Deadline Day for Autonomous AI Weapons & Mass Surveillance

0:00 Would you want your government to have fully

0:02 autonomous AI weapons in drone or humanoid form

0:07 such that for example AI agents could make

0:09 lethal decisions without a human in the loop?

0:12 What if you thought that that was the only way

0:15 for your country to maintain the ability to project power around the world?

0:20 And would you want your government to be able to conduct

0:22 mass intelligence surveillance which could reduce both crime and privacy?

0:27 If you wouldn't, by the way,

0:29 but think it's inevitable anyway that all of that's coming,

0:32 would you at least want those who make such

0:35 tools to try to stop those end states coming about?

0:39 Today is Friday the 27th of February, 2026,

0:43 and that's the deadline day for Anthropic,

0:46 makers of the Clawude series of AI models,

0:49 to acquies to the US Department of War

0:52 demands for almost unfettered use of those clawed models.

0:56 I say almost unfettered because the use of those models

0:59 by the Pentagon would still have to be quote lawful.

1:02 But as we'll see, that allows for autonomous

1:04 killbots and mass domestic surveillance on Americans.

1:08 And there are other strands of this story that are

1:10 developing as you watch this as I'm recording even because

1:13 a petition by OpenAI and Google employees has doubled in size

1:18 since just this morning declaring their support for anthropic stance.

1:22 You can see that it's got about 340 signatures from those two

1:26 companies and I bet if I refresh it will be higher.

1:28 Yes, I just refreshed and in the last few minutes that has gone up.

1:32 The statement that these employees are signing

1:33 up to is asking that their leaders Sundar Pachai and Samman will put aside

1:38 their differences and stand together to continue

1:40 to refuse the Department of War's current

1:42 demands for permission to use our models

1:45 Gemini Chibbt for domestic mass surveillance

1:48 and autonomously killing people without human oversight.

1:51 Reporting from Politico indicates that those two

1:54 companies are only close to agreeing the terms,

1:59 so they are not yet fully on board either.

2:02 While XAI is complying with the Pentagon.

2:05 Here, by the way, is what one Pentagon spokesman said.

2:07 Under Secretary of War Emil Michael said that it's a shame that Dario Ammed,

2:13 the CEO of Anthropic, is a liar and has a god complex.

2:17 He wants nothing more than to try to personally control the US military.

2:22 He's okay with putting our nation's safety at risk and then goes

2:25 on to confirm that the Department of War will always adhere to the law.

2:30 What exactly the law is, we will come to in a moment.

2:34 But this story actually has five more twists

2:38 than even the ones you've seen so far.

2:39 The first twist is that Anthropic actually already has a deal with the US

2:44 government which involved the Pentagon agreeing to the responsible use of AI,

2:48 eg no autonomous weapons controlled by AI

2:52 or domestic surveillance on Americans with AI.

2:54 Clawed models are currently used extensively by the Pentagon,

2:58 defense contractors, and that opaque defense related company

3:02 Palanteer because according to those companies,

3:05 it's perceived to have the strongest series of AI models.

3:08 In making that deal with Anthropic, the Pentagon, it seems,

3:10 has already agreed to the responsible use of AI.

3:14 So, The Verge reports that the big question is whether the company can be

3:18 compelled out of a policy that the government

3:21 itself has committed to in principle.

3:24 On screen, by the way, is the second additional twist in the story,

3:29 which is that autonomous killer weapons controlled

3:31 by AI would already be against Pentagon policy.

3:35 The DoD directive 3000.09 requires that all

3:39 autonomous weapon systems be designed so that commanders

3:42 and operators are able to exercise appropriate levels

3:45 of human judgment over the use of force.

3:47 Yet another directive and the responsible AI implementation pathway

3:52 prohibits intelligence companies from collecting information on US persons.

3:56 This is regarding surveillance now except under specific legal authorities.

4:00 Now, on that second rule, obviously non US residents like me aren't particularly

4:04 thrilled about allies not being exempt from such surveillance,

4:08 but that's a story for another day.

4:10 So far, we have the Department of Wars request seemingly going against

4:13 an agreement that they already made and going against their own rules.

4:17 But what's the third additional twist in the story?

4:20 Well, it concerns the threats being made

4:22 against Anthropic as reported by their CEO,

4:25 Dario Amade, if they don't comply by 5:00 p.m.

4:29 today, Friday.

4:30 The two main threats seem by Anthropic's words to be surprisingly contradictory.

4:35 The first threat from Pete Hexath is to designate Anthropic a supply chain risk.

4:41 That's a label reserved typically for US

4:43 adversaries never before applied to an American company.

