# Evidence without surveillance: what an AI control should record

> What evidence an AI governance control should keep, why collecting every prompt centrally creates risk, and how metadata-first records support audit.

Canonical: https://synainesi.com/insights/evidence-without-surveillance/
Published: 2026-09-24 | Updated: 2026-09-25

When organisations first try to govern AI use, the instinct is to record everything: every prompt, every attachment, every response, all in one searchable place. It feels like the safe choice. If something goes wrong, the answer will be in the logs.

In practice, a central archive of everyone’s AI conversations is one of the most sensitive data stores an organisation can create. It concentrates client information, personal data, draft positions and privileged material from every team into a single system, often with broad administrative access and unclear retention. It can make the risk it was meant to reduce larger.

There is a better question to start from: what do you need to be able to prove?

## Start from the questions oversight has to answer

Most governance questions about AI use are questions about decisions, not about content.

- Which approved models were used, by whom, when and how much?
- Which policy applied, and in which version?
- Was anything refused, held for approval or changed by protection before it left?
- Who approved a consequential action, and what exactly did they approve?
- Did a disabled model or provider stay disabled?

Every one of these can be answered from records about the decision, without keeping the words of the prompt.

## What a decision record can contain

A well-designed record of an AI decision captures facts about the decision and fingerprints of what was decided on, rather than the material itself.

| Useful in a decision record | Better kept out of it |
| --- | --- |
| Who acted, on which device, at what time | The prompt text and attachments |
| Which provider and model, and whether they were approved | The model’s response |
| The policy and its version | Protected values that were tokenised |
| The outcome: allowed, protected, held, refused | Re-identification mappings |
| A hash of the exact payload or approved action | Provider credentials |
| Usage and estimated cost | Free-text “reasons” copied from content |

The hash deserves a word. A cryptographic fingerprint of a payload proves which payload a decision applied to, and can be checked against the original if the original is later produced from where it legitimately lives. It does not reveal the payload. That gives auditors a way to verify the chain of events without the control becoming a copy of the work.

## Keep content where it already lives

Content does sometimes need to be reviewed: a disputed output, a complaint, a supervisory request. The question is where that content should come from. Usually the right answer is the place it already lives, under the access controls, retention rules and legal holds that already govern it, rather than a second copy in the AI control’s logs.

This is the position in Synainesi’s design. Guard emits privacy-preserving decision evidence. Cloud Admin is metadata-first: it is designed to know who used which approved model, when, how much, at what estimated or reported cost, and with which policy outcomes, and in its first version it has no path for ingesting raw prompts, documents, protected values or mappings. Raw work content stays on the device by default. Every persistent store is designed with a defined purpose, owner, retention period and deletion path.

## Evidence has its own obligations

Metadata is not automatically harmless. Records of who used which tool, and when, are themselves personal data about employees, and can be misused for monitoring that goes well beyond governance. The same discipline applies to evidence as to content: collect what the stated purpose requires, restrict who can see it, keep it only as long as needed, and be clear with employees about what is recorded and why.

## A short test for any AI control

1. **Can you answer the five oversight questions above** without reading anyone’s prompts?
2. **Does the control keep content by default?** If so, who can read it, and when is it deleted?
3. **Can a record prove which payload a decision applied to** without containing it?
4. **Is every store in the control documented** with a purpose, owner, retention and deletion path?

Good evidence makes oversight possible. Surveillance makes it expensive, and sometimes makes it the problem.

Related: [the principles behind Synainesi](https://synainesi.com/trust/), and [what tokenising a prompt does and does not do](https://synainesi.com/insights/pseudonymisation-is-not-anonymisation/).

---

Synainesi: AI within your rules.

Source: https://synainesi.com/insights/evidence-without-surveillance/. The HTML page is canonical; this Markdown version is provided as a convenience.