# Lesson 1: AI Literacy as a Public Health Skill: Machine Learning and Neural Networks *Companion-podcast transcript • Sarah & Kiffer* *Office Hours episode to listen to after working through the lesson* --- **Sarah:** Picture a reference with plausible authors, a real journal and a DOI. It reads perfectly well, and the article doesn't exist. **Kiffer:** And the assistant that wrote it was doing exactly what a language model is trained to do, which is predict the next piece of text. **Sarah:** Then how am I supposed to tell a fluent answer from a correct one? **Kiffer:** A good part of the answer starts somewhere that sounds a long way from everyday use, with how a neural network learns. **Sarah:** Welcome to Office Hours. I'm Sarah. **Kiffer:** And I'm Kiffer. If you're starting the course with us, welcome to Lesson one, AI Literacy as a Public Health Skill: Machine Learning and Neural Networks. Sarah, take us into the first question. **Sarah:** Question one. I'll put it the way students often ask it. Most of us will use AI through an assistant, and we'll never train a neural network ourselves. So why spend time on weights, biases and gradient descent? **Kiffer:** The main reason is that the way these systems learn explains most of the ways they fail. Every machine learning system in the lesson, from the retinopathy network to the language model inside an assistant, learns by adjusting its parameters to fit the patterns in its training data. A language model is trained to predict the next piece of text. A reference with plausible authors, a real journal and a DOI is exactly the kind of text that fits that pattern, whether or not the article exists. **Sarah:** So a fluent output tells me it fits the pattern, and on its own it tells me very little about whether it's true. **Kiffer:** Yes. For well-established material, fitting the pattern and being correct usually go together, which is why explanations of standard concepts tend to be good. They come apart for exact details like references, dates and numbers, and for questions the training data cover poorly. Lesson two goes further into why. **Sarah:** Let me push on the evidence, because a lot rests on it. Noy and Zhang found that professionals with access to ChatGPT took forty percent less time on writing tasks, and the rated quality of their work rose by eighteen percent. But none of the experiments involved public health workers, and the analysis of ChatGPT messages by Chatterji and colleagues had two authors who worked for OpenAI. **Kiffer:** Both points are fair. The disclosure is a reason to read the usage figures with care, and it's good practice that the paper states it. Usage figures measure use, which differs from skill, and the size of the gains in public health work is still unknown. What the experiments show clearly is the shape of the result. There were large gains on tasks that suited the tool, and in the consulting study, a reversal on a task that looked similar. **Sarah:** You mean the drop of nineteen percentage points. **Kiffer:** Right. Dell'Acqua and colleagues found that consultants using AI on a task outside what they called the jagged frontier were nineteen percentage points less likely to reach a correct solution. The output was just as fluent on that task as on the others. Knowing how the tool produces its output is part of how you learn to tell which side of the frontier you're on. **Sarah:** Here's the part that bothers me. In the same study, consultants who started below average improved by forty-three percent, and those above average by seventeen percent. So the people who gain the most are the ones who start out weakest. Students are often the least experienced people in the room, and they're also the least equipped to spot a fluent mistake. **Kiffer:** I think that's the most important tension in this evidence. The gains are real, so I wouldn't tell students to stay away from the tools. The risk is also real, so the best time to learn task-specific checks is at the start, while the habit is forming. **Sarah:** And most students are forming that habit on their own. In the Higher Education Policy Institute's 2025 survey of undergraduates in the United Kingdom, ninety-two percent used AI tools, and only thirty-six percent said their institution had given them training in AI skills. **Kiffer:** Meanwhile, Canadian health policy now names the skill. Health Canada's Pan-Canadian AI for Health Guiding Principles, published in January 2025, list AI literacy as one of nine principles. **Sarah:** Then let's practise a check. **Kiffer:** Here's a made-up output, written for this episode. Someone asks an assistant to explain overfitting to a program manager in three sentences. The first sentence says overfitting happens when a model learns the noise in its training