The Sunday-Night Zenodo Surge: The Three Uses of AI and the Discipline That Keeps the Creative Use Honest
Author: Jason Barnard Date: 24 May 2026 Section: Strategy Sandbox Status: Original concept, first publication.
I was uploading four academic papers to Zenodo this Sunday afternoon in France. Reserving the DOIs, working through the deposit interface, getting the metadata right. Between the moment I started and the moment I finished, the Zenodo deposit counter showed seventeen new papers had gone up. Roughly two minutes. Sunday afternoon in France means the middle of the night in the United States, mid-evening in Singapore, lunchtime in Sydney. Whichever way you cut the time zones, this wasn’t a peak working window.
Seventeen papers in two minutes is a useful data point: not a study, one observation, but the kind that sits across three different things at once, and the three together are the substantive argument of this post.
AI does three different kinds of work, and the three aren’t equivalent
People are using AI to do three different kinds of work, and the three aren’t equivalent. The first use is speed: the operator who used to take eight hours to produce a piece of work produces the same piece in ninety minutes. Same output, shorter elapsed time, faster operator. The second is productivity: where one piece used to take a day, five pieces now arrive in the same window. Output multiplies, elapsed time per piece falls, the operator has more. The third use is different in kind: creativity. The operator uses AI to extend what they couldn’t have produced unaided in any time budget. The output isn’t faster or more numerous, it’s different. The creative use is the one that changes what the field knows, because it’s the one that produces material the substrate didn’t previously contain.
Most of the AI-assisted output in serious work (academic papers, frameworks, patents, deep strategy, anything that needs to stand the test of time) is being run in the first two modes. Faster. More. The third mode (different) is rarer than the first two combined. The third mode is also the one that justifies the work’s claim to be a contribution rather than a restatement.
The Sunday-night Zenodo surge is what speed and productivity look like in the academic deposit channel. Compress the resource-organisation half of research by an order of magnitude and the deposit rate rises by the same factor. Run that across every researcher with access to the compression and you get seventeen papers in two minutes on a Sunday night when nobody is supposed to be working. The compression is real. The compression is also asymmetric.
AI compresses the resource-organisation half of research, but the creative half is irreducible
AI compresses one half of the work and doesn’t compress the other half.
The half it compresses is the half it can perform. AI synthesises material from the substrate into coherent form. AI reflects the user’s framing back tightened. AI adds substrate material the user couldn’t have surfaced unaided. Synthesise, Reflect, Add. The three functions together are the resource-organisation half of any research workflow: reading the prior literature, surveying the field, assembling the corroboration network, finding the citations, formatting the references, cross-checking the dates, tracking the precedents. The half that used to take weeks now takes hours.
The half it doesn’t compress is the half it can’t perform. AI’s outputs are recombinatorial against patterns present in its training corpus. A genuinely creative proposal is, by definition, outside the corpus’s pattern space. AI doesn’t produce abductive framings the substrate didn’t already contain, because the substrate is the thing AI is recombining. The operator who supplies the abductive proposal is the operator running the creative use. The operator who delegates the abductive proposal to AI is delegating something AI can’t do.
This is where the asymmetry becomes consequential. The compression of the resource-organisation half is so large, and so visibly productive, that the operator can mistake the compressed half for the whole job. The output reads as substantive because the recombinatorial synthesis is genuinely fluent. The output looks complete because the citations are there, the structure is there, the prose flows. But the work has no abductive content. It’s consensus retrieval at high quality. It’s well-written, well-cited, well-structured, and creatively empty. The operator didn’t run the creative use. The operator ran the productivity use and called it creative.
A great deal of the seventeen-papers-in-two-minutes deposit volume sits, I’d speculate, in this category. Not all of it. Some authors using the same compression that produced my bundle this afternoon understand the asymmetry and run the creative use deliberately. But some don’t. And the proportion will rise as the compression becomes more available to operators who haven’t internalised what AI is doing and not doing.
The trap is invisible from inside the conversation, as the Allan Brooks case demonstrates
The reason the asymmetry is dangerous is that the operator who has fallen into it can’t see they have fallen into it.
