Microsofts Six AI Principles
Source: https://www.microsoft.com/en-us/ai/responsible-ai?activetab=pivot1:primaryr6
Fairness
Possible problems in AI systems that are used to allocate or withhold opportunities, resources, information in domains like criminal justice, employment and hiring, finance, credit.
Possible problems:
- AI systems can reinforce existing societal stereotypes.
- Cultural denigration
- Over- and underrepresentation.
There is no single "measure" for fairness that can be integrated in all AI systems.
Another important feature is that fairness is not just about the technical components of AI systems, but also about the context in which they are deployed. This means that fairness in AI is a fundamentally sociotechnical challenge.
This means that it is also beneficial to have a more diverse team of people working on AI systems.
Reliability and safety
Systems needs to be consistent with their design ideas and values and principles. Models should not create harm in the world. So before you use a system in a real-world setting, you need to have quantified and well-understood risks. The potential risks are also shared with the users.
Examples of safety:
- Physical systems such as self-driving cars
- Systems making predictions that may have a physical impact on people, e.g. health systems
Reliability
- Small mistakes have big effects when used many times in systems used by large groups of people
Privacy and security
Privacy is a fundamental right. (Well, that's not as obvious as it sounds!) With machine learning comes an increased reliance on data. This also leads to more requirements concerning the security of applications that rely on data.
Innovative solutions: keep as much data as possible on the devices of people themselves and run machine learning models locally.
- Data origin and lineage
- Data use: internal vs. external
- Data corruption considerations
- Anomaly detection to detect adversaries
Inclusiveness
Making sure MS is intentionally inclusive and diverse in approaches to AI.
"If we figure out how we can design for the 3%, then we can also solve for the 97% at the same time". I don't think I get this reasoning.
Representativeness of underserved minority communities. Make sure they are involved in the design and testing process.
Make sure to not take an "ableist" perspective. Wikipedia: "Ableism is discrimination in favor of non-disabled people."
Transparency
Diverse range of goals:
- Mitigating unfairness in ML systems
- Helping developers debug their AI systems
- Gaining more trust from users
Two sides to transparency:
- The people that create AI-systems should be open about
- why and how they should use AI
- the limitations of those AI systems
- Understand the behavior of AI-systems
- Interpretability
- Intelligibility
Accountability
Even though models are becoming increasingly complex, organizations are still accountable for the effects they have in the world.
The structure we put in place so that we are consistently enacting our principles in what we do.
Help partners and customers be accountable.
Accountability principles guide 1) development, 2) how products are sold 3) advocating for regularizations e.g. on facial recognition.
W.r.t. facial recognition, there may be many promising uses for facial recognition, but also many which interfere with civil liberties or push AI in a direction MS is not interested in supporting.
Challenge is to take a high-level principle such as fairness, and then translate that to practice such that designers incorporate it in each level of the design process.
How Microsoft embeds ethical AI in their organization
Relevant committees: https://www.microsoft.com/en-us/ai/our-approach?activetab=pivot1%3aprimaryr5
Something like this would be a great job for me:
The Aether Committee was established at Microsoft in 2017. Our senior leadership relies on Aether to make recommendations on responsible AI issues, technologies, processes, and best practices. Its working groups undertake research and development, and provide advice on rising questions, challenges, and opportunities.