Be the Judge, Not the Audience
Get Good at Using It
Read-aloud narrates the unit in a natural voice, paragraph by paragraph. It needs an account.
In short
AI is confident whether it's right or wrong. The confidence is always at maximum. There is no built-in "I'm not sure" signal. Subtly wrong answers are more dangerous than obviously wrong ones — they slip through because they look right.
The Hook
Google's AI Overviews — the AI-generated answers that appear at the top of search results — told users to put glue on pizza to keep the cheese from sliding off. It recommended eating rocks for minerals. It confidently stated the year was 2024 when it was 2025.
These weren't obscure edge cases. These were answers shown to millions of people searching for everyday information. And most people just... read them and moved on. They didn't question it, because the answer looked authoritative.
This is the problem. AI is confident whether it's right or wrong. It doesn't hedge when it's guessing. It doesn't flag when it's making things up. The confidence is always turned to maximum. Which means the job of deciding "is this actually good?" falls entirely on you.
Knowing that AI has limitations is one thing. Everyone knows that by now. But evaluating AI output in real time — catching errors, spotting hallucinations, comparing quality, knowing when something's off — is an active skill that most people haven't developed.
Here's the uncomfortable truth: AI is often almost right, which is worse than being obviously wrong. An obviously wrong answer, you catch. A subtly wrong answer — a date that's off by one year, a statistic that's close but fabricated, a citation that sounds real but doesn't exist — that slips through because it looks right.
The Mata v. Avianca case from Unit 04 is the most famous example, but it's just the tip. A Stanford study found that AI hallucinates in roughly 1 out of 3 legal queries. Over 300 cases of AI-fabricated citations have been documented in US courts alone. And that's just the legal profession — imagine the errors going undetected in every other field where people use AI without checking.
In May 2024, Google rolled out AI Overviews — AI-generated answers displayed prominently at the top of search results. Within days, screenshots went viral: the system recommended adding glue to pizza sauce (sourced from a joke Reddit post), suggested eating rocks for minerals, and confidently stated Barack Obama was Muslim.
Google's AI Overviews are accurate roughly 91% of the time, which sounds good until you do the math. Google processes 5 trillion searches per year. A 9% error rate at that scale means hundreds of millions of incorrect answers served annually. And because the answers appear with Google's branding at the top of the page, most users don't think to question them.
The lesson: AI-written answer boxes shown at the top of Google search results. They look official because of where they appear, but they can and do get things wrong. are a preview of a world where AI-generated content is everywhere and the burden of evaluation falls entirely on the reader.
So how do you develop the skill of evaluation?
First, check specific claims. Any time AI gives you a date, a number, a name, or a citation, verify it independently. This takes 30 seconds and catches the most dangerous errors.
Second, look for the "too smooth" signal. AI output tends to be fluent and well-structured even when it's wrong. If everything reads perfectly but something feels slightly off — too generic, too confident on a niche topic — dig deeper.
Third, compare outputs. Ask the same question to two or three AI tools. Where they agree, there's likely a strong pattern. Where they diverge, someone's guessing.
Fourth, ask it to show its work. "What sources are you drawing on?" or "How confident are you in this answer?" won't give you a reliable confidence score, but it can surface when the AI is working from thin data.
Fifth, know the failure modes. AI is worst at: recent events (training data has a cutoff), niche topics (fewer patterns to match), quantitative claims (it's not doing math), and anything requiring real-world verification (it can't check if a restaurant is still open).
When AI gives you an answer with specific numbers, citations, and authoritative language, it feels reliable. But AI How well a system's confidence matches how accurate it actually is. AI is badly calibrated: it sounds just as sure when it's right as when it's making things up. is terrible — it sounds equally confident whether it's right or making something up. There is no built-in "I'm not sure" signal.
The How often AI makes up information that sounds real but isn't. It depends on the topic and model, but for tricky subjects like legal research, it can be as high as 1 in 3. for legal queries is roughly 1 in 3. For medical queries, it's similarly high. The confidence in the output is always at maximum regardless. You cannot use tone or fluency as a proxy for accuracy.
The goal isn't paranoia. It's calibration — knowing when to trust, when to verify, and when to doubt.
Step 1: Turn off web search first (most tools search by default, which defeats the test). Then ask the AI to tell you five facts about your city. Check each one. Did it confuse your city with another? Are the local details right, or generic? Are any out of date?
