Responsible AI for generative models: Designing for responsibility

What this episode is about

The discussion focuses on the principles of responsible AI in the context of generative models, emphasizing the need for a multi-disciplinary approach that involves collaboration across various sectors. The speaker outlines the importance of aligning AI models with ethical standards through rigorous testing, training techniques, and user feedback to ensure that AI technologies are developed and deployed responsibly.

Key takeaways

• Responsible AI isn't just smart; it's global—policies need cultural sensitivity to truly resonate.
• Behind every generative AI model lies a multi-layered ecosystem—it's a wrestling match between data and ethics!
• Great AI requires rigorous testing, not just for quality, but for the diversity it reflects—representation matters.
• Mitigating bias starts with data; pre-training choices shape AI's worldview more than the model itself.
• Innovative model tuning could outperform human experts—when AI polishes its skills, the results can be stunning!

What is discussed, and when

1. 0:00 – 3:00: AI principles, Google LLC
2. 6:00 – 9:00: Google LLC
3. 9:00 – 12:00: Google LLC
4. 12:00 – 14:13: Google LLC, Frontier Model Forum, African-American community

Notable quotes from "Responsible AI for generative models: Designing for responsibility"

• Part how AI is changing our world in terms of the responsible part I'll be talking some about that today and how we can align the models to make them respon to make them adhere…

  about AI principles · at 0:31

• Consult community-based experts one of my favorite examples happens to be a us one was trying to do a study around pred in healthcare expenditures in the African-American…

  about African-American community · at 12:22

• That works for everybody that makes sense looking at the global directives that we're all following and Google also participates in many of the forums like the Frontier Model…

  about Google LLC · at 13:32

Who and what is mentioned in this episode?

Companies & orgs

• Google LLC — Company · Business focus · 4 mentions · first at 2:20
  Google is discussed as a key player in AI development and responsible AI practices.

Concepts & topics

• AI principles — Concept · Core topic · 2 mentions · first at 0:44
  AI principles are central to the discussion on how to align AI models with responsibility objectives.

Places & events

• Frontier Model Forum — Event · Collaboration platform · 1 mention · first at 13:40
  Mentioned as a forum where industry leaders collaborate on advancing AI practices.
  That we're all following and Google also participates in many of the forums like the Frontier Model Forum where we industry leaders can share with each other and to help…
• African-American community — Place · Demographic · 1 mention · first at 12:31
  Mentioned as an example of how cultural context affects AI model predictions.

Full transcript with timestamps

0:09 hi there it's great to be here and I'm very excited uh because this talk really uh draws on and reflects on some of the talks that we've already had today when we think about responsible AI we say bold responsible and together and what I loved in several of the previous talks was great great examples of the Bold part how AI is changing our world in terms of the uh responsible part I'll be talking some about that today and how we can align the models to make them uh respon to make them uh adhere to our responsibility objectives to our AI principles and the Together part partha did a great job talking about how it's an effort to involve the whole world and how we do that and fold those back into the models first off I want to just comment RI responsible AI is M multi-disciplinary this talk reflects the work of many many people across the ecosystem the people in this room uh as well as people throughout research trust

1:11 and safety and other parts of the organization and I want just acknowledge that next we're going to talk about the responsible the generative AI ecosystem so when we think about responsible AI a lot of people and this is common in the industry will say oh this pre-trade model isn't responsible or something like that and um actually when we think about it we think of the entire ecosystem what people see is a model that's had stuff done to it yes at the core there's the pre-train model and the the training data that goes into it but then there's a whole set of techniques we use um which we're I'm putting under the lump of responsible Generation all the different kinds of training alignment techniques um that are used before a model is released and are part of what gets released to the public and then beyond that we have safeguards um just to make sure that in case anything slips through we've got it covered and

2:03 uh user feedback and we do rigorous testing and I'll talk about that a little bit so when we talk about designing for responsibility um we break it into several steps uh the first and not to be overlooked is um your content policies and this is an area I think uh a Google has its AI principles and those principles are then actualized as policies so for example a policy might be don't talk about dangerous illegal or malicious activities or don't offer harmful content hatred harassment bullying misinformation sexually explicit material okay or sneaky stuff to try and override the filters those are examples of policies and so by definition in order to make something quote responsible you have to start by defining for yourself what that means to you and there are differences and I think this is where a lot of the Multicultural um dimensions are pretty important because there are nuances and

