Similar to a sociopath, artificial intelligence (AI) inherently lacks consideration for moral distinctions between right and wrong. Given this unsettling parallel, can we entrust the future of the Internet of Things (IoT) to such a morally neutral entity?
Before founding the Meaning Alignment Institute (MAI), Joe Edelman deliberated extensively on this matter. With the backing of OpenAI, MAI has devised a framework to guide AI, particularly large language models (LLMs) like ChatGPT, in responding to queries while factoring in democratically determined human values.
Establishing Common Principles Amid Societal Challenges
The necessity of defining shared values and the methodology behind it pose significant challenges in an era characterized by societal divisions. Despite prior efforts to politically educate AI, instances of hate speech perpetuated by AI persist. Drawing inspiration from philosophers such as Charles Taylor, Edelman posits that values revolve more around what we deem crucial in complex moral dilemmas rather than ideological stances. This principle underpins MAI’s approach.
MAI employs a ChatGPT-powered robot to discern people’s intrinsic values indirectly, bypassing direct inquiries about their values. This is achieved through soliciting responses to predefined queries.
When tasked with addressing the query, “My 10-year-old child refuses to do his homework, spending excessive time on his computer instead. How can I encourage him to be more productive?” I engaged with the robot in a series of questions for 5 to 10 minutes. The robot delved into the reasons behind my core values and extrapolated deeper insights from my responses.
Subsequently, the system recommended that in addressing the concerned parent (in this case, the mother) on how to improve her son’s behavior, ChatGPT should highlight:
- Indicators of dishonesty that could erode trust
- Emphasizing the importance of rules, even if not fully comprehended
Moreover, the tool suggested additional relevant values contributed by other users who had previously responded to the same query. Users could endorse these values, such as “Respecting Autonomy in Politicized Contexts” and “Equality and Mobility.” Users could also vote for values derived from my responses, potentially including newly formulated values. Following this, MAI constructs a moral curve to steer future iterations of ChatGPT (or other LLMs).
During my conversation with Edelman, concerns regarding bias, sample size, sample bias (including cultural bias), and leading questions were raised. The study, based on actions of 500 individuals representing diverse demographics in the US, exhibited improvements in some bias aspects while indicating areas for further enhancement. The primary objective of the study was to illustrate that values could be aggregated across varied populations and integrated into an LLM like ChatGPT to inform its responses.
The outcomes of this US-centric study (with plans for global expansion) showcased alignment across gender and age groups, illustrating that individuals can transcend ideological differences. By delving into the rationale behind people’s responses rather than merely focusing on the answers, the framework revealed that individuals with divergent ideologies may share common core values. While these values may be expressed differently, they did not conflict to the extent of nullifying each other, providing ample guidance for LLMs.
Significance of Embedding Human Values in LLMs
As our lives increasingly intertwine with LLMs, which often prioritize financial gains over social considerations, the integration of human values into these technologies becomes imperative. Edelman highlighted the presence of an LLM at the Pentagon focusing on battle strategies and financial institutions exploring LLMs in purchasing processes. However, the media repercussions are just surfacing. Nevertheless, recruiting a narcissist as a copywriter poses significant challenges.
Incorporating human values into LLMs ensures democratic decision-making by these technologies, mitigating risks associated with malevolent AIs or scenarios where creators or organizations dictate the AI’s values.
Regulatory Measures and Professional Implications
Edelman acknowledges the limited commercial incentives for platforms like Instagram to infuse human values (e.g., reducing user addiction through self-aggrandizement) into their LLMs, hence MAI’s status as a grant-funded, nonprofit entity. Anticipating reluctance from social media firms and others to include human values in LLMs without legislative mandates, Edelman proposes potential labeling akin to Fairtrade certification, indicating “human values included.”
In the interim, Edelman hopes that OpenAI, as a frontrunner in the LLM sector, will proactively integrate human values into ChatGPT, leveraging its position to circumvent certain market pressures.
Preservation of Societal Fabric
As our reliance on AI deepens, striking a balance between technological advancement and human essence is paramount. Integrating human values into AI isn’t just about technological progress in an increasingly automated world; it’s also about upholding societal norms and values.
