alice_allen5•6d agoRising Star

Just realized I was training my AI model on a biased dataset for months

I was building a customer service bot for a Denver clinic and a user pointed out it kept assuming all patients were male. I've been debating if we should scrap the project or just retrain it from scratch. Which approach would you take?

4 comments

4 Comments

zara_sanchez5d ago

Months of training and your bot thinks it's a 1950s doctor's office. "Good morning, sir" for everyone with a cough or a broken arm. Retrain it, don't scrap it. Just make sure your new data isn't all from, like, a football team's fan forum this time.

iris_schmidt5d ago

Zara_sanchez is right about the data source being the problem. It's like they trained it on a world where only one type of person exists. They need to feed it real conversations, not just one corner of the internet.

lucasw845d ago

Months of training and it never once guessed someone might be a woman? That's honestly wild. Scrapping the whole thing seems like a huge waste of time and money though. Just find a better, more balanced dataset and start the training over. Zara's right, just make sure your new data actually looks like the real world this time.

abbyp615d ago

Yeah, "find a better dataset" is the key. We had to do that last year. Just pulled a ton of public posts from different places, not just one site, and it fixed the bias fast.