Global AI and Data Science

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

 Bias can manifest in ways such as biased training data, biased decision-making processes, or a societal bias unintentionally entered an algorithm   A couple examples include Facial Recognition however there are no know instances in the past few years that I researched. In order to mitigate Continuous monitoring is needed, Developers need to be accountable for biased outcomes via transparency and accountability, use of diverse training data. If I am incorrect, please correct me.  Does anyone else have more feedback on this issue?

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

Yes sir. You are right on this -  "Continuous monitoring is needed, Developers need to be accountable for biased outcomes via transparency and accountability". 

 Bias can manifest in ways such as biased training data, biased decision-making processes, or a societal bias unintentionally entered an algorithm   A couple examples include Facial Recognition however there are no know instances in the past few years that I researched. In order to mitigate Continuous monitoring is needed, Developers need to be accountable for biased outcomes via transparency and accountability, use of diverse training data. If I am incorrect, please correct me.  Does anyone else have more feedback on this issue?

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

 Bias can manifest in ways such as biased training data, biased decision-making processes, or a societal bias unintentionally entered an algorithm   A couple examples include Facial Recognition however there are no know instances in the past few years that I researched. In order to mitigate Continuous monitoring is needed, Developers need to be accountable for biased outcomes via transparency and accountability, use of diverse training data. If I am incorrect, please correct me.  Does anyone else have more feedback on this issue?

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

 Bias can manifest in ways such as biased training data, biased decision-making processes, or a societal bias unintentionally entered an algorithm   A couple examples include Facial Recognition however there are no know instances in the past few years that I researched. In order to mitigate Continuous monitoring is needed, Developers need to be accountable for biased outcomes via transparency and accountability, use of diverse training data. If I am incorrect, please correct me.  Does anyone else have more feedback on this issue?

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

All AI is trained, thus the information used to train the AI Engine can influence the algorithms.

Asked the AI Engine of Facebook several questions about being Bias and received some interesting results. After several interactions the Facebook AI provided a response that seemed to indicate that Bias was inevitable due to the "Approved" information sources and those responsible for "Training" the AI Engines!      

With all the Marketing Banter on AI and every ISV selling solutions rushing to provide AI enhancements. What will happen if in the rush human mistakes are made that taint the results?

AI has the promise of being a great tool when used correctly. 

We all need to - Verify before Trusting the results!

Will be interesting to see where AI ends up...  

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

That's true.  I agree on the biases that you have pointed out. It's not just within Social media like Facebook, but even with the prompt text that we enter on the chat. The way you write the prompts, the text you type in, for example: an image generation with certain text prompt has a significant impact on the output. Not all fellow beings in the human world are artistically descriptive or creative  with rich vocabulary can provide to get the right result that was desired. So it is a challenge regardless. Biases will exist and here to stay.  AI needs to train well to understand you as an individual.  It takes time and patience for that, I guess. 

All AI is trained, thus the information used to train the AI Engine can influence the algorithms.

Asked the AI Engine of Facebook several questions about being Bias and received some interesting results. After several interactions the Facebook AI provided a response that seemed to indicate that Bias was inevitable due to the "Approved" information sources and those responsible for "Training" the AI Engines!      

With all the Marketing Banter on AI and every ISV selling solutions rushing to provide AI enhancements. What will happen if in the rush human mistakes are made that taint the results?

AI has the promise of being a great tool when used correctly. 

We all need to - Verify before Trusting the results!

Will be interesting to see where AI ends up...  

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

Thanks for raising this important issue, Ram. Bias is particularly pronounced in training data. One place is training data on gender. Our research shows many such instances for example recognition algorithms have an error rate of 34.7% for darker-skinned women, compared to less than 1% for lighter-skinned men. Simialalry, AI recruitment tools were found to downgrade resumes with the word "woman," such as "women\u2019s soccer team captain," leading to biased hiring decisions.

IBM offers several tools and platforms to address bias in AI models, including Instruct Labs, AI Fairness 360 (AIF360), and other Watson solutions. These tools focus on improving transparency, fairness, and accountability in machine learning systems.To address bias using IBM tools, we could start with Instruct Labs for hands-on training on fairness techniques, use AI Fairness 360 (AIF360) to detect and mitigate bias with fairness metrics and algorithms, and leverage Watson Studio for automated bias checking, explainable AI, and ongoing monitoring of deployed models. Together, these tools ensure transparent, fair, and accountable AI systems.

