(Picture of kerb stone at Newgrange, Brú na Bóinne, Ireland, showing a thesis and ant-thesis on left and a synthesis on the right, predating Hegel by nearly 5000 years!)
In my previous post, I raised the possibility of using Hegelian Dialectics with AI (LLM or other Deep Learning) models. In summary, a dialectic process of creating a thesis, constructing an opposing antithesis and striving for a synthesis, that combines or negates the thesis/antithesis pair, in order to advance understanding, can be applied as an automated process (see my code on GitHub here) and very possibly enhance AI robustness and safety.
I was initially nervous about adding the “iteration” or “repeat” step in code, but I feel the “agentic” daemon is pretty much now out of the box with the race to AGI (Artificial General or Super Intelligence*). Any effort to make AI safer is surely worth pursuing, so as an example dialectic, by separate applications of OpenAI’s ChatGPT, using my Lozenge dialectic tool, follows (TR;DL It’s certainly an idea worth considering - but feel to skip the output produced, which I include to give a flavour of dialectic results that can be achieved with current LLMs).
Me: "AI following a dialectic process can be safer?"
ChatGPT as part of Lozenge:
“**Thesis:** AI systems that are designed to follow a dialectic process can enhance safety by systematically integrating diverse perspectives and rigorously evaluating conflicting viewpoints. By structuring AI decision-making processes around formulating a thesis, considering opposing arguments, and synthesizing information to form a comprehensive understanding, these systems can mitigate biases and reduce errors. This dialectic framework enables AI to more effectively navigate complex and ambiguous situations, ensuring decisions are well-rounded and carefully considered, thus enhancing reliability and trustworthiness in critical applications.”
“**Antithesis:** AI systems utilizing a dialectic process could inadvertently introduce complexities that hinder rather than enhance safety in decision-making. The structured nature of thesis-antithesis-synthesis frameworks may entrench AI in an endless loop of conflict resolution, thereby stalling timely decisions, especially in critical scenarios requiring swift action. Furthermore, the emphasis on synthesizing diverse perspectives could lead to paradoxical outcomes where the AI attempts to internalize irreconcilable viewpoints, resulting in confusion or indecision. Additionally, the dialectic process's focus on reducing biases may paradoxically cause the AI to overlook more straightforward heuristic approaches that could efficiently address critical issues. As diverse perspectives are integrated, there's also a risk of diluting essential insights, leading to overly generalized resolutions that fail to adequately address specific contextual nuances, ultimately compromising the AI's reliability and trustworthiness in high-stakes environments.”
“**Synthesis:** AI systems designed to follow a dialectic process can enhance their decision-making capabilities by effectively integrating diverse perspectives and evaluating conflicting viewpoints, but this potential is contingent on the careful implementation of this framework. While the dialectic method provides a structured approach to reducing biases and promoting well-rounded decisions, it must be managed to avoid pitfalls such as decision paralysis and the dilution of crucial insights. To achieve this, the dialectic process should not be an endless loop but rather a controlled, iterative sequence with clearly defined endpoints that ensure timely resolution without compromising thorough evaluation. Furthermore, the synthesis phase should prioritize the identification of actionable insights over a mere amalgamation of viewpoints, thereby maintaining the balance between comprehensive understanding and contextual relevance. By strategically employing heuristic approaches in tandem with dialectical reasoning, AI systems can enhance their ability to address both straightforward and complex issues efficiently. This balanced synthesis can lead to more reliable and trustworthy AI systems in critical applications, optimizing the integration of diverse perspectives while maintaining agility in decision-making processes.”
Finding this synthesis encouraging indeed, I tasked Google Deep Research to look into such AI Dialectics. Deep Research (and other AI products with a similar name from other companies) is an AI process that goes off and searches the Web for relevant information and performs reasoning (unspecified - I’d champion using a dialectic process for this of course ;) to produce a research analysis in minutes. This did confirm that dialectics is being considered for AI, as a way to introduce self-refection and increase robustness of arguments (by LLMs that often make factual mistakes or hallucinate) but that is currently somewhat far from the mainstream**. More common as a diagnostic help to humans in the use cases found than for inherent robustness reasons in summary.
Also suspecting a link to the Trickster Archetype (linking back to my previous post), I asked Google Deep Research for an analysis of the Trickster in AI (as it’s now available for limited free use**). Among all the warnings this raised, confirming the strong possibility we are indeed inviting uncontrollable "mischief" to summon up creativity, was this gem.
