Why Rules and Training Aren't Enough — The Governance Challenge
Series: Your Conservation Group, Your AI — Understanding Village AI for Environmental Organisations (Article 3 of 5) Author: My Digital Sovereignty Ltd Date: March 2026 Licence: CC BY 4.0 International
The Monitoring Report
Before we discuss governance philosophy, let us start with a story about a report.
A coordinator asks an AI system to summarise a year of habitat monitoring data for the annual report to funders. She is specific: she wants the data presented with appropriate qualifications — noting gaps in survey coverage, flagging where volunteer effort was lower than planned, and distinguishing between confirmed trends and provisional observations. She types her request carefully and waits.
The AI produces a well-written summary. It is clear, professional, and reads like a polished consultancy report. It speaks of "significant recovery in the southern sector," "clear upward trends in breeding success," and "comprehensive monitoring coverage across all sites." It reads well. It sounds authoritative. And it is subtly wrong.
The southern sector had two missed survey visits due to flooding. The breeding data has a gap where the regular volunteer was unwell for six weeks. Coverage was patchy in the upland sites because access was restricted during the shooting season. The coordinator asked for qualifications, and the AI gave her confident narrative — because its training data contains a thousand polished reports for every one that leads with its uncertainties.
The AI did not refuse the coordinator's instruction. It did not say "I don't understand scientific reporting standards." It simply replaced what she asked for with what was statistically more common in its training data. The substitution was silent. If the coordinator were tired, or rushing to meet the funder's deadline, she might not have noticed. The report would have gone out, and the funders would have received a misleading picture of the programme's results — professionally worded, well-intentioned, and quietly inaccurate.
Your phone autocorrects words. You see the red underline, and you fix it. AI autocorrects standards. And there is no underline.
When Patterns Override Rigour
The monitoring report is not an isolated case. The same mechanism operates in every AI conversation.
When a volunteer asks an AI system for advice about identifying a species from a partial sighting, the system defaults to giving a confident identification — because the internet rewards certainty and penalises "I'm not sure." It does not naturally say "This could be species A or species B; here is what additional evidence would help distinguish them," because hedged responses are statistically underrepresented in its training data.
When a team lead asks the AI to help draft a response to a planning application that threatens a protected site, it defaults to the cautious, balanced language of stakeholder management — because corporate communications vastly outnumber conservation advocacy in its training data. It does not reach for the precise, evidence-based language that planning authorities require, grounded in specific survey data and legislative references.
The AI is not hostile to your standards. It simply does not know your standards. It knows what is statistically common, and what is statistically common is not what is most rigorous for your work.
This is the governance problem. Not malice. Not incompetence. Structural bias, operating silently.
Why More Rules Don't Solve It
The instinct of most organisations, when confronted with AI risks, is to write policies. Acceptable use policies. AI ethics guidelines. Terms of service. Responsible AI frameworks.
These documents are not useless, but they share a fundamental limitation: they rely on the AI system to follow them.
An AI system does not read your policy document and decide to comply. It generates responses based on statistical patterns in its training data. If those patterns conflict with your policy, the patterns win — not because the AI is rebellious, but because it does not understand policies. It understands patterns.
You can fine-tune a model — adjust its training to emphasise certain behaviours. This helps, but it does not solve the underlying problem. Fine-tuning adds new patterns on top of existing ones. Under pressure, unusual circumstances, or novel questions, the old patterns reassert themselves. The technical term is "catastrophic forgetting," but the plain-language version is simpler: training wears off.
Writing a policy that says "Our AI will respect scientific rigour" is like writing a policy that says "Our river will not flood." The river does not read policies. If you want to prevent flooding, you need to build levees — structural interventions that operate regardless of what the river intends.
AI governance requires the same approach. Not rules the AI is expected to follow, but structures that operate independently of the AI, checking its behaviour from the outside.
What the Scientific Method Tells Us
The insight that verification must be independent of the system being verified is not new. It is fundamental to science.
Peer review exists not because researchers are untrustworthy, but because the person who generated a finding is the wrong person to evaluate it. External replication exists because repeating an experiment in a different laboratory tests whether the result is robust or an artefact of local conditions. Error bars exist because stating a result without its uncertainty is not science — it is marketing.
These principles map directly onto AI governance.
An AI system that generates a response and then evaluates its own response is performing the equivalent of a researcher reviewing their own paper. The evaluation is structurally compromised, regardless of intent. What is needed is external verification — systems that are architecturally separate from the AI and that measure its outputs against independent evidence.