That's true.  I agree on the biases that you have pointed out. It's not just within Social media like Facebook, but even with the prompt text that we enter on the chat. The way you write the prompts, the text you type in, for example: an image generation with certain text prompt has a significant impact on the output. Not all fellow beings in the human world are artistically descriptive or creative  with rich vocabulary can provide to get the right result that was desired. So it is a challenge regardless. Biases will exist and here to stay.  AI needs to train well to understand you as an individual.  It takes time and patience for that, I guess. 

All AI is trained, thus the information used to train the AI Engine can influence the algorithms.

Asked the AI Engine of Facebook several questions about being Bias and received some interesting results. After several interactions the Facebook AI provided a response that seemed to indicate that Bias was inevitable due to the "Approved" information sources and those responsible for "Training" the AI Engines!      

With all the Marketing Banter on AI and every ISV selling solutions rushing to provide AI enhancements. What will happen if in the rush human mistakes are made that taint the results?

AI has the promise of being a great tool when used correctly. 

We all need to - Verify before Trusting the results!

Will be interesting to see where AI ends up...  

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

Thank you Noor that was very informative and useful.  The error rates you mentioned are not acceptable. I am sure the IBM solutions you mentioned can help tackle bias. The likes of IBM's AI Fairness 360 and Google's What-If Tool should and will help developers uncover hidden biases. I feel, it is also the representation. Datasets should represent diverse demographics and scenarios. It is a comprehensive effort. Inclusion and ethics in decision making is crucial. As you rightly pointed out, AI needs focus on improving transparency, fairness, and accountability in machine learning systems.   One unbiased scenario I can think of is when the decision world revolves around numbers, typically in  Finance- involving personal finance and investment decisions. We can build to a certain degree prediction levels on the future trajectory objectively.      

 I will keep this post open in case somebody else wants to jump in with their 2 cents. 

Thanks for raising this important issue, Ram. Bias is particularly pronounced in training data. One place is training data on gender. Our research shows many such instances for example recognition algorithms have an error rate of 34.7% for darker-skinned women, compared to less than 1% for lighter-skinned men. Simialalry, AI recruitment tools were found to downgrade resumes with the word "woman," such as "women\u2019s soccer team captain," leading to biased hiring decisions.

IBM offers several tools and platforms to address bias in AI models, including Instruct Labs, AI Fairness 360 (AIF360), and other Watson solutions. These tools focus on improving transparency, fairness, and accountability in machine learning systems.To address bias using IBM tools, we could start with Instruct Labs for hands-on training on fairness techniques, use AI Fairness 360 (AIF360) to detect and mitigate bias with fairness metrics and algorithms, and leverage Watson Studio for automated bias checking, explainable AI, and ongoing monitoring of deployed models. Together, these tools ensure transparent, fair, and accountable AI systems.

That's true.  I agree on the biases that you have pointed out. It's not just within Social media like Facebook, but even with the prompt text that we enter on the chat. The way you write the prompts, the text you type in, for example: an image generation with certain text prompt has a significant impact on the output. Not all fellow beings in the human world are artistically descriptive or creative  with rich vocabulary can provide to get the right result that was desired. So it is a challenge regardless. Biases will exist and here to stay.  AI needs to train well to understand you as an individual.  It takes time and patience for that, I guess. 

All AI is trained, thus the information used to train the AI Engine can influence the algorithms.

Asked the AI Engine of Facebook several questions about being Bias and received some interesting results. After several interactions the Facebook AI provided a response that seemed to indicate that Bias was inevitable due to the "Approved" information sources and those responsible for "Training" the AI Engines!      

With all the Marketing Banter on AI and every ISV selling solutions rushing to provide AI enhancements. What will happen if in the rush human mistakes are made that taint the results?

AI has the promise of being a great tool when used correctly. 

We all need to - Verify before Trusting the results!

Will be interesting to see where AI ends up...  

