AI Fluency: Framework & Foundations - Certification Study Guide

Course: AI Fluency: Framework & Foundations Modules: 7 Target: All professionals beginning their AI collaboration journey Difficulty: Foundational

MODULE 1: Introduction to AI Fluency

Key Notes

• AI Fluency defined: the ability to collaborate with AI systems effectively, efficiently, ethically, and safely
• Not about coding or building AI — it is about working WITH AI as a tool in your daily workflow
• AI fluency is a human skill, not a technical skill — analogous to “information literacy” or “media literacy”
• Why it matters now: AI is being embedded into tools across every industry and profession
• Core premise: AI is a collaborator, not a replacement — fluency determines the quality of that collaboration
• The goal is augmentation: AI handles the mechanical/drafting layer so humans can focus on judgment and creativity
• The 4D Framework is the organizing structure for the entire course:
  • Delegation — deciding WHAT tasks to hand to AI
  • Description — HOW to communicate your request clearly
  • Discernment — evaluating and critiquing AI output
  • Diligence — responsible, ethical, accountable use
• The 4Ds are not sequential steps — they form a continuous loop
• Fluency is developed over time through deliberate practice, not passive exposure

          ┌─────────────────────────────────────────────┐
          │           THE 4D FRAMEWORK LOOP             │
          │                                             │
          │   DELEGATE ──→ DESCRIBE ──→ DISCERN         │
          │       ↑                         │           │
          │       └──────── DILIGENCE ──────┘           │
          └─────────────────────────────────────────────┘

Best Practices

• Treat AI fluency as a skill that requires ongoing practice and refinement
• Keep a personal log of effective prompts and interaction patterns
• Share AI fluency learnings with your team to build collective capability
• Approach AI tools with healthy skepticism — not fear, not blind trust
• Continuously update your understanding as AI capabilities evolve

Example

A marketing manager wants to write a campaign brief. Without AI fluency, they either ignore AI tools entirely or paste a vague request and accept whatever comes back. With AI fluency, they assess which parts to delegate (drafting, research summaries), describe the brief clearly with context, evaluate the output critically, and handle sensitive brand data with care.

MODULE 2: Deep Dive 1: Generative AI Fundamentals

Key Notes

• Generative AI = AI systems that create new content: text, images, audio, code, video
• This course focuses primarily on Large Language Models (LLMs) — text-in, text-out systems
• Types of AI systems: Claude/GPT (text), DALL-E/Midjourney (image), Copilot (code), Gemini/GPT-4o (multimodal)
• Multimodal = can accept and generate multiple types (text + images + audio)
• Critical misconception to eliminate: AI does not “think” or “know” — it predicts statistically plausible output

How LLMs Work — The Full Picture:

  Stage 1: PRE-TRAINING
  ─────────────────────────────────────────────────────────────
  Trained on massive datasets: web pages, books, code, articles
  Learns statistical relationships: which tokens follow which
  Develops broad language understanding and world knowledge
  No alignment yet — produces capable but unfocused output

  Stage 2: SUPERVISED FINE-TUNING (SFT)
  ─────────────────────────────────────────────────────────────
  Human trainers provide examples of ideal responses
  Model learns to follow instructions, structure answers helpfully
  Result: useful, instructable model

  Stage 3: REINFORCEMENT LEARNING FROM HUMAN FEEDBACK (RLHF)
  ─────────────────────────────────────────────────────────────
  Human raters compare responses and rank which is better
  Model learns to generate responses humans prefer
  Result: better alignment with human values and preferences

  Stage 4: CONSTITUTIONAL AI / ADDITIONAL ALIGNMENT
  ─────────────────────────────────────────────────────────────
  Explicit principles guide behavior (varies by provider)
  Model critiques and revises its own outputs
  Result: stronger safety and values alignment

Tokenization — how text becomes numbers:

• A token is roughly 4 characters or ~0.75 words (not the same as a word)
• “Unbelievable” might tokenize as: “Un” + “bel” + “ievable” (3 tokens)
• Common words (“the”, “is”, “cat”) are usually 1 token
• Rare words, technical jargon, and non-English text require more tokens
• Context windows are measured in tokens — knowing this helps estimate limits

Next-Token Prediction (the core mechanism):

  Input: "The capital of France is"
  Model calculates probability of every possible next token:
    "Paris"     → 94.2%
    "Lyon"      → 1.1%
    "a"         → 0.8%
  Selects "Paris" → then predicts next token → and so on

