Why AI governance
Every rule on AI assumes a record that does not exist. We issue it.
Governments now require AI output to be labelled. Courts, journals, universities and employers are adding rules of their own. Every one of them assumes someone can prove who made a piece of work, how much a machine did, whether its facts are real, and which human answers for it. Nobody could. Editorrah issues that record, signed, and sells it as infrastructure.
About a six-minute read.
01Rules already in force
Five jurisdictions. Twelve months. One shared assumption.
| In force | Jurisdiction | What it requires | Instrument |
|---|---|---|---|
| 1 Sep 2025 | China | AI-generated text, images, audio and video must carry a visible label and an embedded one. | Measures for Labeling AI-Generated Synthetic Content |
| 22 Jan 2026 | South Korea | Users must be told when a service uses generative AI, and its output must be labelled. | AI Basic Act, Article 31 |
| 20 Feb 2026 | India | Platforms must label synthetically generated audio and video and embed identifying metadata. | IT (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules, 2026 |
| 2 Aug 2026 | European Union | AI-generated and manipulated content must be disclosed. Penalties reach EUR 15M or 3% of worldwide turnover. | Regulation (EU) 2024/1689, Articles 50 and 99 |
| 2 Aug 2026 | California | Large generative AI providers must embed disclosures in what their systems generate and offer a free detection tool. | AI Transparency Act, SB 942 as amended by AB 853 |
Checked against the published instruments on 4 October 2026.
Each rule creates an obligation to show what happened. None creates the record that shows it. That gap is the market.
02What the rules leave open
The rules cover what machines generate. They leave out what people hand in.
A label on a model's output says nothing about the essay, the filing, the manuscript or the report in front of a court, a journal or a regulator. Today's AI governance tools govern models. Four questions about the work itself remain open, and each one is already costing institutions money.
-
1
Who made the work?
Authorship can no longer be read off the page. Detectors guess from the finished text and misclassify 61% of essays by non-native English writers.
7,000
proven AI misconduct cases in UK universities in one year, three times the year before.
The Guardian, FOI responses from 131 UK universities, 2023-24. Liang et al., Patterns, 2023.
-
2
How much did the machine do?
Disclosure is self-declared. Nothing checks the declaration against how the work was actually made.
22 million+
papers flagged for AI writing by one detector in its first year, worldwide. In UK universities alone, 7,000 cases were proven in a year.
Turnitin, April 2024. The Guardian, FOI responses from 131 UK universities, 2023-24.
-
3
Are its facts real?
AI invents cases, quotations and references, and they reach courts and journals before anyone looks.
2,100+
court decisions worldwide where a party relied on AI-fabricated material. One publisher retracted 11,300 papers and closed 19 journals.
AI Hallucination Cases database, HEC Paris, October 2026. Wiley and Hindawi, 2022 to 2024.
-
4
Which human answers for it?
AI agents now act inside companies with no record of who approved what. Policy exists. The record does not.
EUR 15M
or 3% of worldwide turnover: the ceiling for breaching the EU AI Act's transparency rules.
Regulation (EU) 2024/1689, Articles 50 and 99.
Detection is sold into every one of these markets. A verifiable record is sold into none of them. That is the company.
03What changes
Trust moves from the document to the record.
For two centuries a finished document was evidence of the mind that made it. Generative AI ended that in a single product cycle. What replaces it is a record made while the work is being done, signed when it is finished, and checked when it is relied on.
| Until now | From here |
|---|---|
| A finished document was evidence of who made it. | Work carries a signed record of how it was made. |
| Using AI was a private choice. | AI use is declared, checked against the process, and tied to a named human. |
| Facts were checked after someone challenged them. | Facts are checked against their source before the work is relied on. |
| Compliance meant having a policy. | Compliance means producing the record when a court, regulator or insurer asks for it. |
Proof of human authorship and verified fact stop being policy questions and become inputs other systems depend on. We build that input.
04Why not a detector
A detector returns a guess. Governance needs evidence.
| Detection | Editorrah |
|---|---|
| Analyses the finished text only | Reads the process that produced the work |
| Returns a probability score | Checks every fact against the source |
| Misclassifies non-native English writers | Matches the author to their own verified history |
| Beaten by the next rewrite | Issues a signed record with timestamps |
| Collapses under appeal | Built to survive dispute and improve with every check |
No single check is proof on its own. The record combines several, and holds up when challenged.
05Where the record becomes a requirement
Issued wherever work is handed in.
One engine. One record format. Sold at each checkpoint where someone must rely on work they did not make.
Where the record becomes a requirement
Public bodies already run the checkpoints where work is handed in. A record can be required at each of them.
| Body | Where the record is required |
|---|---|
| Courts | At intake, for every filing, including those from self-represented parties. |
| Universities and exam bodies | At submission, for assessed work. |
| Research funders and journals | At submission, before peer review starts. |
| Regulators and auditors | At inspection, as a signed export of where AI was used and who approved it. |
Every record makes the next one more accurate: how the engine improves
06How it becomes required
From optional check to condition of doing business.
Courts
Judges already issue standing orders requiring counsel to verify every citation. A signed certificate is the simplest way to satisfy them.
Insurers
Insurers are adding AI exclusions to policy forms and putting the bill on the company that deployed the AI. A record of human oversight is how coverage comes back.
Gallagher Re with MIT and Testudo, Smart Systems, Blind Spots, March 2026.
Regulators
Disclosure obligations are in force in five jurisdictions. The record is the evidence a regulator asks for at inspection.
Cybersecurity became mandatory spend when insurers and regulators demanded controls. The same forces are now demanding a record.