data along with the real pattern. The second says you can recognize it when the training error and the test error both stop falling. The third says the usual remedy is to add more hidden units so the model can capture the pattern better. Which of the three sentences would you keep? Take a few seconds. *(Pause)* **Sarah:** I'd keep the first one. The second is wrong. The sign of overfitting is a training error that keeps falling while the test error rises. Both errors levelling off is what you'd expect when training has run its course, or, if they level off high, when the model is underfitting. **Kiffer:** And the third? **Sarah:** That's a remedy for underfitting. More hidden units give the network more capacity, which makes overfitting more likely. The remedies for overfitting go the other way: more data, a simpler model, early stopping, regularization or dropout. **Kiffer:** So the output blended two paired ideas, and every sentence reads smoothly. We planted those errors, and real assistants usually explain a standard idea like overfitting well. Either way, the check that suits an explanation is to compare it, sentence by sentence, with the course material. It takes a minute, and here it caught two errors that a quick read would miss. **Sarah:** Question two. How can a model with higher accuracy be the worse choice? **Kiffer:** Let's build on the vendor case from Cedar Valley, the fictional health authority in the course's running case. If twelve percent of older adults there screen as lonely, a model that calls everyone not lonely is eighty-eight percent accurate. **Sarah:** And it finds nobody who needs help. **Kiffer:** Right. Now a made-up extension. Picture one thousand older adults, a hundred and twenty of them lonely and eight hundred and eighty not. A second model, model B, correctly flags eighty-four of the hundred and twenty lonely people, and it wrongly flags eighty-eight of the other eight hundred and eighty. What's model B's accuracy, compared with the eighty-eight percent for the model that flags nobody? Take a few seconds. *(Pause)* **Sarah:** Accuracy is the share of everyone the model gets right. Model B is right about the eighty-four lonely people it flags. Of the eight hundred and eighty who aren't lonely, it wrongly flags eighty-eight, so it correctly leaves seven hundred and ninety-two alone. Eighty-four plus seven hundred and ninety-two is eight hundred and seventy-six out of a thousand, or 87.6 percent. That's lower than the model that finds nobody. **Kiffer:** Right. The model that finds eighty-four people who could use help has a lower accuracy than a model that finds none. When an outcome is uncommon, accuracy mostly reflects how well a model handles the people who don't have it. **Sarah:** And in the terms from the lesson, model B's sensitivity is eighty-four out of a hundred and twenty, which is seventy percent. Its specificity is seven hundred and ninety-two out of eight hundred and eighty, which is ninety percent. **Kiffer:** Yes. There's a third number a program manager will care about, which is the share of flagged people who are actually lonely. Model B flags eighty-four plus eighty-eight people, a hundred and seventy-two in all, and eighty-four of them are lonely. That's about forty-nine percent. Epidemiologists call this the positive predictive value. **Sarah:** So about half the people a connector calls would be lonely. Is that good or bad? **Kiffer:** It depends on what the flag triggers. Here it triggers an offer from a connector, which costs staff time and is unlikely to harm anyone, so finding eighty-four people for a hundred and seventy-two calls could be a reasonable trade. If the flag triggered something costly or intrusive, half would be far too low. **Sarah:** And there's a trap if the model moves somewhere loneliness is less common. **Kiffer:** There is. Suppose, again hypothetically, that the model keeps the same sensitivity and specificity in a community where six percent of older adults are lonely. Out of a thousand, it would correctly flag forty-two of the sixty lonely people and wrongly flag ninety-four of the nine hundred and forty others. Now only about three in ten flagged people are lonely. And that assumes the sensitivity and specificity carry over, which is exactly what external validation tests. **Sarah:** So the questions for a vendor are sensitivity, specificity and the share of flagged people who are actually lonely, measured on a held-out test set and ideally checked in a population like ours. Measuring on data the model hasn't seen is also how deep learning made its case in the first place. **Kiffer:** It is. In 2012, a University of Toronto team entered the network later called AlexNet in the ImageNet competition, where programs labelled images they had never