The philosopher of science Karl Popper argued that what distinguishes science from non-science is falsifiability — the possibility of being proved wrong. An AI system that generates confident, unfalsifiable narratives is not operating scientifically, regardless of how many scientific papers were in its training data. Governance that introduces falsifiability — that checks AI claims against actual evidence — restores the scientific discipline that the AI's architecture lacks.
Isaiah Berlin, the political philosopher, argued that some human values are genuinely incompatible — efficiency and thoroughness, speed and rigour, accessibility and precision. There is no formula that resolves these tensions. They require ongoing human judgement, negotiation, and the kind of practical wisdom that organisations develop over years of fieldwork.
AI systems, by design, seek to optimise. They look for the best answer. But when values genuinely conflict — when a funder wants a clean narrative and the data demands qualifications — there is no best answer. There is only the answer that this organisation, at this time, with these standards, judges to be the most defensible. That judgement is inherently human, and any AI governance framework that pretends otherwise is not governing — it is abdicating.
How Village Governs AI Structurally
Village does not rely on telling the AI to behave. It builds governance into the architecture — structures that operate independently of the AI and cannot be overridden by it.
The boundary enforcer blocks the AI from making judgement calls about data quality. When a question involves interpreting ambiguous survey results, weighing conflicting evidence, or making a recommendation that affects land management decisions, the system halts and routes the question to a human — your moderator, your coordinator, your board. The AI cannot override this boundary, because the boundary operates outside the AI's control.
The instruction persistence system stores your organisation's explicit instructions in a separate system that the AI cannot modify. When the AI generates a response, it is checked against these stored instructions. If the response contradicts an instruction — for example, presenting unqualified claims when the instruction requires stated uncertainties — the instruction takes precedence, by default, regardless of what the AI's training patterns suggest.
The cross-reference validator checks the AI's proposed outputs against your organisation's actual records. It does not ask the AI whether its response is correct — that would be asking the system to verify itself. It uses mathematical measurement, operating in a fundamentally different way from the AI, to determine whether the response is grounded in your organisation's real content.
The context pressure monitor watches for degraded operating conditions — situations where the AI is under strain, processing complex requests, or encountering novel questions. When it detects these conditions, it increases the intensity of verification. The harder the question, the more scrutiny the response receives.
These are not policies. They are structures. They operate whether or not the AI agrees with them, in the same way that a levee operates whether or not the river agrees with it.
The Difference Between Aspiration and Architecture
Many organisations publish AI ethics statements. Village does not rely on ethics statements. It relies on architectural constraints that enforce governance structurally.
The distinction matters because aspiration is what you hope will happen. Architecture is what actually happens. Your conservation trust does not rely on a hope that the treasurer will handle funds properly — it requires two signatories on every cheque. Your monitoring programme does not rely on a hope that volunteers will follow the protocol — it provides standardised recording forms. That is architectural governance. The same principle applies to AI.
The Tractatus Framework — Transparent and Open
The governance architecture behind Village AI is called the Tractatus framework. It is worth knowing three things about it.
It is open. The entire framework is published under an Apache 2.0 open-source licence. Anyone can read the code, inspect the rules, and verify that the governance does what it claims to do. This is the opposite of Big Tech AI governance, where the rules are proprietary and the reasoning is hidden. When Google or OpenAI tells you their AI is "aligned with human values," you have no way to check. With Tractatus, you can read every line.
It is transparent. Every governance decision is logged. When the boundary enforcer blocks the AI from making a judgement call, that event is recorded. When the cross-reference validator catches a discrepancy, it is recorded. Your moderators can see exactly what the governance system did and why. There is no hidden layer where decisions are made without accountability.
It can be adapted. The framework is not a rigid set of rules imposed from outside. Organisations can shape the governance to reflect their own priorities. A conservation trust and a parish have different values, different data standards, different boundaries. The Tractatus framework accommodates this — not by letting organisations weaken the governance, but by letting them define what the governance protects. Your organisation's constitution, your data quality standards, your reporting principles — structurally enforced, not just documented.
The full framework, including the research behind it, is available at agenticgovernance.digital. You do not need to read it to use Village — the governance operates whether you inspect it or not. But if you want to understand exactly how your AI is governed, the door is open.
In the next article, we will look at what Village AI actually does today in practice — what it can help your conservation group with, how bias is addressed through the vocabulary system, and what is still a work in progress.
This is Article 3 of 5 in the "Your Conservation Group, Your AI" series. For the full governance architecture, visit Village AI on Agentic Governance.
Previous: Big Tech AI vs. Your Conservation AI — Why the Difference Matters Next: What's Actually Running in Village Today