Artificial
Intelligence
Merit Badge

16 Key Terms

Scouts prepared these definitions before today. Select any term below to see the official Scoutly definition alongside a counselor's perspective. One term at a time; selecting a new one closes the previous.

These Terms in the Real World

A few terms worth grounding in professional context:

• Machine Learning: In insurance, ML models analyze patterns across thousands of claims to identify which ones are likely to become complex, not through rules someone wrote, but through patterns the system found in historical data. The same logic applies wherever there are enough examples of an outcome to train on: which students are likely to struggle before they fail an exam, which equipment is likely to need attention before it breaks down. The pattern recognition is the same. The domain just changes.
• Narrow AI: A system built to read insurance submissions, broker documents, loss histories, and coverage specifications, and extract structured data in seconds would take a human analyst an hour or more per document. That same system has no idea what to do with a question about history or a request to explain a math concept. It is exceptional within what it was built for and completely useless outside it. That is not a failure. That is what narrow means. The capability is real. The boundary is equally real.
• Digital Workers / RPA: In financial services and insurance, bots handle routine processes at a volume no team of people could match: processing renewals, confirming transactions, moving data between systems. These are not AI in the learning sense. They are automation in the rule-following sense. The work they replace was high-volume, repetitive, and procedurally defined, exactly the kind of work automation has always been suited for. The difference now is that software handles it without physical infrasmation.
• General AI vs. Superintelligent AI: We don't have General AI yet. Everything you interact with today, ChatGPT, Siri, Alexa, is Narrow AI. The jump to General AI is one of the biggest unsolved problems in the field.

Scout Presentations (2a, 2b, 2c, 2e)

Scouts prepared these before today. Have each scout share their work and prompt discussion with the questions below.

"AI or Not?", 10 Rounds

Scouts call "AI" or "Not AI" for each scenario. After each answer is revealed, use the discussion prompt to talk through the reasoning. Both the game and the discussion are required.

What "AI in the Workplace" Actually Looks Like

You have probably noticed that autocorrect flags a word and lets you decide whether to accept it. The pattern in large-scale enterprise AI is the same, at much higher stakes.

A system reads incoming documents, thousands of pages of broker submissions, coverage specifications, and loss histories, and extracts the relevant information in seconds. A human analyst would spend hours on the same material. The system surfaces what matters. The analyst decides what to do with it.

That pattern, AI removes the volume work while humans retain the judgment, holds across most serious production deployments. Not because the AI lacks confidence. Because in a regulated environment, accountability for a decision cannot sit with a system. Someone has to be responsible for the outcome. The AI makes that person faster and better-informed. It does not make the person unnecessary.

Scout Presentations (3a, 3b, 3c, 3e)

How Automation Works, Three Core Benefits

Have scouts explain this in their own words, then use the comparison below to anchor the discussion.

What the Stakes Actually Determine

My work in aerospace was in aftermarket supply chain: how parts move through a logistics network, how vendor contracts are structured, how shipping costs are negotiated, how export regulations like ITAR govern what can move where and when. When an algorithm got a routing decision wrong, the consequence was a cost and a delay. Recoverable. You fix it, adjust the model, and move on.

That tolerance for being wrong is not universal.

AI is probabilistic by design. Every output is a best estimate, not a certainty. What changes dramatically between contexts is what being wrong actually costs. In supply chain optimization, a wrong answer is expensive and fixable. In a system making real-time decisions in a flight safety context, the calculus is entirely different.

This is not a technical limitation that better models will eventually solve. It is a structural property of probabilistic systems. Deploying AI anywhere requires a prior decision: what is the acceptable error rate in this context, and who bears the consequences when the system is wrong? That question has to be answered before the model is built, not after it fails.

Some problems are well-suited to AI precisely because variability in outcomes is acceptable: recommendations, routing, forecasting, classification at scale. Others are not, because the cost of a single wrong answer outweighs the efficiency of getting most answers right. Knowing which category a problem belongs to is a judgment call that no algorithm makes for you.

Scout Share-Out (4a & 4d)

"What Would You Do?", 5 Ethical Scenarios

Read each scenario aloud together. Scouts deliberate and share their thinking. Then reveal the ethical considerations and work through the required discussion questions. Both playing and discussing are required.

Scout-Developed Ethical Guidelines

Each scout developed their own ethical guidelines for AI use. Have them present these now. Use the questions below to facilitate discussion.

What Responsible AI Looks Like Inside a Major Corporation

In regulated financial services, moving an AI model from development into production is not primarily a technical event. It is a governance event.

