Vision AI Strengths & Limits
📋 What it is
Vision AI is fast and tireless but can be fooled, and it only knows the objects it was trained on.
🗣️ Coach says
Computer vision beats humans at speed — scanning millions of images without tiring. But it has real limits: it can be fooled by odd angles or stickers, and it can’t recognise anything it wasn’t trained on. Powerful, but not magic.
🧠 Memory hook
Vision AI = fast + tireless, BUT foolable + limited to what it was trained on.
😂 Giggle
What do you call an AI that sings?
A neural note-work!
😲 Whoa!
Researchers fooled a vision AI into calling a turtle a “rifle” just by tweaking its texture in ways humans can’t even notice — proof it matches patterns, not meaning.
✅ Quick check: A vision AI has never seen a unicycle in training. Can it recognise one?
Say your answer out loud first — then reveal.
No — it can only recognise objects in its training data; something new is invisible to it.
A model’s knowledge is bounded by what it was trained on.
🪄 Trick question: If a vision AI is fooled by a sticker, does that mean it truly “understands” what it sees?
Careful — think it through, then reveal.
No — being fooled by a sticker shows it matches surface patterns, not meaning; it doesn’t understand.
Adversarial tricks reveal that vision AI is pattern-matching, not comprehending.
🧪 Try it! (2 minutes)
Notice a face-detector box on a camera app. Try an odd angle — does it still find the face? (Sometimes not.)