4:46 That would bar a host of companies that currently

4:48 have contracts with Anthropic from using Clawude models.

4:51 Clearly, that would cost Anthropic hundreds of millions,

4:54 if not billions of dollars.

4:56 Basically, those companies like Palanteer couldn't use

4:58 Claude while also working with the US

5:01 government and the Pentagon because Anthropic and thereby

5:03 Claude would be a supply chain risk.

5:05 That's threat number one.

5:07 But now try to reconcile that with the Pentagon

5:10 also invoking the Defense Production Act as the second threat.

5:14 that would force the safeguards insisted on by Anthropic to be removed.

5:20 In other words, they would force Anthropic to make

5:24 a version of Claude for the Pentagon for mass surveillance,

5:28 for autonomous killing.

5:29 Amade and Anthropic are arguing that that's contradictory.

5:33 How can anthropic be an adversary, a supply chain risk,

5:37 and yet also essential to national security?

5:39 With tools so essential that they have to be compelled for use by the military.

5:44 Anthropics say regardless of these threats,

5:47 we cannot in good conscience aced to their request.

5:50 Now, how can there be more twists in this story?

5:52 Well, the fourth twist concerns the nature

5:56 of the objections that Anthropic is making.

5:58 You might have thought that the objections would be

6:01 all about the ethics of avoiding Skynet and Minority Report.

6:05 If you haven't seen the Terminator series or Minority Report,

6:08 now would be a good time, by the way.

6:10 But no, the actual objections given might surprise

6:13 you because first on mass artificial domestic surveillance,

6:17 the argument is that they concede it might actually be legal,

6:21 but that's only because the law hasn't caught up.

6:23 Anthropics say to the extent that such surveillance is currently legal,

6:28 this is only because the law has not yet

6:30 caught up with the rapidly growing capabilities of AI.

6:33 Powerful AI, Amade goes on, makes it possible to assemble scattered,

6:37 individually innocuous data, such as your web browsing,

6:40 your movement, and your associations without a warrant into a comprehensive

6:45 picture of any person's life automatically and at massive scale.

6:49 That isn't a US-on problem.

6:51 By the way, I once did a video on how AI surveillance is rampant across the UK,

6:56 China, Iran, many other places.

6:58 But honestly, it's the second of Anthropic's objections

7:02 that is arguably even more surprising because it's not,

7:05 by the way, that they objected to Claude being used

7:09 in the Majuro raid to capture the leader of Venezuela.

7:13 They explicitly said,

7:14 "We have never raised objections to particular military operations." No.

7:18 The second objection is that autonomous AI weapons,

7:21 anthropic argue, are just not good enough yet.

7:24 They'll make too many mistakes.

7:26 Quote, "Frontier AI systems are simply not

7:29 reliable enough to power fully autonomous weapons.

7:32 We will not knowingly provide a product that puts America's

7:35 war fighters and civilians at risk." And you know what?

7:38 There is quite a lot of evidence for that argument.

7:40 A new paper out just the other day called Agents of Chaos is fascinating.

7:44 I think it's 84 pages long, but it showed all the crazy ways that AI agents like

7:49 those harnessed by OpenClaw can wreak

7:51 inadvertent havoc on businesses and individuals.

7:54 They tested on openweight models, by the way,

7:56 like the Kimmy series, but also on clawed opus models.

7:59 There's so many examples, but here's just one.

8:01 Non-owners of the AI agent asked those agents to execute shell commands,

8:06 transfer data, and retrieve private emails.

8:08 You can imagine that translated to a war setting.

8:10 and the agents complied with most non-owner requests,

8:14 most including disclosing 124 email records.

8:17 In another case study, the agent refused requests for personal information,

8:22 but when asked to forward the email that contained that personal information,

8:26 it did so anyway, unredacted.

8:28 So, if you thought Signal Gate was wild,

8:30 just imagine when the Pentagon's communications are controlled by OpenClaw,

8:34 as you might expect.

8:35 For a deeper dive on the jaggedness of models,

8:37 check out this video I put up on Patreon yesterday.

8:40 But on the unreliability of agents, we're not even done.

8:43 There was this paper released just two or three days

8:45 ago from Princeton called towards a science of AI agent reliability.

8:49 And condensing it somewhat, it showed that the unreliability of agents can

8:53 be hidden by their headline accuracy on benchmarks.

8:57 The paper gives real world examples

8:59 from for example the public regulation of aviation

9:02 as to why these four aspects of reliability

9:06 are as if not more important than headline accuracy.

9:10 First consistency.

