The Trustworthiness of Generative AI in Real-Time Decision-Making

um hi everyone I hope you're are doing absolutely fine I am Meth and U final year B engineering student at auu Cambridge and um also the gdg academic ambassadors and my teammate my friend yesos we both will be uh going to uh explain you about trustworthiness of gen AI in decision making uh for iot device hi yes so the is why we are trying to put things in one thing so it's generate Ai and iot so generate AI is nothing just part of the AI artificial intelligence and iot is the inter internal things what it does is it will try to communicate with other devices just like smartw watches smartphones and autonomous Driving Systems so the thing is that these

devices will try to predict in real time if it's real time will it predict right and if it's right will it predict in real time um that's really the big deal because the trust because here the trust is not an optional anymore it's very critical just uh imagine that uh you are having a smart thermostatic at at home and that smart thermostatic is is deciding a summer temperature um in cold freezing winter now first first thing is that why are we trying to put this thing in one thing um right now we have so many iot devices but those are not automatic in my in my thinking those are semi-automatic to make those automatic I

think the genni is the main thing uh we can think that we we have a very personalized own Healthcare device which can predict our health condition based on the inputs and U we can also think about a coffee machine which is try which which will make prepare the coffee before we getting out of the bed yeah totally agree but here the system makes the decision in real time so uh let's take an example of an Healthcare how many of you would agree that should AI be doing accurately uh symptom diagonosis of the human or just imagine that you are in autonomous car and how would you guys trust to reach your destination through that autonomous car

in the worst condition weather the key thing is the trust because if trust erods I think a very big very best device won't be selling in the market so trust trust is not easily achievable so I think trust is the main thing so now we will be talking about the key Challen challenging in the trusting of jni the first we will talk about the reliability can the system consistently do the things what he's supposed to the meant for even in the unpredictable situations um the next key thing is the accuracy um s Genia highly depend on the input data so if the input data is having noise or biasness it might be predicting wrong things suppose a

very a patient having a very bad health condition and its health is being predicted wrong it might be Nightmare and the finally there comes the transparency for many of the people AI here is a black box and the user actually thinks why the system is making the decision on its own um let's say for example how many uh I know many of people be uh will be using the smartphones having the Google and the Siri so uh have you guys noticed that without saying anything does this kind of system do activate M doesn't um then it comes to the technicality but it's not just limited to technicality it's now human thing if people not knowing what is going behind

it people are not going to buy buying just like uh we can say that we are not going to spend money on the on a random thing and and here in ethical uh issue the Privacy comes the main um I I would never share my datas the personal one with the with the AI because this kind of an AI systems actually requires the large number of datas and sometimes this datas will be the personal two I would never share my one with an AI because if you train an Healthcare model or an um any autonomous car or if we talk about the financial situations we require the realtime data which can actually predict the correct

things now let's think about ethics so suppose uh an AI device is trying to make a decision between between two emergency patients and it is it is predicting but we can't say that it has ethics how can we put ethics in it so we still need some sort of uh Clarity between those and uh then it comes to the good news people are trying to improve these things via user responses and U then it comes to the me yeah um looking ahead to an gen uh we we came across an explainable AI which is also known as an X AI what it actually does is that making an AI system in such a way that the human can easily understand

the things just imagine uh your your your your smart fridge actually orders is explaining you why does why he has ordered an oat milk instead of an almond milk and then there is come there it comes to the ai ai is nothing just a device running running and generate UI on the on itself so it will be taking input and predicting the output on the same device your data won't be traveling to the cloud so far and on top of all that cannot forget about the data privacy it is actually the main thing uh here when we talk about the and here when we talk about the Privacy here the developer have to construct or design a

system in such a way that the system should function uh at its peak and also respecting the ethics and the datas of the user but these things comes with the challenges so people will be going to rely highly on these things and people people need to make a difference between what is trustworthy and what is possible and now let's wrap up these things uh people are trying to people are highly trying trying to achieve these things and there is a bridge they are trying to make there they are trying to make a bridge between the trust and these devices so that it can be worked out in the future and exactly uh if we focus on

building and J gen AI model we have to make sure about the privacy and the rules and regulations thank

[Applause] you guys if you have any questions just feel free to yeah any questions stick your hand up thank you bu by the way

thank you both for the uh for the interesting discussion um I'm just curious if if you had any thoughts around um managing latency uh and also the fact that you you talked a lot about um Health Data um one of the challenges today I think is that there's no commonly agreed standard for how Health Data is is gathered uh and and and distributed um and then finally did did you consider any elements of um Federated learning or tokenized uh training data um in your I think I think I can say about the latency the thing where we need latency uh very very low very low latency we need to make those devices with the generative

AI those those data should not be traveling to the cloud instead it should be directly predicting it yep um and for the fertility learning I would say like I have worked on one of the projects what where it actually does is that uh it gathers the data on the same device itself like it's not sharing the data to anyone outside uh my main motive was to train a model locally instead of sharing the datas to the further one like instead of sharing the datas let's train uh the AI in local system only so uh my main concern was the data privacy there so I have trained the data in such a way that it acts

accordingly to the systems we've got probably time for one more question if anyone has one cool one at the back yep yeah so talking about Healthcare and AI do you think there is a legitimate use or need for AI in healthcare um um I think um just yesterday I submitted one project which is working on RR intervals so how it does is it is trying to detect actual fibrillation via ECG s ECG data and from the ECG other inters are be filtered so those conditions can't be we we we can't be actively monitor we we we need to be in the hospital to detect those things so in the real time we can make smart watches or something

like that to implement it and um also for the healthare uh healthare I have worked on two of the projects um where the first one was I have find to in the pal Gemini model where basically what it does is that I take the data from the kle about the lungs and everything and um once the model is been trained uh if you push the datas and it will say that this part of lungs is actually affected so I would say um just fine tuning this kind of things can help in um NHS and also I've uh fine tune a gamma 2 about the chat boards uh if you say that I have this kind of symptoms you just have to enter

the symptoms and press enter it will suggest you like these are possible cases with you so consult your GP than So yeah thank you again oh there's one question one more yeah super quick one thank you uh that was a very good presentation especially when you touch based on Healthcare devices my question is uh these generative AI models risk uh being at use to breach the privacy of healthcare data generated by the devices and system systems um for those if if it's a very severe thing we need to pred it in the device or otherwise we need to make some standards of the security that's all we need thank you thank you everyone thank you again

by another big round of applause