3 Aug 2026
Your Work Is Not Being Judged. Its Machine Version Is.
There is a new hidden evaluator between your work and the person who decides its value.
It may not make the final decision.
But it can retrieve your work, divide it into fragments, extract selected claims, summarise the evidence and construct the version that reaches the human evaluator.
That version may be weaker than the original.
Your contribution may disappear.
Your evidence may become disconnected from the claim it supports.
Your measurable results may be reduced to vague promises.
A decisive fact may remain visible on your website while never entering the answer produced by the AI agent.
This episode of InnovEU, the EU Project Chronicles is built around selected excerpts from my recent live workshop on LLM-assisted evaluation.
Listen to the episode
In the episode, I examine how machine mediation is already affecting:
* academic peer review;
* grant proposals and funding applications;
* websites and landing pages;
* legal and professional documents;
* marketing and product information.
The evidence is not equally strong in every area.
Academic peer review already offers documented examples of AI-assisted reviews influencing the content of evaluations.
AI-assisted grant screening is emerging, contested and increasingly regulated.
Websites and landing pages face a different form of evaluation: AI agents decide whether facts are discoverable, extractable, trustworthy and strong enough to enter a comparison or recommendation.
Different settings.
The same underlying risk.
Your work is not being judged only by what it contains. It is being judged by what the machine manages to extract from it.
The machine version of your work
During the workshop, I demonstrated two versions of the same grant-proposal section.
The project was identical.
The problems were identical.
The activities, outputs and measurable results were identical.
But the first version required the evaluator to infer the relationship between them.
The second made the evaluation architecture explicit:
Criterion → gap → objective → activity → output → target → indicator
The project did not improve.
The evidence did not improve.
The machine representation improved.
That distinction matters because a weak extraction can turn excellent work into a weak verdict.
The Three Extraction Gates
The episode introduces three questions every important claim must survive.
Disclosure
Is the decisive fact explicitly stated?
“Flexible pricing” is not a price.
“This project will create significant impact” is not an indicator.
“This paper contributes to the literature” is not a contribution.
Extractability
Can the system connect the claim to the evidence supporting it?
A problem on page four, a solution on page twelve and an indicator in an annex may remain three unrelated fragments.
Accessibility
Can the system retrieve the information at all?
The evidence may be trapped in a screenshot, scanned PDF, inaccessible webpage, collapsed menu, calculator or ambiguous table.
The information can be present and still be functionally absent.
Present is not the same as extractable.
Writing for machines without sounding like one
The answer is not to remove your voice.
It is not to flatten every paragraph into repetitive machine prose.
It is not to manipulate an evaluator through hidden prompts.
The answer is a two-pass process.
Pass 1: machine legibility
Make the following elements explicit:
* criterion;
* claim;
* evidence;
* measurable result;
* indicator.
Pass 2: human persuasion
Restore:
* narrative;
* voice;
* relevance;
* tension;
* memorable language.
First make the value extractable. Then make it unforgettable.
Read the companion Decision File
The full evidence map, academic references, documented cases, limitations and strategic implications are available in the companion Polis Doxa Decision File:
Your Work Is Not Being Judged. Its Machine Version Is. (https://open.substack.com/pub/fernandocgaspar/p/how-llms-judge-papers-grants-websites-legal-texts?r=9fdfx&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true)
The article distinguishes between:
* documented use;
* experimental evidence;
* institutional signals;
* emerging practice;
* strategic inference.
That distinction matters.
Guru guessing is not strategy.
Related Polis Doxa reading
Marketing to Machines (https://open.substack.com/pub/fernandocgaspar/p/pmarketing-to-machines-grant-writin?r=9fdfx&utm_medium=ios) explains why professional writing now has a second audience.
Algorithmic Rents examines how platforms gain power by controlling visibility and allocation without owning the underlying assets.
Private Sovereignty explores how platforms acquire practical authority over decisions once reserved for public institutions.
The Governability Crisis explains why governments may be regulating visible AI systems while missing the economic structures beneath them.
Apply the method to your own work
How to Write for LLM Evaluation is a seven-step digital experience for papers, grant proposals, websites, landing pages, marketing copy and professional texts.
You finish with one real section from your own work rewritten, tested and prepared for both machine extraction and human judgment.
The First Edition includes:
* seven written lessons;
* four downloadable PDFs;
* six practical worksheets;
* ten copy-and-paste prompts;
* one anonymised before-and-after case;
* the final LLM evaluation checklist.
Enrollment closes on 31 August 2026.
Access opens on 1 September 2026.
Join How to Write for LLM Evaluation (https://payhip.com/b/Mp9Hx?utm_source=chatgpt.com)
Get the weekly Decision File on AI/digital transition, energy shocks/green transition, trade wars/transition to a New globalization, war economy/Transition to a War Economy and housing crisis.
Get full access to Polis Doxa - The Transitions Letter at fernandocgaspar.substack.com/subscribe (https://fernandocgaspar.substack.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4)
