Regulating and Self-Regulating ChatGPT

Hi. In this week's Visual Legal Analytica, I have analysed ChatGPT from a regulatory standpoint.

.        

ChatGPT, a sibling model of InstructGPT, has gained much traction recently. With over 1 million users and their interactions with this interesting chat model, the user reactions have amazed discourses on education, public policy, lawyering and many other things. However, the model's success so far is not a pessimistic development when it comes to technology governance. Let's take it this way: there are such technological models, which are evolutionary from a technical aspect. Achieving the edge obviously may be backed by technology hype cycles.

However, there is no doubt that ChatGPT is a unique innovation, considering their reinforcement learning algorithms. While people may claim that this piece of technology is capable to replace Google, the generic use cases of this technology depends. In fact, it wasn't long ago when Meta had produced an algorithmic system, Cicero, to predict discourses in diplomacy, which they claim to be in the realm of AI Diplomacy. It is known that Cicero has achieved extraordinary scores in the strategy game at webDiplomacy.net. That is praiseworthy, and if the system is analysed at best, a lot could be understood about its generic and non-generic use cases.

This article is dedicated to analyse the use of reinforcement learning via ChatGPT, based on its use cases / distribution relevance, from a regulatory and policy perspective. The article also provides conclusive insights on where the trajectory of reinforcement learning, could invite regulatory oversight, with a sectoral focus.

Reinforcement Learning at a Glance

Reinforcement learning (RL) is a subset of machine learning practices which are followed by developers and data scientists across the globe. It has a special place in the field of artificial intelligence and law, due to its ubiquitous features. RL is a kind of machine learning method, where any AI system (subject to learning the environments of the relevant data subjects for a set of tasks to be achieved expectedly) are reinforced or exhorted in a pattern, in a specific environment to maximise their notion of cumulative rewards. This reminds us of behavioural economics, wherein earning cumulative rewards is essential for an agent to act in a pattern they are ought to be. Now, using reinforcement learning, it is much possible to create proper use cases for the agent to learn an optimal / nearly-optimal, policy that incentivises the "reward function" or other user-provided reinforcement signal that accumulates from the immediate rewards. The products / services discussed in this article, ChatGPT, Cicero and GitHub Copilot are inspired by or based on reinforcement learning.

Looking at policy realities as they stand, RL can be subject to heavy supervision protocols, which may succeed or fail in building the AI system, accordingly. For example, in the case of State-Action-Reward-State-Action (SARSA) model, RL algorithms are subject to cumulative rewarding. Here, cumulative rewarding is experienced by the algorithm when it has to act in line with a policy statement, which represents a probable set of things to be achieved. This is natural to happen in any SARSA setup. However, In the case of Deep Q-networks (which also evolves at the level of neural networks), the RL algorithm has to learn and self-explore to develop those self-reflected "policy considerations" along with the existent RL techniques in place.

In the next section, certain products or services developed through reinforcement learning are analysed, to estimate their impact, purpose and limitations, from a regulatory standpoint.

Regulating Reinforcement Learning

Regulations cannot curb innovation. However, we are at a tipping point in various D9 countries, where at the vicinity of the proliferation of recognisable and usable disruptive technologies, if governments deny or overlook the issues attached with these technologies, then it is concerning. We also understand that multiple classes of artificial intelligence technologies of socio-economic and socio-technical value, have to be observed carefully, with a sector-to-sector regulatory approach. Governments have already started developed generalised and some sector-specific regulatory methods, especially on recommendation algorithms, recognition services, predictive algorithms and other relevant tech products / services of relevant categories including analytics. RL is a unique case but to make it simple (unlike recommendation algorithms, where narrow regulatory outlooks might help), some entrant regulatory breakthroughs could be helpful.

Let's discuss ChatGPT to understand its scope.

Read the rest of the article at vla.digital.