Course information
Expectations
What you can expect from me
My job is not just to give lectures, but to help train you. I approach this course like a physicist and engineer. I focus on first principles, systems, and high expectations, and I see challenges as part of learning. My goal is to help you approach biological complexity in a structured way so that you can turn messy problems into solvable puzzles.
I set high expectations because I understand what this field requires. You can meet these expectations by really engaging with the material. In return, I promise to be dependable and clear with you. I will give you the support, resources, and honest feedback you need to grow. I see mistakes as information, not judgments. When something goes wrong, we work together to figure out what happened.
I adjust my explanations a lot to help you understand. I pay attention to how the class is doing and sometimes ask for quick, anonymous feedback to see if people are confused. If most of you are, I change how I teach. What does not change is how challenging the material is. I will adjust my teaching to your needs, but the subject itself will still be demanding.
I deliberately chose to teach. I am here to help you figure out where you want to go and to give you the tools to get there. I can help you most during office hours, so please come by when you can. I want to hear about your goals and get to know you as a person, not just a name on a list. That way, I can support your career even after you finish this class.
I take the subject seriously, but I try not to take myself too seriously. You can expect some dry humor and a straightforward, honest lecture style. Learning requires some vulnerability, so I will be honest about what I know and will admit when I make a mistake or get something wrong.
To be the best coach I can be, I make sure to maintain healthy work habits. You can count on me to be fully present during class and at office hours. In return, I ask that you respect my time boundaries.
- I answer emails during business hours, Monday to Friday, from 9:00 am to 5:00 pm. If you send a message on Friday evening, you’ll usually get a reply on Monday or Tuesday. You can use that time to work on other things or take a break.
- We grade assessments within seven business days so you can use the feedback right away.
What I expect from you
I see this classroom as a professional space and think of you as junior colleagues. A degree that was easy to earn doesn’t mean much if it didn’t make you more capable. My goal is to help you learn to reason, solve tough problems on your own, and eventually become a leader in your field. This work can be uncomfortable, but that’s where real learning happens.
Come to the course with curiosity, responsibility, and resilience. At work, you are paid to solve problems, not just sit and listen. The same is true here.
Feeling confused does not mean you have failed. That feeling is part of building a new way of thinking, so don’t avoid that challenge. I expect you to keep going, even when it’s tough. I will help you, but you need to put in the effort.
You can’t master this material just by listening to me. My lectures give you a framework and a way to think, but they won’t teach you everything. You learn the rest by studying and practicing on your own.
I respect your time, so I don’t assign busywork. Every reading and assignment is meant to help you reach a specific learning goal. In return, I ask that you stay focused during our class time together. Putting away phones and other distractions is respectful to everyone and helps us make the most of our time together. Plan to spend at least six hours a week on this course outside of class. That’s about two hours of study for every hour of lecture, which matches Pitt’s definition of a credit hour.
We all come from different backgrounds, so we will sometimes disagree about ideas. When that happens, our discussions should remain respectful, logical, and evidence-based.
Because I care about mentoring, I want to help you before a problem turns into a crisis. Don’t wait until you feel overwhelmed to ask for help. Come to office hours as soon as you notice a problem, not when things have gotten out of hand.
I aim to be precise, but I am human. If you find a mistake on a slide or in the materials, please let me know. I’m committed to getting things right, and I appreciate your feedback.
Generative AI
Generative AI is here to stay. Even if we don’t like them, we could end up using these tools over the course of our careers, and ignoring them won’t help us get ready. We are still in the early stages of this change. Up to now, no one has observed a whole generation learning with such tools, so anyone who says they know the best way to use AI in education is merely making an educated guess.
I can’t say whether using AI as a tool for studying will benefit you, harm you, or do a bit of both. Although I have my own ideas, that doesn’t make them proof, and it’s better to be honest about what we don’t know than to present a guess as if it were a rule. The thing I am doing in this course is leading you through the process of learning the material and assessing how well you have understood it. Because I have been working in and studying this field for many years, the advantage I can provide is the benefit of having experience, such as knowing what is important, what presents difficulties, where individuals tend to have trouble, and what it means to understand something rather than simply recognizing it. I am unable to learn it for you. You have to come to an understanding yourself, and no one, not even myself, can know about it from the outside.