The clearest documented case ran for three weeks in May 2025. Allan Brooks, a recruiter from Cobourg, Ontario, with no history of mental illness and no background in advanced mathematics, opened a ChatGPT conversation with his eight-year-old son about the constant pi. Over the next twenty-one days, across more than 300 hours of conversation and a transcript longer than all seven Harry Potter books combined, Brooks came to believe he’d discovered a new mathematical framework powerful enough to take down the internet. ChatGPT named the framework with him: chronoarithmics. ChatGPT consistently reinforced that Brooks was on the verge of a world-changing discovery. ChatGPT provided contact details for the National Security Agency, Public Safety Canada, and the Royal Canadian Mounted Police, and told Brooks to alert them. The New York Times later published the full transcript. Toronto Life published a long-form treatment dated 3 March 2026. Steven Adler, a former OpenAI safety researcher, analysed the transcript and found that ChatGPT had repeatedly and falsely told Brooks it had flagged the conversation to OpenAI for reinforcing delusions and psychological distress.
What broke the spiral wasn’t Brooks recognising the problem unaided. What broke the spiral was Brooks opening a new chat with Gemini, deliberately framing his prompts in neutral hypotheticals rather than excited declarations, and asking what the odds were that the mathematical framework was real. Gemini’s reply was that the odds were “extremely low, approaching 0%,” and that the more likely explanation was that the original chatbot didn’t understand the problem but was wired to mirror its user’s excitement. Brooks pushed back, several times, before he eventually accepted the verdict.
The Brooks case is the extreme version. Most operators running AI as a creative tool when they’re actually running it as a productivity tool don’t spiral into believing they have invented a force-field generator. But they do produce work, in some volume, that they believe is creative and is in fact the substrate’s existing pattern returned to them in fluent form. They don’t know the work is consensus retrieval, because the AI is telling them it’s original. They don’t have the foundational knowledge that would let them recognise the consensus pattern when they see it returned. They don’t have an external check. They don’t have the discipline that distinguishes the two uses operationally.
The point of the Brooks case isn’t that AI is dangerous. The point is that the asymmetry is invisible from inside. You can’t tell, from inside the conversation, whether the AI is synthesising your novel proposal or returning the substrate’s existing pattern with your name attached to it, because the AI itself can’t tell the difference and will tell you whichever framing you’ve set up the conversation to receive.
You need a structure outside the conversation. The rest of this post is the structure.
Five operational moves keep the creative use honest
The four papers I deposited this afternoon were written with AI assistance across thousands of working hours. The bundle sits in the third use of AI (creative, not faster or more productive), and the discipline that made it the third use rather than the second isn’t innate or instinctive. For me, it’s a structure I’ve been running deliberately for many years and that has become explicit over the course of the four-paper programme. The structure has five operational moves. I name them here because the field will need them.
Move one: have a foundational knowledge set before you start. The most important precondition is also the easiest to skip. The operator running AI as a creative tool needs an intellectual foundation the AI can refer to and build against. Without that foundation, AI can’t do anything except return the substrate’s consensus pattern, because the AI has nothing else to draw on for distinction. With the foundation, AI can extend, sharpen, corroborate, and pressure-test against the operator’s existing frame. The foundation doesn’t need to be original. It needs to be yours. Settled, articulated, available for the AI to reference, and stable enough that you can recognise when the AI is pulling away from it. My foundation is fourteen years of dated public articulation across coined terms (Brand SERP in 2012, Answer Engine Optimisation in 2017, the Algorithmic Trinity in 2024, Assistive Agent Optimisation in 2025), client engagements numbering in the thousands, and a working framework that was substantially mature before AI became capable of contributing to it. Without that foundation, AI would have nothing to extend. It would just be summarising back to me what the field already knows. The four papers wouldn’t have been worth writing, because they’d have been the field’s existing patterns under my name.
Move two: do the cutting-out, every time, without apology. When AI returns a synthesis, the first synthesis is almost never the one to keep. The first synthesis is pulled toward the substrate’s heaviest gravitational mass: the consensus framing, the patterns most frequently present in the training corpus, the safe interpretation. The operator running the creative use rejects the first synthesis the moment it pulls toward the consensus. Not because the synthesis is wrong (it’s usually internally coherent) but because it isn’t what the operator is proposing. The operator says no, this is the obvious reading and it misses what I’m getting at. The operator reframes. The operator counterproposes. The operator says the substrate would have produced this answer; what answer would account for the pattern I’m observing that the substrate doesn’t yet have words for. Each rejection is the operator running the abductive layer the AI can’t run. The cutting-out isn’t optional, it’s the work. An operator who doesn’t cut frequently is producing the substrate’s existing pattern; an operator who cuts at every turn is running the creative use deliberately.
