AI for Everyone study guide · Lesson 13 of 14

What AI Can—and Cannot—Do

Strengths and the Reliability Gap · Energy, Water, and Physical Scale · A Regulation Snapshot

This is the lesson's study guide from the FlashFeed app, plus its key facts and a self-test.

What you'll be able to do

  • Choose reviewable AI uses, interpret footprint claims, and recognize jurisdiction-specific governance.

Strengths and the Reliability Gap

Current generative AI is often useful where work is pattern-rich, well specified, reversible, and reviewable. Common examples include:

  • drafting and rewriting;
  • summarizing supplied material;
  • extracting fields from documents;
  • classifying or clustering text;
  • translating with human review;
  • generating alternatives or examples;
  • assisting with code, tests, and documentation.

Those are task categories, not guarantees. Performance changes with the model, prompt, source material, language, domain, and evaluation method.

The most important limit is the reliability gap: an output can sound better than it is. Language models are optimized to produce plausible continuations, not to make confidence in the prose equal confidence in the facts. A hesitant sentence can be correct; a decisive paragraph can be invented.

Good deployment design does not ask “Is AI smart?” It asks:

  • How often is this system correct on representative examples?
  • Which failures matter most?
  • Can a person detect an error before harm occurs?
  • Is the action reversible?
  • What evidence or tool can independently check the result?
  • Does the system know when to stop or escalate?

Begin with a low-consequence draft or analysis that a knowledgeable person can review. Treat high-stakes advice, rights-affecting decisions, financial transfers, and unsupervised physical actions as different risk classes—not as ordinary chat.

Energy, Water, and Physical Scale

AI feels weightless at the interface, but it runs on physical infrastructure: chips, servers, networking, storage, cooling systems, buildings, power plants, and manufacturing supply chains.

The International Energy Agency projects that total global data-center electricity demand could reach about 945 terawatt-hours in 2030, a little more than double its 2024 level. AI is the largest driver of that growth, but the projection covers all data centers. Calling the entire figure “AI electricity” would erase conventional servers, storage, networking, cooling, and other digital services.

Water accounting is even more context-sensitive. Data centers may consume water directly for cooling and indirectly through electricity generation. Chip manufacturing also has a water footprint. Results vary by cooling design, climate, time of day, electricity mix, and local water stress. A single “water per prompt” number can conceal more than it explains.

Two ideas can be true together:

  • data-center demand is growing quickly and can create serious local grid or water constraints;
  • data centers remain one part of total global electricity demand, and efficiency, grid investment, siting, cleaner power, and workload design can change the impact.

Good footprint claims name the measurement boundary, location, period, and uncertainty.

A Regulation Snapshot

This snapshot is current as of July 27, 2026 and is educational, not legal advice.

The European Union AI Act is a binding cross-sector framework organized largely by risk and use. It prohibits certain practices, imposes obligations on high-risk systems, sets rules for general-purpose AI, and requires transparency for some chatbots and generated or manipulated content. The AI Omnibus entered into force on July 27, 2026, extending key high-risk dates: certain stand-alone high-risk rules are scheduled for December 2, 2027, and rules for AI embedded in regulated products for August 2, 2028. Some transparency obligations begin August 2, 2026.

The United States has no single federal counterpart with the same structure. The July 2025 federal AI Action Plan emphasizes innovation, infrastructure, adoption, national security, and international leadership. Governance also comes from existing sector laws, agency actions, procurement requirements, technical standards, courts, and state laws.

The comparison is about structure, not “regulated versus unregulated.” Both jurisdictions govern AI, but they assign duties through different layers and priorities. A real compliance question depends on location, sector, use, system role, risk, and effective date. Always check the current primary rule that applies to the actual deployment.

Sources

Key terms

AI governance
The rules, roles, controls, and oversight used to direct how AI is developed and used.

Key facts

Water Use Depends on Place and Time

Data centers may use water directly for cooling and indirectly through electricity generation and chip manufacturing. The impact varies with cooling design, climate, grid mix, and local water stress, so one global per-prompt number can mislead.

The U.S. Uses a More Distributed Approach

The July 2025 U.S. AI Action Plan emphasizes innovation, infrastructure, national security, and adoption. Federal agencies, sector laws, procurement rules, courts, and state laws still create additional obligations; there is no single EU-style federal AI Act.

Matched pairs

Each item on the left goes with the one on the right.

Confident tone
A writing style, not evidence that an answer is correct
AI output
A digital result produced by physical computing infrastructure
AI regulation
Rules that vary by jurisdiction, sector, and use

Step by step

Verify an important AI-generated output

  1. Identify the claims or decisions that matter
  2. Locate independent primary or authoritative sources
  3. Check each material claim against those sources
  4. Test calculations, citations, and edge cases separately
  5. Record uncertainty and decide whether human approval is required

Polish and confidence are not substitutes for independent evidence.

Self-test

Answer each question from memory, then check your work against the answer key.

  1. 1. Which task is the best beginner pilot?

    1. AAn unsupervised compliance filing
    2. BAn automatically approved customer refund
    3. CA final résumé-screening rejection
    4. DA reviewable first draft
  1. 2. What most improves trust in an important output?

    1. ARepeating the same model until answers agree
    2. BIndependent verification
    3. CRequesting a more detailed explanation
    4. DRaising the model's self-reported confidence
  2. 3. Why can one universal water-per-prompt number mislead?

    1. AEvery facility uses the same cooling design
    2. BToken count is the only variable that matters
    3. CSystems and locations differ
    4. DIndirect electricity-related water is always excluded
  3. 4. What sector does the IEA's 945 TWh projection cover?

    1. AAll data centers
    2. BData centers in advanced economies only
    3. CAI-optimized data centers only
    4. DCryptocurrency mining centers only
  4. 5. What organizes the EU AI Act's obligations?

    1. ARisk and use category
    2. BModel parameter count alone
    3. CThe developer's headquarters country
    4. DCompany revenue and employee count

Go deeper

Answer key

  1. 1. D A reviewable first draftA reversible draft creates value while keeping the consequence of an error low.
  2. 2. B Independent verificationChecks against primary evidence or tested data address correctness directly.
  3. 3. C Systems and locations differModel, hardware, cooling, weather, grid, and accounting boundaries all affect an estimate.
  4. 4. A All data centersThe figure is a sector-wide data-center projection for 2030.
  5. 5. A Risk and use categoryThe Act distinguishes prohibited, high-risk, transparency, and minimal-risk uses.

Sources

Download all 14 lessons: PDF · EPUB