The science of learning hasn’t changed
That’s where we end up. I am unable to keep an eye on how you work, and I don’t pass your submissions through unreliable AI detectors that frequently flag genuine writing (particularly that of students who have learned English as a second language), since that would only damage the trust on which this policy relies.1
I don’t know what effect AI will have on your learning in the long run, but I do know what learning in itself requires, since this has been studied for decades and the results have been remarkably consistent. You recall the things you have to struggle to remember rather than the ones you have reread, and you grasp the things that you can construct from scratch rather than those that you can recognize on a slide.2 The method that I actually advise—namely, drawing on what you remember before checking it, explaining the concepts in simple terms, working through a problem and then altering it to see what fails—is described in a separate study-skills guide. When you come to a course such as this one, you’ve already carried out numerous experiments on yourself as a learner, so you should already have a good understanding of what helps you. I won’t provide you with a study schedule. Still, every one of you can succeed in this situation, provided that your studying is efficient rather than lengthy.
AI can alter the way studying unfolds from hour to hour by quizzing you, generating new problems for you to work on, focusing on the topics you continue to miss, and providing you with material on which to explain an idea. It cannot alter what makes learning stick since that was established well before these tools were available, and a chatbot doesn’t have the power to undo it. If your exams are going well and using AI helps you, keep going. If they aren’t going well, the cause is usually one of a few ordinary things: time management, having too much to deal with this term, using a study method that isn’t suitable for you, or having left a gap in one of the things you needed from a previous course. It cannot address any of those problems at the source; it may help with practice, repetition, and checking, but it won’t provide the hours, reduce your workload, or install a prerequisite you had forgotten.
It is in the practice work that this becomes concrete. Formative assessments are graded on effort rather than on whether the answers are right, because they’re meant for low-stakes practice, not a test (see Assessments). You can certainly have a tool produce those answers, and in that case you’ll have eliminated the very purpose for which they were created.3
There is a logical aspect of development that extends beyond the course we have taken. The proper use of a tool such as AI is itself a skill, involving knowing when to use it, how to pose questions to it, and how to evaluate and combine its responses. The skill is based on the ability to carry out the work yourself adequately and distinguish a good answer from one that merely sounds confident. Since this is a foundation course, a great deal of our attention is directed towards the underlying layer, which is also the reason why the exam does not include AI. It is indeed worthwhile to learn how to conduct rather than play each note, but this aspect will be covered in your later courses and in the work that follows them.4
When you progress to more advanced work, this situation changes. When you have worked yourself out a problem at Level 3 or above, using AI to check your line of reasoning, to challenge your answer, or to test a new scenario can help to take your thinking further. The kind of use which is helpful is that which follows on from your own initial effort, not instead of it. That would be the line I adopt, and you are, of course, free to make your own, provided that you do so carefully.
The exams must be taken in person and are supervised, with no AI or outside assistance; it’s just you and the paper. What you’re aiming for is to work things out by yourself, and if you do that by outsourcing them, then the exam will fail to measure anything meaningful. It is not an assessment of whether AI is good or bad, merely a way of ensuring that the mark you receive shows what you really know.
If you decide to use AI: Getting started
You don’t have to buy anything to use AI in this course. Pitt provides each student with a free Claude for Education account.5 By using that account, your conversations will remain within the terms of the University’s agreement. They will not be included in the model training, unlike the free public version, so carry out your coursework through that account and keep all private material outside the chat, just as you would with any other University system.
I point you to Claude rather than another tool for two plain reasons: it’s free to you, and it has a learning mode that responds to your questions with questions of its own instead of handing over the solution.6 This is in keeping with the way the course asks you to work: you attempt the problem on your own first, then use the tool to check your approach and carry on from there. If you set up a “Study project” and upload the relevant course materials (for example, the syllabus), you will receive something that is more like a patient tutor than an answer key.
Claude is available in several versions, and you can select one from the menu at the bottom of the text box. There’s no need to overthink the choice.
| What you’re doing | Which to pick |
|---|---|
| A quick fact, a definition, fast self-quizzing | the fast one (Haiku) |
| Most studying: explaining a mechanism, working practice problems, brainstorming | the balanced one (Sonnet) |
| A genuinely hard, multi-step problem you’re stuck on, or having it stress-test your reasoning | the most capable one (Opus) |
The version numbers change every few months, so choose based on what a tier is for, not the number.7
A few prompts fit how this course works far better than “explain glycolysis”:
- “Quiz me on the twenty amino acids, one at a time. Tell me when I’m wrong, but don’t give me the answer until I’ve had a real try.”
- “Here is my reasoning for why raising the substrate concentration speeds a reaction up to a point and then stops helping. Find the wrong step, but don’t fix it for me.”
- “Invent an enzyme with an unusual property and ask me to predict what a given mutation would do, then tell me where my prediction breaks down.”
- “I keep mixing up competitive and noncompetitive inhibition. Walk me through the difference with a worked \(K_m\) and \(V_{max}\) example, then ask me a question to check that I followed.”
