Contribution to AI Basics - Prof. Inhwa Kim

This is the official contribution to the AI Basics course by Prof. Inhwa Kim, Assistant Professor of Teaching for Business at Oregon State University, and scientist in AI-Human Interaction, humanoid robotics, and AI in retail.

1.2 AI Failure Modes

• The probabilistic nature of AI: why errors are structural, not bugs?

Unlike traditional software that operates on deterministic logic and explicit code, Large Language Models (LLMs) function on a probabilistic architecture. They do not retrieve verified facts from a relational database, nor do they possess human-like reasoning. Instead, LLMs generate text dynamically by predicting the most statistically probable next token based on pattern recognition from massive training datasets. Consequently, AI errors are structural—inherent to the core mathematical design of the system—rather than traditional programming bugs or software glitches.

Four Pillars of Structural Vulnerability

• Probability vs. Veracity The model optimizes for statistical likelihood, not objective correctness. While the most probable continuation of a sentence often aligns with reality, it offers no mathematical guarantee of accuracy.
• Absence of Internal Verification LLMs lack an intrinsic mechanism to cross-reference statements against external databases or to self-audit their outputs. Without integration into external retrieval-augmented generation (RAG) tools, the model cannot differentiate between a plausible-sounding fabrication and empirical truth.
• Statistical Patterning vs. Knowledge Storage During the training phase, the architecture maps statistical relationships between semantic concepts rather than indexing a structured knowledge base. The output is a fluid reconstruction of observed linguistic patterns, not a retrieval of stored factual records.
• Optimization for Linguistic Fluency The underlying training objective rewards coherence, syntax, and stylistic naturalness. Because confidence and articulateness are merely products of this fluency optimization, they function as deceptive indicators that frequently mask factual inaccuracies.

Key Takeaway for Practitioners: Because these errors are structural, they cannot be completely "patched" out of existence like traditional software bugs. Managing AI failure modes requires systemic guardrails, such as external retrieval tools and human-in-the-loop validation.

• How can we define the 3 failure modes hallucination, stale knowledge & confabulation (with clear contrasts)?

In Short:

• Hallucination breaks reality by creating fake facts.
• Stale Knowledge breaks time by using expired facts.
• Confabulation breaks context by injecting unfounded assumptions.

• Where does each failure mode come from (training and output cutoff, tokenization, plausibility optimization)?

1. Hallucination

• The Cause: Plausibility Optimization + Tokenization
• The AI is trained to sound natural and fluent, not necessarily to tell the truth (plausibility optimization). Because it builds words piece-by-piece using math codes (tokenization), it sometimes pieces together real-sounding tokens to invent fake facts, fake URLs, or fake book titles. It cares that the sentence looks good, not that it is real.

2. Stale Knowledge

• The Cause: Training Data & Knowledge Cutoff
• This is a timeline problem. Once an AI finishes its training, its memory is frozen at that exact date (knowledge cutoff). It has no live internet connection to see today's world. If a company gets a new CEO or a law changes after that cutoff date, the AI will confidently give you the old information.

3. Confabulation

• The Cause: Plausibility Optimization + Missing Context
• This happens when your prompt has missing information or gaps. Because the AI hates leaving a job half-done (plausibility optimization), it hates saying "I don't know." Instead, it uses its statistical guessing power to fill in the blanks, making up logical-sounding details or motives to give you a complete story.

In Short:

• Stale Knowledge happens because the AI's clock stopped.
• Hallucination happens because the AI values smooth writing over real facts.
• Confabulation happens because the AI makes wild guesses to fill in your missing details.

• How do we deal with hallucination through output validation (e.g. source checking)?

Since hallucination is a structural property of how LLMs generate text, we cannot eliminate it at the model level alone. Instead, we manage it through how we handle, verify, and constrain the output after it is generated.

1. Require sources and grounding
Ask the model to cite where a claim comes from, then independently confirm that the source actually exists and says what the model claims.

2. External fact-checking
Cross-check factual claims—especially names, dates, statistics, and citations—against authoritative external sources rather than accepting the AI's output at face value.

3. Retrieval-Augmented Generation (RAG)
Connect the model to real documents or live search so it generates answers grounded in retrieved evidence, rather than relying purely on internal, unverifiable pattern completion.

4. Cross-verification
Rephrase the question or query multiple models/approaches. Inconsistent answers across attempts are a warning sign of hallucination.

5. Prompting for calibrated uncertainty
Explicitly instruct the model to say "I don't know" when evidence is insufficient, rather than defaulting to a confident-sounding guess.

6. Requesting specificity
Ask for precise, checkable details (exact quotes, page numbers, links) rather than vague summaries—specific claims are easier to verify or falsify than general ones.

7. Human-in-the-loop review
In high-stakes domains (law, medicine, finance, academia), a human must review and validate AI output before it is treated as final.

Key Takeaway for Practitioners: AI output should be treated as a draft requiring verification, not a final answer. No matter how fluent or confident the model's response sounds, fluency is not evidence of truth. This is the single most important mindset for managing hallucination at the output stage—the burden of verification always remains with the human user.

