │ Makes "stupid" mistakes obvious to │ │ humans │ │ ↓ │ │ Example: Proposes chemically impossible │ │ molecules, physically absurd experiments│ └─────────────────────────────────────────┘

FULLY AUTOMATED RESEARCH SYSTEMS

THE VISION: ┌─────────────────────────────────────────┐ │ AI + Robotics = Autonomous scientist │ │ ↓ │ │ System: │ │ 1. Generates hypotheses (AI) │ │ 2. Designs experiments (AI) │ │ 3. Runs experiments (robots) │ │ 4. Analyzes results (AI) │ │ 5. Iterates (loop back to step 1) │ │ ↓ │ │ No human in loop │ └─────────────────────────────────────────┘

EXAMPLES ALREADY EXIST: ┌─────────────────────────────────────────┐ │ ADAM (2009, Aberystwyth University): │ │ • Robot scientist in biology │ │ • Discovered gene functions in yeast │ │ • Autonomously designed experiments │ │ ↓ │ │ EVE (2020, Manchester): │ │ • Successor to ADAM │ │ • Drug discovery against malaria │ │ ↓ │ │ CLOUD LABS (Emerald Cloud Lab): │ │ • Fully robotic labs │ │ • Scientists submit experiments remotely│ │ • Robots execute │ └─────────────────────────────────────────┘

MATERIALS SCIENCE AUTOMATION: ┌─────────────────────────────────────────┐ │ AI designs material compositions │ │ ↓ │ │ Robots synthesize samples │ │ ↓ │ │ Automated testing │ │ ↓ │ │ AI analyzes, suggests next composition │ │ ↓ │ │ Loop runs 24/7, no human needed │ │ ↓ │ │ Speed: 100x-1000x faster than human labs│ └─────────────────────────────────────────┘

THE IMPLICATIONS: ┌─────────────────────────────────────────┐ │ Science becomes industrial process │ │ ↓ │ │ Humans: Set goals, interpret findings │ │ ↓ │ │ Machines: Do research │ │ ↓ │ │ Bottleneck: No longer human effort, but │ │ computational/experimental resources │ └─────────────────────────────────────────┘

UNDERSTANDING vs. PREDICTION

THE TRADITIONAL VIEW: ┌─────────────────────────────────────────┐ │ Science = Understanding nature │ │ ↓ │ │ Not just predicting, but explaining │ │ ↓ │ │ "Why does X happen? What mechanism │ │ produces X?" │ └─────────────────────────────────────────┘

AI CHALLENGE: ┌─────────────────────────────────────────┐ │ Neural networks predict accurately │ │ ↓ │ │ But: Can't explain HOW │ │ ↓ │ │ Internal workings = incomprehensible │ │ ↓ │ │ Question: Is prediction without │ │ explanation still science? │ └─────────────────────────────────────────┘

THE DEBATE: ┌─────────────────────────────────────────┐ │ POSITION 1: PREDICTION IS ENOUGH │ │ • Science's goal: Predict phenomena │ │ • Understanding = nice, but optional │ │ • If AI predicts perfectly, job done │ │ ↓ │ │ POSITION 2: UNDERSTANDING IS ESSENTIAL │ │ • Science seeks explanations, not just │ │ predictions │ │ • Must know WHY, not just WHAT │ │ • Black-box AI isn't real science │ └─────────────────────────────────────────┘

HISTORICAL PARALLEL: ┌─────────────────────────────────────────┐ │ Babylonian astronomy: │ │ • Predicted eclipses accurately │ │ • No understanding of mechanisms │ │ ↓ │ │ Was that science? │ │ ↓ │ │ Most say: No, pattern recognition only │ │ ↓ │ │ Is AI science same? Pattern recognition │ │ at massive scale? │ └─────────────────────────────────────────┘

THE MIDDLE GROUND: ┌─────────────────────────────────────────┐ │ AI discovers patterns │ │ ↓ │ │ Humans interpret, build theory │ │ ↓ │ │ Division of labor: │ │ • AI = Discovery engine │ │ • Humans = Meaning-makers │ └─────────────────────────────────────────┘

IMPACT ON SCIENTIFIC CAREERS

SCENARIO 1: DISPLACEMENT ┌─────────────────────────────────────────┐ │ AI automates: │ │ • Data analysis │ │ • Literature review │ │ • Hypothesis generation │ │ • Experiment design │ │ • Paper writing (already happening) │ │ ↓ │ │ What's left for humans? │ │ ↓ │ │ Fewer science jobs needed │ └─────────────────────────────────────────┘

SCENARIO 2: AUGMENTATION ┌─────────────────────────────────────────┐ │ AI = Tool, not replacement │ │ ↓ │ │ Scientists 10x-100x more productive │ │ ↓ │ │ Focus on: │ │ • Big questions │ │ • Interpretation │ │ • Theory building │ │ • Creativity │ │ ↓ │ │ Different skills, but still need humans │ └─────────────────────────────────────────┘

SCENARIO 3: BIFURCATION ┌─────────────────────────────────────────┐ │ Science splits: │ │ ↓ │ │ TIER 1: Elite scientists using AI │ │ (Small number, highly productive) │ │ ↓ │ │ TIER 2: Everyone else │ │ (Displaced or relegated to routine work)│ │ ↓ │ │ Inequality in scientific labor market │ └─────────────────────────────────────────┘

THE TRAINING PROBLEM: ┌─────────────────────────────────────────┐ │ If AI does science, how do we train │ │ scientists? │ │ ↓ │ │ Traditional: Learn by doing experiments,│ │ analyzing data │ │ ↓ │ │ But if AI does that, what do students │ │ learn? │ │ ↓ │ │ Risk: Generation that doesn't understand│ │ foundations │ └─────────────────────────────────────────┘

GOVERNANCE AND ACCESS

QUESTION 1: PROPRIETARY AI ┌─────────────────────────────────────────┐ │ AlphaFold: Open (DeepMind released) │ │ ↓ │ │ But many AI tools: Proprietary │ │ ↓ │ │ Only companies with resources can build │ │ powerful AI │ │ ↓ │ │ Science becomes corporate? │ └─────────────────────────────────────────┘

QUESTION 2: BIAS IN AI ┌─────────────────────────────────────────┐ │ AI trained on existing data │ │ ↓ │ │ Data contains historical biases │ │ ↓ │ │ AI perpetuates biases │ │ ↓ │ │ Example: Medical AI trained on mostly │ │ white male data → doesn't work well for │ │ women, minorities │ └─────────────────────────────────────────┘

QUESTION 3: ACCOUNTABILITY ┌─────────────────────────────────────────┐ │ If AI designs drug that harms patients: │ │ ↓ │ │ Who's responsible? │ │ • AI creator? │ │ • Scientist who used it? │ │ • Hospital/company? │ │ ↓ │ │ Legal/ethical framework unclear │ └─────────────────────────────────────────┘

QUESTION 4: DUAL USE ┌─────────────────────────────────────────┐ │ AI can design beneficial drugs │ │ ↓ │ │ Same AI can design bioweapons │ │ ↓ │ │ How to enable good uses, prevent bad? │ │ ↓ │ │ Governance challenge │ └─────────────────────────────────────────┘