Before a model touches a real decision, pricing, underwriting, claims routing, it goes through bias testing across demographic groups, validation that it performs as documented in the conditions where it will actually run, documentation of training data sources and known failure modes, and review by legal, compliance, and IT security. The process exists because a system that produces discriminatory outcomes, even unintentionally, even as a byproduct of patterns in historical data, carries real legal and financial exposure for the organization that deployed it. Regulators in insurance and financial services watch model behavior closely. This is not a formality.

A Scout is Trustworthy. The governance process that takes teams months to complete is an attempt to make that value operational at scale: to demonstrate, through documentation and testing, that a system does what it claims to do, for the people it claims to serve. The ethical guidelines you drafted earlier in this session are the informal version of the same exercise. The principles are not different. The accountability structure is.

How Deepfakes Actually Work

The technology behind deepfakes is worth understanding, because it is one of the clearest examples of a powerful idea in modern AI: two systems training against each other.

Picture a forger and a detective learning together. The forger creates a fake image and shows it to the detective. The detective says real or fake. The forger adjusts based on what got caught, and tries again. The detective gets better at spotting fakes. The forger gets better at making them. Repeat this millions of times, automatically, at computer speed. After enough rounds, the forger has become precise enough that most people, and even some software, cannot reliably tell the difference.

That is the actual mechanism. It has a technical name, a Generative Adversarial Network, or GAN, and the word "adversarial" just means the two systems are competing. The concept was published in 2014 and changed what AI could create.

Detection is hard for a specific reason: the system was mathematically optimized to fool detectors. It did not stumble into being convincing. That was the training objective. Researchers who build detectors look for patterns the generator leaves behind: a blink rate that is slightly off, lighting that does not match the edge of a swapped face, color values that are statistically inconsistent at seam lines. When detectors improve, generators adapt. Both sides are machine-learning systems improving on each other.

The same underlying technique, two systems training against each other, is also used for legitimate purposes. In financial services and insurance, related methods are used to generate synthetic data: realistic but entirely fictional records that preserve the statistical properties of real data without exposing any real person's information. This allows AI systems to be trained on sensitive data that could not otherwise be shared. The mechanism is the same. The application and the consent structure around it are what determine whether it serves the person or exploits them.

The broader point: every photo or video you have seen carries an assumption that it records something that actually happened. That assumption is now mathematically breakable. For your generation, verifying the authenticity of digital media is not a solved problem. It is an active area of research.

Scout Presentations (6a, 6b, 6c)

Live Prompt Lab, 3 School-Related Tasks

Each slot shows a weak prompt and a strong prompt for a school-related task. Scouts type their own improved version in the text box, then compare with the strong example. This demonstrates the requirement to write clear instructions for school tasks.

Prompt Engineering in Professional AI Systems

The structure that makes a prompt reliable works the same way at any scale.

The difference between that and what a production AI system uses is output specificity. Instead of three short paragraphs, a system might require a structured format that a computer can parse without ambiguity: specific fields, specific labels, no variation permitted. The constraint is tighter because the output is not going to a person who can make a judgment call. It feeds directly into another automated process. Vague instructions produce variable outputs. Variable outputs break automated processes. The four-part structure is how you prevent both.

Three Questions That Shape Every Career in AI

Before researching a specific career, scouts should understand that "AI jobs" is not one thing. Every career in this space sits at the intersection of three distinct dimensions. Defining all three changes what you search for and what you find.

One Path Into AI, Many Ways In

There is no single required path into AI. Here's mine:

Reading the Labor Market with Primary Sources

A March 2026 Washington Post interactive analysis, worth bookmarking and having scouts explore directly, examined over 350 occupations for AI exposure and worker adaptability. A few findings worth discussing:

• Jobs in computer programming, marketing, financial analysis, and customer service show high overlap with current AI capabilities. But many workers in those fields also have the skills and flexibility to adapt, including higher education, varied experience, and access to strong job markets. Exposure does not automatically mean replacement.
• The most at-risk group identified: approximately 6 million clerical and administrative workers who are both highly exposed to automation and have fewer pathways to transition. This skews heavily toward female-dominated occupations, a dimension of AI's labor market impact that rarely makes the headline summaries.
• Previous predictions that ATMs would eliminate bank tellers, that early AI would decimate radiologists, and that a specific 2013 study's estimate that nearly half of jobs could be automated by computers would all have led scouts to make poor career decisions. Economists consistently note: "We do not have a good track record of predicting how technological change will play out in the labor market."
• As of the analysis date, there was no measurable evidence that AI is putting Americans as a whole out of work. That does not mean the transition will be painless, but it does mean that sensational headlines in either direction should be evaluated against the primary data, not accepted at face value.

All 8 Requirements

Confirm each requirement has been met before closing the session.

If You Want to Go Deeper