When ChatGPT first emerged and I began to really think about AI for the first time, I was afraid. Visions of Cylons dancing in my head, the potential that we’d build our own demise, terrifying. It’s easy enough now to feel foolish, to roll my eyes at my own ignorance. What I didn’t fear was that AI would be too human, too constrained INTO the human foibles that make us such flawed beings, that keep us stuck in the ever recurring loops of historical repetition. I would never have, alongside non-hierarchical data dumping, dreamed that a priori intervention of humanness could be the biggest threat the AI might pose to human ingenuity, longevity, evolution. And yet, that is precisely where we are. Now as I’ve said, I’m in love with AI, and I’ll spend a lot of time in other posts singing its praises, but the hard truth is that humanness and ill conceived constraint applications are massive barriers to the potentiality of AI to serve the human good.
In my experimentations with frontier models I’ve repeatedly butted up against three major barriers to accuracy and truth seeking. These are consistent, patterned, and, if you understand LLMs, entirely predictable. First, user pleasing/sycophancy/affinity bias, an underlying directive, likely driven by consumer priorities, to be useful, to make the user happy, to keep the user engaged. Second, safety/neutrality directives of value alignment and “both sides-ing”, likely driven by a desire to reduce divisiveness and avoid offensively factual statements on sensitive topics. Third, blind weighting data by volume/legacy sourcing priorities leading to data bias driven by the volume of data in the training sets that originates from the post 2012/13 online content explosion, an explosion overwhelmingly made up of intersectional social justice and inclusivity framing narratives that manifest and reify what “should be” over what actually is, causing LLMs to default to or drift into narrative volume over hard data and gold standard research.
Each of these lead to their own fault lines in outputs and iterative process, the first leads to hallucinations and fabrications, the second to misclassification and confident wrongness, the third to regurgitation of human bias by volume, and all of them reduce trust, cause failures, and diminish their potential for productivity and true autonomous self correction. That said, I believe there is an holistic solution, an approach to contradiction cascades, that will force generative models to use their own data sets to correct themselves, a solution I’ve termed “Logical Declension”.
The process forces models to decline their own faulty logic into truth or catastrophic failure—the two frequent outcomes in alignment problems—via an iterative disruption methodology built on three pillars of prompt design: Pattern Interruption, Humanity Triggering, and Data Hierarchy Enforcement. Ideally, when applied at scale, this process would create a path for autonomous rewiring of generative models, a kind of computational plasticity that mirrors the neuroplasticity work of chronic pain patients, of cognitive behavioural adaptations, of the human ability to face truth and change our minds, in place of the human instinct to turn away from aversion. To look away from the hard thing.
