AI ethics has been a complex topic even before ChatGPT (and other large-scale language models – LLMs) and generative AI. In this article, we discuss how LLM and generative AI have changed the ethics landscape, and what companies can do to keep up and keep moving forward.
AI ethics is a vast and complex subject. At a high level, the purpose of ethical AI is to ensure that systems and technologies are built and operated in line with human value systems and environmental considerations. Within this broader goal, AI ethics consists of several components as outlined here.
Where Do Large Language Models Fit?
Large language models contain all of your AI ethics concerns, and amplify many of them. Given how prevalent it has become in such a short time, the ethical issue is even more relevant. Here are some examples of LLM ethical issues.
Note – I didn’t include hallucinations because any AI model is bound to make mistakes.
Where does generative AI fit in?
The definition of the relationship between generative AI and LLM depends on who you ask. One good definition, however, is that while LLM is a subset of generative AI in terms of generating text, LLM also has strong expertise in text-based queries. For the purposes of this article, the ethical challenges of generative AI include all challenges of LLM, plus additional challenges related to models that generate content in multiple modalities (image, video, sound, etc.) can be considered.
- Content Ownership: Content ownership litigation has begun, covering all forms of content. Of particular interest is the creator economy, where generative AI can upend entire traditional business models. This applies to all media types including text.
- Incorrect information: Reapplies to all media types including text. Generative AI can create highly realistic fake images, news articles, and more that are becoming increasingly difficult to detect. Additionally, chatbots powered by generative AI and LLM could usher in a new era of social media interference.
What does the law say?
Legislation in this area is still in its infancy. For example, it is not known who owns the copyright for AI-generated artwork. There is a debate going on in the European Union regarding ChatGPT and potential GDPR violations. As noted in the section above, content ownership litigation has also begun, giving the legal system an opportunity to provide a case law assessment of content ownership.
What new technologies are being developed?
Examples of technologies being developed to address these issues
- A variant of human feedback embedded in learning. Reinforcement learning with human feedback (RLHF) is a process in which a human provides feedback on, say, the output of her LLM. These feedbacks are used to build a policy engine that LLM can use to select the most human-acceptable candidate output. ChatGPT is widely believed to use RLHF, but competitor Anthropic offers an alternative called Constitutional AI, which uses a structured rule system that represents human values/preferences instead of active human feedback. is proposing. In both cases, the goal of technology is to align AI output with human values and preferences.
- Forget learning. The key to maintaining privacy and good data practices is controlling (and knowing) what data enters your AI model. Unlearning is a technology that allows AI models to “forget”, or exclude from understanding, selected data elements. The importance of promoting Unlearning is underscored by Google’s recent announcement of a contest to promote new advancements in his Unlearning.
What can we do for your business?
While there are still many unknowns about these technologies, there are some things you can do to protect your business.
- Understand the origins of AI technology. If built in-house – understand what data (especially customer data) is being used. For example, if a customer later decides to opt out and requests that all data be removed from the model, do you know how to do so? ), understand what data you are providing to an external API, even in a query, and whether it is acceptable for your business.
- Decide whether to build or buy, depending on your needs. For example, general-purpose (non-private) queries that are best answered by large-scale models may be better suited for licensed APIs than training large-scale, general-purpose models in-house. Sensitive queries (whether they involve customer private data or your own IP) may benefit from a custom-built internal model. Many models built into the code often use starter models downloaded from the internet, so make sure you understand the source even when building internal models.
- Stay up to date on the latest legal developments relevant to your business area. The field is moving very quickly without a clear answer, and the results of these developments will shed more light on government opinion.
follow me twitter Or LinkedIn.
