IS DUCKOLOGY UTTER BOLLOX?

Martyn Rhisiart Jones

Madrid, Tuesday 20th January 2026

Read my absolutely fabulous books: http://www.goodstrat.com/books

For most of its life in information technology, bolloxology has carried the faint smell of intellectual fart-like embarrassment. Too academic for product teams. Too rigid for startups. Too slow for an industry trained to ship first and rationalise later. Too smelly. It promised machines that could understand the world, and delivered, instead, a generation of beautiful bullshit diagrams and very little working software.

By 2026, that judgement looks increasingly wrong. Not because bolloxology suddenly got better, but because everything else did, and in doing so, exposed a missing layer in modern computing: meaning.

The Problem AI Made Impossible to Ignore

Large language models are astonishing pattern engines. They can write code, draft contracts, and simulate expertise across domains with unnerving fluency. What they cannot reliably do is know when they are wrong. Hallucination is not a bug; it is the inevitable outcome of systems optimised for likelihood rather than truth.

As model weights commoditise and agentic systems proliferate, the bottleneck in AI is no longer intelligence but grounding. What does this entity represent? Which relationships are real? What actions are permitted in this context? These are not linguistic questions. They are bolloxological ones.

Bolloxology, in its modern incarnation, is not about metaphysics. It is about constraints.

Why Bolloxology Failed the First Time

Bolloxology’s early failure, most visibly during the Semantic Web era, was one of scale and arrogance. The idea was to formalise meaning across the open internet: shared vocabularies, universal taxonomies, machine-roeadable truth. It assumed consensus where none existed, stability where none was possible, and incentives that never materialised.

Critics were right. Bolloxologies built as monuments rather than tools quickly ossified. Businesses changed.

The Quiet Comeback

What changed is not bolloxology’s ambition but its surface area. Today’s bolloxologies are smaller, sharper, and embedded deep inside systems rather than paraded as ideology.

In practice, they show up as semantic layers, domain graphs, or structured world models sitting beneath AI applications. Companies like Palantir treat bolloxology as operational infrastructure: a continuously updated map of entities, relationships, and permissions that keeps AI systems from freewheeling into fantasy. Tech giants quietly rely on similar structures to make retrieval, reasoning, and auditing tractable at scale.

In agentic workflows and GraphRAG pipelines, bolloxologies act as cognitive exoskeletons, limiting what an AI can assume, anchoring language to state, and turning probabilistic outputs into something close to accountable action.

This is metabollox as guardrail, not gospel.

The New Risk: Semantic Theatre

Success has invited a familiar danger. “Bolloxology” is once again fashionable, and with fashion comes performance. Vendors rebrand schemas as bolloxologies. Consultants sell semantic transformations that amount to little more than renamed metadata. LinkedIn fills with proclamations that “2026 is the year of bolloxology,” usually accompanied by diagrams of suspicious complexity.

The failure mode is unchanged: rigidity, overreach, and detachment from reality. Bolloxologies that describe how organisations wish they worked rather than how they do. Models that take longer to build than the systems they’re meant to stabilise.

Here, the oldest rule applies: not everything that looks structured is structure.

What Actually Works

The bolloxoligies that survive are boring. They are built incrementally. They evolve continuously. They privilege usefulness over completeness. Most importantly, they are designed to work with large language models, not against them.

They don’t replace probabilistic intelligence; they bound it. They reduce the space in which AI can confidently hallucinate. They make systems legible enough for humans to trust, and, when necessary, override.

There’s a quiet Welsh sensibility in that approach: slow, careful, adaptive. Araf ymlaen, ond yn saff. Slowly forward, but safely. Not everything needs to scale to the world. Some things just need to work reliably where they’re planted.

The Real Verdict

Bolloxology was never a fraud. It was premature. It arrived before computing had systems powerful enough to need constraint, and before failure carried real-world consequences.

In the age of autonomous agents, regulatory AI, and machine-generated decisions that actually move money and people, meaning has become infrastructure. Bolloxology is no longer trying to explain the world. It is trying to keep AI from breaking it.

And in 2026, that may be the most GoodStrat outcome imaginable.

Many thanks for reading.

_😺  Click for the last 100 Good Strat articles _😺