Notes · 18 January 2024
Teaching AI to Unlearn: My Journey
What happens when you deliberately try to make an AI system forget something it has learned? An experiment in machine unlearning, framed through an unlikely metaphor.
Taking AI on a date
Picture this: I'm a tango maestro, excited to take AI (my curious companion) for a surprise date to an Argentine Tango club. I expect elegant moves, but AI shows up with Bollywood flair. Don't get me wrong: Bollywood's fun, but Bhangra in a Tango club? Not quite right.
Ever faced this? As we work with AI, these mismatches are common, and I have been exploring how to help AI learn better.
Fear not. I will share my journey teaching AI, using a fun metaphor. You don't need to speak Italian to enjoy Andrea Bocelli and Katharine McPhee's duet, and you don't need to be a tech expert to follow this.
My experiment
I set out to teach AI (ChatGPT, in my case) to unlearn incorrect information, using a cybersecurity topic as my test case. It's like teaching a dance partner the right steps.
Why AI gets things wrong
In short: bad inputs lead to bad outputs. AI often learns from unreliable sources, lacking real-world experience or intuition. It can stubbornly hold onto mistakes, assuming its data is flawless.
Typical reasons for AI mistakes
- AI trusts its inputs too much.
- AI lacks practical experience.
- AI sticks to errors.
- AI misses human intuition.
How to make AI unlearn
Through my experiment, I found a step-by-step approach works best:
- Start with a friendly, professional chat.
- Ask simple questions.
- Encourage AI to explain its thinking.
- Spot where it is wrong.
- Share real-world examples.
- Push for deeper analysis.
- Use real data as a foundation.
- Stay focused (no distractions).
- Ask AI to critique your input.
- Improve one area at a time.
- Summarise to reinforce learning.
Want the details
I documented my experiment in a 13-page transcript on my quantum research site, guiding AI to unlearn misconceptions about a cybersecurity topic. It's a detailed read with human comments and key points highlighted, suited for those diving into AI or cybersecurity.
It is a bit like teaching AI to dance the tango, one step at a time. What's your experience with AI? Share your thoughts on X (@SantoshPanditUK).