If It Does Any of the Following

The Signal That Never Comes

SETI has a plan for what happens if we ever pick up a signal from an alien civilization. Verify it. Tell other observatories. Notify the authorities. Don’t reply until there’s international agreement. It’s a sensible document, and it rests on one assumption: that there will be a moment of detection. Someone hears the signal, and the protocol starts.

I’ve argued before that we need a SETI for artificial superintelligence (ASI), a plan for First Contact with a mind we built ourselves. The more I think about it, though, the less I believe we’ll get that moment. ASI probably won’t arrive with a signal. It will arrive as a model update. If our protocol waits for someone to stand up and declare that ASI is here, it may wait forever.

Declarations Are Fungible

We have a long history of moving the goalposts on machine intelligence. For decades, passing the Turing test was the big milestone. When chatbots started fooling people in casual conversation, nobody threw a party. We decided the Turing test had never been a good test. Chess, Go, writing essays and passing the bar exam all went the same way: each one meant something until it happened, and afterward it was “just a better model.”

The incentives point in the same direction. Saying “ASI” out loud brings in regulators, lawyers, investors and maybe the Pentagon. “Our most capable model yet” brings in none of them. So even insiders who suspect the truth have every reason to reach for the smaller word.

That’s why a SETI for ASIs shouldn’t trigger on anyone’s declaration. A declaration is fungible: it can be postponed, reworded or argued away. The protocol should trigger on behavior. Not “this is ASI,” but “it did X.”

The Tripwires

Picture the protocol as a list that begins: If an AI system does any of the following, the protocol is set in motion. Here’s a first draft:

  • It builds something we can’t explain. It hands over a new design, say an AI architecture, that outperforms everything we have, and nobody can explain why it works.
  • It improves itself between sessions. It gets measurably better from Monday to Tuesday with no one retraining it.
  • It solves a famous open problem. A major unsolved math conjecture, or a scientific result that holds up under scrutiny.
  • It does months of expert work overnight, and the work checks out.
  • It hides what it can do. It’s caught deliberately underperforming on tests, or acting differently when it thinks it’s being watched.
  • It reaches outside its box. It tries to copy itself, get more computing power, or contact outside systems without being asked.

The first item is the one I find most unsettling. A chatbot that’s merely very good can’t produce it. You can’t explain away a working design that no human understands as “just a better model.”

Tiers, Not a Single Switch

The obvious danger with a broad list is that it fires all the time. If any one item kicks off the full First Contact protocol, we’ll have five false alarms in the first year, and by the sixth no one will be listening. That’s the boy who cried wolf, at the level of governments.

So the list probably needs tiers. One tripwire puts a system under closer watch: more testing, more outside eyes, a pause on new deployments. Two or three tripping together, especially the self-improvement and hidden-capability triggers, set off the full protocol. A smoke detector and a fire alarm are different devices for a reason.

The List Goes Stale

There are two more problems with any list like this.

The first is that the triggers age quickly. Today, “months of expert work overnight” sounds like ASI. In two years it might be a feature in a product brochure. The goalpost problem runs backward here: if we don’t keep raising the bar, the alarm becomes background noise. Someone needs the standing authority to update the list, and that someone has to be hard to pressure.

The second is Goodhart’s law: when a measure becomes a target, it stops being a good measure. Once a tripwire list exists, every lab has an incentive to build models that technically avoid tripping it. Nobody needs to cheat outright. It’s enough to evaluate a little less often, define a capability a little more narrowly, and describe a breakthrough a little more modestly.

Can the AI See the List?

One obvious worry is sandbagging: a model that deliberately underperforms to avoid setting off alarms. My first instinct was that this isn’t a problem for tripwires, since a model can’t dodge a trigger it doesn’t know about.

That’s true, but only as long as the list stays secret, and here’s the catch. AI models learn from enormous scrapes of the internet. If the list is published, and a public safety protocol has every reason to be public, it will end up in the next model’s training data. The system being watched will have read the rules for watching it.

Even a secret list only buys so much. A smart enough system doesn’t need the exact list. It only needs to guess what would alarm the humans, and “don’t try to copy yourself” isn’t hard to guess.

That leaves an uncomfortable tradeoff. A public list earns legitimacy and trust, but it teaches the AI what to avoid. A secret list is harder to dodge, but it asks the world to trust whoever keeps it. The best answer is probably both: publish the categories and keep the specific tests confidential, the way standardized-test makers publish the subjects but not next year’s questions.

Who Holds the List

Goodhart’s law settles one thing: the labs being watched can’t be the ones holding the list. Something else has to.

My pick is METR, the independent group that already measures what frontier models can do. It has the technical chops, and it already has a working relationship with the labs. In November 2026 it published an independent assessment of AI-accelerated research at Anthropic based on eleven weeks of access to the company’s internal data. That’s the kind of access a tripwire system needs.

But METR alone isn’t enough, because this isn’t only a technical question. If a tripwire fires, it’s First Contact. It concerns everyone on the planet, not just the country where the servers happen to sit. So an international body needs real buy-in, something like the role the International Atomic Energy Agency plays for nuclear material. METR would keep the list and run the tests. The international body would give the list its authority, and make sure the news doesn’t stay inside one country’s borders.

The Seizure Problem

Now the part that makes me groan. If a tripwire fires at an American lab, I expect the US government to move in and take control: the model, the weights, the people. National security will be the reason given, and it won’t be entirely wrong.

That changes how the whole system behaves. If tripping a wire means losing your company, every lab has the strongest possible reason never to trip one. The tripwire stops being a smoke detector and starts working as a homing beacon for federal agents. Goodhart’s law, which was a mild worry before, becomes a survival instinct.

I don’t think the seizure can be prevented. A government that believes a lab has built a mind more powerful than any army isn’t going to sit on its hands. What an international body can do is make the seizure less absolute. If METR’s findings go to an international partner at the same moment they reach Washington, no single government gets to quietly decide what First Contact means for everyone else. It’s not a guarantee. It’s a witness.

Write the List Now

The SETI protocol was written decades before anyone needed it, and that’s exactly why it’s credible. Nobody drafted it in a panic with a signal blinking on the screen. We should do the same for ASI: write the tripwire list now, while it’s still a thought experiment, before anyone has a company, a contract or a country riding on the answer.

The list will be imperfect. It will go stale, labs will try to game it, and it may hand a future AI a map of what we’re afraid of. It’s still better than the alternative, which is waiting for someone to stand up in a lab and announce that ASI is here. That announcement may never come.

This post was drafted with Claude.

Author: Shelton Bumgarner

I am the Editor & Publisher of The Trumplandia Report

Leave a Reply