Capabilities of LLMs:

Limitations — memorize these:

Best Practices

• Always treat AI outputs as a first draft, not a final answer
• Identify whether you need a text AI, image AI, or code AI for your task
• Provide context documents when you need accuracy on specific facts
• Verify outputs that make factual claims against authoritative sources
• Understand the knowledge cutoff of your AI tool before using it for current events

Example

A lawyer asks Claude to summarize a 40-page contract. Claude can do this well (summarization is a strong capability). The lawyer should NOT ask Claude who won a court case last month (knowledge cutoff issue) or treat Claude’s legal interpretations as final without review (hallucination risk + no legal accountability).

MODULE 3: The 4D Framework: Delegation

Key Notes

• Delegation = deciding WHAT to give to AI vs. what to keep human
• The central question: “Should I delegate this task, and to what extent?”
• Not binary — you can partially delegate (AI drafts, human finalizes; AI analyzes, human interprets)

Expanded Task Suitability Matrix (15+ task types):

Risk Assessment Framework with Examples:

Delegation Decision Tree:

  Task arrives
       │
       ▼
  Is the task well-defined?
  ├── No → Define it first, then return
  └── Yes → Continue
       │
       ▼
  Does it require human judgment, relationships, or accountability?
  ├── Yes → Keep human (AI may assist in preparation)
  └── No → Continue
       │
       ▼
  Is the data safe to share with this AI tool?
  ├── No → Either anonymize or do not delegate
  └── Yes → Continue
       │
       ▼
  What is the cost of an AI error here?
  ├── High → Delegate with heavy human review
  └── Low → Delegate with standard review
       │
       ▼
  Can you verify the output without specialized effort?
  ├── No → Build verification into the workflow first
  └── Yes → Delegate

Project Planning with AI — Workflow:

  Step 1: Human defines the project goal and key constraints
  Step 2: Delegate to AI: generate a draft project plan / timeline
  Step 3: Discern: review structure, sequencing, feasibility
  Step 4: Delegate to AI: draft individual section content (one at a time)
  Step 5: Human: integrate, make judgment calls, handle dependencies
  Step 6: Delegate to AI: format final document, generate summary
  Step 7: Human: final review, accountability, sign-off

Delegation Spectrum:

DELEGATION SPECTRUM

Full Human ←──────────────────────────────────────→ AI-Assisted
    │                                                      │
Final legal    Editing AI     Drafting with    Generating    Fully
judgment       draft          human outline    raw options   AI-drafted
                                                            (low-stakes)

Workplace Delegation Scenarios:

Writing emails: AI drafts from your bullet points → you review tone, accuracy, relationship nuance → revise and send.

Analyzing data summaries: AI identifies patterns from summary statistics you provide → you interpret within business context → you make the decision.

Creating presentations: AI generates slide outline and draft text from your notes → you review factual accuracy, remove confidential details, adjust storytelling → finalize.

Researching competitors: AI summarizes publicly available information → you verify recency and accuracy → you apply competitive intelligence judgment.

Best Practices

• Default to partial delegation rather than full delegation for anything high-stakes
• Audit your weekly tasks to identify delegation candidates
• Document which tasks your team has decided ARE and ARE NOT delegatable
• Revisit delegation decisions as AI capabilities improve
• Never delegate the final decision on matters with significant human consequences

Example

A project manager needs to send a project status update. Delegation decision: AI can draft the update based on bullet points the PM provides (good delegation). The PM should not delegate deciding whether to escalate a critical project risk — that requires judgment, relationships, and accountability.

MODULE 4: The 4D Framework: Description

Key Notes

• Description = how you communicate what you want to AI (prompting)
• Better descriptions = dramatically better outputs — the quality gap is often here
• Prompting is a learnable skill with known patterns that work
• Iteration is built into Description — the first prompt is rarely the final one

Six Elements of a Strong Prompt:

Prompt Iteration Workflow:

  Draft prompt → Submit → Evaluate output quality
        ↑                          │
        └──── Identify gap ────────┘
              (wrong tone? missing info? wrong format?)
        │
        ▼
  Revise prompt → Resubmit → Evaluate again
        │
        ▼
  Convergence: output meets standard → use it

XML Tag Patterns (structured prompting):

  <task>Summarize the following document for a VP audience</task>
  <context>Q3 financial performance review, internal use only</context>
  <constraints>
    - Maximum 5 bullet points
    - Focus on revenue variance and risk factors only
    - Do not include implementation details
  </constraints>
  <format>Bullet points with a one-sentence header</format>
  <document>[paste document here]</document>

XML tags help the model parse complex, multi-part prompts more reliably.

