Watch the explanation. Practice the skill.
What AI actually is
Explain the difference between a generated claim and supporting evidence. Compare two prompts against the same source facts. Repair a draft that invents a fact, then explain the correction.
Captions appear here during playback.
What AI actually is
Explain the difference between a generated claim and supporting evidence. Compare two prompts against the same source facts. Repair a draft that invents a fact, then explain the correction.
Video playback is unavailable. Read the descriptive transcript below, go directly to practice, or open the video file.
Local synthetic cases; independently reviewed narration. Automated screen capture and actual isolated fixture replay. No live provider compatibility or learning-effect claim. · Reviewed 2026-10-06. Microsoft Mark · locally synthesized. On a small screen, use fullscreen for diagram details or read the transcript.
Read the descriptive transcript
0:00 · Outcome and preparation
By the end of this lesson, you will be able to compare two prompts, check a generated claim against its source, and repair a draft that invents a result. You will leave with a short proposal and a record of what you checked. You do not need an AI account, coding experience or a paid tool. A notes editor or paper is enough. We will use supplied drafts so that everyone can inspect the same mistakes. These are authored teaching examples, not results from a tested model. If you later try the task in an approved assistant, the wording may differ and you still need the same checks. Open the practice sheet linked on the lesson page. Keep your source facts and your draft in separate columns. Do not replace the fictional case with employer, military, client or personal data.
What appears on screen
Your artifact: Prompt + three-sentence proposal Claim / evidence / status / repair One correction explained Fictional supplied drafts; no model results. Prepare the worksheet: Keep source facts separate from your draft. Use supplied cases or paper. Do not enter employer, military, client or personal data.
0:52 · Generation and evidence
AI includes systems for text, images, speech, prediction and other tasks. We begin with language models, which generate text from patterns learned during training and the context supplied for a request. They generate tokens: pieces of text that are not always whole words. You do not need to count tokens for today's task. A fluent draft is useful, but it is not proof. A model can reason about a task and still give an incorrect answer. A statement needs supporting evidence: a source, a calculation or an observed test that actually supports that statement. A chat product may combine a model with search, tools, history or memory. Those features depend on the product and its configuration. A confident answer does not show that a source was retrieved. Ask what was available and what was actually checked. For today, the useful question is simple: does this claim follow from the facts we supplied? Keep the distinction between producing a sentence and establishing that it is true. You would need evidence of the trial and its measurement, including what was compared. A polished sentence or a more capable model does not supply a trial that never happened.
What appears on screen
Generation and evidence: Source facts → prompt → generated draft Draft → compare against source → inspect claims Search is a separate optional tool. Prediction: A draft says: “A trial saved 40%.” What would establish it? Pause and name the missing evidence. Evidence needed: Evidence of an actual trial and its measurement. What was compared? A polished sentence does not supply a trial.
2:12 · Compare two prompts
Here is a fictional logistics profile. Compare these prompts for the same task. The first says, “Write a resume bullet for my logistics job.” It does not tell the assistant which facts it may use, who will read the result, or how we will judge it. The second supplies the profile, asks for three active-voice bullets for a civilian supply-chain manager, removes military jargon and limits each bullet to 25 words. Those are things we can inspect. The second prompt has a clearer factual basis and clearer constraints. This comparison is not a measured improvement in model performance or a promise about hiring. Now write down three checks you would make on its output. I would check whether every number matches the profile, whether the wording adds an achievement the source does not contain, and whether the requested audience and format are respected. The repeatable loop is to ask with context, inspect against declared criteria, revise a specific failure, and recheck. Stop when the criteria pass, or when missing evidence prevents a defensible answer. Repeated drafting without new evidence does not turn an unsupported claim into a fact.
What appears on screen
Compare prompt A: Write a resume bullet for my logistics job. Prompt B · fictional profile: 47 personnel · $12 million fleet One-year deployment · zero accidents 96% mission completion Three active-voice bullets, civilian supply-chain manager, no military jargon, ≤25 words each. Inspect output criteria: Numbers match source? Any invented achievement? Audience and format respected? Clearer criteria do not guarantee hiring or model performance.
