I was building a chatbot for my church and realized I was training it on bias

For about six months, I volunteered to help my local church in Boise make a simple chatbot to answer common questions. I was pulling text from old newsletters and sermon notes to feed it. Last week, a new member asked it about our stance on community aid, and its answer only referenced programs from twenty years ago, completely ignoring our recent food bank work. It hit me then: I wasn't just giving the bot information, I was accidentally locking in an old, incomplete view of what the church does. I was so focused on the tech working that I didn't think about the ethics of what I was teaching it. That single wrong answer showed me I was building a system that could mislead people based on my own narrow data choices. How do you even start to audit for that kind of hidden bias in a small project?

4 comments

4 Comments

jakewhite3mo ago

Nah, that's overthinking it. @grant_knight79 has a point about old papers, but a church bot just needs basic facts, not a perfect history. If someone wants the latest food bank hours, they should just call the office. It's a free tool made by a volunteer, not some official record. The old newsletters are still true info, just not the newest.

grant_knight793mo ago

Yeah that's a real problem with any kind of database or AI thing. Read an article a while back about a library that digitized its old local paper, and the search results made it look like the town was stuck in the 1950s because that's when the paper was most active. They had to actively go find newer community bulletins and school newsletters to add, otherwise anyone researching would get a totally wrong idea. For your church bot, maybe you need a checklist of all the big areas the church works in now, like the food bank or youth programs, and then go hunt for source material that talks about that recent stuff. It's like you have to force yourself to look for the gaps in your own knowledge, which is super hard lol.

the_max3mo ago

Used to think Jake was right about keeping it simple, but grant_knight79's library example really shows the risk. If the bot only pulls from old newsletters, it'll tell people the youth group meets in a room that got turned into a closet five years ago. You gotta go find those recent bulletins, even if it's a pain.

jamie_webb671mo ago

Dude this is such a real thing. I did a similar project for a small museum my buddy volunteers at, just a bot to answer questions about exhibits. I scraped a few old brochures and thought I was good. But then someone asked about a specific artist and the bot gave info from a 1998 show, totally missing the new installation from last summer. It was embarrassing because the bot looked official but was basically stuck in the past. What helped me was making a list of every question the bot might get and then testing if the sources actually matched todays reality. You really do have to go hunt down the newest stuff manually, even when it feels like overkill.