← Understanding a computer science concept

Evaluating AI effectiveness and training, in depth

Beyond describing how an AI is trained, the standard rewards evaluating it — specifically, understanding that "accuracy" alone is often the wrong measure.

Why raw accuracy can mislead

• FALSE POSITIVE: the AI flags something that is NOT actually a problem (e.g. a healthy patient flagged as having a condition).
• FALSE NEGATIVE: the AI MISSES something that IS actually a problem (e.g. a sick patient cleared as healthy).
• The RIGHT balance between these two error types depends on the CONTEXT and the cost of each kind of mistake — not on which is more common in the training data.
• For a medical screening AI, a false negative (a missed diagnosis) is usually far more costly than a false positive (an unnecessary follow-up test), so the system should be tuned to accept MORE false positives in exchange for FEWER false negatives.

Worked example Evaluating a medical-screening AI beyond raw accuracy An AI diagnostic tool is 95% accurate overall. Evaluate whether this is a sufficient measure of its effectiveness. RAW ACCURACY ALONE IS INSUFFICIENT: 95% accuracy could still mean a high rate of false negatives specifically — e.g. correctly clearing 98% of healthy patients but missing 20% of actually sick ones, while the overall accuracy figure stays high because healthy patients are the majority. SEPARATE THE ERROR TYPES: report false positive rate and false negative rate SEPARATELY, not just the single blended accuracy figure. WEIGH THE COST OF EACH: a missed diagnosis (false negative) can cost a patient their life if untreated; an unnecessary follow-up (false positive) costs time and money but is far less serious — so the tool should be evaluated (and tuned) to minimise false negatives specifically, even at the cost of more false positives. CONCLUSION: the tool's EFFECTIVENESS depends on which error type it minimises, not just its overall accuracy percentage. ✓ The strongest answers explicitly separate false positives from false negatives and argue WHICH matters more for the specific context — a single accuracy number, however high, does not evidence this reasoning. ✕

Treating a high accuracy percentage as sufficient evidence of effectiveness

A single accuracy figure can hide a serious imbalance between false positives and false negatives, especially when one outcome (e.g. "healthy") is far more common in the data than the other. Always ask which error type matters more for the specific use case before judging an AI "effective".

Error type	What it means	Typical cost
False positive	Flags a non-problem as a problem	An unnecessary follow-up — usually lower cost
False negative	Misses an actual problem	A missed diagnosis — often much higher cost

Practice for this standard covers the whole topic — flashcards and quiz.