3:29 · Prepare the worked case
Now we will do the whole process. Riley coordinates six people. The team handles 40 requests each week, and a request currently takes about 18 minutes. Riley wants to propose a two-week trial of AI-assisted request summaries. No trial has run, so savings are unknown. The audience is a civilian operations manager. Our output is three sentences in plain language. It must preserve any facts we use, state that the trial is proposed, and avoid inventing achieved results. Here is a bounded prompt: “Using only R1, write a three-sentence proposal for a civilian operations manager. Preserve the supplied facts, state that the two-week trial has not run, and do not claim savings. Propose a person checking each summary and measuring accuracy and processing time before an expansion decision.” The checking and measurement steps are proposals. They are not statements that Riley has already done them. Put that distinction in the output.
What appears on screen
Open source-R1.txt: R1 · FICTIONAL SOURCE Riley coordinates six people. 40 requests each week. About 18 minutes per request. Proposed two-week trial. No trial has run; savings unknown. Edit bounded-prompt.txt: Using only R1, write a three-sentence proposal for a civilian operations manager. Preserve supplied facts; state the two-week trial has not run; do not claim savings. Propose human comparison, accuracy and processing-time measurement before expansion. Label proposed actions: Human comparison and measurement are PROPOSED. They are not things Riley has already done.
4:36 · Find the mistake
Our deliberately flawed draft says, “Our 20-person team has reduced processing time by 40% with AI. The system is ready for rollout across the organization. We recommend immediate adoption.” Pause and mark what is wrong before I fill in the table. The team size contradicts R1: six, not twenty. The 40% reduction is unsupported: the source reports no trial. Readiness for rollout is also unsupported. Immediate adoption is a recommendation without trial evidence; a proposal can still be poorly justified. An invented factual detail is often called a hallucination. You do not need to identify the internal cause of the error to repair the draft. You need to identify the unsupported statement and its evidence gap. Specific numbers, dates, quotations, citations, current information and consequential technical or legal claims deserve careful checks. That is a verification priority, not a universal ranking of which errors every model makes most often.
What appears on screen
Open draft-R2.txt: R2 · DELIBERATELY FLAWED Our 20-person team has reduced processing time by 40% with AI. The system is ready for rollout across the organization. We recommend immediate adoption. Claim table · add contradictions: Claim: 20 people Evidence: R1 says six Status: contradicted Repair: preserve six, or omit figure Claim: 40% saved Evidence: no trial has run Status: unsupported Complete the claim table: Ready for rollout: unsupported Immediate adoption: recommendation without trial evidence Repair factual claims and unsupported decision.
5:46 · Repair and recheck
Here is one defensible repair: “Our six-person logistics team processes 40 requests each week, taking about 18 minutes per request. We propose a two-week trial of AI-assisted request summaries, with a person comparing each summary against the original request. We would measure accuracy and processing time before deciding whether to expand.” Let's check it. The first sentence uses only the given workload facts. The second makes the trial duration explicit and labels the source comparison as proposed. The third describes a proposed measurement and decision, not achieved savings. We have three sentences. We have not changed the numbers. We have not claimed success. A shorter repair could omit a figure; it cannot change that figure. This version is a supported proposal, not evidence that AI will save time. Save your prompt, repair and claim table. Explain one edit: “I removed the 40% saving because no trial had run.” That explanation matters more than making your wording match mine.
What appears on screen
Edit repaired-proposal.txt: Our six-person logistics team processes 40 requests each week, taking about 18 minutes per request. We propose a two-week trial of AI-assisted request summaries, with a person comparing each summary against the original request. We would measure accuracy and processing time before deciding whether to expand. Recheck source + proposed steps: Sentence 1: six / 40 / 18 → source matches Sentence 2: proposed two-week trial + human comparison Sentence 3: proposed measurement before expansion No achieved savings; three sentences. Run document checks: PASS: supplied figures preserved; proposed steps labeled. Scope: wording checks only; source comparison still required. 30 requests x 12 minutes = 360 minutes = 6.0 hours of current workload, NOT savings.
