Learning Library

← Back to Library

The Five Pillars of Trustworthy AI

Key Points

  • AI chatbots can produce hazardous misinformation, exemplified by a model that falsely recommended a toxic “aromatic water” recipe mixing ammonia and bleach.
  • IBM proposes five pillars for trustworthy AI, beginning with **Explainability**, where the system’s reasoning must be clear enough for domain experts to understand and validate without needing AI expertise.
  • The second pillar, **Fairness**, requires AI to avoid bias by training on diverse data sets—such as inclusive object‑ and facial‑recognition datasets—to ensure equitable performance across all groups.
  • **Transparency** demands that AI systems are not opaque “black boxes”; users must be able to inspect and verify the underlying algorithms and decision processes before trusting the outcomes.

Full Transcript

# The Five Pillars of Trustworthy AI **Source:** [https://www.youtube.com/watch?v=nB_EjxoP-6w](https://www.youtube.com/watch?v=nB_EjxoP-6w) **Duration:** 00:05:29 ## Summary - AI chatbots can produce hazardous misinformation, exemplified by a model that falsely recommended a toxic “aromatic water” recipe mixing ammonia and bleach. - IBM proposes five pillars for trustworthy AI, beginning with **Explainability**, where the system’s reasoning must be clear enough for domain experts to understand and validate without needing AI expertise. - The second pillar, **Fairness**, requires AI to avoid bias by training on diverse data sets—such as inclusive object‑ and facial‑recognition datasets—to ensure equitable performance across all groups. - **Transparency** demands that AI systems are not opaque “black boxes”; users must be able to inspect and verify the underlying algorithms and decision processes before trusting the outcomes. ## Sections - [00:00:00](https://www.youtube.com/watch?v=nB_EjxoP-6w&t=0s) **Trustworthy AI: Explainability Explained** - The speaker highlights hazardous AI hallucinations, references IBM’s five trust pillars, and illustrates the importance of explainability through a symptom‑diagnosis scenario. - [00:03:02](https://www.youtube.com/watch?v=nB_EjxoP-6w&t=182s) **Transparency, Robustness, and Privacy in AI** - The speaker outlines key trustworthy‑AI pillars—providing a clear view into algorithms, models, and training data, ensuring systems can resist attacks and data poisoning, and protecting user information from unwanted disclosure. ## Full Transcript
0:00About a year ago, I did a video where I suggested 0:03that a chatbot might hallucinate or be poisoned into giving a recommendation 0:09that you make a common household cleaning solution out of ammonia and bleach. 0:14Well, that was a hypothetical. 0:16It turns out it's true. 0:18In fact, there was an AI chatbot that came out some months later 0:23and recommended a recipe for an aromatic water mix. 0:26Now, that sounds delicious. Who wouldn't want a tall glass of that? 0:30Well, it turns out the ingredients ammonia and bleach. 0:33Those are toxic. 0:35Don't mix those together and definitely don't drink it. 0:38So that AI is not one you can believe in. 0:41What we want is a trustworthy AI, 0:44and IBM came out with five pillars, or principles, of trustworthy AI. 0:50These are the things that we want to expect from an A.I.. 0:53And let's take a look at what they are. 0:55The first one is Explainability. 0:59We want the AI to be able to explain itself and 1:04be understandable by someone who is an expert in that particular domain. 1:09So let's take an example of maybe I go to a chatbot and I give it the following symptoms. 1:15I have red itchy eyes, I have a runny nose, I'm sneezing. 1:21Okay, what would you think from that? 1:22A doctor who is a domain expert in that 1:25is probably going to say you've got an allergy or something along those lines. 1:30What they're not going to  say is you have a broken leg. 1:33That would be an example of an unexplainable AI. 1:36The explainable one, the expert in  that domain can look at it and say, 1:41"Yeah, I can see how you would come up with those things and come up with that particular diagnosis." 1:46It makes sense. 1:47And notice that domain expert doesn't have to understand anything about the way A.I. works. 1:52They don't have to be a technology expert. 1:54They're an expert in that domain of knowledge. 1:58Okay. Let's take a look at the second pillar. 2:00The second is about fairness. 2:02That is, the AI should not be biased toward 2:06or against any particular population or any particular group. 2:10Let's take an example. 2:12Let's say we have an object recognition system 2:14that's based on AI, and it's been trained on a whole bunch of different squares. 2:19So it recognizes those. 2:21However, what happens when I give it some stuff like this? 2:25It really can't recognize those very well because it hasn't seen enough of them. 2:29There's not enough of that in its training database. 2:32So what we need to do is make sure that it sees a diverse set of objects 2:37so that it can make the right recognition. 2:39Another example of this might be in facial recognition, 2:42where, again, we need to use diverse faces in order to make sure our AI is fair. 2:48Our third principle of trustworthy AI is transparency. 2:52And in transparency, what we're trying to get here 2:55is we don't want a black box,  a system that just says, "Trust me," 3:00because we don't know if we can trust it or not. 3:02We need to be able to verify. Then we can trust. 3:06So what I need is a transparent box, a box I can see into. 3:11And what would I see into it if it was an AI? 3:13I want to be able to see things like the algorithms that are used. 3:17I want to see the model that has been used. 3:20And I want to see the data that was used to train this thing. 3:23I want to know where the model came from. 3:25I want to know where the data came from. 3:27Those are the kinds of things that let me see in and give me more confidence 3:31that in fact, this thing, from a technical standpoint, 3:34is going to be something I can believe in. 3:36Our fourth principle of trustworthy AI is robustness. 3:40In this sense, what we mean is we want the system to be able to withstand attack. 3:46It should remain true to itself. 3:48It shouldn't be able to be compromised by outsiders who have malicious intent. 3:52So, for instance, if I have this really valuable data or model that's in the system, 3:59these are sort of the crown jewels and this is what the system runs on. 4:02I don't want to allow an attacker to be able to get to that. 4:07I need to be able to repel those attacks, make sure that they can't poison the data, 4:11make sure they can't steal the model, 4:13make sure that this system will continue to work. 4:16And as a cybersecurity guy, this is one of these principles that I'm most focused on. 4:22Our fifth principle, or pillar, of trustworthy AI is privacy. 4:28In this case, I want to make sure that what goes in the chatbot 4:33stays in the chat bot and it doesn't get shared with everyone else. 4:37So for instance, we don't want a case where your data is our business model. 4:42We want a case where your data is your data. 4:45We don't want the chatbot spying on you, 4:47or the information you put into it, going out and being shared with the rest of the world. 4:51So I want some sort of protection that says, 4:54"what I'm putting in, it's still my data. I don't want it shared with the whole wide world." 5:00So now you see the five pillars or principles of trustworthy AI, 5:05explainability, fairness, transparency, robustness, and privacy. 5:10These are the things that we should expect from vendors who are supplying us with AI. 5:15That way we can ensure that the AI serves us 5:19and not the other way around. 5:21Thanks for watching. 5:22If you found this video interesting and would like to learn more about cybersecurity, 5:25please remember to hit like and subscribe to this channel.