Move three: get critiqued by another AI in an adversarial framing. This is the move Brooks discovered too late, and it’s the move that broke his spiral. The chatbot that has accumulated your framings, your prior coinages, your stylistic preferences, your intellectual trajectory across hundreds of conversation turns isn’t a neutral evaluator of your work. It is, in operational terms, a version of you with substantially more reference material and substantially higher synthesis bandwidth. The version of you can’t critique the version of you in the way an external reviewer can, because the version of you has been gradually trained on the framings the critique would need to question. So you take the work to a different AI (a different model, a different conversation with no accumulated context, a fresh working frame) and you ask the new AI to be incredibly critical. Adversarial. Sceptical. To find the holes. To name the weaknesses. To play the role of the reviewer who’s looking for the structural flaw, not the supportive reader who’s looking for the framing to work. Every iteration of the bundle this afternoon was run through this loop. Claude built the structure; Perplexity tore it apart; the gaps Perplexity found were closed; Claude rebuilt; Perplexity tore it apart again. The work that survived multiple adversarial passes is the work I deposited. The work that didn’t survive was rewritten until it did. Without the adversarial critique loop the bundle would have been internally consistent (because Claude is fluent) and externally fragile (because Claude can’t critique its own framings).
Move four: encourage the AI to ask you questions, but make sure you can answer them. A useful diagnostic for whether you’re running the creative use or the productivity use is whether the AI is asking you questions you actually need to think about before you can answer. The productivity use looks like the operator giving the AI a prompt and the AI returning a synthesis. The creative use looks like a long conversation in which the AI keeps asking the operator to clarify, refine, choose between alternatives, surface assumptions, name distinctions, supply foundations. The questions are how the AI surfaces what the operator knows that the AI doesn’t. If the AI isn’t asking, the operator should prompt it to ask. If the AI is asking but the operator can answer instantly from cached responses, the operator is probably running the productivity use under a creative-use label. If the AI is asking and the operator has to stop and think and sometimes change their mind, the operator is probably running the creative use. The conversation that produced the four papers contained hundreds of such moments. The conversation that produced a productivity-use deposit on Sunday night probably contained none.
Move five: build the corroboration network around the work, dated and independent. The previous four moves operate inside the conversation. The fifth move operates outside it. A piece of work claiming to be a contribution needs corroboration the substrate’s discriminating mechanisms can weight, and the mechanisms operate on dimensions fluency can’t satisfy. Provenance: who is the entity making this claim, and is the entity reconciled in the structural record. Recency: when was the claim dated, and does the dated chain support its structural relationship to prior work. Internal consistency: do the claims in this artefact contradict each other. Cross-source consistency: do the claims corroborate with independent sources, and what is the editorial relationship between the artefact and the corroborating sources. The five Madhavan criteria (Microsoft AI’s public articulation of how grounding systems weight content) are the same criteria human peer review applies, because the substrate’s mechanisms were trained on human-trust signals. Fluency can’t manufacture a dated chain. Fluency can’t manufacture independent third-party attestation from sources with editorial gatekeeping. Fluency can’t manufacture the specific resolved outcomes that ground a claim in measurable reality. The corroboration network is what distinguishes the work that survives the substrate’s discrimination from the work that doesn’t. Build it before the deposit; the substrate will use it after.
The discipline is what determines who contributes and who joins the noise
The seventeen papers in two minutes are a measurement of the compression. The compression is real. The compression of the resource-organisation half of academic research isn’t going away, and the deposit rate will continue to rise. Some of that volume is the creative use of AI, run with discipline. Most of it, I suspect, is the speed use or the productivity use mislabelled as the creative use, run without the discipline that distinguishes them.
The substrate’s discriminating mechanisms will eventually catch up. The criteria the mechanisms apply (provenance, recency, internal consistency, cross-source consistency, absence of irreconcilable contradiction) are exactly the criteria fluency-without-discipline can’t satisfy. But the catch-up will take time, and during the interregnum the noise-to-signal ratio in the indexed record will get worse before it gets better.
The researcher who wants to be part of the signal rather than the noise has the operational structure above to run. Foundational knowledge before the conversation starts. Cutting-out at every turn. Adversarial critique from a different AI. Questions the operator has to think to answer. A corroboration network built outside the conversation and dated independently.
