GDPAI Lect4-Lab Inclusive and Diverse AI Design | Python Visual Hands-on Tutorial

Indexed: 2026-05-21

[00:00:05]Hey come back student. I hope you are enjoying so far and as we have started the hands-on or visual lectures and some of you might be enjoying this visual lectures and and find some of you after the theory lectures this visual lectures might be giving extra value or making you uh overcome some of the confusion point or help you learn the concept in a better way.

[00:00:39]So now uh this particular is belong to the lecture four and I'm taking some simple simple example to explain the concept which has been mentioned in this uh lecture, lecture four.

[00:00:56]So what we are doing in this notebook is that uh by the way, this notebook will be available in the description of this uh YouTube video.

[00:01:06]You can download and then after downloading you can upload the notebook to your own Google Colab and you can run it and can see and play with it and uh learn better. Okay?

[00:01:22]So this notebook is basically to visually show that how this concept in practice look like.

[00:01:31]And so it this will take you through the inclusive AI design, what are the barriers to inclusion uh what are the practices of ethical AI development, some practical design strategy and some case studies style examples. But I will be using very simple simple example not a complex one.

[00:01:51]So simple three libraries importing NumPy, Pandas and matplotlib.

[00:01:58]And the first thing is that in this example I have taken a speech example that uh here a speech and vision kind of system which is we are testing across four different kind of group.

[00:02:12]Like for vision recognition or a speech recognition or So, in our example there are four groups are there, group A, B, C, and D.

[00:02:21]Now, purposefully it has been maintained that uh uh not equal data belong to different group. So, this design has been done so that I can explain the unequal uh and lack of inclusion impact.

[00:02:38]Okay?

[00:02:39]So, in this you can see that I have already mentioned that group A, B, C, and D accuracy pre pre made 94 91 82 76.

[00:02:52]And then users, number of data in the different training groups are 30 30 to 20 and 10. So, you can see purposefully lower amount of data is there in the group D while higher amount of data is in group A.

[00:03:09]In real life this kind of example could be true and if you visualize it we can see that uh this will how it look like the accuracy group A has higher accuracy while group D has a lower accuracy and B and C are lower than group A.

[00:03:25]Then what to see that model uh performance if you see uh model performance is going to show 88%.

[00:03:36]You see?

[00:03:44]You can see the model performance is 88%. So, good.

[00:03:50]Now, my model might show that overall good 88% accuracy, 89% accuracy, but how it is performing across different group or the groups uh which has less number of data.

[00:04:05]For example, in this example, group D had lower amount of data.

[00:04:10]And this data uh variation or data imbalance can happen in different stages, even in the data data collection stage and during the design stage.

[00:04:20]Now, question comes here that who is getting benefited?

[00:04:28]Our system is giving 88.8% accuracy, 89% accuracy, but uh who is which group is getting the best most benefit?

[00:04:38]And which group is getting left behind?

[00:04:42]Now, in this particular example, what we can also look into that when we were designing the system when we are testing the system, who were missing during that phase as well.

[00:04:57]Now, one major barrier for inclusion age uh biasness in the data as we have designed this data like that or imbalance training data.

[00:05:08]So, this one is unbalanced training data. So, one of the example is unbalanced. Now, you may be wondering that why don't you get the data balanced only?

[00:05:17]I give you you all are using ChatGPT.

[00:05:20]And all of you are aware of ChatGPT or other kind of model like Rock, Claude, Gemini, Copilot.

[00:05:27]Now, those models, the base model GPT, which is a language model, has been trained on uh internet scrolling and any data which has been digitized.

[00:05:40]So, this is a example when the data set condition was in June 2023.

[00:05:47]What we see that among all the data 48 47% data in this LLM training large language model training data belong to English language.

[00:06:00]48% of the data.

[00:06:02]So as a result this model LLM models are more trained on uh English data.

[00:06:11]And while we see the Asian language now Asia has uh quite a large population excluding the Chinese language and Japanese language.

[00:06:22]Other Asian languages, there are 49 other Asian languages they represent only 3.8% among all the data's data set used for this GPT training. So where is one language 48% and 49 language only 3.8% of the data is there.

[00:06:43]When you go for the Middle East language 1.4% data out of four Middle East language.

[00:06:51]And African language is further poor condition.

[00:06:55]51 African and Oceanic and indigenous language but their representation in the LLM training data is all only 0.1% negligible.

[00:07:07]So this data set itself is biased now.

[00:07:12]One way you can say that oh it is biased, but the problem is that most of the developed country are from Europe or from English speaking country.

[00:07:25]So because of development the content and everything are in English language.

[00:07:30]Even I am teaching in English language.

[00:07:32]This lecture I'm not recording in Indian language like Hindi, Bangla which I can speak.

[00:07:38]I can understand couple of other Indian language. I may not be able to speak.

[00:07:43]English and Bangla I can speak.

[00:07:45]But the digital content of other languages are not available on the entire internet.

