Learning Library

← Back to Library

Five Steps to Trusted AI

Key Points

  • The speaker likens building trustworthy AI to a home renovation, emphasizing that both require a careful, step‑by‑step process before the final product can be relied upon.
  • Three major risks of generative AI are highlighted: legal exposure from evolving regulations, damage to brand reputation from mishandled outputs, and operational hazards such as leaking PII or trade secrets.
  • To create trusted AI, the first principle is “know your scope” – clearly define what the model is allowed to do and set guardrails that route out‑of‑scope requests (e.g., pricing queries) to human agents.
  • The second principle is “know your foundation” – understand the underlying data, infrastructure, and constraints of the system, just as a renovator must be familiar with a house’s pipes and wiring before beginning work.

Full Transcript

# Five Steps to Trusted AI **Source:** [https://www.youtube.com/watch?v=mfxgfU5Abdk](https://www.youtube.com/watch?v=mfxgfU5Abdk) **Duration:** 00:09:47 ## Summary - The speaker likens building trustworthy AI to a home renovation, emphasizing that both require a careful, step‑by‑step process before the final product can be relied upon. - Three major risks of generative AI are highlighted: legal exposure from evolving regulations, damage to brand reputation from mishandled outputs, and operational hazards such as leaking PII or trade secrets. - To create trusted AI, the first principle is “know your scope” – clearly define what the model is allowed to do and set guardrails that route out‑of‑scope requests (e.g., pricing queries) to human agents. - The second principle is “know your foundation” – understand the underlying data, infrastructure, and constraints of the system, just as a renovator must be familiar with a house’s pipes and wiring before beginning work. ## Sections - [00:00:00](https://www.youtube.com/watch?v=mfxgfU5Abdk&t=0s) **Trusted AI Like Home Renovation** - The speaker compares establishing trustworthy AI models to a kitchen remodel, highlighting legal, reputational, and operational risks before outlining five steps to secure both generative and traditional AI. ## Full Transcript
0:00AI is everywhere but how can 0:06we trust the models that are there for 0:09us to 0:11consume it's similar to a home 0:19renovation I'm actually renovating my 0:23kitchen and I'm so excited to get brand 0:28new countertops and allwhite cabinets 0:33stainless steel 0:34appliances as well as brand new 0:38lighting but I might want these things 0:41tomorrow but I need to follow a process 0:44to make sure that that kitchen is 0:47delivered to me safely and I can trust 0:50it for years to come it's a little bit 0:53like generative AI models we might want 0:57them tomorrow but we need to take the 0:59step steps to make sure they're trusted 1:01and secure in this video I'll cover five 1:05ways to build trusted AI both generative 1:09and traditional models but first let's 1:12talk about what could go wrong let's 1:15talk about three 1:19risks just like in my home renovation 1:22process there are many risks that can 1:24occur everything from making sure the 1:27people doing the work have legal 1:28Protections in case something happens on 1:31the job to redoing a floor and not 1:35completing the right process or steps 1:38and more money has to be spent to fix it 1:41so let's cover three risks for 1:43generative AI models the first risk is 1:49legal there are a growing list of legal 1:54implications for using AI improperly or 1:59not following all the steps needed for 2:02organizations to use a model there's a 2:04number of growing regulations like the 2:06EU act the New York hiring bias law as 2:10well as the executive order from the 2:12White House on generative Ai and the 2:15number will only grow over 2:18time next we have 2:21reputation 2:25risks that's what you want on your score 2:28everything from 2:30your brand Matters from your reputation 2:34there's an instance of a large 2:36organization deploying a generative AI 2:38chatbot gave a very high value item for 2:42only a fraction of its original cost 2:47finally we have 2:49operational 2:51risks these risks can result in immense 2:55fines or loss of productivity for a 2:58company this could be everything from 3:00regurgitating pii information 3:03unintentionally or exposing Trade 3:06Secrets now that we understand the risks 3:09at stake let's talk about how to 3:12build trust in our AI models here are 3:17five simple 3:19tips 3:22first is know your scope just like in my 3:26home renovation process I want to Define 3:30specifics on what I'm going to be 3:32working on and what contractors can and 3:35can't