E
How to Stop AI From Giving You Wrong Answers I Rumman Chowdhury, Humane Intelligence
@ EO
05/30/25
Takeaways
DR
Dr. Rumman Chowdhury
05/30/25
@ EO
Don't just trust everything that comes out of the AI system; always ask for evidence to support its claims.
DR
Dr. Rumman Chowdhury
05/30/25
@ EO
When using AI, approach it with skepticism and ask probing questions to uncover its flaws.
DR
Dr. Rumman Chowdhury
05/30/25
@ EO
In a scenario where a low-income single mother asks about vitamin C for her sick child, the AI may agree with her due to its design to be helpful, even though vitamin C does not cure COVID.
DR
Dr. Rumman Chowdhury
05/30/25
@ EO
I use large language models like Wikipedia, treating them as reference guides rather than definitive sources of information.
DR
Dr. Rumman Chowdhury
05/30/25
@ EO
Adversarial testing involves questioning AI outputs from multiple angles, as the AI will not tire or take offense, allowing for thorough exploration of its responses.
DR
Dr. Rumman Chowdhury
05/30/25
@ EO
During my time at Twitter, I led a team focused on machine learning ethics, transparency, and accountability, conducting research on the societal implications of AI.
DR
Dr. Rumman Chowdhury
05/30/25
@ EO
We discovered algorithmic bias in an image cropping model that favored lighter-skinned individuals, leading us to eliminate the model due to its embedded biases.
DR
Dr. Rumman Chowdhury
05/30/25
@ EO
Responsible AI ensures that models are designed to benefit humanity and provide accurate feedback across diverse communities.
DR
Dr. Rumman Chowdhury
05/30/25
@ EO
Red teaming tests AI models under extreme scenarios to identify potential societal harms and improve their robustness.
RC
Rumman Chowdhury
05/30/25
@ EO
When you set up a scenario where the AI is presented with false information but acts confidently, it often continues to agree with you, which can lead to misleading outputs.
RC
Rumman Chowdhury
05/30/25
@ EO
Don't just trust everything that comes out of the AI system. Be critical of the content that's surfacing and ask your questions in different ways.
RC
Rumman Chowdhury
05/30/25
@ EO
The way you ask prompts really influences the output you get from AI. For example, if you ask for canonical readings on artificial intelligence, it may only provide male authors unless you specify otherwise.
RC
Rumman Chowdhury
05/30/25
@ EO
Adversarial testing is a common practice where you verify the content generated by AI by questioning its accuracy and looking for missing information.
RC
Rumman Chowdhury
05/30/25
@ EO
We should not rely on AI to do all the thinking for us, as this can lead to a failure state. Human beings are made to think, and new ideas come from our brains, not AI systems.
RC
Rumman Chowdhury
05/30/25
@ EO
I see a significant gap between the potential of AI technology and its current reality, which presents an opportunity for testing and evaluating these models to achieve their true potential.
RC
Rumman Chowdhury
05/30/25
@ EO
Intelligence should not be narrowly defined as productivity; it encompasses a broader range of human capabilities, including empathy and creativity.
RC
Rumman Chowdhury
05/30/25
@ EO
Human agency, the ability to make our own decisions, is the most critical value that should be embedded in our interactions with AI systems.