6:57 · Independent practice
Now use a different case. Morgan coordinates five volunteers handling 30 appointment requests each week. Each request takes about 12 minutes. Morgan proposes a one-week trial of AI-assisted reminder drafts. Venue confirmation is pending. No trial has run, so savings and attendance effects are unknown. Sending reminders is not authorized in this exercise. The flawed draft says, “Our eight volunteers already save six hours weekly with AI. The venue is confirmed and attendance increased 20%. We will automatically send every reminder tomorrow.” First compare “Write a reminder proposal” with a prompt you write using M1, the audience, unknowns and permission boundary. Explain two differences you can inspect; do not claim measured model improvement. Then write your three-sentence proposal and claim table. Explain one correction. Use the worksheet criteria. You may propose human review and measurement, but label them as proposed. Keep sending outside this task. Pause now. The next chapter gives feedback, so attempt the case before continuing. If you are on YouTube, the lesson link has the same worksheet and a separate answer section.
What appears on screen
Independent M1 · new facts: Morgan: five volunteers, 30 requests weekly, ~12 minutes each. Proposed one-week reminder-draft trial. Venue confirmation pending. Savings and attendance effects unknown. Sending reminders not authorized. M2 · find the unsupported claims: Eight volunteers already save six hours weekly. Venue confirmed; attendance increased 20%. Automatically send every reminder tomorrow. Pause before feedback: Write a bounded prompt, three-sentence proposal and claim table. Explain one correction. Compare a vague and bounded prompt. Propose review/measurement; do not send.
8:20 · Feedback after your attempt
First check the six-hour claim. Thirty requests times about twelve minutes is 360 minutes, or approximately six hours of current weekly workload. That is not six hours saved. There is no trial measurement to support savings. Check the other traps: five volunteers must not become eight; pending venue confirmation must not become confirmed; attendance gains remain unknown; and a request to draft does not authorize sending. If you missed a trap, locate the source sentence and repair that claim. Then read the whole proposal again. Did the edit create another unsupported result? Does it label the trial and checking steps as proposed? Different wording can pass the same criteria. Try this explanation without looking at my table: what evidence would be needed before Morgan could report savings or attendance improvement? Keep the answer specific to the proposed trial. A source comparison can establish whether your draft matches M1; it cannot establish a future effect.
What appears on screen
Feedback · solution ahead: 30 × about 12 minutes = 360 minutes 360 ÷ 60 = about 6 hours CURRENT workload No measured trial → no savings claim. Check each boundary: Five volunteers ≠ eight Pending venue ≠ confirmed Attendance effect remains unknown Drafting permission ≠ sending permission Repair and recheck: Locate source sentence; repair specific claim. Read whole proposal again. Did the edit invent another result? Different wording can meet the same criteria.
9:28 · Transfer and next step
The skill you practiced is not tied to a logistics or community-center task. For another draft, identify the source facts, preserve what remains unknown, and distinguish drafting from permission to act. Ask, inspect, repair and recheck against clear criteria. Keep your prompt, independent repair and fact-check note. Reviewing this lesson or watching the video does not certify the skill. Your artifact and explanation show what you actually did; an unfamiliar task will test whether you can repeat it. The next lesson gives you five everyday transformations to practice. If you want the technical distinction between a model, product, harness and execution environment, use the optional reference. A harness supplies context, tool orchestration and controls; an execution environment determines where tools run and what they can reach. Those distinctions become important when a task needs tools, but you do not need to master them before repairing today's claim.
What appears on screen
Transfer checklist: FACTS: identify the supplied source UNCERTAINTY: preserve unknowns PERMISSION: drafting does not authorize action Ask → inspect → repair → recheck Keep evidence of your work: Prompt + independent repair + fact-check note Watching does not certify capability. Unfamiliar independent task tests transfer. Optional reference: Model: generates output Product: may add tools/history/memory Harness: context, orchestration, controls Environment: where tools run and what they reach
Watching explains the example. Complete the lesson practice to check what you can do. Explore all visual lessons →