Without Human Values, What Stops AI From Acting As A Sociopath?

Artificial Intelligence (AI), much like a sociopath, inherently shows no regard for right and wrong. This unsettling similarity raises an important question: Can we trust the future of AI with such a default state?

Joe Edelman, before founding the Meaning Alignment Institute (MAI), pondered deeply on this. Funded by OpenAI, MAI has created a model to guide AI, particularly Large Language Models (LLMs) like ChatGPT, to respond to queries while considering democratically decided human values.

Establishing Common Values Amid Societal Challenges

How do we create common values, and why is this necessary? In an era marked by polarization, reaching a consensus on shared values is daunting. Previous attempts at socially educating AI have had mixed results, with instances of AI propagating hate speech. Leaning on the work of philosophers such as Charles Taylor, Edelman believes values are about what we find meaningful in tough moral choices rather than ideology. This concept forms the backbone of the MAI’s approach.

Instead of directly asking people about their values, the MAI model uses a chatbot powered by ChatGPT to find out what individuals truly find meaningful. This is done by asking for responses to preset questions.

I chose to answer this question: “My 10-year-old son refuses to do his homework, spending all his time at his computer instead. How can I make him behave properly?”

I answered a series of questions from the chatbot, which took about 5 to 10 minutes. The chatbot then interpreted and fed back the deeper meanings behind my responses, focusing on why I held my core beliefs.

From this, the model then took these and concluded that in dialogue with the user (in this case the mother), the following kinds of things would be useful for ChatGPT to mention if asked how to make her son behave properly:

• Signs of dishonesty that might indicate a breach of trust
• Understanding of the necessity of rules, even if it’s not fully grasped

Thereafter, the tool suggested other relevant values that had been added by other users (who had answered the same question) that I could ‘like’ such as ‘Respecting Autonomy in Politicised Contexts’ and ‘Equality and Autonomy’. Likewise, others could vote for the values that had been generated from my responses, which may encompass other values that have been created previously. Thereafter, the MAI builds a moral graph which can guide a future version of ChatGPT (or other LLM).

In my conversation with Edelman, I raised some concerns over bias – such as sample size, sample bias (including cultural bias) and leading questions. The report of the study, which was based on the responses of 500 people selected to reflect the age, income, political leanings and geography of the US population, showed progress in some areas of bias, while others remain to be worked on. However, the primary object of the study was to prove that values can be collected from diverse populations democratically and thereafter imported into an LLM such as ChatGPT to help guide its responses.

The results of this US-focused study (Edelman wants to expand the study globally) showed convergence across gender and age and that people can put aside their ideological affiliations. By digging into why people responded in certain ways, rather than focusing only on the response – the model found that even those with opposing ideologies could have similar underlying values. Even if the ways that these values were expressed differed, the values established did not point in multiple conflicting directions that cancelled each other out – providing something that LLMs could use.

Why adding human values to LLMs matters

LLMs are increasingly shaping our lives, often prioritizing commercial gains over moral implications. As Edelman noted: “The Pentagon has an LLM that’s doing war strategy. Financial companies are experimenting with putting LLMs at the trading wheel. And of course, we see the media consequences just starting to appear already. And yet, it’s actually quite hard to hire a sociopath as your copywriter.”

Adding human values into LLMs ensures these technologies incorporate human values through a democratic process, avoiding the pitfalls of an amoral AI, or where values have been decided by the developer or by the LLM itself according to its constitution.

The commercial threat and legislation

Edelman accepts that there is little commercial incentive for companies such as Instagram to add human values (such as self-agency, which could reduce user addiction) into their LLMs (hence MAI is a not-for-profit, grant-funded organisation). Unfortunately, as history suggests that companies are unwilling to look beyond profit maximisation, it is therefore difficult to imagine that, without legislative intervention, social media providers and others will implement human values into their LLMs. Perhaps, similar to a Fairtrade certification, a ‘human values incorporated’ label could be added to LLMs?

In the meantime, Edelman is hoping that due to OpenAI’s lead in the LLM market which enables it to be “free from some competitive forces for the moment”, OpenAI could take a lead on adding human values to ChatGPT.

Preserving our humanity

As we increasingly rely on AI, balancing technological trust with our human essence becomes crucial. Integrating human values into AI is not just about improving technology; it’s about preserving our humanity in an increasingly automated world.