.png)
Module 1
How AI Systems Work
AI systems generate outputs by learning patterns from vast amounts of training data. You provide a prompt — a question, a task, an instruction — and the system responds with text that matches the pattern of a correct-looking answer. These systems can produce genuinely useful output: language, code, summaries, classifications, and analysis. They do not understand information the way people do, nor do they verify that their answers are true. A well-formed, polished answer is not evidence the system verified anything. AI can be useful without being reliable. Fluency is not understanding, and confidence is not correctness.
What AI Is Good For
Used within its strengths, AI is genuinely useful. These are the tasks it reliably does well — where speed and fluency are the point, and a human still reviews the result.
| Use | What AI does well |
|---|---|
| Drafting | Produces a solid first version fast, so you start from something to refine rather than a blank page. |
| Summarizing | Compresses long or dense material into its key points, giving you a quick way in before you read the source. |
| Rewriting | Adjusts tone, structure, and clarity on demand, and can offer several versions to choose between. |
| Brainstorming | Generates a wide range of options, angles, and starting points to react to and build on. |
| Transformation | Converts content between styles, structures, and formats — outline to prose, notes to table, one format to another. |
| Pattern assistance | Surfaces possible themes, groupings, or inconsistencies across text for you to confirm. |
What This Looks Like in Practice
Two documented examples of AI capability at its most consequential — not everyday drafting, but work with real, verifiable impact.
AlphaFold2 predicted the structure of nearly all 200 million known proteins, now used by more than two million researchers across 190 countries — work that won the 2024 Nobel Prize in Chemistry.
Source: The Nobel Prize in Chemistry 2024, Nobel Prize Outreach, October 2024; underlying research: Jumper, J. et al. “Highly accurate protein structure prediction with AlphaFold.” Nature 596, 583–589 (2021).
Be My Eyes’ “Be My AI,” built on GPT-4, gives blind and low-vision users real-time visual description of daily surroundings — reading a label, describing a room, checking what’s in a fridge.
Source: Be My Eyes & OpenAI, “Introducing Be My AI,” joint announcement, March 2023.
These are landmark, well-documented cases — but they’re not the only shape AI capability takes, and they won’t be the last ones written.
What AI Is Weak or Risky For
| Risky Use | Why It Needs Control |
|---|---|
| Factual claims | May fabricate or misstate, confidently. See: Hallucination → |
| Legal, medical, financial, or HR tasks | Stakes are high; context matters; AI cannot take responsibility. See: Context Collapse → |
| Complex calculations and formal logic | Can look well-reasoned while the steps are wrong; verify independently. See: False Precision → |
| Security-sensitive tasks | Can leak, mishandle, or be manipulated. See: Data Exposure → |
| Autonomous action | Can act on bad assumptions at machine speed. See: Security Failure → |
| Long planning chains | Can drift, omit constraints, or contradict itself. See: Automation Drift → |
Is AI Appropriate for This Task?
Before using AI on any task, work through these questions:
- Does this require factual accuracy? Verify independently before using the output.
- Does it involve sensitive or private data? Do not paste until organizational policy is clear.
- Could someone be harmed by a wrong answer? Require human review before action.
- Could money, access, employment, safety, or legal status be affected? Treat as High or Critical risk.
- Will AI take action automatically? Require formal controls before proceeding.
- Would you be comfortable explaining this decision later? If no, stop.
Before Module 2: run the questions above on a task you’re actually facing.