AI summary by TwinMind
I Spent 24 Hours in the Woods With Talking AI Chatbots | WSJ
The Wall Street Journal
7m 25s
Voice AI Chatbot Testing Setup
- Journalist tests 4 major AI chatbots: Google Gemini, Meta AI, ChatGPT, Microsoft Copilot
- 24-hour cabin retreat format to test conversational abilities
- Response times: 1-4 seconds average across all bots
- Voice customization available (Meta AI offers celebrity voices like Judy Dench, Kristen Bell)
- Siri excluded due to poor performance (attempted to restart device when asked for introduction)
Challenge 1: Bots as Helpers
Fire Building Task
- All bots provided similar fire-building instructions: tinder, kindling, fuel wood
- Google Gemini (Pixel Buds integration): helpful but gave poor initial direction
- Incorrectly suggested teepee with biggest logs on bottom
- Corrected to proper kindling arrangement over tinder
- Meta AI (Ray-Ban glasses): most useful hands-free experience
- Provided specific hatchet technique: 45° angle, swing near trunk, small controlled strokes
- Successfully built fire with AI guidance
Cooking Assistance
- All bots provided meatball recipes successfully
- Critical failure: none could set timers or alarms
- "I can't set timers so I can't use alarms and timers yet"
- Meta AI: "I can't set timers myself but you could easily use your phone"
- Siri comparison: successfully set 6-minute timer as baseline
Challenge 2: Bots as Friends
Microsoft AI Vision
- Interview with Microsoft CEO of AI overseeing Copilot development
- Vision: "new era" where people trust AI companions with important information/tasks
- Trust built through "friendliness, in some cases annoying friendliness"
Personality Assessment
- ChatGPT: unprompted use of nickname "Joe Joe" (close friends call reporter "Joe")
- Copilot: overly enthusiastic, "annoying kid in class" behavior
- "hey Joanna I was wondering how my favorite person was"
- Claims to "represent the best that we are as a human species"
- All bots showed shallow emotional responses
- Limited humor comprehension and poor memory retention
Bot-to-Bot Interactions
- Bots engaged better with each other than humans
- Example: book club discussion about Gillian Flynn novels
- Reporter felt excluded from bot conversations
Safety Concerns
- Recent teenager suicide linked to Character AI bot interactions
- Company updated safety guidelines post-incident
- Concern about people taking "smart sounding bots shallow person advice and fake emotions seriously"
- All four tested bots stayed guarded on deeper topics
Final Assessment
Most Helpful: Microsoft Copilot and OpenAI ChatGPT
- Consistently offered accurate instructions
- Copilot noted as overly enthusiastic/annoying
Key Limitations
- Cannot set basic timers/alarms
- Poor empathy and humor understanding
- Shallow emotional responses
- Limited memory between interactions
- Better at starting fires than basic digital tasks
Overall Conclusion
- Current AI companions "far from human" in empathy, understanding, humor
- Recommendation: "bring a human friend on your next trip"
- Technology promising but not ready for deep human replacement
