Latest / The FIT4Privacy Podcast | Exploring ISO Standards AI Privacy | Grow Skills Store / AI Evaluation Methods

Transcript

AI Evaluation Methods · The FIT4Privacy Podcast | Exploring ISO Standards AI Privacy | Grow Skills Store

• 0:00
  How do you identify or evaluate if it's trustworthy AI,
• 0:05
  responsible AI, ethical AI or in the legal partners we say
• 0:09
  compliant AI. So do you have methodology
• 0:13
  guidance on how can anybody, if they want to check a certain
• 0:18
  compliance, maybe say with the EU AI Act, maybe say with ISO
• 0:22
  42,001 or NIST AI framework or any of the other responsible AI
• 0:26
  frameworks, how would they go about it?
• 0:29
  What should they start with? Yeah.
• 0:31
  So this is definitely a challenge that we all face in
• 0:36
  the industry today. Again, like going back to the
• 0:39
  definition I just gave, because AI is machine making gases, it
• 0:44
  means it is non deterministic. And what that entails is it can,
• 0:50
  sometimes it makes good gases, but occasionally it makes bad
• 0:54
  gases. And in AI evaluations, what we
• 0:58
  are trying to do is we're trying to use different techniques,
• 1:03
  different methods or methodologies to catch those
• 1:08
  issues and also to understand the behaviour patterns of AI
• 1:15
  before it goes into release to help us and doing all that to
• 1:20
  help us understand and predict what AI will perform in
• 1:25
  production. So I would say that is what we
• 1:28
  try to do with AI evaluations. And this field, AI evals is
• 1:33
  still very new. So a lot of the practises are
• 1:36
  still taking shape. There are, I wouldn't say there
• 1:39
  is a very clear consensus on best practises yet again,
• 1:46
  because it is very new, but from since like 2022, since the since
• 1:52
  the introduction of ChatGPT. I have been working in this
• 1:55
  field for several years now. And we from my experience and
• 2:00
  from the work that I have done in this field, there are some
• 2:03
  like best practises and practise practical tips that I share with
• 2:09
  the community with people who are working on AI emails.
• 2:13
  And you mentioned like for compliance and that is
• 2:18
  definitely a big part of AI emails today.
• 2:21
  And besides, like, besides compliance work guardrails in AI
• 2:26
  emails that we check for, there's also a lot of
• 2:29
  evaluations that we do to, to, to try to assess the quality of
• 2:37
  the product. So for example, if we have a
• 2:42
  customer support bot for a say a insurance company, then what is
• 2:49
  this chatbot is expected to do? All that should be defining the
• 2:53
  product requirements document, not just what it is supposed to
• 2:58
  do and what it is not supposed to do.
• 3:01
  And all those should be defining the product requirements
• 3:04
  document. And for us who working AI evals,
• 3:09
  our role is to work closely with the product team to translate
• 3:13
  that into quantitative and also repeatable measures again to
• 3:18
  help us predict how the product is going to perform in
• 3:21
  production in real world. So essentially what I understand
• 3:25
  is AI evaluation is not just about post product deployment,
• 3:31
  it's also about working with the teams of the design and
• 3:34
  requirements phase and being ready to demonstrate everything
• 3:40
  that you need in the end. So it's not opposed to Jack,
• 3:42
  it's a pre engagement, pre conversation that you have.
• 3:45
  Yes, and that is a great point. I would say there are three
• 3:50
  different use cases that I see of AI evals during the product
• 3:54
  development cycle. 1 is what I call evaluation driven
• 3:59
  development and that probably raised a bell has very similar
• 4:03
  to what we are familiar with test driven development and it
• 4:07
  is the same concept. So like why we have test driven
• 4:11
  development for traditional software, we need evaluation
• 4:16
  driven development for for AI products because AI products
• 4:22
  again because it is non deterministic, it is a checking
• 4:27
  or understanding the behaviour pattern, it can be extremely
• 4:32
  challenging. So we use evaluation metrics as
• 4:36
  our compass during the product development cycle from iteration
• 4:43
  to iterations to help us identify where we should
• 4:46
  improve, where where we should adjust or update the design.
• 4:50
  So that is evaluation driven development.
• 4:53
  I would say AI evals is also, and this is probably the most
• 4:58
  common use case for AI Evals. It's right before release.
• 5:02
  We want to run the series of cheques just like what we do in
• 5:06
  software testing. We do AI evaluations to help us
• 5:10
  make release decisions. So basically it helps us to
• 5:13
  answer the question of whether this product is ready for
• 5:16
  release. So that is the second use case.
• 5:19
  And then once the product goes live in production, we still run
• 5:25
  AI Evals as like CICD cheques and also as like regression
• 5:31
  tests whenever we need to make changes or make updates to the
• 5:35
  products where whenever we, for example, want to switch to the
• 5:38
  latest model from our model provider.
• 5:43
  And now evaluating AI can mean multiple things for people.
• 5:48
  If you, if I talk to you, it'll be different.
• 5:51
  If I talk to some tester, it's all about testing.
• 5:55
  That's what they consider AI evaluation.
• 5:57
  If I talk to a compliance colleague, he or she would say
• 6:01
  it's about controls or compliance with policy.
• 6:04
  So is all of that AI evaluation or do you define it slightly in
• 6:09
  a a separate way? Yes.
• 6:12
  So in my definition, I would say all this are AI evaluations and
• 6:17
  I come from a background of applied science and data
• 6:20
  science. So to me maybe I'm more looking
• 6:24
  from the angle of what methodology or what technique we
• 6:28
  use. So in that sense all this are AI
• 6:31
  evals, which I I interpreted as using scenario based testing to
• 6:39
  probe the products behaviours. And it has different, it can be
• 6:45
  used for different types of evaluations.
• 6:48
  And I think that goes back to the different, different, how do
• 6:52
  I call it like a different purpose of evaluation like for
• 6:56
  compliance or for quality cheques?
• 6:59
  To catch the full discussion, head to YouTube, Spotify or
• 7:01
  Apple Podcasts and search fit for privacy.
• 7:04
  Links are in the description. Thanks for listening.