do I just want them to focus on 3:38the kitchen just like in my home 3:40renovation process I'm going to Define 3:43my AI model scope by setting guard rails 3:48around that scope I'm going to say 3:50exactly what the model can do and even 3:54more 3:55importantly what it 3:58can't a good example of this is with 4:02chatbots if I create an AI chatbot for 4:05an organization I might not want the 4:07chatbot to answer any questions related 4:10to pricing pricing is outside of the 4:13guard rails in this case I'm going to 4:15send all of those questions outside of 4:18the generative AI model and straight to 4:20an agent second we have 4:25the 4:27foundation know your 4:30Foundation I know all the details about 4:34my house before I get started in the 4:37project right I know the types of pipes 4:40I might have as much as possible about 4:43my house as I can so I don't run into 4:47risks I didn't see coming I want the 4:51same thing for my model I want to 4:55understand the data used to build the 4:58model what it was recommended to be used 5:01or not used for as well as what type of 5:06model it 5:07is open or closed as well as the model 5:12architecture one way to do that is 5:15through 5:16model 5:19cards model cards show all the details 5:23about a Model A large language model for 5:25you to use everything from the data that 5:27was used to build the model from the 5:30architecture about the model how 5:32training was done on the model and how 5:35it can or can't be used so this gives 5:38you the foundation you need to get 5:40started and know the model you select 5:43for 5:44use third knowing and setting your life 5:49cycle 5:52governance in my home and with the help 5:56of a contractor I'm going to document 5:58the entire 6:00home improvement process so that I know 6:03the different steps and safeguards it 6:05takes to move from one stage to the next 6:08as well as who's doing the work the same 6:11thing for my model I want to set 6:13up and document a specific process so 6:18that I know all the steps that are being 6:20used to build the model I know versions 6:23of the model that are being used I know 6:26who's making model changes I know which 6:28version is going going to production and 6:30that should include everything from any 6:33training data that's being used to the 6:35different types of prompts that I'm 6:37using to build my model as well as test 6:41data right that's verifying that I want 6:43to move to the next 6:44step next we have our fourth step which 6:49is 6:50monitoring 6:53risk throughout the home renovation 6:56process I'll need to monitor the risk RS 6:59every step of the way I need to check in 7:02and make sure that the home is still 7:06stable and stand and monitor that 7:09nothing is going to go wrong with my 7:11structure throughout the process and 7:13that might include several tests and 7:16tracking of that 7:17information the same thing for my model 7:21I'm going to want 7:23to track steps over time and metrics 7:27specific to 7:30bias and 7:33hallucination to make sure that the 7:35model is operating in a way that I'd 7:38like both in production as well as 7:40throughout the testing process and this 7:43will ensure that when an issue does 7:46arise I'm able to quickly react and take 7:50action on it even better if you can find 7:53a way to 7:55automate the process so that this can be 7:58done seamlessly and you can move on to 8:00other tasks and be alerted if there are 8:02any issues at all 8:06finally we cannot leave out 8:11compliance it's important to know in my 8:14home renovation process am I up to code 8:17right what code regulations are going to 8:20impact my renovation and to track them 8:23over time and Link them to specific 8:25steps of the process the same thing goes 8:29for my generative AI model I'm going to 8:32link different parts of the model or 8:35different steps to potential legal 8:38regulations as well as use cases so that 8:41if a Law changes or a requirement 8:44changes I can very quickly track that 8:47back to the part or the model if I have 8:51a number of models that is impacted by 8:54that rule so extremely important to make 8:57sure I can quickly react and and adjust 9:00so that I'm not 9:02penalized severely for that error with 9:05my model nothing is as important to a 9:08relationship yet as fragile as trust AI 9:14can truly transform your customer 9:17experience but keep these five tips in 9:19mind to make sure the models that you're 9:22building are models your customers can 9:26trust and remember if you get into 9:28trouble 9:29there's people you can trust who can 9:32help you along the way thanks for 9:35watching before you leave please 9:37remember to like And 9:44subscribe