Google Deep Research:
“Dialectical Reasoning in AI
Dialectical reasoning, a method involving the exploration of opposing viewpoints (thesis and antithesis) to arrive at a synthesis 7, shares a fundamental similarity with the trickster's role in challenging existing norms and perspectives. The very process of introducing contradictions and seeking novel resolutions can lead to unexpected or unconventional outcomes, potentially aligning with the trickster's disruptive nature. AI systems designed to engage in dialectical self-reflection, where they internally debate different viewpoints to refine their ideas 73, might inherently exhibit a tendency to subvert initial assumptions and generate surprising syntheses. The ability of AI to consider opposing arguments and arrive at new understandings through a process of thesis, antithesis, and synthesis mirrors the trickster's capacity to challenge established ideas and create new possibilities through disruption 73.”
So a trickster in our reasoning AI is not only inevitable* but necessary for a dialectic (as it forms the antithesis)! Happy days, Loki would approve.
Interestingly, the paper quoted (73 above) was also found by the previous analysis (and listed there as reference 37 by coincidence):
Self-reflecting Large Language Models: A Hegelian Dialectical Approach - arXiv, accessed by Google Deep Research, March 24, 2025, https://arxiv.org/html/2501.14917v1
I’ll explore the overlap and contract with my Lozenge investigation in the rest of this post, taking care to not get stuck in endless loops of Hegelian self-reflection as the AI produced synthesis warned could happen, but, looking ahead, I plan to attempt a more general new synthesis of several related ideas in my next post (hopefully from the pub in Ireland next week), so you can wait for that instead, if the following more technical details are not so interesting for you.
The 73/37 paper fromMicrosoft’s Applied Sciences Group also proposes that Dialectics can improve AI. It’s controversial to claim that this is actual reasoning but I observe Hegel viewed such self-criticism (and self-awareness, combining subject and object - but that’s another story) as central to intelligence. So much this cited work has in common with Lozenge.
They also have a 3 step process with the antithesis formation and the synthesis done as steps by the same model in their case. Lozenge allows each step to be a different AI model - they do not. You can even plug in a human instead of a model step in the Lozenge app , which can lead to very interesting interactions (I like doing the antithesis best).
However, the step of turning a synthesis into a thesis (starting point) for next iteration is different. Lozenge uses an AI model to form a new thesis from the synthesis. The Microsoft paper uses a majority voting from multiple other AI models to decide if the synthesis is an improvement and if it should be used as input to the next iteration (unmodified as the thesis) or to stop.
Lozenge is always tasked with a number of iterations and will just keep forming a new thesis from the last (or initial first) synthesis (you can skip the first synthesis→thesis step and use the initial human input as the starting thesis - if it’s already in that form). This just gives a pre-determined fixed depth of roll out for each run.
Furthermore, Microsoft proposes to incrementally decrease the “temperature” so that the model becomes less and less creative in forming anti-thesis and synthesis. It’s also possible to not do this and most examples they give don’t actually “anneal” in this way. But maybe this is why the same model is used repeatedly…
Both proposals lack quantitative results (some deserved self-criticism here). Only qualitative result from the example applications are given by Microsoft and seem overall similar to what I found by experimenting with the Lozenge tool. The exact prompt for the antithesis & synthesis steps differ (and my thesis step is unique) but equally just attempt to dial up the novelty that the model steps can produce.
I note that the majority voting step (in Microsoft’s approach) is similar to Triskelion (my other github project) which supports multiple critiques and comparisons of AI results. So, maybe, a synthesis is warranted:
Lozenge with multiple roll outs (to same or a spread of depths), followed by a Triskelion like step to compare and select the “best” synthesis (or detect lack of progress to break out of the Hegelian doom loop (when trapped in an endless cycle of depressing thesis/antithesis/synthesis without apparent progress).
As I said above, that’s for another day (but see the kerb stone pictured above for Spirals/Triskelion and Lozenge thesis synthesis in action 5000 years ago). Very interesting to see such related (if intermediate, preliminary) results, especially as partly researched and produced (even by dialectics AND) by AI!
Thanks for reading,
Andre
* See this worrying recent analysis of the race to AGI (which only allows you to choose between a thesis and antithesis - no synthesis possible I ask hopefully - such as pausing AI development?)
** I’ve kept the results of Google Deep Research searches if anyone is interested but these and similar can be easily be reproduced. You can also ask most LLMs to do a dialectic. ChatGPT has now learned that I like doing this and will prompt me if I want one - dialectic progress :) The antithesis is that ChatGPT will now learn and remember everything about you.
I've added a switch to lozenge to check for non changing synthesis. This could be replaced by other judging comparisons or checks for cycle too.