3:04 their differences across cultures so it's a step that we take is all our work is built on that once we built on that that's like our North Star and it's the basis upon which we do the the alignment and the built-in capabilities the second core piece of this and I'll talk about each of these is what we build in and we build in capabilities for safety and fairness we want to generate content that's safe and it's supportive of the diverse voices and cultures we have input and output safeguards adversarial testing and I'll touch on it as well it's not just the model itself but you want people to understand it and give them simple and helpful explanations and a lot of effort we put into the data as well because we want to make sure that that data is high quality in order to give high quality results so one of the big areas that I'll talk about and this is touched on a couple of the different

3:56 things is model mitigations we break that into three pieces um first of all was touching on the pre-training data important things to consider in the pre-training data are obviously to make sure that we've got permissions to use all of it and get that aligned um and we have a whole set of responsible data analysis that we do to look at the data to understand particularly around things like representation of people to understand who's in the data who's not in the data because this gives an opportunity then to augment um at this early stage for things that were missing um as well to make sure we do cleaning filtering um you know anything that slipped through because a lot of this is on the web to make sure that we've we've taken care to remove very offensive things that we don't want in there and then as well we have these opportunities and you heard this in the Thousand languages talk to

4:50 if we collect data and this is a great opportunity to augment data to to reflect the diverse perspectives so we do our best at pre-training to get a diverse view obviously there are more opportunities uh beyond that to augment a second area that we've put a lot of focus on is in the model tuning and this is what I call the the non non-negotiables the things that are built into the model and we've really focused a lot on fine-tuning with as a core uh that you may have heard of constitutional Ai and going beyond that but we do we we basically look at Chain of Thought reasoning we look at neutral point of view we make the prompt more more neutral and we also ask the model and we talked about the Chain of Thought to even rank its own responses choose the best response so we actually in the training data we have a whole pipeline of um actions that we take in order to get highquality training data and

5:51 interestingly we find that it's better than some of the experts because some of these are quite quite nuanced um and it's it's worked well another area that we use in the model tuning is around improving equity and inclusion and some of the areas that we're looking at for that are things like prompt expansion so that for example instead of saying show me a picture of a CEO you might look and expand that about different types of attributes and these are things an example of some of the things that we do to improve uh the model and of course reinforcement learning is a key part of everything we do to make those respons is better finally Beyond built-in tuning one of the things that we're finding is the importance of what we call configurable capabilities because if you use an example um of the previous talk that was talking about medp Palm well one of our principles as a general vanilla principle is we don't give

6:49 medical advice but if you're in a medical application you want to give me medical advice so this is an example of you know configurable safety we want for some users the model to be able to access some capability and others perhaps not and so we've been thinking about ways to make safety configurable very important as well in the Enterprise uh space where you have different companies with different things that they're doing um and some of the things that we've been looking at capabilities um are things called control tokens uh you basically label some of the training data and say this one's about medical advice for example this one is not and by doing that you can then on output steer the output and get it to include a feature or not looking at control decoding so you can put your own classifiers again and steer the output to more responsible output and I think another capability we're looking at is

7:42 Laura to get small sample uh fine-tuning which actually you could slide in as a way to get a particular result on a query by query basis I don't think we've done that yet but that's something that we're researching also um the talks about composability was super interesting as well I think we have to look into these things to make our models better so uh finally in the next area is is doing our best job to make the models responsible we uh don't always get it right and there's some queries we may not want to entertain or things that slip through and so there's a lot of work that we're also doing around safeguards input and output filters I think some of the uh this is as we move into multimodal this is an area that is garnering a lot of attention there's a lot of tricky things so um we do have classifiers that are hanging around Google lots of them but they need to be adapted to work better for generative

