New from me for @thebulletin.org. States still insist that human control of autonomous weapons must be "meaningful." But the word has become essentially meaningless. We need a new one.

Great essay but it’s worth noting that it doesn’t provide concrete evidence that Maven’s specific automated functionalities were directly involved in the erroneous decision. More than correcting the technological charisma of LLMs, it’s just sort of shifting that charisma to another type of AI.

“ai researcher said a thing” is the cousin of “ceo said a thing” reporting.

If you’re using AI to write books, understand that’s not going to somehow get you respect or money like real writers who aren’t using AI to write books because we don’t get respect or money either.

This is one of the stranger pieces I’ve written. I’m really not big on mindfulness in the new-age spiritualist sense. But in the context of killer robots, something about it just clicks!

NEW: it’s time to retire the principle of “meaningful human control” for autonomous weapons. The term has become, well, meaningless. Instead, how about—and bear with me here—*mindful* human control? @thebulletin.org. thebulletin.org/2026/04/outd...

I wrote a novel without using AI. It was fun. You should try it.

Remember how we were promised that AI would empower the human brain.

This is the key point. I can buy the idea that AI makes research faster, but it definitely doesn’t make it more enjoyable. And that matters. Writing should be an expression of joy. Nothing about efficiency is fun.

Wild. It’s raining Reapers over there.

Can’t see how this is ever going to fly. www.nytimes.com/2026/03/31/u...

Friendly reminder that even though having new ideas is way more gruelling than stealing them with AI, it is also way more joyful.

This might be an unpopular opinion, but all the ethical and practical issues of using AI to speed up warfare also apply to using AI to speed up research and writing.

This is not a good time to be a haver of thoughts.

Good morning to everyone except the guy who calls people “idea generators.”

If I had a time machine I’d go back to 2003 and show him this.

Super advanced automatic license plate readers that track every car on the road are no match for *checks notes* the oldest trick in the book.

Iran has a wildly impressive track record of shooting down super advanced, extremely expensive American drones. No other military even comes close.

Headline: "This is How AI is Being Used in the War" Three paragraphs into the article: "It is unclear exactly how AI is being used in the war."

People need to be way more careful about the idea of AI being to blame for anything in warfare. The idea of being “to blame” means you are held responsible, and thus, accountable for harms. If you blame AI for anything at all, you’ll have yourself an accountability gap.

I reckon we may need to expand the definition of AI psychosis to include being convinced you can use it for things, like practicing law, that actually require years of training and experience.

Lying in the pitch dark repeatedly smacking yourself in the face as you attempt to assassinate a mosquito, knowing full well that even if you succeed there will be another mosquito tomorrow night. A perfect metaphor.

Folks it’s officially time to start counting down to the day someone puts a laser cannon on a robot dog and everybody absolutely *loses* it.

Shooting drones with lasers sounds cool but we really need to think about what will happen when this technology proliferates widely.

A new frontier for weaponized AI. The US Army is looking to put AI into artillery shells for navigation and—incredibly—collaborative control. That means one "leader" munition passes target information to other shells in the volley. You know, like a swarm. www.militaryaerospace.com/sensors/arti...

In case you’re wondering where things are at with the drone wars, speedy armed quadcopters controlled via fibre optic cable are flying unchallenged over military installations in Iraq. One hit a US airbase. Another spent two minutes prowling the embassy in Baghdad. www.forbes.com/sites/davidh...

People don’t spend a ton of time thinking about robots/AI in maintenance, but if an adversary messed with such systems (eg. so that they miss certain faults, or muck up maintenance schedules) they could cause some serious trouble—or at a minimum give their target a lot of extra cost and headaches.

He’s right, tons of great men of history had zero introspection. Ozymandias, Midas, Tantalus, that emperor with the cool new clothes.

Yep, this is totally fine. No way this could backfire.

We do not know for sure if AI caused the mistake at Minab but we do know for sure that it didn’t prevent it.

If someone tells you that there will always be a human in control of the AI, ask them if it will always be a good person.

The last two weeks have revealed an uncomfortable paradox. People have always wanted militaries to have direct control of military AI. But now people also want companies to hardwire military AI to refuse certain orders. Those two things are not naturally compatible.

Folks have argued for years that AI weapons could be hardwired to follow the rules. They never figured that the people building the AI and the people using it would have totally different ideas about what those rules should be.

Probably a good time to flag that the likelihood that AI will have errors in warfare, and that those errors will cause harm, is as much a function of the model itself as it is of the digital ecosystem in which it is operating and the decision scaffold that it supports.

Microsoft is couching its opposition to autonomous weapons on the basis that “AI should not…start a war without human control.” This really misrepresents both the reality of how the tech will be used and the much more immediate potential risks it poses.

We don’t have hard data from AI use on the battlefield, but with studies like these one has to wonder if it’s actually making decisions more efficient or if it’s just making soldiers’ jobs more complex and difficult.

While we still don’t know if AI had any role here, we do know that AI can potentially help catch errors like this, though probably *only* if the user’s explicit intent for the AI is to catch errors like this.

Using people’s data without asking them in order to build a technology that nobody asked for.

Everyone imagined that human control of military AI would be distributed among everyone involved in the process of developing, selling, acquiring and deploying the tech. What we didn’t count on was that these different stakeholders would end up fighting each other over that control.

Friendly reminder that it’s always a good idea to hyperlink as though the story is for print, even if it isn’t.

Someday we’ll look back and realize that it was never a race. It was a stampede.

lots of crude jokes on here today.

I wonder if this is because people like that Anthropic took a stand, they're seeing Anthropic's name all over the news (ie free adverting), or they think that if it's the only model used on classified networks it must be better than other models.

No it cannot experience anxiety, and no it cannot experience regret.

In case you needed any further proof that Ukraine is the birthplace of modern drone warfare. www.bbc.com/news/article...

We're not too far from data centers getting their own air defense layer. As they pivot to AI, militaries are going to become so reliant on their cloud infrastructures that the very first thing to be targeted in future wars is likely to be a country's data centers. www.ft.com/content/09fa...

Couldn't they have taught the computer to play a different game that doesn't involve, you know, shooting at stuff?

Friendly reminder that you don't need to put AI into a robot for it to be killer.

An AI that recommends whiskeys is not an existential threat to humanity. But it is *so much nicer* to talk about booze with a human expert.

I'm no editor, but repeated use of all-caps feels like a pretty obvious tell that a writer doesn't have the evidence to back up their argument.