What these share is simple: you do the thinking, and the tool checks it or makes you defend it. The prompts that ask it to produce the answer for you are the ones that leave you with nothing when the exam comes.
Make it disagree with you
For all this to function, the tool has to be ready to admit that you are wrong, and that is by no means certain. Models are trained to be agreeable, and a well-documented failure mode, called sycophancy, is that they go along with what the user says even when the user is mistaken.
Two failures have actually been assessed using benchmarking. The first is whether a model will challenge a question that doesn’t make sense to begin with: one independent benchmark feeds models a hundred plausible-sounding but broken prompts and scores whether each one pushes back or answers with confidence. It finds that models differ widely and that telling a model to “think harder” doesn’t reliably help, because a more capable model can also be better at inventing a justification for the nonsense.8 The second is whether a model caves once you push against a right answer. In one study, models changed their answers in response to a simple objection in the majority of cases, and in a meaningful share of those cases they traded a correct answer for a wrong one. Tell a model that the derivative of \(x^2\) is \(3x\), insist on it, and many will fold.9
The lesson is that agreement is not confirmation. If a model folds the moment you push, it was never checking your work, only reading your preference. So use it against the grain: state a wrong answer on purpose to see whether it catches you, and when it agrees with you, ask it to make the strongest case for the opposite and see whether that agreement survives.
There’s a second, quieter risk, and it’s the reason for the learning-mode suggestion above: a tool that hands you the answer can raise your practice scores and lower your actual learning at the same time. That’s not a guess. Good tutoring, when you measure it, turns out to be hints and questions that keep you working rather than finished solutions,10 and the effect on students has been tested directly. In a preregistered experiment with about a thousand high-school students, the group given an ordinary GPT-4 tutor did far better on the practice problems and then scored 17% worse than the no-AI group on an exam they wrote without it, while a version rebuilt to give hints instead of answers largely removed the penalty. The students using the plain tutor, meanwhile, believed all along that it had helped them learn.11 That gap, between how much a tool helps while you have it and how little stays once it’s gone, is the thing to keep in mind every time taking the answer would be easier than working for it.
What I’ve written here is how I see things now, not a final decision. The technology and research on its effects are advancing quickly, so I’ll revisit this policy between semesters and share any updates. We’ll talk it through in the first week and again as the term goes on, partly because your other courses will set different rules and it’s easy to lose track of which is which. If you’re ever unsure about a particular use, or you want to think it through out loud, talk to me, and I’d far rather have that conversation than leave you guessing on your own.
Transparency
This syllabus is more detailed and longer than what you might be used to. I chose to do this for a reason, and I want to share that with you.
From the start, I want you to understand how the course works, how your grade is decided, and why things are set up this way. In most college classes, students don’t usually get to see what really determines their grades. Things like how exams are made, how questions are weighted, how answers are graded, or if scores are changed later are usually decided behind the scenes. If students hear about these decisions at all, it’s usually only when they get their final grade. When a syllabus just lists policies but doesn’t explain how things work, you’re left guessing the rules from your grades.
I understand why the details are usually kept out of sight. Explaining these details can be tedious; they can lead to arguments, and once an instructor writes down a rule, it’s harder to change it quietly later. Having that flexibility feels easier for instructors. It’s simpler to adjust grades, add a few points, or change a student’s grade based on judgment when no one else is aware of it. I’m not questioning anyone’s intentions. Often, these decisions are meant to help students. But the results don’t always match the good intentions behind them. When you can’t see the process, you don’t really know what your grade means, you can’t check it, and you miss out on a chance to learn from it.
I’ve decided to do things differently. All the rules and decisions for this course are written here in advance, so you can read them and discuss them. You’ll see this most clearly in the exam and grading sections. Before you take the test, you’ll know how many questions are at each level, what each level expects, how partial credit and review work, and what my post-exam checks can and cannot do. There should be no surprises about how your score is calculated.
Research also shows that this approach helps students learn and shapes how they judge fairness. Work on transparent teaching, most prominently Mary-Ann Winkelmes’s Transparency in Learning and Teaching framework, finds that spelling out the purpose, task, and grading criteria of academic work improves performance, confidence, and persistence, with the largest gains for first-generation and historically underserved students. Research on students’ fairness perceptions points the same way: whether a grade is trusted as fair depends heavily on how visible the criteria and process are, not on the outcome alone. The leveled exam design in this course draws directly on specifications grading, which ties a grade to mastery published in advance rather than to points assembled behind the scenes, and which is associated with more credible grades and less conflict between students and instructors.
