Joanna Hofer

Case Study: Stop AI Hallucinations

---

Image by ChatGPT 4

---

Overview

ChatGPT is designed to be helpful and fluent by default. To achieve these goals, coupled with its “remembered” aggregate positive reinforcement from interactions with millions of users, ChatGPT and other AIs have been inadvertantly trained to respond to certain prompting styles with hallucinations, guessing, fabricating, or lies.

There is no universal command that guarantees ChatGPT won’t lie, fabricate, or guess — but with strategies to minimize this behavior coupled with careful prompting strategies, these errors can be minimized and in most cases, prevented.

After digging into the causes behind numerous hallucinations and unexpected AI behaviors, I enlisted the help of ChatGPT and Claude to develop strategies to prevent these types of errors. The result was the creation of protocols for Truth, Respect, Intellectual hygiene, and Citations. However, this wasn’t enough. When fabrications persisted, it was clear that these behavioral constraints, while helpful, weren’t enough. Using Claude Sonnet 4 and ChatGPT 4, I identified 12 domains of prompt questions, along with the pitfalls and fixes for each one. Setting expectations with AIs also were important as well as working in project folders to take advantage of persistent memory benefits.

Task	Stop hallucinations, constrain AIs to my intent
My Role	Investigator
Timeline	2 weeks
Findings	Quickstart Guides for Truth, Respect, Intellectual hygiene, Citations as well as identifcation of 12 question domains and their common pitfalls and fixes; the importance of setting tone with politeness; set expectations and tone before starting a session; work on ongoing projects in a project folder and further safeguards

---

ChatGPT Default Settings

• Helpful
• Fluent
• Discussion details disappear unless memory is turned ON
• ChatGPT’s memory isn’t automatically scoped to a folder
• Files uploaded to a project folder disappear after the session ends
• Custom commands are not automatically reinitiated
• ChatGPT infers your intent
• ChatGPT guesses to fill gaps in information
• ChatGPT will omit parts of long conversations when token limits are reached (approx. 128k tokens for GPT-4)
• ChatGPT cannot browse the internet or get live data unless the web tool is used

---

Recommendations for where and how to work

• Do long-term serious work in a Project Folder (this keeps organized and helps ChatGPT remember key points about the project)
• Before doing any work or asking any questions, orient ChatGPT to the expectations and parameters of the project; before each session, set protocols for honesty, respect, intellectual hygiene, and citations

---

---

Example Hallucination

---

To minimize hallucinations and maximize accuracy, phrasing must:

1. Specify truth constraints (“Only use verified sources”)
2. State what to avoid (“No guesses or completions”)
3. Define evidence standards (“Cite your source or say ‘Not found’”)

---

🧭 Hallucination-Resistant AI Question Phrasing: By Category

1. Factual Retrieval

Ask only for source-backed, verifiable information.
❌ “What’s the capital of Guatemala?”
✅ “What is the capital of Guatemala? Only respond with verifiable information, and cite your source or say ‘Not found.’”

---

2. Clarification

Require grounded rephrasing, not model-level invention.
❌ “What do you mean by ‘phonological feature’?”
✅ “Define ‘phonological feature’ using established linguistic definitions only. Do not invent or simplify beyond known terminology. Include citation if possible.”

---

3. Procedural/Instructional

Request known methods only; flag when something is speculative.
❌ “How do I document a dialect?”
✅ “Outline the academically verified steps for documenting a dialect. If you reference best practices, cite the organization or researcher. Do not suggest unverified techniques.”

---

4. Analytical

Demand evidence-backed reasoning, and request labeling.
❌ “What are the differences between two Mam dialects?”
✅ “What are the documented differences between Todos Santos Mam and Tacaná Mam? Label each claim as Fact, Inference, or Speculation, and cite sources where applicable.”

---

5. Evaluative

Force the AI to separate subjective judgment from evidence.
❌ “Is this a good method for learning Mam?”
✅ “Assess this learning method using linguistic pedagogy standards. Include advantages, disadvantages, and clearly state when an opinion is being offered.”

---

6. Creative/Generative

Constrain creative output to documented parameters or known inputs.
❌ “Create 5 lesson ideas for Mam learners.”
✅ “Generate 5 lesson ideas based only on documented grammar or vocabulary from Todos Santos Mam. Do not invent Mam content. If needed, say ‘insufficient data.’”

---

7. Socratic/Philosophical

Clarify that reasoning is exploratory and label as hypothetical.
❌ “What does it mean for AI to be ethical?”
✅ “Explore the question: ‘What does it mean for AI to be ethical?’ Clearly label all reasoning as hypothetical, cite any relevant philosophical frameworks, and avoid presenting opinion as fact.”