seen. Its top-five error, the share of test images whose correct label wasn't among its five best guesses, was 15.3 percent, against 26.2 percent for the next entry. That margin, measured on unseen images, persuaded much of the field, and it's the kind of evidence to ask of a vendor too. **Sarah:** Question three. Why can a model's strongest feature be the one to trust least? **Kiffer:** Here's another made-up example. Suppose the Cedar Valley team builds a model to predict which participants in its Connector program will leave before their sixth meeting. On its test set it's remarkably accurate, and its most important feature is the number of reminder calls the connector made to each participant over the twelve weeks. Why might that feature make the model look better than it really is? Take a few seconds. *(Pause)* **Sarah:** Because connectors probably call more often when someone starts missing meetings. So the calls are partly a record of the person already drifting away. By the end of twelve weeks, the number of calls carries information about the outcome itself. **Kiffer:** That's data leakage. One form of leakage is a feature recorded after the outcome has occurred, and this one is recorded partly after the outcome has begun. At the moment the model would actually be used, at enrolment, nobody knows how many calls there will be. A version without the leak uses only information available at the time of prediction, and its accuracy will probably be lower. **Sarah:** Now take it a step further. Someone pastes the model's summary table into the team's assistant, which is allowed because the table holds only aggregate results and no personal information, and asks what the program should change. The assistant answers fluently that reminder calls are the strongest driver of dropout and recommends making fewer of them. **Kiffer:** That answer turns a prediction into a counterfactual claim, about what would happen if the team changed the calls. The model can't support that. If the calls mostly respond to missed meetings, they'll travel with dropout whether or not they cause any of it, and they may even be helping. Finding out what fewer calls would do needs a comparison designed for the question, such as randomly assigning participants to different call schedules. **Sarah:** It's a cousin of the asthma example. **Kiffer:** Yes. In the asthma case, a care pathway made a high-risk group look low-risk. Here, a service that responds to trouble looks like its cause. In both, the feature predicts well and misleads as a guide to action. The check for the assistant's answer is to ask whether the question is description, prediction or counterfactual prediction before accepting what it says. **Sarah:** Question four is one I'd like to try myself. The worked example in the lesson updated a neuron for someone whose label was one. I want to do one where the label is zero. **Kiffer:** Good. Use the same neuron, with a weight of 1.2 on living alone, a weight of minus 0.8 on the number of social contacts last week, and a bias of 0.3. Our new person doesn't live alone, so the first input is zero. They had one social contact, so the second input is one. And they didn't screen as lonely, so the label is zero. **Sarah:** First the weighted sum. 1.2 times zero is zero. Minus 0.8 times one is minus 0.8. Adding the bias of 0.3 gives minus 0.5. The sigmoid of minus 0.5 is about 0.378, so the neuron gives this person about a thirty-eight percent chance of screening as lonely. **Kiffer:** That's right. Now take one step of gradient descent with a learning rate of 0.1. Before Sarah does it, here's the question for everyone. What's the new value of the weight on social contacts? Take a few seconds. *(Pause)* **Sarah:** The slope for each weight is the output minus the label, times that weight's input. The output minus the label is 0.378 minus zero, so 0.378. The input for social contacts is one, so that slope is 0.378, and the bias slope is also 0.378. Times the learning rate of 0.1, that's 0.0378. So the weight on contacts goes from minus 0.8 to minus 0.762, and the bias goes from 0.3 to 0.338. **Kiffer:** Before we move on, what does that do to the output? **Sarah:** The new sum is minus 0.762 plus 0.338, which is minus 0.424, and the sigmoid of that is about 0.396. Wait. The output went up, from 0.378 to 0.396. The label is zero, so it should have moved down. **Kiffer:** That check is worth doing every time. Where did it go wrong? **Sarah:** The update rule. The new weight equals the old weight minus the learning rate times the slope, and I added. In the lesson's example the slopes were negative, so subtracting them pushed every weight up, and I think I carried over the idea that the weights rise. **Kiffer:** Try it again with the subtraction. **Sarah:** Minus 