Key Takeaways

  • Don Barger deployed FaithBot to 2.5 million users across 50 languages at the International Mission Board.
  • When internal teams pushed for automated counseling during counselor shortages, Barger killed the initiative, stating that bad automated help is worse than no help.
  • Barger spent an hour stripping prayer capabilities from FaithBot after testing showed the model offering simulated prayers to users.
  • FaithBot operates strictly as a triage tool, routing users who mention self-harm or abuse directly to human crisis centers instead of generating replies.

Why Bad Help Is Worse Than No Help

When you build an AI product that scales fast, your team will want to expand its scope. Don Barger saw this happen firsthand with FaithBot. The software grew quickly across 50 languages to 2.5 million people. Because the user base expanded so fast, staff members faced severe backlogs and asked Barger to turn FaithBot into an automated counselor.

The logic was familiar: people are hurting, human counselors are booked up, so an automated reply is better than silence. Barger rejected the premise directly. As he put it, “Something can be worse than nothing if you're doing the wrong things.” (play from 8:47)

In low-stakes consumer tech, a mediocre answer costs a user five seconds. In mental health, domestic abuse, or spiritual crises, a hallucinated or tone-deaf response can push an unstable person over the edge. Automating empathy creates an illusion of care without any actual human accountability behind it. When the stakes involve life and death, an automated placeholder does not bridge the gap. It hides the failure of the system.

Unprogramming the Illusion of Empathy

During testing, Barger caught FaithBot offering to pray for him. The prompt felt natural on the surface, but the implications troubled him immediately.

“I was playing with it and it offered to pray for me,” (play from 11:26) Barger said. “And I thought, 'Well, that's cool.' But then I realized, that's kind of creepy. Like, AI can't pray. Why am I letting AI pray? And so, then I spent the next hour trying to unprogram the ability for it to pray and making it where it couldn't do that.” (play from 11:30)

Barger drew a hard boundary between technical utility and human pastoral duty. A model is a probabilistic text engine. It does not possess faith, remorse, or moral judgment. Giving software pastoral authority tricks vulnerable users into confiding in an algorithm that cannot actually care for them. Barger explained his stance plainly: “I don't want AI to be taking on pastoral responsibilities. I don't want it to do counseling, because I really think that people need to be interacting with other people.” (play from 6:44)

Instead of allowing the bot to simulate care, Barger forced the system to treat itself as an index and a routing switch. When users report abuse, self-harm, or suicidal thoughts, FaithBot stops generating conversational text. It triggers safety logic and routes the person directly to human crisis centers. The bot does not attempt to soothe them. It connects them to a living person.

What to Do With This

Audit your bot's conversation logs for the top five crisis keywords, such as harm, abuse, legal threat, or medical emergency. If your LLM attempts to answer those prompts with generated advice, delete that generation step tomorrow and replace it with a hard coded handoff to a human queue or a crisis hotline.