1.3 The structured Prompt

• What would be a good example of a naked prompt vs. structured prompt (the power gap made visible)?

Example 1: Information Retrieval

Naked Prompt

Tell me about the effects of social media on teenagers.

Structured Prompt

Summarize the effects of social media on teenagers based only on peer-reviewed studies published after 2020. For each claim, cite the specific study. If you are not confident a claim is well-supported, explicitly state "Uncertain" instead of presenting it as fact. Present the results in a table with the following columns:

• Effect
• Supporting Evidence
• Confidence Level

Example 2: Image Generation

Naked Prompt

Draw a dog.

Structured Prompt

Draw a Golden Retriever puppy with the following specifications:

• Subject: Sitting upright, facing the camera
• Setting: Autumn park with grass and scattered fallen leaves
• Lighting: Soft natural golden-hour light (around 4:00 PM)
• Camera: Eye-level angle, shallow depth of field, blurred bokeh background
• Style: Photorealistic, DSLR photography
• Expression: Playful, mouth slightly open, looking directly at the viewer
• Quality: 4K resolution, ultra-detailed, sharp fur texture
• Aspect Ratio: 16:9

• Closing the loop: Why is AI's probabilistic nature exactly why structure matters (ties back to 1.2)?

Because AI predicts the most probable answer, not the correct one, structure is what narrows that probability space toward what you actually want.

• What are the core concepts of prompt engineering (role, context, structure, examples, constraints)?

1. Role

Assigning the model a specific persona (e.g., "As a customer support chatbot..." or "As a gardening expert...") narrows its response style and domain framing, producing more targeted, professionally appropriate output.

2. Context

Providing background information or a concrete scenario (e.g., explaining conduction/convection "using a frying pan and boiling water") helps the model ground abstract concepts in relatable terms, improving relevance and clarity.

3. Structure

How the prompt is organized — instructive tone, step-by-step framing, or prompt chaining (breaking a task into sequential sub-prompts: summarize → decide action → draft reply) — guides the model through complex reasoning rather than expecting one-shot completeness.

4. Examples

Few-shot demonstrations or specified formats (e.g., "three multiple-choice questions") show the model exactly what output shape is expected, reducing ambiguity compared to zero-shot requests.

5. Constraints

Explicit limits — length ("three sentences"), format, tone, or scope — bound the model's output space so it aligns with the user's actual requirements rather than defaulting to a generic, unconstrained response.

Core principle tying them together: each element works by reducing the model's guessing room — role and context tell it what domain/perspective to draw from, structure tells it how to reason, examples show it what "correct" looks like, and constraints define the boundaries of an acceptable answer. Together, they convert a probabilistic guess into a targeted, checkable output.

• Your perspective/research on what makes a prompt actually work

A good prompt is not necessarily a long prompt—it is a clear prompt. The following principles consistently lead to better AI responses.

1. Be specific.
   Vague prompts produce vague, generic answers. The more clearly you describe what you want, the more accurate and useful the output becomes.
2. Give clear, consistent instructions.
   Avoid conflicting requirements (e.g., "be comprehensive" and "use only 20 words"). Clear and realistic constraints help the AI produce better results.
3. Show an example when possible.
   If you have a preferred style, format, or tone, providing an example is often more effective than simply describing it.
4. Put the most important instruction first.
   Start with your main goal, then add context, constraints, and formatting requirements. This helps the AI prioritize what matters most.
5. Don't assume the AI knows your context.
   Explain any background information that a human might already know. Hidden assumptions are one of the biggest causes of poor responses.
6. Remember that confidence is not correctness.
   A well-written prompt can improve the quality of an answer, but it cannot guarantee that the answer is true. Important information should still be verified using reliable sources or external evidence.

1.6 First documented low level AI Usage

·        Where do good prompts actually have an impact in professional / educational environment?

1. Effective Work Performance

Good prompts help professionals complete tasks more efficiently and with higher quality.

• Faster, higher-quality work
• Clearer writing and communication
• More efficient problem solving and decision making

2. Education and Learning

Structured prompts make AI a more effective learning partner by providing personalized guidance and meaningful feedback.

• Personalized learning support
• Better feedback and practice opportunities
• Deeper understanding of concepts

3. Team Collaboration and Knowledge Sharing

Shared prompting practices make AI use more consistent, transparent, and reproducible across teams.

• Reusable prompts and workflows
• Consistent and transparent AI use
• Easier collaboration and smoother handoffs

·        Why do we document AI usage in work environment (team aspect, passing downstream, transparency)?

1. Transparency and Accountability

Documenting AI use makes the creation process clear and builds trust.

• Disclose where AI was used
• Explain how AI contributed
• Clarify human review and final responsibility

2. Responsible and Policy-Compliant Use

Documentation helps ensure that AI is used ethically and in accordance with organizational policies.

• Acknowledge AI assistance when required
• Respect copyright and licensing
• Follow institutional or workplace AI policies

3. Team Collaboration and Knowledge Sharing

Well-documented AI workflows make it easier for others to understand, reproduce, and improve the work.

• Enable others to reproduce results
• Share effective prompts and workflows

Support smoother collaboration and handoffs