The second pen.

Edition 9 opened in a commissioning hall in Abu Dhabi. A senior engineer signs the permit to operate at 09:47 on a Tuesday. The AI did everything else. Three editions traced what stands behind her pen: the rulebook that moves slowest, the gap measured five years out.

One promise remains. What would an AI co-signer actually need to be?

Take the question seriously. Do not wave it away with sentiment or fear. Write its specification the way you would for any safety-critical system, requirement by requirement, then read what it describes.

Here is the trace.

Start with accountability, because every other requirement hangs from it. When a signed facility fails, the regulator does not open a ticket. It opens a file on a person.

The European Parliament tested the alternative. In February 2017 it passed a resolution inviting the European Commission to consider "electronic persons". The idea: a legal status for the most sophisticated autonomous systems, able to carry liability for the damage they cause. The resolution was non-binding. The status was never created.

The law that did arrive chose the opposite architecture. The European Union Artificial Intelligence Act requires, in Article 14, that high-risk systems be designed for oversight by natural persons. Those persons must understand the system, catch its anomalies, and resist over-relying on it. They must be able to set its output aside entirely.

Eight years separate the two instruments. The trend is not towards machines that carry accountability. It is towards naming the accountable human more precisely.

A co-signature with no name behind it is a watermark.

A name is necessary. It is not sufficient. The name must be certified competent under a framework a regulator recognises.

The United States engineering profession is explicit. The National Society of Professional Engineers (NSPE) defines responsible charge. Engineering decisions must be personally made by the professional engineer, or by people under that engineer's direction and control. Its ethics board has already applied the test to AI: sealing a document without maintaining responsible charge over the AI tool and its output is unethical.

The society's position on artificial intelligence goes further. People who design and deploy AI systems with direct public-safety impact should be held to the standards of professional licensure.

Note what exists nowhere in that landscape: a licensing pathway for the software itself. No jurisdiction issues one.

And who writes that framework? The committees Edition 10 named the slowest clock in the system. Before an AI co-signer can be certified, a committee must first define certifiable. Then roughly 32 sovereign nuclear frameworks, and their data centre equivalents, would each need to recognise the certificate separately. Signatures do not transfer between jurisdictions for humans. There is no special lane in which they transfer for machines.

Behind every signature sits a policy. Edition 11 named capital as the layer that prices the gap. A signature no insurer will stand behind is not accepted on any project worth permitting.

AI insurance exists. Munich Re has underwritten model performance since 2018. With the specialist insurer Mosaic, it now offers AI vendors performance cover against defined model failures.

Look at the shape of that product. It backs a vendor's promise that a model will hold a measured accuracy level. Claims are paid on performance data. That is real underwriting, and it is the giveaway. What the market insures is a performance number, backed by the company that built the model. What no market insures is a model's own acceptance that a facility is fit to operate. The insurable party is the maker. You cannot claim against a model. It owns nothing.

Aviation is the strongest counter-argument, so it deserves the floor. Autoland is certified autonomy. Aircraft land themselves in fog at airports on every continent. If aviation can certify an autonomous decision, why can infrastructure not certify an autonomous signature?

Two answers, both from aviation's own regulator. The European Union Aviation Safety Agency (EASA) publishes an AI roadmap grading AI into three levels. Human assistance. Human-AI teaming. Autonomy. Full autonomy without human override sits at the far end of a horizon running towards 2050. The rulemaking for the levels below it is still being written.

And autoland never removed the signature. Every certified autonomous function enters service through type certification. That is a chain of named, accountable people signing for the system's behaviour. Aviation did not delete accountability. It relocated it, into more signatures, made earlier, by people.

Tally the specification. A legal name a regulator can visit. A competency certificate from a framework that does not yet exist, recognised separately in every jurisdiction. An underwriter willing to stand behind judgement rather than accuracy. And a capacity for accountability that grows as the model improves, because the better the reading, the heavier the judgement.

Read the list twice. It is not a product roadmap. It is the job description of a competent person.

That is the end of the inversion. The nearest thing to an AI co-signature today is a maker's warranty with an insurer behind it. A company's name. A company's capital. A company's directors. Trace the second pen to its end and a person is holding it after all. Natural or corporate, but a person.

Edition 9 stated it: no AI signs the permit. Four editions later, the reason is traced. Not sentiment. Not protectionism. Not committee inertia. Structure. Signing is not the last step of reading. It is the first step of answering. Whatever signs, answers. Nothing that cannot answer will ever sign.

So the second pen is not waiting for a better model. The work, for as long as the build outruns the rulebook, is making the first pen lighter to lift. Cleaner data underneath the judgement. Sharper exceptions in front of the expert. A heavier signature carried by a better-prepared hand.

The facility is still dark until a person says otherwise. That was never the system failing.

That is the system.

Twelve editions close the written arc. Still Dark launched in March as an eight-edition pilot. More than 700 of you now read it on LinkedIn alone. The argument it set out to make is made. The gap between built and working is by design, not bad luck. Everything still rests on the expert.

What the written page cannot do is let the experts answer back. So the Thursday slot turns to where the experts are. Permit to Operate, the paired podcast, steps forward: the people who carry the gap, in their own voices. Facility Intelligence Newsletter members keep their 48-hour early access to every edition that ships, and Still Dark returns as a written depth piece when an argument demands the page.

The page has made its argument. The voices carry it from here. Watch this space.

If you build, operate, invest in, or regulate infrastructure anywhere in the world, this is written for you. Subscribe to Still Dark.

This newsletter lives in the gap between digital delivery complete and permit to operate. That gap is where value dies, and where it can be recovered.

I also co-author The Vistergy Brief at vistergy.com/archive. Satellite and geospatial monitoring, facility lifecycle intelligence, and standards architecture across LNG, nuclear, data centres, utilities and construction. Subscribe to both for the full picture.

The podcast is live. Permit to Operate is available at vistergy.com/pod, Apple Podcasts, and Spotify.

Permit to Operate.

David Ayeni

P.S. The Facility Intelligence Newsletter is open. £12 a month or £120 a year (17% annual discount).

What you get:

The first member analysis, From visible to decision-worthy, is live. Independent analyst-led publication. Upgrade.

104 matches. 48 teams. 39 days of football. Right now, streaming means logging into a cable account you don't have or paying for a subscription you'll cancel in August.

Norton Neo is a free browser with a free built-in VPN. No sign-up, no credit card, no catch. Private by default, backed by Norton security. Anti-fingerprinting and ad blocking run quietly in the background while you watch.

Download in 45 seconds. Watch every match for free.

Fast. Safe. Intelligent. That's Neo.

Try Norton Neo Now