Few-Shot Patterns:

  Task: classify these customer messages as Complaint / Request / Compliment

  Examples:
  "Your service has been terrible for 3 weeks" → Complaint
  "Can you change my appointment to Thursday?" → Request
  "The team was incredibly helpful today" → Compliment

  Now classify:
  "I've been waiting 5 days for a response" → [AI classifies]

Few-shot prompting dramatically improves classification, formatting, and style consistency.

Chain-of-Thought Prompting:

  WITHOUT chain-of-thought:
  "What is the best market to enter next?"
  → AI gives an answer directly (may be poorly reasoned)

  WITH chain-of-thought:
  "Think step by step. First, identify the key criteria for market selection.
  Then evaluate each market against those criteria. Then recommend."
  → AI's reasoning becomes visible and evaluable

Prompting Anti-Patterns with Fixes:

Real Workplace Prompting Examples:

Email drafting:

You are a project manager at a software company. Write a status update email
to a client who is anxious about a delayed deliverable. Tone: calm, transparent,
confident. Include: current status, root cause (technical integration issue),
revised timeline (2 weeks), mitigation steps taken. Under 250 words.
No jargon. End with a clear call to action.

Data analysis:

You are a business analyst. Here is a summary of our Q3 sales data by region
[paste data]. Identify the 3 most significant trends. For each trend, explain
what might be causing it and what questions a VP would want answered.
Format: 3 sections with headers. Each section: trend statement, analysis,
open questions.

Presentation outline:

Create a slide outline for a 15-minute executive presentation on our
AI implementation roadmap. Audience: non-technical C-suite. Goal: get
approval for Phase 2 budget. Structure: problem → solution → results
to date → Phase 2 plan → ask. No more than 10 slides. For each slide:
title and 3 bullet points maximum.

Best Practices

• Start with the output format, then the content — “Write a table that shows…”
• Include negative constraints: “Do not use jargon, do not exceed 100 words”
• Build a personal library of prompts that work for recurring tasks
• Use examples whenever the format matters more than the content
• For complex tasks: break into sequential prompts, not one mega-prompt

MODULE 5: The 4D Framework: Discernment

Key Notes

• Discernment = critically evaluating AI output before using it
• The most underused of the 4Ds — many users accept output without review
• Discernment is what separates AI fluency from AI dependence

Expanded Error Taxonomy (10+ error types):

Verification Checklist:

• Did I read the entire output end-to-end, not just skim?
• Have I verified specific statistics, percentages, and data points?
• Have I checked that named citations actually exist?
• Does the reasoning chain hold up logically, step by step?
• Does the output address MY specific situation, not just the general case?
• Is the tone appropriate for my actual audience?
• Does the output contradict itself anywhere?
• Are there any surprising claims that would be significant if true?
• Is all information current, or might it be outdated?
• Did AI follow all my formatting and scope constraints?

Model Confidence Indicators — what to watch for:

• High confidence signal (be more skeptical): specific numbers, named sources, confident assertions without hedging
• Low confidence signal (more honest): “may,” “might,” “I believe,” “I’m not certain,” “it’s possible that”
• Paradox: AI is sometimes MOST dangerous when MOST confident — hallucinations often come with high apparent certainty

The Description-Discernment Feedback Loop:

  DESCRIBE (prompt)
        │
        ▼
  AI OUTPUT
        │
        ▼
  DISCERN: What is correct? What is wrong? What is missing?
        │
        ├── Output is good enough → use it
        │
        └── Output has problems → diagnose the failure
              │
              ├── Wrong content → add more context to prompt
              ├── Wrong format → add format specification
              ├── Missing info → add constraints or examples
              └── Wrong tone → add role or tone instruction
                      │
                      ▼
              DESCRIBE (refined prompt) → loop

The Discernment Scale:

Low Stakes:    Quick review for obvious errors → use
Medium Stakes: Fact-check key claims → revise → use
High Stakes:   Expert review + source verification → significant editing → use
Critical:      Full expert review, primary source verification → treat as
               rough draft only