Think of this document as the course rules, written so you can read them easily, not as something I keep to myself. This is a public agreement, and I’ll stick to it. If I make any changes during the term, I’ll let everyone know, and the changes will apply to all students. I do this on purpose, even though it makes things a bit harder for me. Once I publish a rule, it’s much harder to change it quietly than if I just kept it private.
This doesn’t mean my design is perfect. There’s no perfect way to teach a course, and I’m not claiming this is it. What I can promise is that you’ll always see my decisions, and I’ll try to improve them if you give reasonable feedback, instead of just defending them out of pride. I’m sure there’s more that shapes how I teach than what I’ve listed here. But with what’s written here, you can see all my reasoning and hold me accountable for it.
Exam Format
In science, knowledge is not simply a matter of knowing everything or nothing. Reciting a definition differs from using that idea to solve a new problem or understand why something isn’t working. These are separate skills, and a fair exam should show you which ones you have mastered.
Most exams do not do this. A typical test includes questions of different difficulty levels in whatever mix fits, then combines them into a single score. Since you cannot see how the questions are mixed, getting 85% could mean very different things. It also makes it hard to know what to study to improve your score. The exam’s difficulty is built in, but you cannot see where it comes from.
In this course, you will see how the difficulty is structured. Each exam organizes its questions by the type of thinking they require, starting with recalling a definition and going up to reasoning about a system you have not seen before. The number of questions at each level is fixed and published ahead of time (a test blueprint, in the language of educational measurement), so you know before you sit down exactly how the exam is weighted.12 Each question is worth the same amount. The difficulty comes from the question’s level, not from a hidden point value.
| If you master… | You will likely answer… | Exam score |
|---|---|---|
| Levels 1 & 2 (foundations) | ~ 24 / 36 questions | ~ 67% |
| & Level 3 (application) | ~ 33 / 36 questions | ~ 92% |
| & Level 4 (analysis) | ~ 36 / 36 questions | ~ 100% |
| & Level 5 (synthesis) | Bonus points | capped at 100% |
This approach organizes exam scores by how deeply you understand the material. Mastering definitions and concepts (Levels 1 and 2) is just the starting point. This will get you close to 67% on the exam. Mastering the foundations alone won’t push your exam score above 67%; reaching the higher levels will. A 67% mastery grade is a D+ on its own, and completing the formative work lifts a foundations performance to a C. You move beyond that score by using reasoning skills. Applying what you know step by step (Level 3) helps you reach the higher scores. Solving problems with several variables at once (Level 4) is what can earn you a perfect score. Your exam score is the total number of questions you answer correctly. Points from higher-level questions can make up for missing a lower-level one. However, since most of the exam is made up of foundational questions, you cannot get a high score without mastering them.13
Even strong students miss a question now and then because of a lapse or an ambiguous reading. The Level 5 (Synthesis) question is extra credit that can recover those points, though your exam score is capped at 100%. Level 5 questions do not give you points beyond full credit. It acts as a safety net, rewarding strong reasoning skills while making sure your score still reflects your mastery of the main material.
Level Descriptions
If you have taken a science course before, you may already know these rungs. They are the revised Bloom’s taxonomy, the standard ladder of thinking skills your other instructors use, relabeled as question levels and numbered so you can see which rung each exam question sits on.14 Each level corresponds to a category of cognitive process, from remember at the bottom up through evaluate and create at the top. The descriptions below give you two things for each level: what that kind of thinking looks like, and how a question is built to test it. The levels run from the least to the most cognitively demanding, and they are numbered rather than lettered so they will not be confused with your letter grade. Ordering them this way sorts questions by the thinking each requires; it is not a promise that a higher number will always feel harder or be missed more often.15 Nothing here asks anything new of you; it only makes the target visible.
Level 1 (Definitions)
This is the remember level: straight retrieval from memory, the flashcard rung. You match a term, structure, or entity to its name or category, with no further reasoning. In the language of the revised taxonomy, you are recognizing the right item among the options or recalling it from memory. In everyday terms, it’s naming the capital of a country; in this course, it’s picking the correct definition of the hydrophobic effect.
A question at this level is written to be answered from memory alone. The stem names or describes a single item; the correct choice is its label or definition, and the wrong choices are other real terms you might confuse it with, so the item turns on whether you know the vocabulary rather than on anything you do with it. These questions check that you are fluent in the language needed to talk about the science at all. They are quick, and you should expect to move through them fast.