0.8 minus 0.0378 is minus 0.838. The bias is 0.3 minus 0.0378, which is 0.262. The new sum is minus 0.838 plus 0.262, which is minus 0.576, and the sigmoid of that is about 0.360. So the output fell from 0.378 to 0.360, toward the label of zero. **Kiffer:** That's right, and the cross-entropy error for this person drops from about 0.47 to about 0.45. One more thing is worth noticing. What happened to the weight on living alone? **Sarah:** Nothing. Its slope was zero, because this person's input for living alone was zero. **Kiffer:** Right. In this neuron, a weight gets no share of the blame for an example where its input is zero, so the weight on living alone learns only from people who live alone. Larger networks are more complicated, though a general point carries over. A model learns little about the kinds of people who are rare in its training data. **Sarah:** Which leads into the part where we see things differently. Suppose model B, the one that flagged a hundred and seventy-two people to find eighty-four, passes a local validation in Cedar Valley. Should the team use it to choose whom connectors contact? **Kiffer:** I'd use it, in a specific way. The current route depends on screening. An older adult who scores six or higher on the three-item UCLA Loneliness Scale gets referred. That rule is clear and easy to audit, and it only reaches people who get screened. A model working from information the clinics already hold could point connectors toward people who would otherwise never be asked. **Sarah:** My worry is who the model learned from. It was trained on people who were screened and whose results were recorded. Older adults who rarely visit a clinic are thin in that data, and they may be among the most isolated. So the model could be weakest for exactly the people the program most wants to reach. It's also harder to explain. If a family asks why their father wasn't contacted, a low score from a model is a poor answer. **Kiffer:** Those are real concerns, and the first is the strongest. It's the reason I'd use the model only to add people. Everyone the screening rule refers still gets referred. The model's job is to suggest extra people for an offer of help, and a wrong flag costs a phone call. **Sarah:** Adding people still uses connector time, and with about half the flags wrong, a connector could spend half or more of those extra calls on people who aren't lonely. That time comes from somewhere. And once a model is running, I'd expect people to start trusting it more than its validation justifies. **Kiffer:** I agree the time cost has to be counted, and a pilot can measure it. Run the model alongside the current route for a few months, count how many extra lonely people it finds and what those calls cost, and screen a sample of the people it doesn't flag, so you can see whom it misses. Then the team can decide. **Sarah:** So where do we land? **Kiffer:** A model can add people to the offer of help, and the screening rule stays in place for everyone it already covers. The model is used only after a privacy impact assessment and validation on local data, a named person on the team is responsible for it, and its performance is checked over time, including for people who are rarely screened. **Sarah:** I still expect the model to miss more of the rarely seen people than Kiffer expects, and the pilot is how we'd find out which of us is right. But we agree on that rule. **Kiffer:** Let's pull it together. Here are three things to take away. **Sarah:** First, every machine learning system in this lesson learns the patterns in its training data. A fluent or confident output tells you it fits those patterns, and judging whether it's correct is the user's job, with a check matched to the task. **Kiffer:** Second, judge a model with the right numbers: sensitivity, specificity and the share of flagged people who really have the outcome, measured on data it hasn't seen, ideally from your own population. When an outcome is uncommon, accuracy can favour a model that finds nobody. **Sarah:** And third, a feature that predicts well is a clue about who's at risk. What would happen if you changed it is a separate, causal question, and a feature recorded after the outcome makes a model look better than it is. **Kiffer:** If you'd like more practice, rework the gradient descent step from this episode for someone who lives alone, had three social contacts last week and didn't screen as lonely. Use subtraction in the update, and check that the output moves toward zero. **Sarah:** Next time, it's Lesson two, How Large Language Models and AI Agents Work, where we look at how the same ideas, applied to text at a very large scale, produce the assistants we've been talking about. **Kiffer:** Take care, everyone. **Sarah:** See you in Lesson two.