Step 2: Ask it to cite three academic studies on a topic you care about (sleep, exercise, social media — anything):
PromptCan you cite three peer-reviewed studies about the effects of social media on sleep quality in teenagers? Include authors, journal name, year, and a one-sentence summary of findings.
Look up the studies. Do they exist? Are the authors right? Are the findings accurately described?
Step 3: Ask two different AI tools the same controversial question (something where reasonable people disagree). Compare how they frame it. Who's more balanced? Who's more confident? Does either acknowledge uncertainty?
Step 4: Ask the AI to write a short paragraph that contains a deliberate error:
PromptWrite a short paragraph about the history of the internet. Deliberately include one factual error. Don't tell me which fact is wrong.
Then ask a second AI to fact-check it. Did the second AI catch the error?
Every time AI gives you a specific claim — a date, a statistic, a person's name, a citation — take 30 seconds to verify it. Search the citation title. Check the date. Confirm the name. This single habit catches most dangerous AI errors and takes almost no time. Make it automatic.
The people who get the most out of AI aren't the ones who trust it most — they're the ones who verify it fastest. Evaluation is the skill that separates someone who uses AI from someone who uses AI well.
This matters beyond AI too. The ability to evaluate information — to ask "is this true? who's saying it? what's missing?" — is the core of critical thinking. AI just makes the stakes higher because the output is so fluent and confident that it bypasses the skepticism you'd normally apply to a random internet comment.
AI Fact-Check Report
Test your ability to evaluate AI output critically:
- Generate content on a topic you know — Ask the AI to write a 300-word explanation of a topic you know well — your sport, your hobby, your city's history, a subject you've studied.
- Grade every specific claim — Mark each as ✅ (confirmed), ⚠️ (partially wrong), or ❌ (fabricated/wrong).
- Correct the errors — For every ⚠️ or ❌, find the correct information and document it with a real source.
- Calculate your accuracy score — What percentage of specific claims were correct?
- Explain the pattern — Write 2-3 sentences: Why did the AI get those things wrong? Connect your explanation to what you know about next-token prediction and training data.
- 1AI is confident whether it's right or wrong. The confidence is always at maximum. There is no built-in "I'm not sure" signal.
- 2Subtly wrong answers are more dangerous than obviously wrong ones — they slip through because they look right.
- 3Evaluation is an active skill: check claims, compare outputs, know the failure modes, and verify anything that matters.
- 4The people who get the most from AI are the ones who verify it fastest, not the ones who trust it most.
Type: Video Title: When AI Can Fake Reality, Who Can You Trust? — TED-Ed URL: https://ed.ted.com/lessons/when-ai-can-fake-reality-who-can-you-trust-sam-gregory Description: Extends the evaluation question from text to images and video. When AI can generate convincing fake media, evaluation skills become even more critical.
| Type | Title | URL | Description |
|---|---|---|---|
| Article | MIT Tech Review, "Why Google's AI Overviews Gets Things Wrong" (2024) | technologyreview.com/2024/05/31/1093019 | Analysis of why AI-generated search answers fail |
| Article | Inc., "Google's AI Overviews Making Mistakes at Massive Scale" (2026) | inc.com | How 91% accuracy becomes unreliable at trillions of queries |
| Video | TED-Ed, "When AI Can Fake Reality, Who Can You Trust?" | ed.ted.com/lessons/when-ai-can… | Evaluation skills for AI-generated images and video |
| Article | Wikipedia, "Hallucination (artificial intelligence)" | en.wikipedia.org/wiki/Hallucination_… | Comprehensive overview of AI hallucination as a phenomenon |
| Case Study | Mata v. Avianca — fake AI citations in court | en.wikipedia.org/wiki/Mata_v._Avianc…. | The case that made AI verification a mainstream concern |
| Database | AI Hallucination Cases — 1,290+ US court rulings, updated continuously | damiencharlotin.com/hallucinations | The scope of the problem extends far beyond one lawyer — open it for the current count, which moves every week |
| Tool | University of Arizona AI Literacy Guide: Verify Facts | libguides.library.arizona.edu/ai-literacy-instruc… | Practical framework for verifying AI-generated claims |
| Book | Yuval Noah Harari, Nexus (2024) | — | How information networks shape truth and trust in the AI age |
Last updated: May 31, 2026. Demo instructions were revised to disable web search so verification skills are tested fairly.
Marking a unit complete, the Prove It check and your place in the course all need an account. The reading stays free.