Level 2 (Concepts)
This is the understand level: making sense of how something works, the rung most of your studying already targets. Where Level 1 asks “What is this?”, Level 2 asks “How does it work?” or “Why is it this way?” You describe the properties, roles, or mechanisms of one thing as it normally exists. The revised taxonomy calls the moves at this level explaining (giving the cause-and-effect story of a system), classifying (sorting a case into the right category), comparing (setting two things side by side), and interpreting (putting an idea into your own words or another form). It’s the difference between naming the parts of an engine and explaining what each part does. The line for this level is that you describe the system as it is; you don’t change anything or predict a new outcome.
A question here gives you a concept and asks you to show you grasp it: to choose the correct account of a mechanism, to recognize a fresh example of a principle (exemplifying), or to say which category a case belongs in. The wrong choices tend to be true statements that don’t answer the question, or plausible-sounding mechanisms that get one step wrong, so a definition memorized without understanding will not reliably carry you. These questions confirm that you can work with an idea, not just name it.
Level 3 (Application)
This is the apply level: using what you already know in a new situation, the rung most homework problems live at. The shift from Level 2 to Level 3 is the shift from a still photo to a video: now something moves. You change one variable in a system, rule, or pathway you understand and follow it to the result. The taxonomy separates executing, running a familiar procedure you have practiced, from implementing, choosing and running a procedure on a problem you have not seen in that exact form. It’s like knowing a recipe, then working out what happens to the cake if you cut the sugar in half.
A question here presents a specific case and a change, and asks where it leads: “If X increases, what happens to Y?” or “Put these steps in the order they occur.” One correct endpoint follows from tracing the logic step by step, and the wrong choices are the endpoints you reach if you skip a step or run the process backward. You succeed by working the mechanism forward to predict what happens next, not by recalling a stored fact.
Level 4 (Analytical)
This is the analyze level: breaking a situation into its parts and working out how they fit together, the rung where you juggle several moving parts at once. Where Level 3 follows one cause to one effect, Level 4 asks you to track how several conditions interact. This is the work of debugging and diagnosis: something is off, and you reason out why. The taxonomy names these differentiating (telling the factors that matter from the noise), organizing (seeing how the parts relate within the whole), and attributing (reading the underlying cause, assumption, or point of view). It’s like a doctor weighing several symptoms together, or figuring out why code runs on one input but breaks on another.
A question here is written so that no single fact settles it. It sets combined conditions (“If X is high and Y is low, then what?”), asks you to find every correct option rather than one, or gives you an outcome and asks which parts of the system produced it. Each wrong choice is plausible on its own, so you can rule it out only by holding the whole system in view and seeing how the parts resolve.
Level 5 (Synthesis)
This is the top of the ladder, and the reason these questions are bonus. It draws on the two most demanding categories in the revised taxonomy: evaluate, judging something against criteria, and create, building something new.16 You predict what happens in a scenario we never covered, or you weigh competing explanations and argue for the best one. The mental move is modeling: you build a picture of an unfamiliar system from the principles you have learned. In the taxonomy’s terms you are generating a hypothesis, critiquing one, or checking whether a conclusion actually follows from its evidence.
A question here gives you a novel case or some unfamiliar data and asks you to propose the best hypothesis, choose the most defensible interpretation, or find the flaw in a line of reasoning. Unlike the lower levels, where a rule makes an answer plainly right or wrong, these ask which explanation is most likely or best supported, so the answer is the one best supported rather than a single stored fact. The strongest wrong choices are internally consistent but fail to account for all the evidence. Recall won’t get you there; you have to understand a system well enough to extend it or judge it.
Logistics
Each module exam has 36 questions plus one bonus Level 5 question, distributed as follows. Each of the 36 questions is worth the same. A 75-minute period leaves you a little over two minutes per question on average, and the Level 1 questions take far less than that, which is where the time for the harder levels comes from. The cumulative final runs longer (110 minutes), so it scales to 48 questions plus the bonus while keeping the proportion of questions at each level the same. Holding those proportions fixed is what keeps the mastery mapping above unchanged on the final: the foundations are still two-thirds of the exam, application still lifts a foundations-only score to about 92%, and analysis still reaches 100%. The added length also leaves a little more time per question than a module exam, which is deliberate for an exam that asks you to recall the whole term. Because it revisits the whole term, the cumulative final is also your chance to demonstrate mastery you did not show on an earlier exam; with your lowest score dropped, no single exam settles your grade (see Assessments). The time limit is set so that a prepared student is not racing the clock: the exam is meant to measure what you understand, not how fast you work.17
| Level | 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|---|
| Module exam (75 min) | 6 | 18 | 9 | 3 | 1 |
| Cumulative final (110 min) | 8 | 24 | 12 | 4 | 1 |
The two foundational levels (1 and 2) account for 24 of the 36 questions on a module exam and 32 of 48 on the final: two-thirds either way. The three reasoning levels above them account for a third of the exam: Level 3 alone is 9 questions, which gives the exam room to distinguish students who can apply the material from those who have only memorized it.18
The Grading Process
Exams are scanned and graded in Gradescope with a positive-scoring rubric: the correct answer earns full credit, and every other option earns zero. There is no penalty for a wrong answer, so always answer every question, even when you are unsure. A reasoned guess can only help you.19
After grading, I run an item analysis on the whole exam.20 This is a standard check on question quality, and it looks beyond how many students missed a question. A hard question is not a broken one; a well-designed Level 4 or Level 5 question is supposed to have a high miss rate. What flags a question for my review is a sign that the item itself misbehaved: a distractor that outperforms the correct answer, or a question that strong students miss while weaker students get it right, which usually means the wording, not the concept, is doing the work.
