The video explores the substantial impact that generative ai has had across various domains, such as industry transformation, legal document summarization, and customer engagement. generative ai has enhanced productivity and shortened time to delivery for enterprises. However, it also introduces risks like copyright infringement, lack of transparency, and most notably, bias. The video delves into different types of biases, such as algorithmic bias, cognitive bias, confirmation bias, outgroup homogeneity bias, prejudice, and exclusion bias, providing real-world examples to elucidate these concepts.

Addressing bias in AI requires a systematic approach involving ai governance. This includes setting up policies, practices, and frameworks to manage AI activities responsibly, ensuring fairness, equity, and inclusion. The video discusses how enterprises can use specific governance tools to detect and mitigate bias, ensuring that AI's benefits reach consumers and the corporate entity effectively. Methods to minimize bias involve stakeholders in selecting diverse training data, employing fairness tools, building varied AI teams, and continuous data monitoring to adjust to evolving perceptions and real-world data changes.

Main takeaways from the video:

Please remember to turn on the CC button to view the subtitles.

Key Vocabularies and Common Phrases:

1. generative ai [ˈdʒɛnərətɪv eɪˈaɪ] - (noun) - Artificial intelligence that can generate new data based on learned patterns from existing data. - Synonyms: (AI generation, data generation AI, creative AI)

generative ai has had a wide ranging impact on the world today.

2. algorithmic bias [ˌælɡəˈrɪðmɪk ˈbaɪəs] - (noun) - Bias that occurs when AI algorithms produce systematic and unfair results. - Synonyms: (systematic bias, computational bias, model bias)

AI bias or machine learning bias, also known as algorithmic bias, refers to AI systems that create and produce biased results.

3. cognitive bias [ˈkɒɡnɪtɪv ˈbaɪəs] - (noun) - A type of bias where human thinking patterns influence judgments and decision-making processes. - Synonyms: (thinking bias, perception bias, mental bias)

There's also cognitive bias. Now we must remember humans design AI systems.

4. outgroup homogeneity bias [ˈaʊtˌɡruːp hoʊməˈdʒiniəti ˈbaɪəs] - (noun) - A cognitive bias where people perceive out-group members as more similar to each other than in-group members. - Synonyms: (group bias, similarity bias, stereotyping bias)

That's outgroup homogeneity bias. It's based on an assumption that people outside your diverse group are all similar.

5. exclusion bias [ɪkˈskluːʒən ˈbaɪəs] - (noun) - Bias that occurs when certain groups are inadvertently left out from a data set or study. - Synonyms: (omission bias, absence bias, sampling bias)

The last kind is the exclusion bias. This is where you leave out inadvertently data that was important for the sampling.

6. proprietary information [prəˈpraɪətɛri ˌɪnfərˈmeɪʃən] - (noun) - Information that is owned by a company and gives the company a competitive edge. - Synonyms: (owned information, confidential information, trade secrets)

These come from proprietary or private or personal information coming out as outputs of the LLMs.

7. ai governance [eɪˈaɪ ˈɡʌvənəns] - (noun) - Methods to direct, manage, and monitor all AI activities within an enterprise for responsible development. - Synonyms: (AI oversight, AI management, AI control)

Identifying and addressing bias requires ai governance.

8. supervised learning [ˈsuːpərˌvaɪzd ˈlɜrnɪŋ] - (noun) - A type of machine learning where the model is trained on labeled data. - Synonyms: (labeled learning, directed learning, guided learning)

But when we are making decisions between supervised and unsupervised learning models, we need to be a little bit more careful.

9. unsupervised learning [ʌnˈsuːpərˌvaɪzd ˈlɜrnɪŋ] - (noun) - Machine learning using data without labeled responses. - Synonyms: (undirected learning, autonomous learning, exploratory learning)

For unsupervised learning models where AI alone identifies bias, you need to leverage tools.

10. fairness indicators [ˈfɛrnəs ˈɪndɪˌkeɪtərz] - (noun) - Tools or metrics used to assess fairness in AI models. - Synonyms: (equity measures, impartiality metrics, unbiased indicators)

These tools basically use fairness indicators and there are several tools out there from Google, we've got Google Toolkit, AI Fairness, 360 from IBM, OpenScale and what have you.

Unraveling AI Bias - Principles & Practices

generative ai has had a wide ranging impact on the world today. And it's for all of us to see. Starting from the economic impact, the impact that it's had on industry transformation, legal document summarization, customer engagement, cost savings and so many more. The reason why we've had this impact is because of the benefits that generative ai has provided. I like to talk about three our ability to perform complex tasks using generative ai, the increase in productivity that we can see. And for enterprises, it's the shorter time to value to get products and services out.

But as with all new technology that has risks, generative ai also has its associated risks. Some of the emerging risks come from downstream based model retraining. It could be copyright infringement. There's also traditional risks that we've seen with AI. These come from proprietary or private or personal information coming out as outputs of the LLMs. It could also be the lack of transparency that the model offers in explaining the results that it gives out. But the most amplified risk that we see today is the one that comes from bias. And that's the topic of discussion.

In this video we're going to be looking at the types of bias, the principles of how to avoid them, and actually the methods that we can use to avoid and create bias free AI systems. So let's look at the different types of biases, but before that, let's define what a bias is. AI bias or machine learning bias, also known as algorithmic bias, refers to AI systems that create and produce biased results.

