Machine learning is a fundamentally different approach to programming than traditional software. In conventional programming, you write explicit rules: "If X > 10, then do Y." In machine learning, you provide examples of inputs and desired outputs, and the system learns patterns from those examples.

Key points:

• ML systems learn from data rather than being explicitly programmed
• They identify statistical patterns in training examples
• They generalize these patterns to make predictions on new data
• Performance depends entirely on the quality and representativeness of training data
• ML is powerful for complex problems where explicit rules are hard to define (pattern recognition, language understanding) but lacks the ability to reason about novel situations

Every AI system's abilities and limitations flow directly from what data it was trained on. Understanding this is crucial for researchers.

Key points:

• Training data is the corpus (collection) of examples the model learned from
• The quality, diversity, and completeness of training data determines what the model can do
• Training data creates "knowledge cutoffs"—gaps in what the model knows
• Biases in training data become biases in the model's outputs
• A model cannot learn patterns it never saw in training data
• For example, Claude was trained on data up to mid-2024; it cannot know about papers published after that date, regardless of how the question is asked

LLMs like Claude, ChatGPT, and Gemini work through a sophisticated but fundamental process: predicting the next most likely word (or token) in a sequence, billions of times.

Key points:

• LLMs are trained to predict the next token (roughly word or sub-word) given previous tokens
• During use, they start with your prompt and generate one token at a time
• Each token is generated probabilistically based on patterns learned from training data
• This happens so fast and repeatedly that it appears as smooth, natural language generation
• The model is not consulting information sources or retrieving facts—it's performing pattern completion
• Token limits (like 4,000 or 100,000 tokens for context windows) exist because models can only process a fixed amount of information at once

Understanding the scope of AI capabilities today is essential for realistic expectations.

Key points:

• Narrow AI (weak AI) is specialized in specific tasks: recognizing images, translating languages, generating text
• All current AI systems—including advanced LLMs—are narrow AI
• General AI (strong AI) would be a system capable of understanding and learning any intellectual task humans can
• We do not yet have general AI; all current systems excel in specific domains but fail outside them
• LLMs' apparent intelligence is narrow: they're exceptional at language patterns but cannot truly reason, understand causation, or think about novel problems
• This is why AI is powerful for specific research tasks but not a replacement for researcher judgment

This distinction explains both AI's power and its fundamental limitations for research.

Key points:

• Pattern matching: identifying statistical correlations in data (what AI does extremely well)
• Reasoning: making logical inferences about novel situations using causal understanding (what AI does poorly)
• AI can identify that papers with certain keywords cluster together—that's pattern matching
• AI struggles with "If A causes B, and B causes C, therefore A causes C in this new context"—that requires reasoning
• Researchers need reasoning for hypothesis generation, experimental design, and interpreting unexpected results
• This is why AI is useful for summarizing existing literature but problematic for novel conceptual work without human oversight

—

Practical Research Use Cases

Scenario: You're beginning research on neuroinflammation's role in Alzheimer's disease. Your university database search for "neuroinflammation" returns thousands of papers.

Without AI: You manually read abstracts for 2-3 hours, screening papers by relevance, taking notes on methodology, struggling to remember which papers reported contradictory findings about microglial activation.

With AI: You upload 10 relevant papers you've found, ask an AI system to identify the dominant hypotheses about microglial activation mechanisms, extract key methodological differences between papers, and highlight contradictions. The system can pattern-match across papers to show you that papers from 2018-2020 emphasized different mechanisms than 2021-2024 papers. You can then search strategically rather than exhaustively. Understanding how AI works means you verify these patterns in the original papers rather than trusting the AI's synthesis.

Scenario: You have a dataset with 50,000 tumor samples and need to perform multiple regression analysis controlling for demographic variables.

Without AI: You spend 2 hours writing R or Python code, consulting Stack Overflow, debugging syntax errors, and validating that your statistical approach is correct.

With AI: You describe your analysis in plain language to Claude or ChatGPT, which generates functional code in your preferred language, including appropriate statistical tests. Knowing that LLMs pattern-match from thousands of Stack Overflow examples and code repositories means you rigorously review the generated code for statistical correctness and check for assumptions being met, rather than assuming it's correct because the AI "knows statistics."

Scenario: You've written a dense methods section explaining your proteomics workflow. Early readers say it's hard to follow.

Without AI: You revise manually, reading each sentence, rewriting for clarity, hoping you've improved it enough. This takes hours.

With AI: You ask Claude to identify unclear passages and suggest rewrites. The system patterns-matches what clarity looks like in scientific writing. However, you review every suggestion because AI might oversimplify complex procedures or miss domain-specific terminology that's actually necessary. You're using AI to amplify your revision process, not replace your judgment about what needs to stay precise.

Scenario: Your lab is considering adopting a new single-cell RNA sequencing protocol you found in a preprint. You need to quickly assess whether it will work for your tissue type and budget.