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

I am involved in creating AI product quality guidelines in Japan, and I agree that bias in AI is a critical issue.
https://www.qa4ai.jp/

At the same time, I came across a research paper discussing how AI was used to teach Go to students, which resulted in eliminating the performance gap between male and female players. This makes me wonder if there are cases where, even with inherent biases, AI might perform better than humans in certain contexts.
Can Artificial Intelligence Improve Gender Equality? Evidence from a Natural Experiment

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

I am involved in creating AI product quality guidelines in Japan, and I agree that bias in AI is a critical issue.
https://www.qa4ai.jp/

At the same time, I came across a research paper discussing how AI was used to teach Go to students, which resulted in eliminating the performance gap between male and female players. This makes me wonder if there are cases where, even with inherent biases, AI might perform better than humans in certain contexts.
Can Artificial Intelligence Improve Gender Equality? Evidence from a Natural Experiment

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

I am involved in creating AI product quality guidelines in Japan, and I agree that bias in AI is a critical issue.
https://www.qa4ai.jp/

At the same time, I came across a research paper discussing how AI was used to teach Go to students, which resulted in eliminating the performance gap between male and female players. This makes me wonder if there are cases where, even with inherent biases, AI might perform better than humans in certain contexts.
Can Artificial Intelligence Improve Gender Equality? Evidence from a Natural Experiment

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

Totally agree - there are also greater issues we have to collectively consider as humanity 

https://community.ibm.com/HigherLogic/System/DownloadDocumentFile.ashx?DocumentFileKey=5d671f4d-beed-170e-a890-a0f4b417c9ca&forceDialog=0

a short note I wrote/published yesterday

I am involved in creating AI product quality guidelines in Japan, and I agree that bias in AI is a critical issue.
https://www.qa4ai.jp/

At the same time, I came across a research paper discussing how AI was used to teach Go to students, which resulted in eliminating the performance gap between male and female players. This makes me wonder if there are cases where, even with inherent biases, AI might perform better than humans in certain contexts.
Can Artificial Intelligence Improve Gender Equality? Evidence from a Natural Experiment

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

Thank you for sharing such an excellent report! The insights in Proactive AI Governance are incredibly thorough and relevant.

Your analysis and recommendations are proving to be highly valuable as we work on developing our group's AI governance guidelines. It's great to see such a well-rounded perspective on tackling the risks and complexities of generative AI.

Looking forward to seeing more of your work in the future. Thanks again for this amazing contribution!

Totally agree - there are also greater issues we have to collectively consider as humanity 

https://community.ibm.com/HigherLogic/System/DownloadDocumentFile.ashx?DocumentFileKey=5d671f4d-beed-170e-a890-a0f4b417c9ca&forceDialog=0

a short note I wrote/published yesterday

I am involved in creating AI product quality guidelines in Japan, and I agree that bias in AI is a critical issue.
https://www.qa4ai.jp/

At the same time, I came across a research paper discussing how AI was used to teach Go to students, which resulted in eliminating the performance gap between male and female players. This makes me wonder if there are cases where, even with inherent biases, AI might perform better than humans in certain contexts.
Can Artificial Intelligence Improve Gender Equality? Evidence from a Natural Experiment

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

Thank you for sharing such an excellent report! The insights in Proactive AI Governance are incredibly thorough and relevant.

Your analysis and recommendations are proving to be highly valuable as we work on developing our group's AI governance guidelines. It's great to see such a well-rounded perspective on tackling the risks and complexities of generative AI.

Looking forward to seeing more of your work in the future. Thanks again for this amazing contribution!

