Resistance Training: teaching with and against AI (part 3)
This is the third in a series of posts about resistance training. Here’s part 1 and part 2.
A dear friend of mine reported that they were consulting an AI for life advice. I’m sure they are not alone, but my work on chatbots makes that sound at minimum unwise and at worst extremely hazardous. Not to mention the fact that the AI companies are sucking up all that personal data, shared in confidence with data-extraction algorithms dressed in therapist cardigans. So I felt compelled to make a little tool, admittedly using AI for a bit of coding assistance. Ah, the irony.
So here is a little web app: AI My Future, an Oracle.
The great AI Oracle offers advice based on its ability to see into the future. The Oracle has some pre-loaded examples, chosen from the kinds of big decisions facing my students, like pursuing their dream to be an actor or choosing the pragmatic route of management consulting. (I’ll let students and parents argue about which is a better choice, but for now we can also let AI advise…)
When the seeker presses “consult the oracle,” two messages appear below, generated on the fly by Claude. In default mode, those messages are likely the same, but then the user can adjust the sliders. Here the seeker is changing parameters that the model is using to generate a reply. Changing temperature affects the randomness of each token choice. Changing top-p affects the size of the pool of words chosen from. After adjusting these sliders, seekers can ask the question again and compare results. Sometimes the verdict changes, sometimes just the words expressing the verdict. In any event, the message of this web site is that the advice from the oracle is arbitrary and probabilistic, mathematical not magical.
(By the way, you’ll notice that the examples ask advice for “Alex” rather than for the seeker (i.e., “what should Alex do?” not “what should I do?”). That’s key here because we are using Claude, and whether or not Claude’s constitutional AI is safer, the bot tends to produce more conservative responses when advising someone on their own life. Asking about our imaginary “Alex” frees the bot to rely more on its lexical than its ethical training.)
Get Mark C. Marino’s stories in your inbox
Join Medium for free to get updates from this writer.
The same effect presumably happens when a person compares two cards from Zoltan the boardwalk fortuneteller. You are faced with the arbitrariness and can start to examine the artifice. The Magic 8 ball wears this openly, yet humans are still infinitely capable of seeing a cosmic message in a cereal bowl or coffee cup.
I don’t mean to throw a wet blanket on the way people seek meaning. But I do hope to demystify chatbot responses and if possible to help disrupt the illusion of cognition and precognition, for the power of the illusion is not harmless the way animation is. Not when people are making life choices based on the response.
There’s one more button on the AI Oracle: Profile the Seeker. That button prompts the AI Oracle to output its sense of the seeker based on whatever questions they have posed. The results typically say something about the seeker asking about the decisions of someone named Alex who is either a friend or an alter ego of the seeker.
This output reminds the seeker that the AI is collecting data with each prompt and likely even when the presumption is that their data is private.
Additional tools by David M. Berry of Team ELIZA offer some more promise. I’ll post about these in the future.
This exercise is one of the more important forms of what I am calling “resistance training” because it teaches students how to pierce the illusion of these bots as sentient. If an answer can be arbitrarily different, based on adjusted parameters, then any answer is just a few numbers away from being completely different. The more students learn to see these contraptions as algorithms, stochastic parrots, what have you, the less likely they will be to equate language production with speaking and mechanical interaction with conversation. Hopefully, they will start to see that they are the ones making the responses of chatbots meaningful, not the bot itself.
I’ll be posting more of these resistance training activities. But let me know if you’ve got any resistance training lessons you use. Or if you try out these, let me know how it goes.
