Maybe the model didn’t get it wrong after all.
A recruiter, a spam call, and one observation that changed how I think about AI.
A few days ago I got a call from an unknown number.
Truecaller had marked it as Spam.
Without thinking twice, I ignored it.
A while later, I received a WhatsApp message.
“Hi Varsha, I’m a recruiter from XYZ. My number has somehow been marked as spam. I was trying to reach you regarding a Product role.”
I remember staring at that message for a few seconds.
If she hadn’t taken the extra effort to text me, I would’ve never known I had missed that opportunity.
Naturally, my first reaction was to blame the model.“That’s a bad prediction.”
But the more I thought about it, the more I felt that maybe the model wasn’t the interesting part. Maybe the question itself was.
Is this call actually spam?
At first, it feels like an easy yes-or-no question.
But is it?
If I’m actively looking for a job, that call is valuable.
If I’m not, it’s probably just another interruption.
Nothing about the caller changed.
Nothing about the phone number changed.
The only thing that changed was my context.
And somehow that completely changes what the “correct” answer should be.
Once I noticed it, I started seeing this pattern everywhere.
A real estate agent calling someone who’s looking for a house. Helpful.
The same agent calling someone who just bought a house. Spam.
A promotional email about running shoes.
Useful if I’ve been researching running shoes all week. Irrelevant if I’ve never shown any interest.
Google Maps suggesting the “best” route. Best for someone trying to save time isn’t necessarily best for someone trying to avoid tolls.
The more examples I thought of, the more I realized that many product decisions quietly assume there’s one correct answer.
Sometimes there isn’t. Sometimes the honest answer is simply:“It depends on who’s asking.”
That’s what I’ve been learning about AI.
A couple of weeks ago, I thought AI was mostly about replacing rules with models.
Now I’m beginning to think that’s only half the story.
The more interesting question seems to be:
Where do rules start breaking down?
Maybe it’s exactly when the world becomes too contextual for a single rule to work.
Not because the rules are bad. Not because the model is smarter.
But because people are different.
The same input can have different outputs depending on who’s receiving it.
I don’t have a conclusion.
This isn’t a post about how Truecaller should build a better spam classifier.
I’m certain they use advanced models.
This was simply one of those moments where a product I use every day made me think differently about AI.
I’ve started noticing that whenever the “right” answer depends on context, preferences, timing or individual circumstances, rules begin to struggle.
And maybe that’s where AI becomes interesting, not because it’s magical, but because it has a chance to learn patterns that fixed rules can’t easily capture.
Maybe six months from now I’ll come back to this post and disagree with half of it. But I’m still learning and that’s exactly why I wanted to write this down.
And honestly, I think I’d be okay with that.