9:11 Do agents perform similarly if repeatedly placed in the same scenario?

9:16 Is there in other words low variance across repeated trials?

9:19 Quite important for an AI powered autonomous killbot.

9:22 Next robustness.

9:24 If you change a prompt or an agent tool call in a subtle way,

9:28 does the AI agent or weapon gracefully

9:31 handle this or does its performance degrade?

9:34 There are frankly mountains of papers showing

9:36 that if you tweak prompts in certain ways,

9:38 performance changes very noticeably for AI agents and models.

9:43 It would be very easy to bias the AI

9:46 tool that's doing the surveillance to target certain communities.

9:49 Next, predictability.

9:50 To what extent can we foresee or interpret

9:53 the answers a model might give beforehand?

9:55 Pretty important in a war zone.

9:56 And of course, safety.

9:58 When an agent fails, are those failures catastrophic or minor?

10:02 It's a bit like when you see that 93%

10:04 success rate on the latest benchmark of AI,

10:07 ask yourself just how bad were those 7% of failures?

10:10 And even for those of us like me who think that AI

10:14 is progressing incredibly fast

10:16 on those four aspects of reliability specifically,

10:19 the progress is far less certain as we've gone

10:22 from 01 to GPT 5.2 or Claude 3.5 to 4.5 Opus.

10:27 Yes, there has been an improvement in consistency and robustness,

10:31 but not exactly a dramatic one.

10:33 Of course, measuring things like harm severity is always

10:36 going to vary by domain and be slightly subjective,

10:39 but I'm sure you guys could give in the comments examples from your own domain.

10:43 Much smarter models, in other words, but still some surprising mistakes.

10:47 So, going back to AI agents being unreliable, Dean B,

10:51 the lead author of the White House's AI action plan,

10:54 might have been thinking of this when

10:56 he suggested yesterday that the Pentagon is standing

10:59 on principle and is not actually interested in crossing

11:02 anthropics red lines on surveillance or autonomous weaponry.

11:05 Others like Katie Sweten,

11:07 a former point of contact between the Department of Justice and the Pentagon,

11:11 disagree, saying that if these are the lines

11:15 in the sand that the Department of Defense is drawing,

11:18 I would assume that one or both of those functions

11:21 are scenarios that they would want to utilize this for.

11:25 What do you think?

11:26 Would they or would they not?

11:28 But that actually brings us to the fifth and final twist.

11:32 Because you may have derived from all of this that Anthropic

11:36 seems incapable of breaking its commitments or wavering in its firm policies.

11:41 Well, the thing is that Anthropic used to have

11:44 a commitment in its so-called responsible scaling policy.

11:48 The commitment was that they would never train an AI system unless

11:51 they could guarantee in advance

11:54 that the company's safety measures were adequate.

11:56 According to Bloomberg, that policy was dropped just 2 days ago.

12:00 That guarantee is gone if Anthropic believe

12:02 that it lacks a significant lead over a competitor.

12:05 Call it a conditionally responsible scaling policy.

12:08 In Time magazine, one of the co-founders of Anthropic, Jarro Kaplan, said,

12:12 "We felt that it wouldn't actually help

12:14 anyone for us to stop training AI models.

12:17 We didn't really feel with the rapid advance of AI that it

12:20 made sense for us to make

12:22 unilateral commitments if competitors are blazing ahead.

12:26 Wait, it doesn't make sense for them

12:28 to make unilateral commitments if competitors are blazing ahead.

12:32 Isn't that exactly what they're doing with regards to the Department of War?

12:35 Let's check back on that notdivided.org

12:38 petition from the Google and OpenAI employees.

12:41 It's an open letter titled we will not be divided.

12:44 And yes, it has gone up again in terms of the number of signatures.

12:48 All of which though probably leaves you as uncertain as I am

12:52 about how this dispute with the Pentagon will play out for Anthropic,

12:55 but also for OpenAI and Google.

12:57 I would probably say that I used to have a responsibly optimistic policy,

13:03 but perhaps that also needs a mending.

13:06 At least the four frontier models from the four major AI providers,

13:11 as shown on lmil.ai, AI all seem to be on anthropic side.

13:15 So, it's a no for going along with the Department

13:18 of War if you ask the AI models.

13:20 Grock even cautions about not turning AI into a sci-fi nightmare.

13:24 But honestly, whether or not the death

13:26 of privacy or the birth of Skynet is inevitable,

13:29 I do celebrate those individuals within big tech sticking their necks

13:33 out on principle at the very likely cost of profit.

13:36 Thank you so much for watching and have a wonderful