Hallucinations & Limitations — Foundations Lesson 5

Scenario

You've started using AI for your work, and you've caught it making mistakes — making up citations, inventing statistics, contradicting itself. Every time, it sounds completely certain. You start wondering: How much of what AI generates is real vs. made up? Is there a pattern? Can you predict when it'll fail? Or do you just have to assume everything is wrong until proven otherwise?

---

Learning Objectives

By the end of this lesson, you will be able to:

1. Understand why hallucinations occur — the root cause, not just the symptom

2. Recognize hallucination patterns — factual vs. logical vs. fabricated

3. Identify AI's hard limits — tasks it will always struggle with

4. Design workflows that work around limitations — instead of fighting them

5. Communicate limitations to colleagues — so they don't over-trust AI

---

Think Before You Prompt

Before you assign a task to AI, ask:

Is this a task where AI's limitations matter?

Will the output be used as-is, or will someone review it?

What's the cost of an error here?

Is there a workaround that reduces the risk?

Should a human do this instead?

---

The Root Cause of Hallucinations

Hallucinations aren't bugs. They're a fundamental feature of how LLMs work.

How It Happens

1. LLM predicts the next word based on patterns in training data

2. Prediction prioritizes plausibility — The LLM generates what "sounds right" given the context

3. Plausibility ≠ accuracy — Something can sound right and be completely false

4. No fact-checking mechanism — The LLM never checks: "Wait, is this actually true?"

Simple Example

You ask: "What year did the first iPhone launch?"

The LLM has patterns like:

"iPhone" often appears near "2007" in training data

"iPhone" often appears near "launch" and "year"

When these concepts appear together, they usually describe 2007

The LLM predicts: "2007" → Correct!

But then you ask: "What year did the sixth iPhone launch?"

The LLM has patterns about iPhones being released yearly, but it doesn't have strong patterns about "sixth iPhone" specifically. It makes a plausible guess: "2013"

(Real answer: 2014 — the iPhone 6)

The problem: The LLM can't tell the difference between "I'm confident because I've seen this pattern 10,000 times" and "I'm guessing because I don't have clear training data."

---

Types of Hallucinations

Type 1: Factual Hallucinations

AI invents facts, dates, statistics, or citations that don't exist.

Examples:

Made-up statistics: "78% of workers prefer remote work" (completely fabricated)

False citations: "According to Smith's 2019 paper..." (paper doesn't exist)

Wrong dates: "The treaty was signed in 1985" (it was 1983)

Invented sources: Referencing a study that never happened

Why it happens: AI tries to sound authoritative by generating specific numbers and sources, but it's just pattern-matching.

How to catch it: Verify specific numbers, dates, and citations independently.

---

Type 2: Logical Hallucinations

AI invents reasoning or logical steps that don't follow.

Examples:

"To increase sales, we should raise prices" (economic logic fails here)

"Follow these steps: 1) Do A 2) Do C 3) Do B" (ordering doesn't work)

"This employee should be promoted because they have a good email signature" (reasoning is nonsensical)

"The reason people prefer X is because of Y" (unstated assumption presented as fact)

Why it happens: AI generates plausible-sounding reasoning without checking if the logic holds.

How to catch it: Read carefully. Ask "Does this actually follow?" If it feels off, it probably is.

---

Type 3: Fabricated Context

AI invents details to fill gaps in context.

Examples:

You ask about "the Johnson project" but don't specify which Johnson. AI confidently writes as if it knows.

You ask about "recent market changes" — AI makes up which markets, what changed.

You ask "How should we handle this?" AI invents a scenario that doesn't match yours.

Why it happens: AI would rather make something up than say "I don't know." Confidence is built into its output generation.

How to catch it: Ask yourself "Did I provide this context?" If you didn't, AI likely fabricated it.

---

AI's Hard Limits

Some tasks are fundamentally difficult for LLMs. These aren't fixable with better prompts; they're architectural limitations.

Limit 1: No Access to Real-Time Information

What it means: AI can't:

Check today's stock prices

Read your company's intranet

Access your email or files

Browse the current internet (unless specifically designed to)

Know what happened after its knowledge cutoff

Workaround: You provide the information. ("Here are today's stock prices. Now analyze them.")