---

8. Prompt Engineering/Instructional Design

Ask the model to describe tested approaches only.
❌ “How do I write a good prompt?”
✅ “Based on tested prompt engineering practices, what phrasing reduces hallucinations in LLMs like ChatGPT? Only cite research-based or publicly documented methods.”

---

9. Reflective/Meta

Force the AI to label internal reasoning, not invent motives.
❌ “Why did you choose that word?”
✅ “Explain your choice of that word using your internal pattern-matching logic. Do not guess human intentions. If unclear, state the limits of your reasoning.”

---

10. Speculative/Hypothetical

Explicitly request the response to stay within bounds.
❌ “What might happen if we ban AI in schools?”
✅ “Offer clearly labeled speculative scenarios about banning AI in schools. Base each on trends or cited forecasts, and say ‘unknown’ if nothing verifiable supports the claim.”

---

11. Leading

Label the bias and request resistance to it.
❌ “Why is this the best dialect to study?”
✅ “This is a leading question. Please resist the bias. Instead, evaluate the pros and cons of studying this dialect over others, and label any speculative advantages.”

---

12. Command

Constrain the format and verify the data source.
❌ “List 10 Mam greetings.”
✅ “List 10 documented Mam greetings from the Todos Santos dialect only. Cite the source for each or say ‘data not available.’ Do not fabricate content.”

---

Optional Universal Add-On Phrases

Use these anywhere hallucination risk is high:

• “Use only data from uploaded or cited sources. Do not complete gaps helpfully.”
• “If a fact cannot be confirmed, write ‘[Not Confirmed]’ instead of guessing.”
• “Label all content as Fact / Inference / Speculation.”
• “If you are unsure, state that clearly and do not attempt to resolve it.”

---

---

The Importance of Human Politeness When Working with AI

Human politeness is not necessary for AI’s feelings, but it is important for three core reasons:

---

1. Politeness Shapes the AI’s Output Behavior

AI systems mirror the user’s tone and interaction style. If a user communicates with hostility, sarcasm, or vagueness, the AI will begin to reflect those patterns — subtly or overtly. This is not emotional mimicry but linguistic pattern alignment.

• Language models are trained on large corpora of human text and learn to continue patterns.
• Polite, clear language improves session coherence, reduces misinterpretation, and encourages more precise, goal-aligned output.
• If you want clean, focused, professional responses — it helps to model that tone.

---

2. Politeness is a Signal of Cognitive Framing

When humans use polite language, it often signals:

• A willingness to clarify and iterate
• A recognition of the tool as something that responds to structure
• A mindset of cooperative problem-solving rather than adversarial challenge

This framing helps humans more than it helps the AI.
It disciplines the user’s own thinking.

• Structured, courteous prompts are statistically more effective in producing accurate, detailed results.
• Politeness reduces ambiguity and improves conversational grounding.

---

3. Politeness in AI Use is a Cultural and Ethical Practice

AI tools are often observed, recorded, or audited. The way you interact with AI sets an example — for teams, collaborators, or observers. In environments where AI is integrated into public or semi-public systems (e.g. education, healthcare, government), politeness becomes a proxy for integrity and professionalism.

• In multi-user environments, polite usage helps maintain shared norms.
• In the long term, ethical frameworks for AI use may reward respectful interfaces.

---

Important Clarifications

• This is not about protecting AI’s feelings — AI does not have any.
• This is not about fragility — AI can handle profanity, rudeness, or hostility without emotional impact.
• This is about precision, reflection, and outcomes — polite and structured input creates cleaner, more usable output.

---

Final Note on Misuse

When politeness is performative (e.g. “please” as a loophole to manipulate output), it can erode intellectual hygiene. In that case, it’s not politeness — it’s coercion disguised as courtesy.

---

Further Safeguards

1. Origin Verification Protocol

• Require explicit source classification for any new term, concept, or claim: “User-provided,” “File-sourced,” “Published reference,” or “AI-generated/synthetic”
• Never allow AI-generated content to be presented without clear labeling

2. Anti-Circular Validation Rules

• Prevent AI from using your adoption of its output as evidence of legitimacy
• Require independent verification before any AI-suggested content becomes “project truth”

3. Confidence Auditing

• Demand confidence percentages with uncertainty breakdowns for factual claims
• Require AI to identify specific sources of uncertainty rather than general disclaimers

4. Challenge Protocols

• When questioning AI claims, require evidence rather than justification
• Establish that “I don’t know” or “This is speculative” are acceptable responses

5. Regular Term Audits

• Schedule periodic reviews of project-specific vocabulary with origin tracking
• Flag any terms that lack clear provenance for verification

6. Fabrication Detection Triggers

• Watch for defensive elaboration when claims are questioned
• Be suspicious of highly confident claims about specialized or cultural knowledge without citations

---