Every AI/ML interview loop at a company like Google eventually reaches the same place: a whiteboard-free, thirty-minute stretch where someone asks you to explain what overfitting is, or what a p-value actually means, and then keeps asking why until you either land on the mechanism or run out of road.
The questions below come from the ML Fundamentals set in squizzu's question bank — the ones that keep resurfacing in Google AI/ML and Data Scientist interviews. None of them are exotic. That is precisely what makes them dangerous: they sound easy enough to skip while you prepare, and they get asked anyway.
Five of them are playable right here in this article. Pick an answer, see whether you were right, and read the full explanation. The remaining 15 are listed in full below.
Try 5 of them right now
No sign-up, no email — just answer. Each question gives you the explanation and an in-depth breakdown of why the other options are wrong, which is usually the part that matters in the follow-up.
If you got fewer than four of those, you're in normal company. Precision vs. recall, the direction of the bias-variance trade-off, and the definition of a p-value are the three that trip up experienced candidates most often — and all three come up constantly.
The other 15 questions
These are the remaining questions in the set. Read them as a self-check: if you can't give a two-minute answer to each — definition, mechanism, trade-off, example — that's your revision list.
- What is supervised learning?Core ML concepts
- What is unsupervised learning?Core ML concepts
- What is cross-validation and why is it useful?Model validation
- What is gradient descent?Model training & optimization
- What is the difference between normalization and standardization?Data preprocessing
- What does it mean when two features are correlated?Feature engineering
- What is the difference between L1 and L2 regularization?Regularization
- What is Recall?Metrics
- What is a confusion matrix?Metrics
- What is the AUC-ROC curve used for?Metrics
- What is Bayes' Theorem?Statistics & probability
- What is entropy in decision trees?Tree-based models
- How do random forests improve over decision trees?Ensemble methods
- What is boosting?Ensemble methods
- What is A/B testing?Experimentation
The 5 answers, written out
If you'd rather read than click, here are the five you just played — definition first, then the mechanism that the follow-up question is usually aiming at.
What is overfitting?
A model overfits when it learns the noise and quirks of the training set on top of the real signal, so it scores well on data it has seen and poorly on anything new. You detect it as a widening gap between training and validation error, which is exactly why cross-validation is the standard check. The levers for fixing it are more data, a simpler model, regularization, and early stopping. Its mirror image is underfitting: a model too simple to capture the pattern, which performs badly on both sets.
What is Precision?
Precision is the share of true positives among everything the model predicted as positive — TP / (TP + FP). It answers "when the model says yes, how often is it right?", so it punishes false positives. Recall keeps the same numerator but changes the denominator to all actual positives, TP / (TP + FN), so it punishes false negatives instead. Which one you optimize is a product decision: spam filtering leans on precision because burying a real email is expensive, while cancer screening leans on recall because a missed case costs more than a false alarm.
What's the trade-off between bias and variance?
Bias is the error you get from a model too simple to represent the underlying relationship — it underfits. Variance is the error you get from a model flexible enough to fit the noise in this particular sample — it overfits. Lowering one typically raises the other, so the goal isn't zero of either but the point where total error on unseen data is smallest. A linear model on a curved relationship is the high-bias case; an unpruned decision tree is the high-variance one.
What is bagging?
Bagging — bootstrap aggregating — trains many copies of the same model on different bootstrap samples, each drawn with replacement so it contains roughly 63% unique rows, then averages the predictions or takes a majority vote. Because the models' errors are largely uncorrelated, averaging cancels them out and reduces variance. Every model trains independently and in parallel, which is precisely what separates bagging from boosting: no model in a bagged ensemble knows what the others got wrong. Random forests are bagging plus random feature selection at each split.
What is a p-value?
A p-value is the probability of observing results at least as extreme as the ones you got, assuming the null hypothesis is true. It is a statement about the data given a hypothesis, not about a hypothesis given the data — and that reversal is the single most common way this question goes wrong. A p-value of 0.03 does not mean there's a 97% chance your effect is real; it means data this extreme would appear 3% of the time if there were no effect at all. Below your significance level you reject the null, and at or above it you fail to reject — which is not the same thing as proving there is no effect.
What a Google Data Scientist interview actually tests
The 20 questions above have a shape worth naming. There's no transformer architecture question in the set, no "design a recommendation system," no distributed training. What's there is fundamentals: definitions you can state precisely, metrics you can choose between, and statistics you can apply to a live experiment.
That tracks with how AI/ML and Data Scientist loops are generally structured. Coding and system design get their own dedicated rounds; the fundamentals conversation is checking something narrower — whether you understand the standard tools well enough to pick the right one and defend the choice when someone keeps asking why.
Two clusters carry more weight than their share of the list suggests:
- Evaluation — precision, recall, confusion matrix, AUC-ROC, cross-validation. That's five of the twenty. If you can only prepare one area, make it the ability to say which metric you'd optimize for a specific product, and what it costs you.
- Statistics and experimentation — p-values, Bayes' theorem, A/B testing. At a company that runs experiments continuously, answering "I'd A/B test it" without being able to define a p-value does not survive the follow-up.
How to answer these in the room
The gap between a passing answer and a strong one isn't knowledge, it's structure. Almost every question above can be answered in four beats:
- Definition. One clean sentence. "Overfitting is when a model learns the noise in the training data, not just the signal."
- Mechanism. Why it happens. "It happens when the model has enough capacity to memorize the training set relative to how much data there is."
- Trade-off or diagnostic. How you'd detect or manage it. "You see it as a widening gap between training and validation error — which is why cross-validation is the check, and regularization, more data, or a simpler model are the levers."
- Concrete example. From your own work if possible. "On a churn model I trained, an unpruned tree hit 0.99 AUC on train and 0.71 on validation."
Interviewers are not checking whether you memorized a definition. They're checking whether you've ever had to make the trade-off in practice.
The four traps that catch most candidates
- Precision and recall, reversed. Precision is over predicted positives; recall is over actual positives. Under pressure, people swap the denominators. Anchor it to a product decision — spam filtering cares about precision, cancer screening cares about recall — and the definitions stop sliding.
- Bias and variance, backwards. High bias → underfitting. High variance → overfitting. The complex model is the high-variance one. This gets reversed constantly, and it's an instant credibility hit because everything downstream depends on it.
- Bagging described as boosting. If your answer mentions "learning from the mistakes of the previous model," you have described boosting, not bagging. Bagging is parallel and independent; boosting is sequential and error-focused.
- P-value as "probability my hypothesis is true." It's the probability of the data being this extreme given the null hypothesis, not the probability of the hypothesis given the data. Getting this wrong in a Data Scientist interview at a company that runs thousands of experiments a year is expensive.
What to do next
Reading questions is not the same as answering them. The five above gave you the feedback loop — you committed to an answer before you saw the explanation, which is the only way to find out what you actually know versus what merely looks familiar.
Answer the remaining 15 in quiz mode on squizzu — same format, with an explanation and an in-depth breakdown on every question, plus the rest of the ML Fundamentals set to keep going.
Still working on the application that gets you into the loop in the first place? How to write an AI/ML resume in 2026 covers what to name explicitly so a screening model can match it.