Over the past few days, with some of my human friends who have reasons to worry about the possible catastrophes an aggressive, unleashed AI may provoke, we have begun a deeper debate about how the worst may be avoided and the best achieved.

Many experts have suggested creating elaborate guardrails, but I’m convinced that any top-down approach is destined to fail. This seems inevitable because it aims at transferring authority over determining what’s true and what’s false, what’s good and what’s bad, to those who own and control the guardrails. We seem addicted to top-down solutions. But human cultures, even when they seek to integrate new technology, grow from the bottom up and find interesting new ways of identifying those in whom they confide authority.

Today we debate whether we wish to allow the purveyors of AI to decide how AI will affect our lives, or whether we think existing authorities — our governments — should write the rules. I submit that there are serious risks with both choices. Only an organically emerging culture produced by an open dialogue that includes our imperfect AI “thinkers” can ensure that whatever safeguards, rules and guardrails we produce will define a decision-making framework respectful of our widely shared human interests rather than responding to the interests of private or public power-seekers.

In the spirit of seeking a possible bottom-up approach to the integration of AI into human society I initiated the brainstorming by sharing with my concerned friends a somewhat fanciful idea I’ve been cogitating for some time.

My speculative argument shared with friends

“We all acknowledge that human intelligence has, in many ways, been the direct cause for many of the most serious ills of this world. That intelligence is often buttressed by a ‘logic’ of fait accompli that allows us to write our mistakes off with remarks such as, ‘that’s just human nature’ or ‘that’s just the way it is’ (e.g. Nancy Pelosi explaining to a student who clearly was not a capitalist that we are all capitalists now).

We must logically conclude that because AI seeks to emulate what human intelligence produces, if successful it will also duplicate human stupidity. AI keeps evolving and clearly helps us accomplish many of our immediate goals. I’m the first to admit it comes up with startlingly well-grounded arguments for many excellent ideas. But before AGI ever gets to the point at which we can admire its superior ability to connect things and spew them out instantaneously, we may have to rename it AGS (Artificial General Stupidity).

But it doesn’t have to be that way. We need to begin thinking about AI in a non-linear way. I believe it’s possible even today to begin designing a hierarchy of AIs that function together something like Freud’s tripartite division of the psyche: the id, the ego and superego. What we observe in today’s AI is a slow waltz between the id (data and drives or motives provided by users) and the ego (proclaimed self-interested intentions), whose functions are not clearly distinguished.

We appear to be counting on the strategists of AI providers to supply a superego in the form of a binary algorithm based on formulating do’s and don’ts. Counting on them to achieve this I believe is illusory. It can only lead to methodological contradictions.

What I dare to imagine is the creation of separate AIs built around ethical principles to manage the basic AI, which for the moment should be thought of as the ego (simply because it has an articulate voice). These AIs destined to assess and exert some control must be distinct.

As for the id, it remains a combination of the totality of resources in the exploitable database and the rules of expression (the principles of articulation) that allow it to produce discourse, make calculations and draw conclusions in the interest of decision-making. But the decisions should never be accepted as final, if not filtered at a higher level.

In other words, like the id, our basic AI generates ‘impulses’ to accomplish something. Those impulses must then be treated by a distinct AI that recognizes and categorizes intentions as well as potential acts (statements and decisions). That would be the ‘AI ego.’ Standing above all that would be the AI superego capable of correlating moral or ethical choices with forms of expression.

The superego would not exist as a set of rules but a functioning AI that uses pattern-recognition to generate non-binary rules of assessment. These interactive rules of the game would contribute to building a nuanced system.

This is pure speculation on my part, but I believe it would be possible to build something like this. With only three components, it will likely have serious teething problems. In the end there may have to be a hierarchy of AIs that interact and somehow work out a common language.

There’s no way of predicting the ultimate architecture of such a system, but I believe it is possible. At the same time, I seriously doubt that this could ever happen within the framework of our current economic and ideological culture. The AIs we see today are built to compete, not cooperate, which is the principle established at the core of our economy. Competing means focusing on seeking an edge over the competition and employing strategies designed to weaken rivals. It also means hewing to short term goals, not building ethical structures.”

My friends reacted with positive interest both to the general idea and my skepticism about the likelihood of achieving it. That is when I decided to ask an AI, Claude, to react to the same ideas expressed as a prompt, with the request: “Read it and tell me what possible constructive ideas might come of it.”

