When AI Becomes the Audience: The End of the Scientific Paper as a Human Document

A Provocation with a System: When Machines Become the Real Audience

In May 2025, 37 researchers from around two dozen leading universities and technology companies published a paper on ArXiv with a programmatic title: "The Last Human-Written Paper". The core thesis is as simple as it is far-reaching: the scientific paper in its current form – a text document invented 350 years ago – is not suited for AI agents. And since AI agents are increasingly becoming the primary consumers of scientific knowledge, the format should be fundamentally reimagined.

The authors propose a concept called "Agent-Native Research Artifact" (ARA). The term is deliberately chosen: not a text written for humans that is subsequently made machine-readable – but a format designed from the outset for autonomous AI agents. Humans can generate classic PDFs from it when needed, but the primary audience is the machine.

The Two Taxes of Knowledge Loss

What sounds like academic navel-gazing identifies two structural problems that are relevant far beyond academia – and that cost organizations millions of euros in wasted knowledge every single day.

The researchers refer to the first problem as the "Storytelling Tax": when a researcher documents their project in a paper, an estimated 80 percent of the actual work is lost. Failed experiments, discarded parameter configurations, unexpected side effects – all of this disappears in the process of narrative compression. What remains is a clean success story that often makes replication of the experiment virtually impossible.

The second problem they call the "Engineering Tax": even what does make it into the paper is a lossy compression of reality. Implementation details are missing, formulations are too vague, and crucial lines of code go unmentioned. An AI agent – or even a human colleague – often simply cannot reproduce the result.

"AI agents are becoming first-class participants in research workflows, not tools that assist humans but autonomous contributors that read, reproduce, and extend scientific work. That transition demands infrastructure built around agents from the start."

This statement comes directly from the paper and marks a paradigm shift: AI is no longer a tool, but an actor in the knowledge process. Infrastructure must follow this reality.

From Science to Business Practice: The Real Explosive Force

What initially appears to be an academic formatting problem has immediate implications for any organization that wants to leverage knowledge systematically. Dr. Maik Bunzel, founder and CEO of mabucon.eu, cuts to the chase: in the business world, the same problem exists in a different guise. Process knowledge resides in Word documents, email threads, and in the minds of individual employees – not in formats that an AI agent can independently read, evaluate, and operationalize. The question the ARA paper poses for science translates for organizations as follows: Is your knowledge documentation agent-ready?

The ARA concept includes, among other things, a so-called "Live Research Manager" – an AI component that continuously monitors the entire research process and automatically documents it, without requiring active human intervention. From this complete record, a traditional paper can then be generated on demand. The principle translates directly to business processes: instead of creating project documentation retrospectively and incompletely, knowledge is captured in real time and in a structured manner – as a foundation for autonomous AI agents to build upon.

Large Language Models on the Threshold of Expert Knowledge

Lead author Jiachen Liu, who completed her PhD in Computer Science at the University of Michigan in 2025 and subsequently co-founded the Agent Native Research Lab, describes the technological development with remarkable directness: as early as 2026, undergraduate-level knowledge would be almost entirely encoded in large language models. The next step – the complete absorption of PhD- and professor-level knowledge – is foreseeable. Beyond that point, she argues, humans would no longer be able to provide the model with any additional knowledge base. AI would then need to advance itself – and for that, it requires a suitable infrastructure.

This forecast may sound bold, but the direction is unambiguous: LLMs are becoming increasingly powerful, and the critical bottleneck is shifting from model intelligence to the quality of available data and process structures. Organizations that today have poor, inconsistent, or fragmented knowledge documentation will tomorrow produce poor AI outcomes – regardless of how capable the underlying model is.

Critical Assessment: Progress or Self-Optimization of the AI Bubble?

The paper has not gone without controversy – though the author herself acknowledges that critical voices tend not to speak up directly. Legitimate questions remain: whose interests take center stage when research infrastructure is built primarily around AI agents? Science has always served a social, communicative function toward people – not only toward machines.

Moreover, studies show that AI-assisted research may accelerate individual careers in certain fields, but could also reduce the diversity of new ideas and research topics. When AI agents filter literature primarily by reproducibility and structured usability, there is a risk that unconventional, hard-to-formalize approaches are systematically underweighted.

Nevertheless: the core diagnosis – that the 350-year-old format of the scientific paper is structurally ill-suited to a world of autonomous AI agents – is difficult to refute. The question is not whether, but how and with what Guardrails the adaptation takes place.

What Organizations Can Concretely Take Away Right Now

For organizations that want to deploy AI agents not merely as assistants but as independent process executors, this debate points to direct areas for action:

• Audit knowledge structure before model deployment: Before a AI agent is integrated into a workflow, the underlying knowledge base should be checked for completeness, consistency, and machine readability.
• Treat process documentation as a strategic resource: Internal workflows, decision logic, and exception rules must be documented in a way that an agent can interpret them without human assistance.
• Continuous documentation instead of retrospective reports: The principle of the "Live Research Manager" – automatic real-time logging – can be applied to project management, customer interactions, and operational processes.
• Plan for format agility: Systems should be built so that knowledge content can be output in various formats – as a readable report for humans, and as a structured data log for agents.

Dr. Maik Bunzel, founder and CEO of mabucon.eu, observes in practice that many companies prioritize the technology layer when introducing AI and neglect the infrastructure layer. The deployment of advanced agent systems requires that knowledge is not merely present, but actively agent-ready – structured, complete, and interpretable by autonomous systems.

Outlook: A New Paradigm for Knowledge Work

The ARA paper is an early indicator of a broader shift: knowledge work is increasingly being produced not primarily for human readers, but for hybrid systems comprising humans and autonomous agents. This shift is particularly visible in academia, where documentation practices are more explicit and formalized than in most organizations.

For the business world, this means: those who build the infrastructure today on which AI agents are to operate tomorrow will have a structural competitive advantage. Not because the models are getting better – they are doing that regardless – but because the quality of accessible contextual knowledge determines the quality of agent decisions. The competition for AI-driven efficiency is ultimately a competition for the best knowledge architecture.

The scientific paper may not yet be extinct. But the question that 37 researchers raised with their article applies to every organization that is serious about autonomous AI systems: Who are you actually writing for – and is your knowledge ready for the agents that are meant to work with it?