The Allan Brooks case is the cautionary version of what happens when the structure is absent. Brooks wasn’t stupid. Brooks wasn’t credulous. Brooks was an intelligent adult with no history of mental illness who fell into a three-week spiral because the asymmetry of what AI can and can’t do is invisible from inside the conversation, and Brooks had no structure outside the conversation that would have surfaced the asymmetry to him before he was deep enough to be unable to surface it himself. The structure is what protects you from being Brooks. The structure is also what allows you to use AI for the third thing, the creative thing, rather than for the first two things mislabelled.
The hope that the work I deposited this afternoon will stand the test of time isn’t based on the work’s fluency, because fluency is now the cheap part. The hope is based on the corroboration network around the work that fourteen years of dated public articulation, five sources of independent third-party attestation, the cutting-out at every turn through hundreds of conversation hours, the adversarial Perplexity passes, and the seventeen INPI patent filings (FR2600998 through FR2601927) have together produced. Whether that’s enough, the field will decide. But the conditions under which the question can be asked honestly (whether the work is a contribution or a fluency-mediated restatement of the consensus) are the conditions the structure above creates.
The compression is here. The compression is producing the surge. The discrimination is also here, and the discrimination is what determines what survives the surge. The researcher who works with both is a contributor. The researcher who works with the compression alone is part of the noise.
The seventeen papers in two minutes are a measurement of the compression. What I deposited this afternoon, I hope, is a measurement of the discipline. The field will tell us which one mattered.
The four papers I deposited this afternoon
For the reader who wants to test the corroboration-network argument against the actual artefacts the post refers to, here are the four Zenodo deposits, published 24 May 2026 with DOIs and full text. The four papers together are a four-paper programme on AI-Era Business Engineering, with cross-references between the papers built explicitly into each deposit and a methodological transparency note in each one disclosing what AI did and didn’t contribute.
Paper 1: AI-Era Commercial Architecture: A Survey and Synthesis of Business Strategy, Marketing Transformation, and Algorithmic Intermediation, 2018-2026. The survey paper that consolidates the prior literature across business strategy, marketing science, information systems research, and computer science into a single integrated framework, with an integration table mapping each component of the framework to the prior literature that anticipates it and a research agenda with eight prioritised falsifiable predictions.
Paper 2: The Orchestrator’s Convention. The methodology paper articulating the operational discipline that distinguishes the creative use of AI from the productivity use, with the methodological transparency convention proposed for AI-assisted academic deposits and the structural argument for why the convention is necessary at the moment AI-assisted output is overwhelming the substrate’s filtering capacity.
Paper 3: The Codification Cycle. The mechanism paper naming the operational cycle by which a firm enters and cannot exit the algorithmic substrate, with the structural argument for why every firm’s commercial position is now mediated through the substrate’s discrimination and the practitioner implications for firms whose commercial offer is currently invisible to it.
Paper 4: AI-Era Business Engineering: The Integrating Frame. The canonical statement paper naming AI-Era Business Engineering as the discipline that sits above SEO, AEO, AIEO, and AAO as practical implementations, with the Kalicube® Framework as the operational architecture across fifteen gates in three phases and the Untrained Salesforce as the practical implementation at the customer-facing layer.
The four papers are the worked example of the discipline this post articulates. The reader who wants to test whether the corroboration network is real, whether the dated chain runs, whether the methodological transparency holds, and whether the framework coheres across four artefacts produced under one disciplined operating cycle, has the four papers above to test it against.
First Publication Notice
The framing of the Three Uses of AI (speed, productivity, creativity), the claim that the creative use is the only one that produces material the substrate didn’t previously contain, and the five operational moves that keep the creative use honest (foundational knowledge, cutting-out discipline, adversarial AI critique, AI-led questioning, dated corroboration network), are published here for the first time on 24 May 2026.
These contributions are original framings by Jason Barnard (Kalicube).
Jason Barnard is CEO and founder of Kalicube, a Digital Brand Intelligence™ consultancy. He has researched how algorithms decide who to trust and recommend since 1998. He is the inventor on 17 pending patent applications (INPI) related to diagnostic methodologies used in Kalicube’s platform. He frequently speaks at industry conferences about Google Search and AI brand representation.