[00:07:51]So, when they were collecting the data, they only got English language and later on uh European's language and by other languages are low. So, data collection stage itself is biasness there. So, as a result outcome of the large language model is going to be poorer.

[00:08:14]Uh I was working with one language, which is called Sami language, and data was not present in this data set.

[00:08:23]I have to collect the data.

[00:08:25]And after putting over 2 months of effort, whatever digital content available on Sami language By the way, Sami has only 20,000 speakers in the world.

[00:08:37]And out of 20,000 people, it has 10 dialects. And when I collected the data, the data set size it has only uh few million tokens only.

[00:08:51]So, not a trillions of tokens. So, if we go by comparison, it will be fraction of 0.0001% of the this data set condition.

[00:09:03]So, biasness can be there. So, one of the barrier is the bias data.

[00:09:09]Other could be that uh team is homogeneous. So, when you are making this uh LLM data set collection or any kind of data set collection, you do not have the people from different diversity group.

[00:09:24]Like when I was also my team was also making the Sami language data set, my team who were working, none of us speaks Sami language.

[00:09:34]None of us knows Sami language. So, different intricacy of those that language we were not able to capture it, definitely.

[00:09:44]So, the data set I didn't release it public then. We did the research. We learn it how the foundational model can be trained for sign language.

[00:09:54]But, there are many challenges were there that we were not able to make the guardrails, etc. on sign language. So, we didn't release the model as well as we did release the data because the data set collection was uh the team was uh not having diversity.

[00:10:15]What other is that we try to uh right now since English language is there, so everyone is making the GPT creating the English language. That is one size.

[00:10:26]We don't care other language. English language is work with it. So, one size fits all design assumption.

[00:10:33]Lack of accessible testing environment.

[00:10:36]So, there are also not enough testing environment for including diversity.

[00:10:42]So, in this I have just uh given the uh criteria that risk score of bias training I have just given example. Don't take it uh literally.

[00:10:54]If you have a bias training data, the risk score is nine.

[00:10:58]Having homogeneous team, not having diversity, risk score is eight.

[00:11:04]We are making system one site for all or we are making a system for uh one UX design, user interface.

[00:11:15]And we didn't include it people with visually impaired or other kind of in impairedness.

[00:11:22]So, that assumption often we do. So, the risk of biasness is seven.

[00:11:28]And not been enough tested, that is uh six.

[00:11:32]Now, if we see visually, so we can see that the bias training data if is there the the bias the barrier to inclusive AI design is higher.

[00:11:46]This is example.

[00:11:47]Uh there is no guideline which has been there, but I just uh for uh explanation purpose.

[00:11:55]Now, these barriers are often uh reinforce one another. So, like lack of uh means with a bias training data can impact other score.

[00:12:07]Like a one-size-fits-all and then the non-inclusive uh not having enough testing.

[00:12:15]Now, narrow data may produce narrow system behavior. If you have a smaller data, like I have we were building Sami language. So, that language trained model we're giving not proper result like English language GPT model.

[00:12:29]So, because data is narrow, so the system behavior was narrow. So, this billions of uh parameter LLM model was not been trained with few million token.

[00:12:40]Uh that was one uh problem.

[00:12:43]Now, narrow testing can fail to catch different kind of exclusion. So, uh those could be there there and that can will come up when the model is deployed.

[00:12:55]So, I take a toy example where what I do is I take access accessibility is excluded. Like we take a standard interface user who are visually okay and users are screen reader and users with the speech accessibility needs.

[00:13:15]Right?

[00:13:19]Now, uh standard interface the What has happened that the task completion rate is 0.93 but screen reader user task completion rate was found 0.71 and people with speech issues their task completion rate of the system was found 0.68. So we see that performance are varying for different type of users.

[00:13:49]Now So to need to know that interface can be functional for many user while people with visual impairment or speech impairment this interface was not useful and that's why inclusivity in AI design is important so that we can consider the accessibility issue right from the beginning.

[00:14:16]Now Concept comes like ethical AI there's a compliance is one issue but Any other than compliance with regulatory people don't say anything. Do we overlook the other ethical aspect of AI system design?

[00:14:34]So compliance says what is allowed but ethics says what is right to do means what is moral to do what is right to do.

[00:14:43]So Let's say we take the example that Fairness principle Is the system disadvantaging some other groups?

[00:14:56]Score is uh zero.

[00:14:58]Transparency can important decision can be explained is there explainability is there in the system?

[00:15:05]Accountability Can we make people responsible when some harm happen which stage of the which group in this whole AI system development and usage were responsible for uh harm?

[00:15:22]Privacy is the private data, personal data is protected when you have designed the system. Uh are they getting identified?

[00:15:31]Uh Now, inclusivity have we included diverse communities uh while designing the system?

[00:15:40]So, now, if you take the scorecard for this hypothetical system, uh scorecard of this hypothetical system, then what we see that uh privacy uh has not been uh In the ethical review of the system, we see the privacy got eight score.

[00:16:04]While inclusivity got uh less than five score.

