The Ethical Frontier: Navigating the challenges of open-source large language models

Over the years, several high-profile ethical AI horror stories have emerged highlighting the challenges faced by the rapidly advancing field without appropriate policies and guardrails in place.

Microsoft’s Tay – an AI chatbot released on Twitter – quickly spiraled out of control, adopting racist and homophobic language within hours of its debut. Amazon faced controversy when its hiring algorithm exhibited gender bias, disadvantaging female applicants for technical positions.

The infamous Cambridge Analytica scandal revealed how AI-driven microtargeting techniques were used to manipulate political opinions and sway elections. In healthcare, AI algorithms have been found to discriminate against certain racial and ethnic groups, exacerbating existing health disparities. The COMPAS recidivism risk assessment tool, designed to help courts make parole decisions, faced backlash for perpetuating racial biases by disproportionately denying parole to people of colour.

These and many cautionary tales serve as stark reminders of the potential ethical pitfalls that lie at the intersection of AI and society, emphasising the need for greater scrutiny and responsibility in the development and deployment of AI technologies.

As we trailblaze in the world of ChatGPT, recognising the immense potential large language models (LLMs) such as GPT-4, ChatGPT, PaLM, and LaMDA hold in transforming various industries, these models also present us with a series of ethical dilemmas that warrant careful consideration.

The challenges including the creation of harmful content, the reshaping of labour markets, the risk of generating misleading information, the potential for disinformation campaigns, their role in weapon development and cyberwarfare, privacy concerns, unpredictable emergent behaviours, and the possibility of an unwarranted acceleration of innovation in potentially problematic areas.

The LLMs are trained on internet data, and I don’t have to convince you about the toxicity of the internet contents. To control toxic output from LLMs, companies like OpenAI employ a combination of grounding, prompt engineering, and bias control measures. Prompt engineering involves carefully crafting user inputs to guide the model towards desired and safe responses.

For example, a control prompt may explicitly request non-toxic or non-biased information which automatically gets appended with your requests and run in context such as “Please provide a neutral and inclusive summary of the topic without using any discriminatory language” or “Present an argument that is well-reasoned and evidence-based, without resorting to discriminatory language or stereotypes.” This isn’t a silver bullet but, along with post processing filters, covers a wide range of use cases to mitigate the risks associated with harmful content generation and enhance the responsible use of LLMs.

However the proverbial genii came out of the bottle with Meta’s LLaMA (Large Language Model Meta AI) model leak. The model’s weights were intended to be made available only for academics and researchers on a case-by-case basis, but one of the recipients leaked the code on GitHub. This unexpected leak provided programmers and developers worldwide with open access to their first GPT-level LLM, sparking a frenzy where developers have optimised, expanded, and discovered new use cases and derivatives – including Alpaca from Stanford, Lora, Dolly, BigScience’s BLOOM, EleutherAI’s GPT-J, GPT-NeoX, Polyglot, and Pythia, as well as Cerebras-GPT from Cerebras Systems, and Flamingo and FLAN by Google.

This expansive range of LLMs demonstrates the potential for increased innovation and collaboration in the AI community but we also find ourselves grappling with the ethical implications and potential misuse of these powerful tools.

The barrier to entry on building and running LLMs has significantly decreased. However this lowered barrier presents ethical challenges since now anyone can run these models without any guardrails and cause chaos and mayhem by generating potentially harmful or misleading content.

To address these issues, several guardrail initiatives, including Nemo Guardrails, have been introduced to help ensure the responsible use of LLMs. To address this fully, we need self-governance, guardrails in tech-spaces, and regulations to ensure these powerful models don’t get exploited.

The author is has recently completed a book on the topic of Responsible AI in the Enterprise with Heather Dawe, which will be published in summer 2023.