Everyday Guidebooks
Cover of When Artificial Intelligence Gets It Wrong

More Books

When Artificial Intelligence Gets It Wrong

A Beginner's Guide to Bias, Bad Predictions, and Questionable Automated Decisions

by Todd Michael

Available now

An AI system can sound certain and still be wrong. It can also produce an answer that seems useful while hiding a flawed assumption, an incomplete dataset, or a goal that does not match what people actually need.

When Artificial Intelligence Gets It Wrong offers artificial intelligence for beginners through a question-driven exploration of how automated systems make mistakes. Todd Michael explains how training data, prediction, measurement, and human choices influence outcomes—and why even a system that works well for some people may fail others.

Short case studies bring the ideas into everyday situations, from recommendations and search results to hiring and customer service. Questions throughout the book help you examine who designed a system, what it was asked to optimize, whose experiences may be missing, and what happens when an outcome is challenged.

You do not need a technical background. You only need a willingness to look beyond the confident answer.

Contents
  1. 1. The Confident Answer
  2. 2. What AI Systems Actually Do
  3. 3. Where the Data Comes From
  4. 4. When History Becomes a Pattern
  5. 5. The Goal Is Not the Whole Job
  6. 6. The Human Choices Behind the Machine
  7. 7. Made-Up Answers and Missing Context
  8. 8. Search, Feeds, and Recommendations
  9. 9. Customer Service and Automated Screening
  10. 10. Different People, Different Error Rates
  11. 11. When the World Changes
  12. 12. Following a Decision Through
  13. 13. Asking Better Questions
  14. 14. Living With Imperfect AI
  15. Other Books by Todd Michael

Read a sample: 1. The Confident Answer

Maya is planning a weekend trip and wants to know whether a small museum will be open on Sunday. She types a question into an AI assistant. A response appears almost immediately:

“The museum is open Sundays from 10 a.m. to 5 p.m.”

The answer is exactly the shape Maya wanted. It names a day, gives clear hours, and saves her the trouble of opening several websites. It does not sound uncertain. If the assistant also produces a tidy paragraph about the museum’s collection, the whole response may feel even more trustworthy.

But what does Maya actually know?

She knows that a system produced a sentence. She does not yet know whether the hours came from the museum, whether they are current, or whether the museum has made an exception for a private event. The answer may be right. It may also describe an old schedule, confuse the museum with another one, or simply provide a plausible-looking time without a dependable source.

This imagined exchange captures a central problem with artificial intelligence in everyday life: an answer can be easy to read before it has earned the right to be believed.

That does not mean Maya must investigate every sentence an AI system produces. If she is browsing for possible trip ideas, a mistake may cost little. If she is about to buy a nonrefundable ticket or arrange transportation around those hours, a quick check on the museum’s official site becomes more worthwhile. The important skill is not constant suspicion. It is knowing when a polished answer needs support and what kind of support would count.

People often use the word confident in two different ways. A person may be confident because they have examined strong evidence. Or they may merely sound confident: fluent, decisive, and free of hesitation. Those qualities sometimes travel together, but they are not the same.

AI outputs can create a similar confusion. A chatbot may use a firm tone. A search tool may place an answer at the top of the page. A recommendation may arrive with a label such as “best match.” A workplace system may display a score with two decimal places. Each presentation suggests a degree of authority. None, by itself, proves that the output is correct or appropriate.

Consider the number 92.4. Its precision makes it look carefully measured. Yet its meaning depends on questions the number cannot answer for itself. What was measured? Compared with what? Was the system tested on situations like this one? Does the number represent a probability, a ranking, or an internal score? Was it calculated from complete information?

Even an explicitly stated probability needs interpretation. If a system says there is an 80 percent chance of rain, the useful question is not whether this particular forecast sounds confident. It is whether forecasts made under comparable conditions tend to be reliable over time. A single outcome cannot tell us everything about the quality of a probability: rain can occur when the forecast probability is low, and it can stay dry when the probability is high.

Many everyday AI outputs provide less information than a weather forecast. They may not explain how a score was produced, what uncertainty it represents, or whether the system has been tested for the task at hand. In those cases, the polished surface is especially tempting—and especially incomplete.

Imagine that Jordan opens a streaming service after a long day. The service recommends a suspense series under the heading “Picked for you.” Jordan enjoys the first episode. The recommendation has done something useful.

Now imagine that the next several suggestions are also suspense series. Jordan begins to wonder whether those are truly the best choices or simply the kinds of programs the service expects Jordan to watch. The distinction matters because “best” can mean different things. Best for a relaxed evening? Best according to reviewers? Most similar to something Jordan finished last week? Most likely to keep Jordan watching?

A recommendation is not necessarily wrong because it fails to match every preference. Tastes change, and any suggestion involves uncertainty. But a recommendation can be misleading if its presentation encourages Jordan to assume it serves a goal that has not been established.

The same issue appears in shopping. A product labeled “recommended” may be a good fit, but the label alone does not tell a shopper whether the selection reflects price, availability, past browsing, paid placement, or some combination. Different services use different methods, and the details may not be visible.

Jordan does not need to solve the entire recommendation system to make a sensible choice. A few questions can restore perspective:

- What does this label appear to promise? - What information might the service be using? - What alternatives am I not seeing? - What would make this suggestion useful for my purpose?

For a television show, sampling an episode may be a perfectly reasonable test. For an expensive purchase, Jordan might compare specifications, read independent reviews, and check the return policy. The recommendation can be a starting point without becoming the final word.

Now consider a different imagined situation. Lena tries to make a purchase with her bank card. The transaction is declined, and her banking app says it detected unusual activity.

This message has a credible purpose. Automated systems can help organizations notice patterns that deserve attention. A genuine fraud warning may prevent further trouble, and Lena may be grateful…

More More Books

Cover of Dog Breeds for Real LifeDog Breeds for Real LifeAvailable nowCover of How to Find the Right Dog Breed for Your HouseholdHow to Find the Right Dog Breed for Your HouseholdAvailable nowCover of Digital Nomad Ideas for Accessible Remote WorkDigital Nomad Ideas for Accessible Remote WorkAvailable nowCover of The Bathhouse Keeper's Day OffThe Bathhouse Keeper's Day OffAvailable nowCover of A Hurricane Guide for CaregiversA Hurricane Guide for CaregiversAvailable nowThe Fountain Pen FamiliarComing soon

About Todd Michael

Todd Michael is an American author and the founder of ToddMichael.tv. He writes practical, plain-English books on money, health, preparedness, technology, games, humor and more.

About the author Todd Michael · Todd Michael’s author site

Part of the ToddMichael.tv network.