What do I mean by biased results? Biased results reflect and perpetuate human biases within a society, normally including historical and current social inequalities. They're pretty harmful. We've seen a lot in the news where companies and enterprises have been questioned on the kind of biases that they have in the data that they've trained their models on.

Let's look at the different types of biases that we have, starting with algorithm bias. This is a systematic and erroneous behavior of an AI system to always produce an unfair outcome. How do I explain this? Let's look at an example where an AI developer developed a loan application system and in that system it automatically prevents applicants born before 1945. You've created age bias and it's just automatic and systematic over time.

There's also cognitive bias. Now we must remember humans design AI systems. There's always a human intellect and input when you actually create these systems. Let's take for example the tendency of a human brain to think in a certain way. An Example would be recency bias, which is a type of cognitive bias. You're kind of influenced by recent events. What would be a good example? Spread of COVID in 2020. Or it could be an ongoing war. It skews your thinking and it builds that bias into the systems that you are building.

Confirmation bias. The next type of bias depends on and it's a related bias to cognitive bias. It relies on pre existing beliefs. Beliefs such as left handed people are more creative, right handed people are less. That can easily creep into the way you think about the data that you're using.

The next one is outgroup homogeneity bias. Now this is a little tricky to explain. Let me show you. Let's assume that you've created a data set of training data that you believe is a diverse training data set. You probably have made a larger assumption by where you think that the group outside of the diverse group are all similar. That's outgroup homogeneity bias. It's based on an assumption that people outside your diverse group are all similar.

Prejudice the next type. This is societal, faulty societal assumptions. The most popular example would be all nurses are female, all doctors are male, and so on and so forth. It's very easy for this bias to come in into the AI systems that you're developing.

The last kind is the exclusion bias. This is where you leave out inadvertently data that was important for the sampling. What do I mean by that? Let's say you send out a survey to a set of individuals that incidentally were the smartest employees in your enterprise. You've left out an entire group that could represent the less or average set of employees who, when it comes to performance, skewing your results.

Typically, at the start of any AI innovation within an enterprise, it's very easy to get enamored by the wonderful cool things that Genai can do for you. You have a successful prototype, a poc, a pilot for a small set of users, and you're quick to announce success until AI starts getting tested by internal and external users and you start seeing biased results. That's when you know that you need to step back and relook at your AI initiative with a little more effort.

Identifying and addressing bias requires ai governance. Now, what ai governance offers is a method to direct, manage and monitor all AI activities within your enterprise. It's a set of policies, practices and frameworks that enable you to do responsible development of AI. It typically engages tools and technologies that detect fairness, equity and inclusion. It is the best tool for enterprises to use to ensure that the benefits of the governance goes directly to the customers, to the consumers, employees as well as the enterprise.

Avoiding bias may sound harder than it is, but there are proven methods that one can use to ensure your enterprise is bias free in all your AI applications. Let's talk about a few selection of learning models. Now it's obvious that you're going to choose a learning model that aligns with the business function that you want to achieve and it scales in the appropriate way. But when we are making decisions between supervised and unsupervised learning models, we need to be a little bit more careful.

For supervised learning models, the stakeholders select the training data. It's then important for us to ensure that the set of stakeholders are a diverse set. Most often I see enterprises use data scientists to select the data. It's important to get entities from all the different business functions within your enterprise to ensure that the right input has been provided to select training data.

For unsupervised learning models where AI alone identifies bias, you need to leverage tools. These tools basically use fairness indicators and there are several tools out there from Google, we've got Google Toolkit, AI Fairness, 360 from IBM, OpenScale and what have you. It's important that you invest the effort in understanding the capabilities of these tools, how they can be used by your applications to ensure you have a bias free application.

The second one is creating a balanced AI team. What do I mean by balanced? Essentially what I'm meaning here is it should be a varied set of team members. They should be different racially from economic status, education levels, gender. We should also include innovators, the ones that are sponsoring the AI initiatives, the creators of the AI and the consumers of the AI. Having a varied AI team allows you to ensure that you have bias free decisions from ground up. The selection of data, the selection of algorithms is done across the entire team.

The third method that I want to talk about is data processing. The general notion that we usually have is, well, bias is all in the data. Once you have selected the proper data, which is bias free, you're good. But that's not the case. Data needs to be processed. There's pre processing that you do with your data, there's inline processing or in processing, and there's also post processing of data. You have to be mindful to ensure that while you have selected bias free data, bias is not creeping in in any of these stages of data processing.

The last one that I want to talk about is monitoring. One must understand that biases evolve over time. Do we think of EVs the way we think about them now. Was it the same kind of thinking that we had 20 years back? Probably not. We're more favorable towards EVs today than we were before, so it's important to continuously look at trends and real world data to ensure your AI systems are not stagnant and are moving along evolving along with real world data now.

I've also seen many companies employ third party assessment teams that would assess all of your enterprise applications for detecting bias. It's a great and optional way to ensure that your AI systems are built fairly, they are bias free and also endorsed by third party assessment.

ARTIFICIAL INTELLIGENCE, TECHNOLOGY, INNOVATION, BIAS IN AI, AI GOVERNANCE, MACHINE LEARNING BIAS, IBM TECHNOLOGY