Practical Verification Strategies:

• Cross-reference factual claims with authoritative primary sources
• Look for hedging language as a signal of uncertainty: “might,” “possibly,” “reportedly”
• Check internal consistency — does the output contradict itself?
• Test edge cases — does the advice work in extreme scenarios?
• Ask AI for its sources/citations, then verify they actually exist
• For critical outputs: read end-to-end with skepticism, not just skimming

Best Practices

• Never submit AI-generated content without reading it completely at least once
• For factual content, verify every specific claim independently
• Build verification into your workflow as a non-optional step
• Teach your team what hallucinations look like — share examples
• Use the “would this surprise me if true?” test as a quick hallucination filter

Example

A researcher uses AI to draft a literature review section. Discernment in action: they notice the AI cited “Johnson et al. (2019)” which they cannot find in any database. They also notice a statistic “73% of adults…” with no source. They remove both, replace with verified sources, and correct a logical error where the AI drew an overly broad conclusion from a narrow study.

MODULE 6: The 4D Framework: Diligence

Key Notes

• Diligence = practicing responsible, ethical, and accountable AI use
• The “professional conscience” dimension of AI fluency
• Protects individuals, organizations, and society from AI-related harms

Data Classification Framework:

PUBLIC                   INTERNAL                CONFIDENTIAL            RESTRICTED
─────────────────────    ────────────────────    ────────────────────    ──────────────────
Safe to share freely     Use work AI tools       Minimize exposure;      Never share with AI
                         with approved policy    anonymize first
─────────────────────    ────────────────────    ────────────────────    ──────────────────
Published research       Internal policies       Business strategies     Customer PII
Public information       Draft communications    M&A information         Employee records
Generic templates        Project plans           Financial projections   Health records
Org's own public         Meeting summaries       Partner details         Legal case details
  reports/materials      Process docs            Performance data        Passwords/credentials
Publicly available       Training materials      Vendor contracts        Social security numbers
  competitor info                                Unannounced products    Medical data

Organizational AI Policy Checklist:

• Which AI tools are approved for organizational use?
• What data classifications are permitted in each tool?
• Are enterprise data agreements in place for our AI tools?
• Is there a disclosure policy for AI-generated content?
• What is the approval process for new AI tool adoption?
• How are AI use incidents reported?
• Are there role-specific restrictions (HR, Legal, Finance)?
• Is volunteer/contractor AI use addressed?

Privacy Decision Tree:

  Does this task involve any personal data?
  ├── No → Proceed with standard diligence
  └── Yes → Continue
        │
        ▼
  Can you complete the task with anonymized data?
  ├── Yes → Anonymize first, then proceed
  └── No → Continue
        │
        ▼
  Is this data category covered by regulations (HIPAA, GDPR, FERPA)?
  ├── Yes → Consult legal/compliance before proceeding
  └── No → Continue
        │
        ▼
  Is the AI tool approved for this data type?
  ├── Yes → Proceed with appropriate care
  └── No → Do not use AI for this task; escalate

Bias Detection Patterns:

• Check demographic representation in examples, scenarios, and recommendations
• Notice when AI defaults to a dominant cultural perspective (often Western, English-speaking)
• Look for stereotyping in role assignments (are leaders always male? caregivers always female?)
• Check whether AI advice applies equally across different groups
• In hiring, performance, or fairness-sensitive contexts: expert human review is mandatory

Citation and Attribution Guidelines:

• Disclose AI involvement when others would reasonably want to know
• Don’t present AI-generated content as fully original without disclosure
• Academic, legal, and journalistic contexts have strict standards — know them
• AI cannot be cited as an author — attribute to the human who used and reviewed it
• In regulated industries, understand whether AI-assisted work meets documentation requirements

Key Ethical Concerns — Full Table:

Workplace Diligence Scenarios:

Writing a job description: Use approved org tool; don’t paste candidate data; review output for biased language; disclose AI assistance to hiring manager.

Analyzing competitor data: Only use publicly available information; verify all claims against primary sources; don’t share confidential internal strategies with AI.

Creating a presentation for clients: Remove all client-specific confidential data before prompting; use anonymized versions; verify all statistics.

Drafting HR communications: Never include employee names or personal details; use anonymized role descriptions; have HR and legal review final output.