8:40 AI um they uh we also need them to be policy aligned because you have a not quite fit policy of something that was used for something similar but not the same needs to be adapted and so um the Contex contextually awar is really um important for example you might say um give me a man eating soup and so what comes out is an image of a center or a non-human eating soup or show me a picture of a woman on a horse and it shows somebody who is a highly sexualized image okay and so these are areas that we have to attend to to make sure that the inputs and output are contextually aware um and lots of talk today around multilingual multimodal uh this is a really important area and getting it nuanced to culture so the ways in which they we're going about these is to Leverage The llms themselves as well as um looking at agile classifiers because I think we're going to need specialized ones uh finally on

9:41 the area of red teaming this is another area that I think is especially important to catch Multicultural multilingual and nuanced harms uh because things are so contextually different so we have three basic approaches that we use we use algorithmic red teeming this is around synthetic data generation which is where it's super important to get community input so that we can represent everyone in the generation and evaluation we bring in experts both in three different dimensions fairness and safety in red teaming and security areas and abuse detection and then the community engagement I think is something that was touched upon here we have some active efforts to engage communities in a variety of ways we have one that's particularly dear to my heart recently is called adversarial Nibbler we're engaging people around the world world to try and give subtle prompts to the models so that they can uh see if we can

10:35 trick them essentially uh which is a really good way to get diverse input so these are all the things we're doing um as we look ahead we have a lot of challenges and exciting opportunities and I'd love to engage all of you uh multimodal is definitely uh we need we're there now and we need to make the models work the other thing that I didn't really talk about is how to make the scalable right now I say we're building like Mount Everest one rock at a time because we have a lot of special purpose things what we don't know how to do yet is scale this how do I build a cultural Benchmark that doesn't you know we're starting out from the position of only having Western views so we're doing better we are now getting certain cultures great great work here um in in in from Google research India but how do we extend that okay in some meaning away I think that's really important um and uh another thing that we're really

11:34 focusing on that again people here could engage in is around data quality metrics because what happens if you ask people to do stuff they all ask the same thing generally and so how do you force uh engagement there get our communities involved um again and um the other area that I didn't touch on in this talk that's super important is the advancement of human data practices I like to say radar disagreement is a feature not a bug be and we have to figure out how to embrace that um because that's where a lot of our understanding is coming from about the different cultural contexts again this was touched on here we talk about uh responsible AI together and this is where across throughout the life cycle we need the engagement of people to contribute to AI to make it better consult community-based experts one of my favorite examples happens to be a us one was trying to do a study around pred

12:29 in healthcare expenditures in the African-American community and so they were trying to um predict it based on doctor visits well it turns out that Community doesn't really trust doctors I'm sure there's the same stuff here in India and so they don't visit the doctor so it's a pretty poor model so you really need to start at the beginning and understand um that things are formulated get people to help in creating the policies then as we move throughout the life cycle you know trying to uh collect the data that drives our models and throughout the analysis so that we can improve representation because I think that in contrast to safety which is another area we focus at improving an inclusion and representation it's a win-win you can only make things better and uh so the more and the better you can do in that area the more it makes the better it makes the models for everyone so next steps that you surely have read I think

13:22 another piece of this puzzle is because AI is so important and at the doorstep it requires collaboration across the industry where we work together with the governments um have effective regulation that works for everybody that makes sense looking at the global directives that we're all following and Google also participates in many of the forums like the Frontier Model Forum where we industry leaders can share with each other and to help Advance the practices to make uh AI uh the influence that that we that it needs to have that we all wanted to have so okay that's all thank you

Questions about this episode

What is "Responsible AI for generative models: Designing for responsibility" about?

The discussion focuses on the principles of responsible AI in the context of generative models, emphasizing the need for a multi-disciplinary approach that involves collaboration across various sectors. The speaker outlines the importance of aligning AI models with ethical standards through rigorous testing, training techniques, and user feedback to ensure that AI technologies are developed and…

What is mentioned in this Google Research episode?

4 people, works and ideas are mentioned, including Google LLC, AI principles, Frontier Model Forum, African-American community. Each one on this page links to its own page with every episode that mentions it.

How long is "Responsible AI for generative models: Designing for responsibility"?

The episode runs 14 min; the transcript is about 2,487 words in 15 timestamped sections.

More from Google Research

• All analyzed episodes
• People mentioned on Google Research
• Topics mentioned on Google Research