AI Safety
An technology company uses an AI hiring system to screen applicants, which is trained on past data of successful hires. The model tends to reccomend applicants with male names. This is an example of...
What is algorithmic bias?
This is the term for an organism's complete set of DNA.
What is a genome?
This range gives you a plausible range of values for some variable or parameter, and is reported alongside the main result.
What is a confidence interval?
This type of machine learning uses labeled data (input-output pairs) to train a model, such as imaging samples labeled "cancer" or "no cancer."
What is supervised learning?
This type of machine learning processes information from many formats, such as text, images, audio, video, and sensory data to make decisions.
What is multimodal learning?
This principle means a clinical AI model should be able to explain or justify its predictions to a physician.
What is interpretability?
Name the type of biomarker that predicts disease progression or outcome.
What is a prognostic biomarker?
This value tells you the likelihood of observing your result if there was truly no effect, and is a measure of statistical significance.
What is p-value?
This term describes when a model performs very well on training data but poorly on new data because it memorized noise rather than learning general patterns.
What is overfitting?
This law protects the privacy of patient health information used in training or deploying medical AI systems.
What is HIPAA?
This evaluation metric for AI means that benefits and harms are distributed acceptably across groups and contexts.
What is fairness?
This common file format stores raw DNA sequencing data, and is often used as the starting point for genomic analysis.
What is a FastQ file?
Your friend sees a study claiming "AI diagnosis correlates strongly with a specific gene expression pattern," and claims that that gene expression causes the diagnosis. What would you caution him about?
What is inferring causation from correlation?
This is the portion of your dataset used to actually fit/adjust the model's parameters, as opposed to the portion used to evaluate it.
What is training/validation data?
This process removes or masks personal identifiers (like name, address, SSN) from a dataset before it's used for research in order to protect patient privacy.
What is data de-identification / anonymization?
A hospital's AI model recommends against a treatment for a patient, but neither the physicians nor developers can explain why the model suggested this. Name this well-known problem in medical AI, and explain why it matters for patient trust.
What is the "black box" problem? This matters because doctors can't verify the reasoning behind a decision that affects patient care, making it harder to catch errors and build patient trust.
Testing multiple hypotheses and running more tests in a biomedical study increases the risk of this kind of error.
What is a false positive?
A study reports that patients who took fish oil pills lived longer, but the patients who took it were also younger and healthier. Name the reason why this comparison might be misleading.
What is a confounding variable?
This technique reduces a dataset with hundreds of correlated features down to a smaller number of combined variables that capture most of the original variation.
What is Principle Component Analysis (PCA)?
This mathematical formula represents the distance between the model's performance and the ground truth, and is typically a measure that is minimized in model training.
What is the loss function?
This describes a scenario where a chest X-ray model performs well but may have learned to recognize hospital markers or scanner artifacts rather than actual disease biology. What is this failure called?
What is shortcut learning?
Used to determine a good hypothesis for biomedical research, this metric might ask whether the biomarker improves real-world treatment outcomes.
What is clinical utility?
A screening test for a disease has a sensitivity of 95% and a specificity (true negative rate) of 90%. A patient tests positive. What piece of information would you use to prove that even with seemingly high accuracy, the patient might still be more likely to not have the disease? How would you prove this?
What is a disease's low prevalence in the population? / What is using Bayes' theorem?
Two hospitals both train models to detect pneumonia from chest X-rays. Hospital A's model has 99% training accuracy but 70% accuracy on new patients. Hospital B's model has 85% training accuracy and 83% accuracy on new patients. Which model is better for real deployment, and why?
What is Hospital B's model? (The small gap between training and testing suggests it generalizes well, while Hospital A's large gap suggests overfitting.)
A hospital trains a model to predict which sepsis patients will recover, using data pulled only from patients who were discharged (data from patients who died in the ICU before discharge are not in the dataset). Name the bias that explains why this model might predict overly optimistic outcomes.
What is survivorship bias?