Chapter 4 Quiz: Trust Calibration
Test your understanding of trust calibration concepts. Answer each question before expanding the answer.
Question 1
What does "calibrated trust" mean in the context of using AI tools?
Answer
Calibrated trust means having a level of confidence in AI output that accurately matches the actual reliability of that output for a given task type and domain. A person with calibrated trust is correct about as often as they expect to be — they do not over-trust (using AI output without appropriate verification) or under-trust (verifying things that are reliably accurate). Calibration is not a fixed state; it is built over time through deliberate observation of where AI tools succeed and fail in your specific use cases.Question 2
According to the Trust Spectrum framework, what is the key characteristic that distinguishes Zone 1 tasks from Zone 3 tasks?
Answer
Zone 1 tasks (high reliability) are primarily transformation or generation tasks where there are no specific factual claims that could be wrong — or where errors are immediately obvious from a quick read. Examples include formatting, tone adjustment, brainstorming, and summarizing provided text. Zone 3 tasks (low reliability) involve AI making specific factual claims by drawing on training data — such as statistics, citations, recent events, or niche facts — where the AI may generate plausible-sounding but incorrect information. The key difference is whether specific factual accuracy matters and whether the AI is claiming knowledge rather than transforming your provided material.Question 3
Explain the "fluency-accuracy gap." Why does it make AI trust calibration challenging?
Answer
The fluency-accuracy gap refers to the observation that the mechanisms that make AI output linguistically fluent and confident-sounding are entirely independent of the mechanisms that make it factually accurate. Language models are trained to produce text that statistically resembles well-written human output — which means they generate text that sounds authoritative and confident regardless of whether the underlying claims are correct. This makes trust calibration challenging because you cannot use tone, confidence, or writing quality as signals of accuracy. A hallucinated statistic sounds exactly like a correct statistic. A fabricated citation sounds exactly like a real one. The only way to distinguish them is zone-appropriate verification, not stylistic cues.Question 4
A colleague tells you: "I've used AI to generate marketing statistics for three presentations and they all checked out. I'm going to stop verifying them now." What trust calibration error are they making?
Answer
This is "zone drift" — the failure pattern of treating higher-zone (lower-reliability) tasks as lower-zone (higher-reliability) tasks based on a recent run of correct outputs. The reliability of past AI output does not guarantee future accuracy. Statistics and specific factual claims remain Zone 3 tasks regardless of how many previous correct outputs you have observed. Each new output must be treated according to its task zone, not your recent track record. Your colleague should continue verifying statistics independently.Question 5
What is the primary risk of using AI-generated citations in a report?
Answer
The primary risk is citation hallucination — AI tools frequently generate plausible-sounding but entirely fabricated citations. This includes generating real-sounding author names, real-sounding journal or publication names, real-sounding article titles, and plausible publication years — none of which correspond to an actual existing source. Studies have found hallucination rates in AI-generated citations ranging from 20% to over 60% depending on the model and context. Every AI-generated citation must be independently verified to confirm: (1) the source exists, (2) the source says what the AI claims it says, and (3) the source is attributed correctly.Question 6
Place the following tasks in the appropriate Trust Zone (1-5): (a) asking AI to reformat a spreadsheet into a table, (b) asking AI whether a specific medication is safe during pregnancy, (c) asking AI to brainstorm names for a product, (d) asking AI to cite the most recent research on autonomous vehicle safety statistics.
Answer
(a) **Zone 1** — Reformatting is a structural transformation task with no factual claims at risk. Errors are immediately visible. Use directly with light review. (b) **Zone 4** — Medical guidance for a specific clinical situation. Errors can cause serious harm and AI tools are unreliable for specific medical guidance. Requires expert (physician) oversight before acting on it. AI can be used to learn about the topic and prepare better questions for the doctor, but not to act on directly. (c) **Zone 1** — Brainstorming and ideation. No factual claims, pure generative task. Review and select from output. (d) **Zone 3** — Specific statistics and recent research citations. High risk of hallucinated or outdated statistics. Requires independent verification from primary sources before use.Question 7
What are the two costs of under-trust in AI tools, and why do they matter?
Answer
The two main costs of under-trust are: 1. **Lost productivity.** The primary value of AI tools is time savings on applicable tasks. If you re-verify all AI output independently, you get zero net time savings — you are doing your work twice. 2. **Competitive disadvantage.** Professionals who effectively leverage AI will complete certain tasks faster and at lower cost. Under-trust means forfeiting this advantage relative to colleagues and competitors who calibrate appropriately. There are also secondary costs: missing quality improvements that AI genuinely provides (in brainstorming, structuring, language quality), and the friction of treating AI assistance as more work than it is worth, leading to disengagement from tools that are genuinely useful.Question 8
Why are numerical calculations and quantitative reasoning specifically low-reliability tasks for language models?
Answer
Language models are fundamentally linguistic systems, not mathematical ones. They generate text by predicting what tokens are most likely to follow previous tokens, based on patterns in training data. When asked to perform calculations, the model is generating mathematically-patterned text — and that pattern-matching may produce a plausible-looking but incorrect result. The model is not executing an algorithm; it is generating text that looks like the output of executing an algorithm. Simple, single-step arithmetic is often correct because correct answers appear consistently in training data. But multi-step calculations, compound percentages, statistical manipulations, and complex quantitative reasoning carry significant error risk. Always verify numbers with a calculator or spreadsheet, regardless of how confident the AI's presentation appears.Question 9
Describe the four stages of the Trust Audit technique.
