AI Top Misconceptions
Misconception 1 — "More Accuracy = Better AI"
A vendor says their classifier is 99% accurate. It sounds great. But accuracy alone is meaningless without knowing how rare the target actually is. Imagine 100,000 items classified, with 1% actually positive. A 99% accurate model flags 990 true positives — but also 990 false positives from the clean items. Precision is just 50%: for every real hit, there is one false alarm. Same accuracy headline, very different real-world experience depending on the base rate.
Why the Base Rate Trap Catches Everyone
This is the base rate fallacy: people instinctively focus on the accuracy number and ignore how rare the target actually is. It explains disappointment with AI tools across medical screening, fraud detection, content moderation, recruitment — anywhere the thing being detected is rare. The fix is not "higher accuracy" — it is asking about positive predictive value at the real-world prevalence. That is the number that predicts the actual experience.
Misconception 2 — "The AI Knows Things Like a Human Does"
A trained model is a frozen set of weights, not a knowledge base. It produces statistically likely outputs based on patterns seen during training. It does not "understand" your business, has no memory of previous conversations unless explicitly given, and cannot tell when it is wrong. Confident-sounding wrong answers are not a malfunction — they are the predictable behaviour of a probabilistic system being asked to act certain.
Misconception 3 — "AI Keeps Learning From How I Use It"
Most deployed AI models do not learn from your usage. They are frozen snapshots, retrained on a schedule the vendor controls. Feedback loops — when they exist — flow into the next training run, not into the live model you are using right now. The model you are using today reflects what the world looked like at training time. Ask vendors: retraining cadence and what signal drives it.
Three Questions to Ask Any AI Vendor
Instead of "how accurate is your AI?", ask three sharper questions: (1) What is the base rate of the thing you detect, in a real deployment environment? (2) What does the model output when it does not know the answer — does it say so, or does it guess confidently? (3) How often is it retrained, and on what signal? Vendors who answer all three cleanly are typically the ones worth a deeper conversation. Vendors who deflect any of the three usually have something to hide in that answer.
Try It Yourself
The base rate trap lands differently when the numbers are your own. This takes five minutes, no AI tool — just rough arithmetic on a system you already live with.
Pick one detection or screening system from your own life — spam filtering, a fraud alert on your card, a smoke alarm, a medical screening you have had. Estimate how rare the real event actually is (say 1 in 100 or 1 in 1,000), assume the system is 99% accurate, and work out roughly how many false alarms it produces for every real catch.
- You wrote down an actual base rate estimate, not just "it is rare"
- You can say roughly how many false alarms come with each real hit
- You can explain why "99% accurate" sounded better before you did the math
Put in your own volumes and see the false-alarm arithmetic.
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.