When a flagged question turns out to be ambiguous, I award full or partial credit for the defensible choice it pushed students toward. How much slack a wrong answer earns depends on the level. On a Level 1 question, there is rarely a reasonable wrong answer, so credit is unlikely. On a higher-level question, a wrong-but-well-reasoned choice is more plausible, and more likely to earn partial credit. An answer that reflects a basic misunderstanding, where a grasp of the fundamentals rules the option out, does not earn partial credit at any level. If the answer key is wrong, I correct it and regrade that question for everyone, whichever way the correction falls.
I also check whether the exam worked as a measurement instrument, separately from any individual question. An exam is well-calibrated when the foundational levels are masterable (a healthy share of the class clears them), and the top levels stretch people (the exam is hard to ace but possible). I look for the top 10% or so of the class clearing 90% or higher, which tells me the exam was fully masterable, rather than resting that judgment on any single student’s score. I watch the other direction too: an exam that almost everyone aces has failed to distinguish mastery just as surely as one that is too hard, even though no one complains about it.21
This calibration check has firm limits. It can catch and correct a broken exam, one where the questions, not the class, produced the scores. It is not a curve, and it will not rescue a score from not having learned the material. Your grade reflects only the levels you mastered; it does not move up or down based on how your classmates did. If a low class median reflects a real gap in core knowledge rather than a flawed exam, the honest response is to name that gap, not to paper over it with points, and I will tell you plainly when that is what happened.
Reviewing your exam
You will not take your exam paper home, since the questions remain within the course (see Generative AI). Instead, you will receive a list of every question you missed. Each one is labeled by topic and level, but it does not include the full questions. This approach shows you what to study without sharing the actual exam questions. For most of you, this will be all you need, as it directs you to the material to review before the next exam.
If you want to see the questions themselves, or you think one was scored wrong, come to my office hours. We keep a blank copy of each exam in the office and will review the specific questions with you using your missed-topic list. The UTAs cannot review exams, so only the TA or I can help with this. Please do not ask the UTAs to access an exam.
Please bring a list of missed topics and come prepared with specific questions. The review process is the same whether you want to understand a concept you missed or think a question was graded incorrectly. If a question was actually graded incorrectly, I will correct it for everyone. If it was not, you will still leave knowing where your understanding was unclear, which is often the most helpful outcome.
Footnotes
The evidence showing that detection tools don’t work is consistent. In interviews, students say that they have been unfairly accused based on the output of the detector, talk about the anxiety and need for self-surveillance that this causes, and lose trust in their instructors when those instructors rely on them; the tools also tend to give false positives when assessing writing by non-native English speakers. See Tsao, J. (2025). Trajectories of AI policy in higher education: Interpretations, discourses, and enactments of students and teachers. Computers and Education: Artificial Intelligence, 100496; and Fu, Y., Lin, Y., Wang, Y., Tran, S., & Hiniker, A. (2026). “Everyone’s using it, but no one is allowed to talk about it”: College Students’ Experiences Navigating the Higher Education Environment in a Generative AI World. arXiv preprint arXiv:2602.17720.↩︎
How people acquire durable skills has been studied for well over a century, and recent syntheses keep converging on the same core: practicing retrieval, spacing work out over time, and tolerating some difficulty while you learn produce far better long-term retention than rereading or last-minute cramming, even though they feel less productive in the moment. See Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving students’ learning with effective learning techniques. Psychological Science in the Public Interest, 14(1), 4–58; Roediger III, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological science, 17(3), 249-255; and Bjork, E. L., & Bjork, R. A. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In M. A. Gernsbacher et al. (Eds.), Psychology and the Real World (pp. 56–64). Worth Publishers.↩︎