Totally agree - there are also greater issues we have to collectively consider as humanity 

https://community.ibm.com/HigherLogic/System/DownloadDocumentFile.ashx?DocumentFileKey=5d671f4d-beed-170e-a890-a0f4b417c9ca&forceDialog=0

a short note I wrote/published yesterday

I am involved in creating AI product quality guidelines in Japan, and I agree that bias in AI is a critical issue.
https://www.qa4ai.jp/

At the same time, I came across a research paper discussing how AI was used to teach Go to students, which resulted in eliminating the performance gap between male and female players. This makes me wonder if there are cases where, even with inherent biases, AI might perform better than humans in certain contexts.
Can Artificial Intelligence Improve Gender Equality? Evidence from a Natural Experiment

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

I am involved in creating AI product quality guidelines in Japan, and I agree that bias in AI is a critical issue.
https://www.qa4ai.jp/

At the same time, I came across a research paper discussing how AI was used to teach Go to students, which resulted in eliminating the performance gap between male and female players. This makes me wonder if there are cases where, even with inherent biases, AI might perform better than humans in certain contexts.
Can Artificial Intelligence Improve Gender Equality? Evidence from a Natural Experiment

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

Thank you for your comment. I couldn't agree more. In engineering-especially in AI-diversity is essential for spotting any anomalies or biases in data, generated outputs, and classification results.

I am involved in creating AI product quality guidelines in Japan, and I agree that bias in AI is a critical issue.
https://www.qa4ai.jp/

At the same time, I came across a research paper discussing how AI was used to teach Go to students, which resulted in eliminating the performance gap between male and female players. This makes me wonder if there are cases where, even with inherent biases, AI might perform better than humans in certain contexts.
Can Artificial Intelligence Improve Gender Equality? Evidence from a Natural Experiment

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

Isn't every model based on documents written by humans? So yes we can filter on irregulaties. But how do we know that the content of the document is with true or reliable facts?  A clear screening of each document which is embedded in the model should be necessary, but is that possible? 

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

Investigation is critical.

Isn't every model based on documents written by humans? So yes we can filter on irregulaties. But how do we know that the content of the document is with true or reliable facts?  A clear screening of each document which is embedded in the model should be necessary, but is that possible? 

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

It's entirely possible for training data or documents to be both true and biased.

Many organizations that generated data gathered of the last few decades have been operating with social or institutional biases in place, so the bias ends up embedded into the data.

Question is, how do you remove the bias from the data without also removing its value as training data?

The point of training it to get the AI to make fair decisions based only on legitimate grounds.  But how many elements in the data are truly unbiased?  Names? Culturally biased. Addresses? Often culturally and/or socio-economically biased? Gender? Maybe, although non-conforming values will be scarce (and thus probably biased)  Income? Socio-economically biased. Strip all of that out and you're left with 'someone asked or a loan', which means you might as well make the decision at random...

------------------------------
Mik Clarke

Original Message:
Sent: Tue January 14, 2025 07:14 PM
From: J A Hansen Esq IV
Subject: Is it even possible to build AI models without human bias.

Investigation is critical.

------------------------------
J A Hansen Esq IV

Original Message:
Sent: Mon January 13, 2025 04:48 AM
From: Oscar Dubbeldam
Subject: Is it even possible to build AI models without human bias.

Isn't every model based on documents written by humans? So yes we can filter on irregulaties. But how do we know that the content of the document is with true or reliable facts?  A clear screening of each document which is embedded in the model should be necessary, but is that possible? 

------------------------------
Oscar Dubbeldam
CEO / CTO
Migrato
Strijen
+31650734796

Original Message:
Sent: Thu December 26, 2024 11:33 AM
From: Ramkumar Yaragarla
Subject: Is it even possible to build AI models without human bias.

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

------------------------------
Ramkumar Yaragarla
------------------------------

Fantastic observation Mik - which leads us to the core of any "predictive" AI system.  The AI system which is used to predict things (like loan applicability) is SUPPOSED to be biased.  It's supposed to choose or predict which group a particular individual will fall into.  The key is in eliminating biases that do not (or should not) impact the prediction, and only considering the things that DO bias a prediction.  These are the factors that most heavily influence a particular prediction.  In your AI model predicting loan payment satisfaction, there is a bias towards individuals with larger incomes, since those people tend to have more available resources to pay off a loan.  A Data Scientist may call this a feature of the model, and they will look for a combination of features and weights that will combine to make the model more accurate - more biased - towards choosing individuals who can pay back a loan.

So be careful when discussing bias in AI models.  Some of it may be inherent in the data, some of it may come from organizations or societal institutions, and some of it just "human" bias (I like people who read my blog posts and articles....).  The point here is that you should worry about, and always consider, bias in your data and models.  You just shouldn't follow that path to it's very vanilla and un-useful end.  You want fairness and a lack of discrimination - what I might call a "known set of acceptable biases".