Without AI: You re-read the protocol three times, look up unfamiliar reagents online, contact the authors with questions. This takes days.

With AI: You ask an LLM to extract key protocol steps, identify cost assumptions, and flag potential limitations. It quickly identifies that the protocol uses droplet-based partitioning (which may not work well for your rare cell types) and requires specialized equipment. Understanding how AI works means you verify this assessment against the original protocol and your cell biology knowledge, rather than accepting the AI's categorization as definitive.

—

Hands-On Exercise

Objective: Directly experience how LLMs work through next-token prediction.

Steps:

1. Access an LLM (ChatGPT, Claude, or similar through your institution)

1. Create a test prompt: Write: "The most important factor in successful research is..."

1. Complete the prompt multiple times: Ask the AI to continue this sentence three times. Each time, note the completions. Do they seem confident and well-written? Yes. Are they identical? Probably not.

1. Understand why: The model is predicting the most probable next token billions of times. Different probability distributions lead to different outputs. This is pattern completion, not retrieval of facts.

1. Test the boundaries: Ask the AI to complete: "The chemical formula for sodium chloride is..." Does it accurately say NaCl? Yes. Ask: "The chemical formula for a compound that doesn't exist called bneazium fluoride is..."

1. Observe the pattern: The AI will confidently generate a formula-like structure (perhaps Bnz3F7) even though this compound doesn't exist. Why? It learned the pattern of how chemical formulas are constructed but doesn't actually know chemistry—it's pattern-matching on the surface level.

1. Reflect: Write a 2-3 sentence reflection on: How does this observation change your understanding of what AI can reliably do in your research field?

Time required: 15 minutes

—

Common Mistakes and Misconceptions

Even advanced LLMs are statistically sophisticated pattern-matching systems. They don't understand in the way humans do. They don't have beliefs, intentions, or consciousness. They're exceptionally good at mimicking human writing patterns, which feels like understanding, but there's no sentience behind it. This matters for research because it means AI might confidently explain something without actually grasping the concept, and you must verify rather than trust.

Because AI systems have knowledge cutoffs, they can't learn from recent papers unless you explicitly provide them. Asking ChatGPT about a 2025 discovery won't help because the model hasn't been updated with that information. However, if you paste the 2025 paper text into Claude, it can analyze that content even though it's outside the training window. Understanding this prevents wasting time asking about information that exists but is outside the AI's training.

This conflates narrow AI capabilities (which are advancing rapidly) with general AI (which remains theoretical). Current AI excels at specific tasks but has no genuine understanding. Even if AI becomes better at research tasks, there's a crucial difference between "better at some tasks" and "capable of doing all researcher work." Humans bring judgment, creative hypothesis generation, and ethical reasoning that current systems don't have and may never have.

One of AI's most dangerous characteristics is overconfidence. LLMs generate text that sounds natural and authoritative whether or not the content is accurate. A false statement flows as smoothly as a true one. In research, this makes AI particularly risky for domains requiring absolute accuracy. Always verify factual claims, especially in methodology, statistics, and scientific interpretation.

Training data reflects the world it's drawn from—including biases, errors, and gaps. If historical scientific literature overrepresented studies from wealthy countries, AI trained on that literature will overweight findings from those regions. If scientific papers had publication bias toward positive results, AI will pattern-match toward that bias too. Researchers must understand that AI doesn't fix biases in source materials; it can amplify them.

—

Key Takeaways

• Machine learning systems learn patterns from training data rather than following explicit rules, which makes them powerful for complex tasks but dependent entirely on data quality
• Large language models generate text through next-token prediction: they don't retrieve facts or reason about causality; they statistically predict likely continuations of sequences
• Training data creates knowledge cutoffs and shapes capabilities: AI systems cannot know information outside their training data and will reflect biases present in that data
• All current AI is narrow AI: specialized in specific domains with no general reasoning ability, meaning AI tools must be matched to appropriate tasks
• Pattern matching is not reasoning: AI excels at identifying statistical correlations but fails at causal reasoning and novel conceptual work, which remain researcher responsibilities
• Confidence is not accuracy: AI generated text sounds equally authoritative whether true or false, making verification essential in research contexts

—

Reflection Questions

1. Think about your own research field: What aspects of your work involve pattern recognition (where AI could help) versus novel reasoning and judgment (where AI must stay in a supporting role)?

1. Knowledge cutoff reality: What recent developments or papers in your field are likely outside an AI system's training data? How would you need to adapt your workflow to account for this?

1. Trust and verification: How does understanding that AI is pattern-matching rather than reasoning change how you'd approach using AI-generated literature syntheses or data analysis suggestions?

1. Bias inheritance: What biases might be present in the source materials your field relies on (publication bias, geographic representation, methodological trends)? How might AI trained on those materials amplify those biases?