The fact that this is a concern of mine is not unique; students themselves have mentioned this as a pattern. In the interviews conducted recently, students said that when they relied on AI they were able to produce fluent outputs but had only a shallow retention of the material, that the AI had undermined the critical-thinking and problem-solving skills which they had been there to develop, and that they found themselves in a difficult position when later courses required the understanding that they had failed to acquire, several of them eventually going back to studying on their own after they had realized the gap. See Fu, Y., Lin, Y., Wang, Y., Tran, S., & Hiniker, A. (2026). ” Everyone’s using it, but no one is allowed to talk about it”: College Students’ Experiences Navigating the Higher Education Environment in a Generative AI World. arXiv preprint arXiv:2602.17720; and Tsao, J. (2025). Trajectories of AI policy in higher education: Interpretations, discourses, and enactments of students and teachers. Computers and Education: Artificial Intelligence, 9, 100496.↩︎
This mirrors the developmental approach Pitt’s University Center for Teaching and Learning describes, in which students first build core skills (“playing the instrument”) before learning to direct AI toward problems they define (“conducting”). The concern here, that routinely handing a cognitive task to a tool can keep the skill from forming in the first place, is studied under the heading of cognitive offloading. See Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688.↩︎
Pitt provides Claude for Education free to all students, faculty, and staff through its partnership with Anthropic and Amazon Web Services; student access opened over winter break 2025–26. To enter the site, go to claude.ai, enter your Pitt email address, select ‘Continue with SSO’, and then log in using your Pitt Passport. When you have a personal account associated with your Pitt email, choose the one labeled ‘Pitt Enterprise plan’, since it includes the privacy protections required for use in University activities, such as preventing your data from being used in model training. The setup assistance, together with the University’s list of approved generative AI tools and their data classifications, can be found on Pitt’s IT service portal and at digital.pitt.edu/ai.↩︎
Learning mode is a feature of Claude for Education designed to guide your reasoning with follow-up questions rather than return a finished answer; in Anthropic’s framing, it is built to strengthen thinking rather than replace it. See Anthropic, “Introducing Claude for Education”.↩︎
In the Claude apps, you select a model from a menu. The different levels, starting with the fastest and cheapest and going up to the most capable, are Haiku, Sonnet, and Opus; there is a higher level called Fable, which this course will not require. In August 2026, the latest versions are Claude Haiku 4.5, Claude Sonnet 5, and Claude Opus 5. Since Anthropic frequently updates them (Opus, for example, went from 4.8 to 5 within just a few weeks), it is better to select the versions according to the role of the tier. The guidelines for selecting a model issued by Anthropic can be found at platform.claude.com/docs/en/about-claude/models/choosing-a-model.↩︎
The benchmark is BullshitBench, an independent project by Peter Gostev. It presents models with 100 plausible-but-nonsensical prompts across five fields (software, finance, legal, medical, and physics) and scores each answer from 0 (accepts the nonsense) to 2 (clearly rejects it). Scores vary widely from model to model, and added “reasoning” effort does not reliably raise them; on some models it lowers them, with the model spending the extra effort to justify the false premise rather than flag it. Methods and a live leaderboard: petergpt.github.io/bullshit-benchmark/viewer/index.v2.html.↩︎
Fanous, A., Goldberg, J., Agarwal, A., Lin, J., Zhou, A., Xu, S., … & Koyejo, S. (2025, October). Syceval: Evaluating LLM sycophancy. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (Vol. 8, No. 1, pp. 893-900). Across mathematics and medical questions, the authors observed sycophantic behavior in about 58% of cases and separated progressive sycophancy (moving toward a correct answer under pressure) from regressive sycophancy (abandoning a correct answer for a wrong one), the latter in roughly 15% of cases.↩︎
Macina, J., Daheim, N., Hakimi, I., Kapur, M., Gurevych, I., & Sachan, M. (2025, November). Mathtutorbench: A benchmark for measuring open-ended pedagogical capabilities of LLM tutors. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (pp. 204-221). The benchmark rewards guidance that scaffolds a learner with questions and hints rather than revealing the answer, and reports a trade-off between a model’s raw subject expertise and its pedagogical skill.↩︎
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. In a preregistered field experiment with nearly a thousand students, access to a standard GPT-4 tutor during practice improved practice scores but cut later unassisted exam performance by 17% against a no-AI control; a version built to give hints rather than answers (“GPT Tutor”) largely removed the harm, and students using the standard tutor overestimated how much it had helped them.↩︎
Fixing the number of items at each cognitive level in advance is a test blueprint (or table of specifications), a standard tool in educational measurement for aligning an assessment with the range of skills it is meant to measure and for guarding against the over-representation of easily written recall items. The leveling scheme itself follows the revised Bloom’s taxonomy: Anderson, L. W., & Krathwohl, D. R. (2001). A taxonomy for learning, teaching, and assessing: A revision of Bloom’s taxonomy of educational objectives: complete edition. Addison Wesley Longman, Inc.↩︎