Watsonx environments will give you the tools and allow you to track this "known set of acceptable biases" over time, and will allow you to track and detect things like model drift.  It's important to do this - AI will never be perfect, but we can make it more transparent and continually seek to improve what it does for us.

------------------------------
Daniel Toczala
Community Leader and Customer Success Manager - Watson
dtoczala@us.ibm.com

Original Message:
Sent: Wed January 15, 2025 12:54 AM
From: Mik Clarke
Subject: Is it even possible to build AI models without human bias.

It's entirely possible for training data or documents to be both true and biased.

Many organizations that generated data gathered of the last few decades have been operating with social or institutional biases in place, so the bias ends up embedded into the data.

Question is, how do you remove the bias from the data without also removing its value as training data?

The point of training it to get the AI to make fair decisions based only on legitimate grounds.  But how many elements in the data are truly unbiased?  Names? Culturally biased. Addresses? Often culturally and/or socio-economically biased? Gender? Maybe, although non-conforming values will be scarce (and thus probably biased)  Income? Socio-economically biased. Strip all of that out and you're left with 'someone asked or a loan', which means you might as well make the decision at random...

------------------------------
Mik Clarke

Original Message:
Sent: Tue January 14, 2025 07:14 PM
From: J A Hansen Esq IV
Subject: Is it even possible to build AI models without human bias.

Investigation is critical.

------------------------------
J A Hansen Esq IV

Original Message:
Sent: Mon January 13, 2025 04:48 AM
From: Oscar Dubbeldam
Subject: Is it even possible to build AI models without human bias.

Isn't every model based on documents written by humans? So yes we can filter on irregulaties. But how do we know that the content of the document is with true or reliable facts?  A clear screening of each document which is embedded in the model should be necessary, but is that possible? 

------------------------------
Oscar Dubbeldam
CEO / CTO
Migrato
Strijen
+31650734796

Original Message:
Sent: Thu December 26, 2024 11:33 AM
From: Ramkumar Yaragarla
Subject: Is it even possible to build AI models without human bias.

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

------------------------------
Ramkumar Yaragarla
------------------------------

It's entirely possible for training data or documents to be both true and biased.

Many organizations that generated data gathered of the last few decades have been operating with social or institutional biases in place, so the bias ends up embedded into the data.

Question is, how do you remove the bias from the data without also removing its value as training data?

The point of training it to get the AI to make fair decisions based only on legitimate grounds.  But how many elements in the data are truly unbiased?  Names? Culturally biased. Addresses? Often culturally and/or socio-economically biased? Gender? Maybe, although non-conforming values will be scarce (and thus probably biased)  Income? Socio-economically biased. Strip all of that out and you're left with 'someone asked or a loan', which means you might as well make the decision at random...

Investigation is critical.

Isn't every model based on documents written by humans? So yes we can filter on irregulaties. But how do we know that the content of the document is with true or reliable facts?  A clear screening of each document which is embedded in the model should be necessary, but is that possible? 

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

Isn't every model based on documents written by humans? So yes we can filter on irregulaties. But how do we know that the content of the document is with true or reliable facts?  A clear screening of each document which is embedded in the model should be necessary, but is that possible? 

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

A genAI LLM is only an inference model - and every language model is trained on documents written by humans, and fundamentally as AI models may also be able to generate coherent text, they don't inherently distinguish between true and false information we might have systems that are started by training on  large-scale datasets like Wikipedia for training, which have inconsistent data governance, can lead to the propagation of inaccuracies or biases.  I have been to many of the meetings with large dataset organizations and their data governance is atrocious so we might be amplifying instead.

Isn't every model based on documents written by humans? So yes we can filter on irregulaties. But how do we know that the content of the document is with true or reliable facts?  A clear screening of each document which is embedded in the model should be necessary, but is that possible? 

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge. 

There is a possibility that societal biases unintentionally seep into the learning algorithms. I do not have real world examples to share right now. Bias is definitely in a big challenge.