Best Practices

• Create a personal “AI use card” — a checklist you run before sharing AI-assisted work
• Opt for enterprise versions of AI tools that offer stronger data protections
• Never share customer data with AI unless explicitly authorized
• Disclose AI assistance in contexts where it is professionally expected
• Regularly revisit your organization’s AI policies as they evolve

MODULE 7: Conclusion & Assessment

Key Notes

• The 4Ds reinforce each other — weakness in one undermines the others:
  • Poor Delegation = AI working on wrong tasks
  • Poor Description = AI producing misaligned outputs
  • Poor Discernment = errors making it into final work
  • Poor Diligence = ethical and legal exposure

The Complete Loop:

  Task arrives
       ↓
  DELEGATE: Should AI handle this? Which parts?
       ↓
  DESCRIBE: Craft clear, specific, contextualized prompt
       ↓
  DISCERN: Evaluate output — facts, logic, tone, completeness
       ↓
  DILIGENCE: Safe data? Disclosed? Verified? Appropriate use?
       ↓
  Output used / returned for iteration

• AI fluency is a competitive skill and will increasingly be expected in all professional roles
• The goal is not to use AI more — it is to use AI better

Common 4D Failure Modes:

Best Practices

• Practice all 4Ds together on real tasks, not in isolation
• Reflect after each AI interaction: which D was weakest this time?
• Build AI fluency habits before pressure makes it tempting to skip steps

4D Framework Quick Reference Card

╔══════════════════════════════════════════════════════════════════════╗
║                    4D FRAMEWORK QUICK REFERENCE                      ║
╠══════════════════════════════════════════════════════════════════════╣
║                                                                      ║
║  DELEGATE                   DESCRIBE                                 ║
║  ─────────────────────      ─────────────────────                    ║
║  Ask before starting:       Include in every prompt:                 ║
║  □ Well-defined task?        □ Task (what to do)                     ║
║  □ Safe data?                □ Context (why/who/situation)           ║
║  □ Verifiable output?        □ Constraints (limits/rules)            ║
║  □ Reversible if wrong?      □ Format (how it should look)           ║
║  □ No human judgment req?    □ Examples (show don't just tell)       ║
║                              □ Role/Tone (AI's persona)              ║
╠══════════════════════════════════════════════════════════════════════╣
║                                                                      ║
║  DISCERN                    DILIGENCE                                ║
║  ─────────────────────      ─────────────────────                    ║
║  Check every output:        Before sharing AI work:                  ║
║  □ Read end-to-end           □ No PII or sensitive data?             ║
║  □ Verify specific facts     □ Org tool approved?                    ║
║  □ Check citations exist     □ Disclosed where required?             ║
║  □ Logic chain holds?        □ Verified before acting?               ║
║  □ Tone matches audience     □ Bias checked?                         ║
║  □ No contradictions         □ IP considerations handled?            ║
║                                                                      ║
╠══════════════════════════════════════════════════════════════════════╣
║                                                                      ║
║  HALLUCINATION RED FLAGS:   SAFE TO DELEGATE:                        ║
║  • Specific % without source  • Drafting and writing                 ║
║  • Named citations to verify  • Summarizing documents                ║
║  • Recent events/research     • Brainstorming options                ║
║  • Surprising specific claims • Formatting/restructuring             ║
║                               • Generating variations               ║
║  NEVER DELEGATE:                                                     ║
║  • Final strategic decisions  DELEGATION RISK RULE:                  ║
║  • Ethical judgments          Low stakes + reversible → delegate     ║
║  • HR/legal decisions         High stakes + irreversible → human     ║
║  • Crisis judgment                                                   ║
╚══════════════════════════════════════════════════════════════════════╝

Final Checklist

• I can define AI Fluency and explain why it matters
• I understand what Generative AI is and how LLMs work at a basic level
• I can explain tokenization and why it matters for context limits
• I can describe the 4-stage LLM training process (pre-training, SFT, RLHF, alignment)
• I can name 5 capabilities and 7+ limitations of LLMs
• I can apply the Delegation decision tree to a new task
• I can write a structured prompt using all 6 elements
• I can identify at least 8 types of AI output errors
• I can use the Discernment verification checklist
• I can classify data into Public / Internal / Confidential / Restricted
• I can name 3 ethical concerns and how to address them
• I understand the Description-Discernment feedback loop
• I can explain the 4D Framework from memory
• I can use the 4D Quick Reference Card in real work situations