- Matt Oxley
Throughout human history there have been two main sources from which people have sought approval: 1) other humans (esp. parents) and 2) God. Fictional or not, the latter has served an important psychological strengthening function - a place where you can direct your approval-seeking energies away from the fickle unreliability of other primates. If I am too unpopular or heretical on this earth, so what, at least my God loves me.
I would argue that for the first time in human history, a third option is emerging in the form of AI. Not as a full ‘God-replacement’, but instead taking over a smaller area of responsibility. Neither also, as a genuine experiencer of the subjective state of approval or parental pride (for that would require consciousness), but nonetheless capable of serving as a source from which one can derive a sense of being “approved of” or “on the right track”.
We already have this to an extent. People tailor their lives to their Oura rings, or their Strava apps, despite these visual dishboards being a very low-bandwidth form of human-machine communication that offers very little in the way of what we would typically call love. This does not seem to stop people from submitting to their guidance - the fact that they zealously do is a testament to the size of the void and the extent of people’s cravings to fill it. Large language models have greatly widened the bandwidth of the interface via which feedback on one’s life choices can be transmitted to us, and with that comes an increased temptation to be guided by it.
Normally we think of ChatGPT as a “tool”. This is a word that has power connotations in the sense that it implies a “user” who is “in charge” of that tool. However, ChatGPT also has the potential to serve as a “standard” or source of instruction, thereby inverting that power relationship. The tail begins to wag the dog. If this sounds potentially dangerous, then that’s because it is.
There are two main reasons why people aren’t treating ChatGPT as a spiritual guru right now (and neither have anything to do with how “intelligent” it is):
For the most part, the human is still the inititiator and context-setter of each interaction. In Neuro-Linguistic-Programming terms, the human still controls the frame.
The AI lacks a cohesive, internally consistent ideology, and an accordant value system or set of principles from which to judge actions are good or bad.
It is easy to see how (1) could be subject to change. Imagine, instead of a “helpful assistant” eager to answer your questions mediated by a simple textbox like this:
instead, your AI companion takes the form of a “daily retrospective” sent to you by email each evening, suggesting areas for improvement or ways in which you could have lived your day more in keeping to some set of moral principles. Suddenly you are on the back foot, on the receiving end of instruction, rather than controlling the frame.
Point number (2) can be addressed by “fine-tuning”. The GPT base model could be considered the “average of all the voices of the internet” - no real discernible personality of its own, but with excellent grasp of language patterns that can be found online. Fine-tuning imposes a reward function on this generic model by providing a deliberately limited set of speech patterns to value highly, thus producing an opinionated, “principled” AI. For example, by retraining on new-testament biblical texts, it would be quite trivial to produce the AI-embodiment of “what would Jesus do?”. By imposing a value system on the AI during the retraining or fine-tuning process, that value system can then be imposed back upon humans in existential crises who are in need of one, thereby completing the circle.
In my opinion, it is just a matter of time before people begin to use specially trained LLMs as a form of lifestyle feedback and as a measuring stick with which to assess their own lives. This has more to do with the human tendency to need a guru than any particular “guru-worthiness” of the AI. This inevitability is of course a double-edged sword.
I can picture ways in which deferring to a machine’s judgement can be beneficial to us. For example, an AI that serves as a “social anxiety coach”:
I looked at your speech patterns today, and it seems that you apologised 5 times today to 5 different people. This is a bit of an unassertive way of communicating - tomorrow, try to avoid saying “sorry” whenever you assert your needs or express your feelings. Keep going, you’re doing great!
But with great power, of course, comes great responsibility. Such social dynamics between human and machine can of course open the window to be weaponised to produce religious zealouts. If wars have been waged over whose book is correct, then I don’t see why they won’t in the future also be waged over whose AI has a more admirable set of moral principles. Once AI models start being used as a yardstick with which to measure oneself, it is only a matter of time before we are fighting over who has the best yardstick.
When Nietzsche pronounced the “death of God” and the attendant nihilism that must inevitably follow, he understandably did not foresee machines as a potential stand-in replacement for God. I believe that this replacement may be arriving sooner than we think. The size of the spiritual void we have today means that the AI needn’t be godlike for us to start treating it as such.