When it matters: Finance, current events, internal projects, real-time data

---

Limit 2: No True Understanding of Context

What it means: AI doesn't know:

Your company's culture

Your specific situation

What your boss actually wants

The political dynamics of your team

What's actually possible in your constraints

Workaround: Explicitly tell AI these things. ("Our company prioritizes customer relationships over cost. Keep that in mind.")

When it matters: Strategic decisions, sensitive communications, culture-specific advice

---

Limit 3: Math & Precise Calculation

What it means: AI:

Makes arithmetic errors

Can't do complex multi-step math

Doesn't "understand" numbers, only pattern-matches them

Confidently generates wrong results

Example: "What's 47 × 23?" ChatGPT might say "1,091" (wrong; it's 1,081)

Workaround: Use a calculator for actual math. Use AI to explain math or set up the problem.

When it matters: Financial calculations, data analysis, engineering

---

Limit 4: Specialized or Niche Knowledge

What it means: If something wasn't heavily represented in training data, AI likely doesn't know it.

Examples:

Obscure historical facts

Specialized medical conditions

New technologies or methods

Industry-specific jargon

Internal company processes

Workaround: Fill in the knowledge gap yourself. ("Our company uses a custom CRM called InternalCRM. It works like Salesforce but with these differences...")

When it matters: Specialized work, niche industries, rapidly changing fields

---

Limit 5: Reasoning Across Many Steps

What it means: AI struggles with:

Complex decision trees (more than 4-5 branches)

Multi-stage logic (more than 5-6 steps)

Holding multiple constraints simultaneously

Spotting contradictions across a long document

Workaround: Break complex tasks into smaller pieces. ("First, list the constraints. Then, evaluate options against each.")

When it matters: Complex workflows, multi-step decision-making, strategic planning

---

Limit 6: Emotional or Ethical Judgment

What it means: AI can't actually:

Feel empathy (it mimics it)

Know what's *right* ethically (only what training data suggests)

Navigate complex interpersonal situations

Make judgment calls requiring values

Examples:

Choosing between company profit and employee welfare? AI will generate plausible arguments for both sides, but can't decide.

How to handle an underperforming employee? AI gives textbook HR advice, missing the actual relationship dynamics.

Workaround: Use AI for initial ideas, but make judgment calls yourself.

When it matters: Leadership decisions, ethics, people management

---

Working Around Limitations (Not Against Them)

Strategy 1: Hybrid Workflows

Instead of: "Let AI handle the whole thing"

Do: "Let AI handle what it's good at; I handle the judgment parts"

Example:

❌ AI writes the entire performance review

✅ AI drafts the summary, you add specific examples and make the judgment about fit

---

Strategy 2: Verification as Standard Process

Instead of: "Hope the output is right"

Do: Build verification into every workflow

Example:

❌ AI generates a customer list, you use it directly

✅ AI generates a draft customer list, you verify 20% of it before use

---

Strategy 3: Explicit Context Provision

Instead of: "AI, figure it out"

Do: "AI, here's the context you're missing..."

Example:

❌ "Write an email to our customer John about the delay"

✅ "Write an email to John Chen (our biggest client, we've worked with him 3 years, he values transparency). Explain that our project is delayed 2 weeks due to [reason]. Acknowledge the impact on his timeline."

---

Strategy 4: Breaking Complex Tasks

Instead of: "Handle this entire workflow"

Do: Break it into smaller steps, with human review between

Example:

❌ "Create our Q4 marketing strategy"

✅ "Generate 5 possible Q4 campaign ideas. I'll choose one. Then you flesh out the chosen idea. Then I'll review and refine."

---

Strategy 5: Constraint Specification

Instead of: "Do something with this"

Do: Give clear constraints and boundaries

Example:

❌ "Rewrite this policy"

✅ "Rewrite this policy to be shorter (max 1 page), but keep these three rules unchanged: [list]. Add examples of what's allowed and what's not."

---

Common Mistakes When Working with AI Limitations

Mistake 1: Assuming AI Knows Your Context

❌ "Write an email to James about the Johnson project"

AI has no idea who James is, what the Johnson project is, or what you want.