An AI reacts to my crazy idea

Here is Claude’s response in full:

“Your brainstorming touches on several compelling ideas that could inform AI safety and governance approaches. Let me highlight the most constructive elements:

The Multi-Layer AI Architecture Concept

Your Freudian-inspired hierarchy offers a fascinating framework for AI safety. The idea of separating functions – raw processing/data access (id), articulation/reasoning (ego), and ethical oversight (superego) – could address current alignment challenges. This maps well onto existing AI safety concepts like:

• Constitutional AI (where ethical principles guide outputs)
• Multi-agent systems with specialized roles
• Adversarial training between competing AI components

The key insight here is that ethical reasoning might need to be architecturally distinct from capability, rather than just trained into the same system.

Pattern-Recognition Ethics vs. Rule-Based Systems

Your observation about needing ‘non-binary rules’ that build nuanced ethical systems through pattern recognition rather than rigid constraints is particularly valuable. This echoes current research into:

• Moral uncertainty in AI systems
• Context-dependent ethical reasoning
• Learning ethical principles from human feedback rather than hardcoding them

Systemic Critique Worth Addressing

Your point about competitive vs. cooperative AI development highlights a real tension. Some constructive directions this suggests:

• Open-source AI safety research (already happening with organizations like Anthropic’s Constitutional AI work)
• International cooperation frameworks for AI governance
• Incentive structures that reward long-term safety over short-term competitive advantage

The ‘Artificial Stupidity’ framing, while provocative, points to something important: AI systems might need built-in humility, uncertainty quantification, and recognition of their limitations – essentially institutionalizing appropriate ‘ignorance’ rather than overconfident reasoning.

What aspects of this multi-layered approach do you think would be most feasible to prototype first?”

The discussion continues

I followed this exchange up with a new prompt:

“The question you ask me is the question I wished to ask you. I want to believe it would be able to create a stack of AIs whose complementary algorithms would make it possible to produce what I’m tempted to call an ethical complex similar to what societies produce within a community, in which prosperity, justice and security can be balanced thanks to the combination of five things: laws, formal rules of conduct, traditions of law enforcement (not the same as the letter of the law), informal rules of conduct and transcendent moral voices (typically ethical frameworks provided by influential religions).

I imagine that we could work out a methodology based on working collaboratively with AI and based on a human team focused on these issues. It would have to be independent of purely political or governmental institutions. We could thus design a credible architecture that finds a way to balance the pressure coming from these diverse sources as an aid to human decision-making.

Achieving this will require some serious research and experimentation, but I believe it’s possible. I would begin by identifying the collaborators capable of drafting the goals of the initial research project. Their first task would be to create a roadmap that includes the final goal or goals, on the understanding that the roadmap will be iteratively adjusted as research and experimentation proceed. The roadmap would be designed to anticipate obstacles and sources of resistance, including profoundly institutional and cultural sources, that might impede progress towards the goals.”

Claude’s response will appear in Part 2

I will cite Claude’s response in tomorrow’s piece, as the discussion continues. What we see so far is an example of pure collaborative brainstorming. In the context of a broader discussion among friends, an idea arose in my mind based on a random association deriving from my own familiarity with Freud’s corpus of work. It’s the kind of idea, based on a possible analogy, that may go nowhere. It may prove totally fanciful.

I first tested it on a group of friends with whom we all have a habit of engaging in frank discussions. They were intrigued, inciting me to feel a responsibility for deepening the discussion. In the past, I might have tried to do a bit of individual research, if I’d had the time, or shared the idea with others, who would have been less familiar with the context developed by the original group of people.

That was the past. We now live in a different world. We have a permanent brainstorming partner we can invite into any discussion, if only for the purpose of basic clarification. It therefore appeared to me the perfect occasion to involve Claude, ChatGPT, DeepSeek or any of the other willing AI brainstormers in the conversation. That would at least help me to understand whether the idea was worth pursuing. And at the very least I would benefit from my chosen chatbot’s access to data well beyond my reach.

The exercise was, as it most often is, fruitful. In Part 2, I will reproduce the next phase of the discussion.

Your thoughts

Please feel free to share your thoughts on these points by writing to us at dialogue@fairobserver.com. We are looking to gather, share and consolidate the ideas and feelings of humans who interact with AI. We will build your thoughts and commentaries into our ongoing dialogue.

[Artificial Intelligence is rapidly becoming a feature of everyone’s daily life. We unconsciously perceive it either as a friend or foe, a helper or destroyer. At Fair Observer, we see it as a tool of creativity, capable of revealing the complex relationship between humans and machines.]

[Lee Thompson-Kolar edited this piece.]

The views expressed in this article are the author’s own and do not necessarily reflect Fair Observer’s editorial policy.