Hypothesis Generation & Exploration

The Ideation Bottleneck

The hardest part of research is not testing hypotheses -- it is generating good ones. A good hypothesis is specific, testable, grounded in existing evidence, and ideally non-obvious. Most researchers generate hypotheses from a narrow band of literature they happen to have read, constrained by their training and disciplinary norms. AI expands this band. An LLM trained on the breadth of scientific literature can surface connections that a specialist might miss: patterns from adjacent fields, historical precedents, analogies from distant domains. It is not generating knowledge -- it is generating candidates for your scientific judgment. The critical distinction: AI generates hypotheses. You evaluate them. The scientific contribution is the evaluation, not the generation. A hypothesis has no value until someone designs an experiment to test it and interprets the results.

Structured Hypothesis Brainstorming

Unstructured prompts ("give me research ideas") produce generic output. Structured prompts produce testable hypotheses. Here is a framework: ``` HYPOTHESIS GENERATION PROMPT: Context: - Field: {your_field} - Current knowledge: {brief summary of what is established} - Open question: {the specific gap you want to address} - Constraints: {equipment, budget, timeline, ethical limits} Generate 10 hypotheses that could explain or address the open question. For each hypothesis: 1. STATEMENT: One-sentence testable prediction 2. MECHANISM: Proposed causal mechanism 3. EVIDENCE: Existing evidence that supports or contradicts this 4. TEST: How would you test this? What experiment would confirm or refute it? 5. NOVELTY: How is this different from existing hypotheses in the literature? 6. RISK: What would make this hypothesis wrong? Prioritize hypotheses that are: - Testable with {constraints} - Non-obvious (not direct extensions of existing work) - Specific enough to be falsifiable ``` The key is specificity in your context. The more precisely you describe what is known, what is unknown, and what resources you have, the more targeted and useful the hypotheses will be.

Cross-Domain Analogy

The most powerful hypothesis generation technique is cross-domain analogy: finding structurally similar problems in unrelated fields and importing their solutions. ``` CROSS-DOMAIN PROMPT: I study {phenomenon} in {your_field}. The key properties of this phenomenon are: - {property_1} - {property_2} - {property_3} What analogous phenomena exist in completely different fields (physics, biology, economics, computer science, ecology, sociology)? For each analogy: 1. What is the analogous phenomenon? 2. What has that field learned about it? 3. What solutions or explanations have they found? 4. How could those solutions translate to my domain? 5. What are the limits of this analogy -- where does it break down? ``` Example: A researcher studying information cascades in social media might find analogies in epidemiology (viral spread), ecology (invasive species propagation), physics (phase transitions), and economics (bank runs). Each analogy suggests different mechanisms and testing strategies that may not be obvious from within the communication studies literature alone.

Evaluating AI-Generated Hypotheses

Not every AI-generated hypothesis deserves investigation. Apply these filters: **Filter 1: Testability.** Can you design a concrete experiment that would confirm or refute this hypothesis? If the answer requires technology that does not exist or data that cannot be collected, it is not currently testable. Set it aside for future consideration. **Filter 2: Novelty.** Search for the hypothesis in your literature database. Has it already been proposed and tested? If yes, what were the results? A hypothesis is only worth pursuing if it is either untested or tested with inconclusive results. **Filter 3: Mechanism.** Does the proposed mechanism make physical, biological, or logical sense? AI can propose plausible-sounding mechanisms that violate basic domain constraints. Your expertise is the filter here. **Filter 4: Feasibility.** Given your resources (lab equipment, compute budget, timeline, team size), can you actually test this hypothesis? Brilliance is irrelevant if the experiment costs $10M and you have $10K. ``` EVALUATION MATRIX: | Hypothesis | Testable? | Novel? | Mechanistic? | Feasible? | Score | |------------|-----------|--------|--------------|-----------|-------| | H1: ... | Yes | Yes | Plausible | Yes | 4/4 | | H2: ... | Yes | No | Strong | Yes | 3/4 | | H3: ... | Unclear | Yes | Weak | No | 1/4 | ``` Hypotheses scoring 3-4 go to experiment design (Lesson 5). Hypotheses scoring 1-2 get documented for future reference but are not pursued now.

The Contribution Boundary

Where is the line between "AI helped me think" and "AI did my thinking"? This is important for intellectual honesty and for your career. AI-assisted hypothesis generation is scientifically legitimate when: - You provided the domain context, constraints, and evaluation criteria - You evaluated, filtered, and selected the hypotheses using your expertise - You designed the experiments to test them - You interpreted the results - You disclosed the AI's role in your methods section AI-assisted hypothesis generation is problematic when: - You accepted hypotheses without evaluation or domain expertise - You cannot explain why a hypothesis is plausible in your own words - You treated the AI's output as evidence rather than as a suggestion The test is simple: if a colleague asked "why do you think this hypothesis is worth testing?" could you defend it without mentioning AI? If yes, the contribution is yours. The AI was a brainstorming tool. If no, you need to go deeper before pursuing it.