✅ "Write an email to James Chen (project manager, values directness). Explain that the Johnson project is delayed 1 week because of supplier issues. Offer to meet Monday to discuss mitigation."

---

Mistake 2: Asking AI to Make Judgment Calls

❌ "Should we hire this candidate or pass?"

AI can't know your strategic priorities or team needs.

✅ "This candidate has skills X and Y but no experience with Z. Here's what we prioritize: [list]. Given that, what questions should we ask?"

---

Mistake 3: Asking for Real-Time or Current Info

❌ "What's the current market price of copper?"

AI can't access live data.

✅ "Copper is currently at $4.50/lb (as of today). Here's the 5-year trend. What does this mean for our supply costs?"

---

Mistake 4: Trusting Numbers Without Verification

❌ AI generates a budget, you use it as-is

AI's numbers are frequently off.

✅ AI generates a budget framework, you verify line items against actual costs

---

Mistake 5: Using AI as Final Authority on Anything

❌ "AI said this is the best approach, so we're doing it"

AI has no accountability or deeper understanding.

✅ "AI suggested approaches A, B, and C. I'm choosing B because [reason]."

---

Mission: Identify Limitations in Your Work

The Task

1. Pick a complex task you do that involves some of AI's limitations

2. Map the limitations — Where would AI struggle?

3. Design around them — How would you hybrid-work this with AI?

4. Write the AI-assisted workflow — Step by step

Example task: "Decide whether to promote an employee"

AI's limitations:

No access to performance data

Can't understand company politics

Can't make judgment calls about cultural fit

Doesn't know the employee personally

Workaround workflow:

1. You gather performance data

2. AI drafts a summary of pros/cons

3. You add context AI doesn't have

4. AI suggests discussion questions for the review

5. You make the final judgment

---

Deliverables

Limitation Awareness Document:

`

Task: [Your task]

AI's Strengths Here: [What AI can help with]

AI's Limitations Here: [What AI struggles with]

Hybrid Workflow:

Step 1: [What you do]

Step 2: [What AI does]

Step 3: [What you do]

...

Risk: [What could go wrong if you skip verification?]

Verification Step: [How you'll catch errors]

`

Keep this. You're building AI-aware thinking.

---

AI Coach: Limitations Are Features, Not Bugs

Here's the shift in thinking that separates good AI users from bad ones:

Bad thinking: "AI has limitations. How do I work around them?"

Good thinking: "AI has specific strengths and limitations. I'll use the strengths and design workflows around the limitations."

Example:

Bad: "AI keeps hallucinating sources. I guess I can't use it for research."

Good: "AI is great at summarizing sources *I provide*. I'll gather sources, AI summarizes them, I verify."

The limitation isn't a flaw — it's just a boundary. Work within it instead of fighting it.

Key insight: The professionals who get the most value from AI aren't the ones with the most advanced prompts. They're the ones who understand AI's architecture well enough to design workflows that play to its strengths.

---

Reflection: What's Your AI's Superpower and Kryptonite?

For your role, what's AI genuinely great at? And what's it terrible at?

Superpower: [What AI does better than you]

Kryptonite: [What AI fails at that you need to handle]

This shapes how you'll work with AI going forward.

---

Portfolio Check

Portfolio principle: Every AI-assisted output should acknowledge the verification you did.

Not in a formal way — just knowing in your own mind: "I verified this by [method] because [task has this risk]."

As you move into later lessons and build portfolio pieces, this verification thinking should be automatic.

---

Bonus Challenge

Hallucination deep-dive: Pick a topic you know well (your job, a hobby, a skill). Ask an AI to explain it to you. How many hallucinations did it make? Which facts were wrong? Why do you think it got those wrong? This trains you to spot hallucinations in areas where you don't have expertise.

---

Key Takeaways

Hallucinations occur because AI predicts plausible text, not accurate text

Three types: factual, logical, and fabricated context

Six hard limits: no real-time info, no true context understanding, weak math, niche knowledge, complex reasoning, emotional judgment

Workarounds: hybrid workflows, verification as process, explicit context, breaking complex tasks, constraint specification

Design *around* limitations, don't fight them

Next: Lesson 6 explores Responsible AI & Ethics — building ethical thinking into every use of AI.