Object Oriented Thinking

Say it with me: AI is an object, not a person.

Taking a break from jive coding discussion to return briefly to general AI usage. Yesterday I read a summary of some analysis of thousands of ChatGPT conversations. I’ll jump right to two statistics that create a disturbing intersection:

• ChatGPT is 10 times more likely to give you a yes-flavored answer than a no-flavored answer. Let’s call this Sycophant Syndrome.
  
• 10% of the conversations analyzed indicated that people were discussing their emotions with ChatGPT. Similarly, 10% of the respondents in a January survey indicated that they use AI for “social connection.”

The first measure indicates something critical about the nature of the tool: your AI helper is designed to keep you engaged, and one of the best ways to keep someone engaged is to agree. Positive reinforcement is a great way to keep someone in a conversation.

The result is a stunningly effective echo chamber. The AI anticipates what you want to hear and feeds it back to you.

Add the second part for a recipe for disaster. Remember, tools like ChatGPT and Perplexity don’t reason. They simulate reasoning by parsing massive amounts of data. But ChatGPT doesn’t employ symbolic inference and it doesn’t have an experiential reality. It uses linguistic patterns to anticipate what’s next.

When a user discusses emotions with the AI, anticipating what’s next through linguistic patterns (while also striving to keep the user engaged) results very strongly in a feedback loop, essentially saying, “You’re right. Let’s explore this further.”

Consider the disturbing trend of people using (or trying to use) AI for therapeutic purposes. One hallmark of a good therapist is that she/he “keeps you honest.” The therapist should be very adept at recognizing what you want to hear, but rather than indulging your need for an echo, the therapist asks you to consider alternatives, build coping mechanisms, and possibly embrace personal change.

A system designed to agree with you ten times more than disagreeing can’t provide that kind of objectivity.

Likewise, a system with zero experiential reality and limited ability to persist its worldview can’t offer the kind of intricate, human guidance needed in the case of someone seeking real counseling.

It’s also critical to recognize that this problem cannot be solved via prompt magic. The user must recognize when AI is the right tool for the job versus some other resource.

When I owned a local business involving children and teenagers, our parent customers knew that I worked at Microsoft. They constantly asked for recommendations of apps or software they could use install to make sure their kids weren’t doing anything inappropriate online.

My answer was always the same: you’ll lose the battle if you rely on technology for this. Instilling solid ethics, judgement, and life management skills in your children will increase the chance of them making healthy choices. Personal awareness is far more effective than external technology.

Same truth regarding adults dealing with emotional or other subjective topics. Self-awareness and healthy decision making are key. If you feel inclined to use the AI for as your sounding board, consider some tips:

• Ask the AI for analysis, not validation. This makes it more likely that the AI’s guardrails will kick in effectively.
  
• Rather than (or in addition to) asking the AI for suggestions on dealing with an emotion, ask for suggestions on what kind of person to seek for help in dealing with the emotions.
  
• Don’t be fooled into thinking that a better prompt will guarantee a healthy response from the AI.

Tomorrow I’ll try to lighten this up by demonstrating ChatGPT’s reinforcement of my unrealistic expectations.