Insight of the Week: Why bias in AI is a GOOD thing ??
How do we prevent AI from being biased?
This is a question we kept coming back to in a fascinating lunch meeting recently, where we got deep into the weeds of AI tech, impact and risks.
It's a great question, which gets to the heart of what modern AI technology is, and the role us humans have in making it a force for good.
The bottom line is that bias is?good. But only if it aligns with your own vision and values.
Let's start with the tech, and build up to the bias.
At their core, GPT models, like ChatGPT, are auto-complete on steroids. Give it a sequence of words, and it'll predict the next most likely word. I went into more detail in my email on:?What's the GPT in ChatGPT?
But autocomplete isn't much use, outside of a smartphone keyboard or spell checker. What made GPT into ChatGPT, was an army of human reviewers, who asked questions and graded (and corrected) the responses. This 'human feedback' was used to tune the GPT model so that it didn't just predict the next most?likely?word, but the next most?satisfying?word, according to the human reviewer.
And how did those reviewers decide? Partly their own intuition. And partly a set of rules laid out by OpenAI, the developers of ChatGPT.
This process meant that ChatGPT does not produce the next most likely word based on all of its training data. It produces the next most likely word that OpenAI and the human reviewers it employs, decided was?better?than the average of it's training data.
ChatGPT became useful?because?OpenAI introduced bias!
Once you understand this, two important consequences follow:
领英推荐
1)?Different individuals, communities, countries.. frankly, any grouping of people that share a common Vision and set of Values, will prefer an AI that aligns with their Vision and Values. So we can expect a multiplicity of AIs.
2)?AI is a force multiplier for ideas and information. I likened it to Electricity for knowledge workers in?last week's email?so if you want your Vision and Values, or those of your community or country to maintain relevance, you?must?contribute to building AI that is aligned with, or shall we say, 'biased' towards those values.
This is getting very philosophical, so let's bring this back down to earth. What does this mean for Customer Service, and Customer Experience?
Your company has a collective Vision and set of Values. These are typically expressed via mission statements, and brand guidelines. But ultimately, they're delivered by your staff. And AI, increasingly, is going to be part of your workforce. So you need to bias the AI you use to align with the Vision and Values of your business, and the expectations of your customers.
So you can't just plug ChatGPT into your customer service channels. You need to do the work to 'bias' it so that it talks the way your brand talks. Serves the way your best customer service agents serve. And leverages your human agents, back office processes, data and APIs in a way that will deliver on your brand promise. On the Vision and Values of your business.
AI is pretty cheap now. It costs a fraction of a cent to get a response from OpenAI's GPT3.5 model (the model that powers ChatGPT). All that power of translation, reasoning and conversation, for less than half a cent.
But it takes a whole lot more to make that response right for your business, and your clients.
If you're wondering how to get started, our Applied AI playbook is a great place to start. I outlined that in my email introducing the?Applied AI Playbook??and YouTube series on?Getting Started with Generative AI.
And if you want to see an example of extremely biased AI, check out my AI replacement,?Terry??he's so 'biased' the only thing he (or it?) will talk about is Conversational AI!
Kerry Robinson?is an Oxford physicist with a Master's in Artificial Intelligence. Kerry is a technologist, scientist, and lover of data with over 20 years of experience in conversational AI. He combines business, customer experience, and technical expertise to deliver IVR, voice, and chatbot strategy and keep Waterfield Tech buzzing.