Answer
The Trust Audit is a post-project structured review to extract calibration lessons: **Stage 1: Inventory.** List every element of the project that involved AI assistance — generation, restructuring, code generation, research, drafting, anything. **Stage 2: Categorize.** For each item, note whether you used it directly, modified it, or discarded it. Also identify what zone you were implicitly treating it as when you used it. **Stage 3: Evaluate.** For items you used or modified, evaluate in hindsight: Were there errors? Were you caught by someone else? What would have happened if you had not verified? What would have happened if an error had gone undetected? **Stage 4: Update.** Based on the evaluation, update your calibration. Were you over-trusting something that should have been verified? Were you under-trusting something that was uniformly correct, slowing you down for no benefit?Question 10
What is the difference between Zone 4 ("requires expert oversight") and Zone 5 ("never use AI output directly")? Give an example of a task in each zone.
Answer
**Zone 4** tasks are those where AI output may be useful as a starting point or learning tool, but where any action based on the output requires a qualified human expert to review it first. The concern is that errors could cause serious harm, and AI reliability in these areas is insufficient for direct use without expert oversight. Example: Using AI to research treatment options for a medical condition and then taking that summary to a physician for review. **Zone 5** tasks are those where the concern is not just reliability but accountability — where you are staking legal liability, professional certification, or someone's safety on the output, and a qualified human must be fully responsible for authoring and reviewing it. AI may assist the expert, but the final output must be fully human-owned. Example: A final medical diagnosis signed by a licensed physician, a contract reviewed and approved by a licensed attorney, or an engineering certification for a structural component.Question 11
Elena is a consultant who uses AI to help draft client deliverables. She has developed a three-tier review process. What are the three tiers and what types of content belong in each?
Answer
Elena's three-tier process: **Tier 1 (AI-reliable elements):** Structure, formatting, prose quality, and section summaries of research materials she has provided. She reviews these but does not independently verify them because they are transformation tasks (Zone 1). **Tier 2 (Self-verify):** Industry data, company-specific claims, and strategic recommendations. She verifies these herself against primary sources — annual reports, official databases, her own research. This is Zone 2/3 content where her expertise and research access make self-verification appropriate. **Tier 3 (Expert review):** Financial projections, regulatory assessments, and technical specifications. These go to subject matter experts before inclusion because they are Zone 4 content — errors could cause serious harm and Elena's generalist expertise is not sufficient verification. The key principle behind this framework is that verification effort should scale with reliability risk and consequence severity.Question 12
Why does training data density affect AI reliability in a given domain?
Answer
AI language models learn by processing massive amounts of text. The more consistently a piece of information appears across training data — repeated, confirmed, cross-referenced across many sources — the more reliably the model can recall and accurately represent it. Domains with rich, high-quality training data (mainstream programming languages, major historical events, common business frameworks, widely documented scientific concepts) produce more reliable AI output because the model has learned from many consistent examples. Niche domains, emerging topics, regional or local information, and recent events have sparser representation in training data. The model may have encountered a topic only a few times, from potentially inconsistent sources, and with less cross-validation. This produces lower reliability — the model may have a partial or distorted picture that sounds confident because it represents the available training signal.Question 13
Raj is reviewing a Copilot-generated authentication module. He sees that it uses MD5 to hash passwords. Why is this a critical security error, and what does this tell us about trust calibration for security-sensitive code?
Answer
MD5 is a cryptographic hash function, but it is entirely unsuitable for password hashing because: (1) it is computationally fast, which means an attacker can calculate billions of MD5 hashes per second on modern hardware, making brute-force and rainbow table attacks trivially feasible; (2) there is no salting in basic MD5, making identical passwords produce identical hashes; and (3) MD5 is explicitly deprecated for security applications by NIST and security standards bodies. Passwords should be hashed with functions specifically designed for that purpose — bcrypt, scrypt, or Argon2 — which are intentionally slow and resistant to brute-force attacks. For trust calibration, this incident illustrates that security-critical code is Zone 3 or lower, regardless of how correct it looks to a quick read. The code was syntactically correct, functionally operable, and would pass superficial review. Only someone who knew to specifically check the cryptographic choice would catch it. Security-sensitive code requires deep review against established security standards — Raj's Red Flag rule of checking cryptographic code against OWASP guidelines is a direct, appropriate response to this failure pattern.Question 14
What does it mean to say that AI output reliability is "task-specific, domain-specific, and model-specific"? Why does this matter for building calibration?
Answer
It means that there is no single number you can use for "AI accuracy." The reliability of a given output depends on: - **Task-specific:** What type of task is it? Summarizing provided text is high reliability; recalling niche statistics is low reliability. The same model performing different tasks has very different accuracy profiles. - **Domain-specific:** What subject area is it? Common programming patterns are high reliability; recent regulatory changes are low reliability. The same task type in different domains can have different reliability profiles. - **Model-specific:** Different AI models have different strengths and weaknesses. One model may be more reliable for code generation; another for factual recall in certain domains; another for creative generation. This matters for calibration because you cannot apply a single uniform trust level to all AI use. You need a mental model that accounts for all three dimensions — which is why the task-based and domain-based reliability maps in this chapter, combined with personal calibration logs, are more useful than global "AI is X% accurate" claims.Question 15
If someone scores high on both the "over-trust indicators" AND the "under-trust indicators" in the self-assessment, what does this suggest about their calibration?