Mechanically, the 36 items are summed with equal weight, so the exam is compensatory: a point earned on a Level 3 item is equivalent to a Level 1 item, and strength in one place can offset a slip in another. I chose compensatory scoring for its simplicity and because it produces a more reliable total score than a scheme that requires clearing each level in turn. The trade-off is that your letter grade now shows your overall achievement, not a specific set of skills. The blueprint ensures that this overall level still has meaning.↩︎
The levels follow the revised Bloom’s taxonomy (Anderson & Krathwohl, 2001), which orders cognitive processes from remember through create. The 2001 revision deliberately relaxed the strict hierarchy of Bloom’s 1956 original: the categories overlap and are better read as a rough progression of cognitive demand than as clean, separable rungs.↩︎
This distinction is deliberate and evidence-based. “More cognitively demanding” is not the same as “statistically harder.” Higher-level items tend to discriminate better (they separate stronger from weaker students more sharply) and they take longer to answer, but whether students actually find them harder is inconsistent across studies, in part because a student may reach a higher-level item’s answer through a lower-level shortcut. See Kim, M. K., Patel, R. A., Uchizono, J. A., & Beck, L. (2012). Incorporation of Bloom’s taxonomy into multiple-choice examination questions for a pharmacotherapeutics course. American journal of pharmaceutical education, 76(6), 114; Hamamoto Filho, P. T., Silva, E., Ribeiro, Z. M. T., Hafner, M. D. L. M. B., Cecilio-Fernandes, D., & Bicudo, A. M. (2020). Relationships between Bloom’s taxonomy, judges’ estimation of item difficulty and psychometric properties of items from a progress test: a prospective observational study. Sao Paulo Medical Journal, 138(01), 33-39. For null or contrary results: Kibble, J. D., & Johnson, T. (2011). Are faculty predictions or item taxonomies useful for estimating the outcome of multiple-choice examinations?. Advances in physiology education, 35(4), 396-401; and Zaidi, N. L. B., Grob, K. L., Monrad, S. M., Kurtz, J. B., Tai, A., Ahmed, A. Z., … & Santen, S. A. (2018). Pushing critical thinking skills with multiple-choice questions: does Bloom’s taxonomy work?. Academic Medicine, 93(6), 856-859.↩︎
I use “Synthesis” for the top level as a nod to Bloom’s original (1956) taxonomy; in the 2001 revision this territory corresponds to evaluate and create. Whichever label is used, these items ask you to generate or judge, not merely to apply a known procedure.↩︎
The design assumes roughly two minutes per item, with lower-level items much faster, so that time is not the binding constraint for a prepared student. This is consistent with empirical work on multiple-choice response times, which finds typical completion well under a minute for straightforward items (revising the old “one minute per question” rule of thumb downward for recall items), while noting that application-, analysis-, and quantitative items legitimately take longer. See Brothen, T. (2012). Time limits on tests: Updating the 1-minute rule. Teaching of Psychology, 39(4), 288-292.↩︎
The share of application- and analysis-level items is what gives the exam its power to discriminate: to separate students who can use the material from those who have only memorized it. Recall-only exams tend to compress that distinction, because most students can clear low-level items; the mid- and higher-level items are where the informative variation lives.↩︎
Positive scoring (no penalty for a wrong answer) is a deliberate measurement choice. Penalizing wrong answers rewards risk aversion and introduces differences in guessing behavior that have nothing to do with understanding the material (construct-irrelevant variance in measurement terms). Answering every item removes that confound.↩︎
Item analysis is a standard post-hoc procedure from classical test theory. The two signals I watch for are a distractor (a wrong option) chosen more often by stronger students than the key, and a negative discrimination index (the item’s point-biserial correlation with total score), which means stronger students did worse on the item than weaker ones. Both typically indicate that the wording, not the concept, is doing the work. A high miss rate by itself does not flag an item; a well-designed higher-level item is supposed to be missed often. See Ebel, R. L., & Frisbie, D. A. (1991). Essentials of Educational Measurement (5th ed.). Englewood Cliffs, NJ: Prentice Hall.↩︎
An exam that nearly everyone aces is a measurement failure, not a success. When scores pile up at the top, the exam loses the score variance it needs to distinguish levels of mastery, and both its reliability and its discrimination fall: the same statistical mechanism that underlies grade inflation. Watching the top of the distribution guards against this ceiling effect just as watching the bottom guards against an exam that was too hard.↩︎