RFC: A Pathway to AI Governance

Bharath_Reddy · 1 December 2023 12:11

Hello, some of us at the Takshashila Institution have been working on a discussion document on A Pathway to AI Governance.

While the narratives around AI governance globally primarily focus on ethics and safety, we believe that it is also important to consider it from a national interest perspective. This involves examining how AI adoption can help humans flourish, strengthen democracy, and promote a stable global order. It also looks at the need for sustainable practices in AI adoption and the importance of ensuring competition in the AI ecosystem.

These considerations must be examined across these stages of the AI supply chain - data, computation, model, and application - to envision the desired outcomes at each stage.

We are a bunch of policy nerds who have put this together, and we would love to hear from other stakeholders. Please let us know what you think of this and how we can improve this.

You can either share your feedback on this forum or as a pull request on the markdown version of the document on the GitHub repo.

rushabh · 7 December 2023 17:25

Maybe its just me, but I have a feeling that AI at this point is just hype and “hammer looking for a nail”. Its been a year since ChatGPT was released but where are the AI driven mass layoffs we have been worried about? Other than all those shitty images on social media (which look eerily like NFT images) what have we got? Just a contrarian thought, that this might end up being vapourware just like “blockchain”. A lot of solutions are being built in image recognition, character recognition, but they have steadily improving for a while.

Anything a heuristic can’t do, AI probably won’t either.

Bharath_Reddy · 8 December 2023 02:00

The term Artificial Intelligence is probably a misnomer, and they are more accurately described as “stochastic parrots”. I agree with @rushabh that they are probabilistic pattern recognition systems that have been steadily improving. But that does not negate the fact that they can vastly increase human productivity. A lot of the work that we do requires skills which tire us out but can be done better by an algorithm.

The debate around AI is muddled with theories about large-scale impact on jobs, existential risks and such. These are not immediate concerns, given the current state of AI. But we feel there are real concerns around the concentration of market power, access to the tech, bias, and misuse. These are some of the concerns we attempt to address in the document.

rushabh · 8 December 2023 04:16

Don’t you think we lost those battles the day we invented machines? Today we are writing in English on a computer invented by Americans. If you want to win those battles from a nationalistic perspective, then there is a lot of catching up we have to do.

Bharath_Reddy · 8 December 2023 06:02

Not at all. I think technology has improved our lives and can do much more. I don’t need to think we need to think from a nationalistic perspective but we might need to think about it from a national interest perspective. The two can be quite different.

rushabh · 8 December 2023 06:29

Yes they have improved no doubt (I am not saying we dial it back), but also created concentration of power and access, bias etc. Somehow both are inter-related.

knadh · 12 December 2023 18:26

github.com/TakshashilaInst/a-pathway-to-ai-governance

A few comments

opened 06:25PM - 12 Dec 23 UTC

knadh

Thank you for opening up the document for comments! Sharing a few quick thoug…hts: 1. There is good focus on marketplace, industry, state capacity, but community and civil society as stakeholders seem to be missing. The impact of these technologies on disaparate communities is going to be significant. It's also important to note outside the industry, the lion's share of incremental breakthroughs and innovation is happening in tinkering communities. 2. In the same vein, imagine a Wikipedia-style, Wikipedia-scale non-profit AI community service. How would that fit into the overal thesis here? > Ownership of Data > A similar framework could also be applied to unlock the value of various other data locked in silos for specific purposes approved by the user. Under such a framework, users will also have the option to opt out of their data being used for training purposes. > Data portability to unlock value for users across markets and also prevent undesirable platform lock-in effects. For example, a seller on one e-commerce platform should be able to transfer their authentic product reviews to another e-commerce platform. Likewise, a user's health data from a fitness tracker can be shared with their chosen healthcare provider to facilitate the creation of a personalised health plan. Data request templates for different use cases need to be established for various use cases in consultation with stakeholders across industry and academia. > Enable data provenance and usage transparency. A data aggregation framework also offers information about the origin and lineage of a dataset available within the market. In addition, it can also serve as an audit trail for the downstream models where a dataset is being used. 3. I think this entire section may be introducing significant scope-creep to the context of the document (where the scope is training data that goes into AI systems). Concepts like universal data portability, provenance etc. fall outside this realm and are entire complex fields in themselves (conceptually and technically, and also philosophically). This is applicable to the use of user data in any context, be it training AI systems, mathematical models, or advertisement targeting systems. > Individuals should be able to dictate if, how, and by whom their data can be used. While it is common practice for firms to aggregate and transform the data they capture for internal and external use, individuals typically lack the means to do the same. India's Data Empowerment and Protection Architecture (DEPA) and the Account Aggregator Framework built on top of it illustrate a consent-based intermediary system14 to facilitate such a transaction. 4. AA provides a structured, smooth system for users to share data with requesters. However, firms aggregating and transforming data from various sources and signals, that is mutually exclusive of organisations using a system like AA to request data. Those practices can continue regardless. > We propose the creation of a sector agnostic entity staffed with data scientists and cybersecurity experts to ensure data uniformity and compliance with best practices. Its mandate would include creating a data engineering plan, conducting audits of different state entities, and enforcing data compliance standards. 5. In addition to adopting and following universal data standards within the govt., the enforcement of a "public money, public code" / "public money, public data" approach seems pertinent here. To cite an example, there are various govt. bodies that use tax payer money to create language training corpora, dictionary datasets etc. and keep them locked away. Such datasets should be in the public domain. > For instance, CSPs could be compelled to share or disclose such data with foreign governments or agencies, and they might also be susceptible to cyberattacks or sabotage by malicious actors. 6. Cyberattack and sabotage can happen to data stored anywhere (especially in a networked system) is suspectible. CSPs, foreign or domestic, is irrelevaant. > ... as quantum computing, ... and edge computing, which can enhance the performance and efficiency of AI applications This is not correct.