[00:16:09]And that shows that how poor this uh system in terms of including the inclusivity and in terms of explainability also, the transparency score shows five, so which is not so important point. So, if this system can be redesigned to include fairness, transparency, accountability, privacy, and inclusivity, so that they get a higher score, then that is the best uh better inclusive AI system design.

[00:16:38]Now, responsible AI just development is required multiple dimension of evaluation, not one kind of score. So, inclusive design versus ethical development. Uh inclusive design focusing on using the strong representation of different types of group and participation, uh including accessibility for different types of group like visually impaired or speech impaired, and real-world usability.

[00:17:05]While ethical development will require that are we making responsible AI? Are we making the transparent AI?

[00:17:13]Is our system outcome can be explained?

[00:17:16]Accountable system, privacy we have uh keeping we are have tried to reduce the harm to be which AI system can do if in the future.

[00:17:28]So, we see that the example is that if we have done the inclusive design emphasis, then we see the score for representation, participation, accessibility is higher.

[00:17:42]And if we have done the ethical development, then what we see that the score for transparency and accountability and then privacy is higher. So, ethical development and uh response and inclusive development, both are have a different meaning.

[00:18:02]So, we need to have uh responsible as well as ethically AI development.

[00:18:07]So, let's see let's see this example with an uh this concept with two an example like uh inclusions What is inclusion? Inclusion say that who is considered in the system and ethics says that how responsible uh how res- responsibility has been exercised while building the system.

[00:18:28]So, there are five practical strategy that can work to build an inclusive and ethical responsible system.

[00:18:35]First is make get uh better data.

[00:18:40]And after that, do audit the model for bias detection and correction.

[00:18:46]Keep human meaningfully involved while building the system and after the deployment also.

[00:18:53]Test the system with different kind of diverse users. And once it is deployed, monitor the system for real-world impact.

[00:19:03]So, uh as per the implementation priority, we can design that uh better data is 10 out of 10.

[00:19:12]Audit for bias is nine. Now, monitor after the impact after deployment is 10.

[00:19:18]So, these are the different priority can be put for implementation. And this way we can build a better system.

[00:19:26]Now, one small example is that assistive inclusion or historical bias in recruitment and broader linguistic inclusion.

[00:19:36]So, one case study like assistive AI for accessibility. And so, in this what we see that if we do the inclusion, the impact is going to be nine for assistive AI for accessibility for visually impaired or speech impaired or any other kind of impairment.

[00:19:53]The ethical quality of the system is eight. Now, if you are making a system AI trained for hiring because of bias history, the impact of inclusion is going to be quite low.

[00:20:10]Because the data itself is biased, so inclusivity will be very low. And similarly, if you are building the system with biased historical data, ethically, we are not building suitably high-score system.

[00:20:25]Now, we are making our speech AI with including many languages, then we are including other languages and that score is getting higher.

[00:20:37]And ethically, ethical score also is getting higher.

[00:20:42]So, we saw that data is what played a role in reducing the AI trained on historical data for hiring practices.

[00:20:54]So, in this three example, what we saw that we have included a design which improved the accessibility, that improved the score for assistive AI system.

[00:21:06]But we used the data which has historical bias that increases the So, what happened? The system learned the unfairness of the past and then final system is also uh having lower score in inclusivity and ethical AI development.

[00:21:24]So, what are the different checklists are there for responsible AI development? So, before the development, we need to ask a question that have we included any excluded any people while >> [snorts] >> collecting the data or designing the data system?

[00:21:42]During the design, we need to also see that uh whose needs uh are are shaping the decision of this development.

[00:21:51]And our errors and harms are distributed fairly across different groups or harms is only happening to one type type of group or two type of groups.

[00:22:02]After the deployment, are we monitoring and is the system is uh putting people accountable and revision has been maintained? Uh yeah. So, that can give you the responsible AI metrics. And so, if we see that responsible AI metrics for hypothetical this group wise metrics, we can see that for group D, the responsibility was quite poor.

[00:22:30]Uh while group A, it was higher. So, we may need to think of getting a better system design to improve it.

[00:22:39]So, the what we see is that inclusive AI design begins with representation, participation, and accessibility into account. While ethical AI development goes beyond the EU AI Act compliance.

[00:22:55]Uh we have to be uh including more barriers of uh inclusive AI design can come from not having good representation of data or not having diverse team or not developing interface for different kind of accessibility and not testing with a diverse group and diverse people. Uh that can lead to the barriers of inclusive AI system design.

[00:23:26]Responsible AI requires practical kind of strategy, not just okay, we are doing responsible AI.

[00:23:34]Uh good system on an average may still fail uh for important community. Means we no matter how much we do some something we may miss. So, we need to be constantly upgrading it.

[00:23:48]So, point is that the inclusive and ethical AI development is not achieved by accident. We have to plan from the beginning.

[00:23:58]Consciously and deliberate choice in data collection, designing accountability, and uh representing different kind of voices right from beginning the system.

[00:24:09]So, I hope you enjoyed this uh visual walkway of the lecture four and hopefully uh we will uh in the next lecture five, we will go through the uh visual walkway for lecture