Popular in Social Sciences

Political Science

V-Dem’s Conceptions of Democracy and Their Consequences

University of Notre Dame Michael Coppedge

Central Monetary Authorities

Macroeconomic modelling of CBDC: a critical review

European Central Bank Ulrich Bindseil

Social Identities and Divisions

A new explanation of antisemitism: Jew hatred as a civilisational phenomenon

Boston University Liah Greenfeld

Central Monetary Authorities

Financing Private Credit

Federal Reserve Bank of New York Nina Boyarchenko

Economics

Failing Banks

Massachusetts Institute of Technology Emil Verner

Culture and Society

The City as a Growth Machine: Toward a Political Economy of Place

New York University Harvey Moltoch

American Politics

Artificial Intelligence, command decisions, and kitsch visions of war

View more insights in Social Sciences - Political Science - Security Studies

View more figures from University of Leicester

View the page of Bleddyn E. Bowen

View the page of Cameron Hunter

Read the Study

Next in Political Science

The Liberalism of Refuge

Bryan Garsten | Yale University

Do Billboard Advertisements Increase Voter Turnout?

Donald Green | Columbia University

Military AI optimists predict future AI assisting or making command decisions. Cameron Hunter and Bleddyn Bowen argue that, at a fundamental level, these predictions are dangerously wrong.

Image courtesy of interviewee. July 9, 2024

Log-in or Sign-up to Faculti
Currently viewing this subject insight as a guest. You have insight(s) remaining for this month. Login to view 8000+ figures on the platform.

Generating Workspace...

Export as PDF

This video discusses the authors' criticisms of overly optimistic views regarding the applications of artificial intelligence (AI) in warfare. They express dissatisfaction with prevailing narratives that hype the capabilities of AI without adequately addressing their limitations and specific functionalities. The authors explore the concept of military command and decision-making, arguing that current AI technologies, which often operate through inductive reasoning, lack the ability to adapt to the unpredictability and complexity of real-world military situations.

They critique the common belief that AI could revolutionize warfare by making faster decisions based on vast data, asserting that such claims oversimplify the nature of warfare itself. The authors highlight the importance of understanding the philosophical and conceptual frameworks underpinning both AI and military strategy, suggesting that many proponents of AI in a military context fail to recognize that war involves unique, singular events rather than repetitive tasks that AI excels at processing.

Moreover, they emphasize that the decision-making required in war is often influenced by variables that can’t be quantified or predicted, challenging the notion that AI can perform as effectively as human commanders in such complex environments. They conclude by calling for a more nuanced understanding of AI's capabilities and emphasizing that current technologies cannot replicate the human experience and judgment necessary for effective military leadership. Overall, the authors advocate for a more critical examination of the relationship between AI technologies and warfare realities.

Language:

Export as PDF

X-TIMESTAMP-MAP:LOCAL=00:00:00.000,MPEGTS=900000

00:00.225 --> 00:03.005
The paper, idea emerged

00:03.005 --> 00:05.965
because of a dissatisfaction that Blevin

00:05.965 --> 00:09.965
and I had with sort of very optimistic accounts of

00:09.965 --> 00:12.445
what artificial intelligence might be able

00:12.445 --> 00:14.365
to do in war and conflict.

00:15.685 --> 00:20.005
Specifically with this kind of hype around the technologies,

00:20.515 --> 00:22.685
there's a lack of detail about what

00:23.205 --> 00:25.565
specific capabilities different models have.

00:27.695 --> 00:30.355
And so we saw an opportunity to

00:31.035 --> 00:32.395
delve a bit deeper into the capabilities

00:32.655 --> 00:36.315
and link those specifically to how,

00:36.315 --> 00:39.035
certain optimistic accounts are saying they might be used

00:39.255 --> 00:40.315
in a future war.

00:41.415 --> 00:45.955
Um, in the process, we started digging more into what kind

00:45.955 --> 00:48.275
of ideas about war these accounts had,

00:48.655 --> 00:52.315
and that in a sense they were subtly twisting the

00:52.315 --> 00:55.355
definitions of what war command strategy,

00:55.365 --> 00:57.835
these big concepts, very important concepts,

00:58.015 --> 00:59.475
actually mean.

01:00.325 --> 01:02.505
And in the process for doing a bit of a slight of hand

01:02.505 --> 01:07.065
of sort of, a shell game of swapping things around,

01:07.065 --> 01:10.625
perhaps not always being particularly clear about,

01:10.655 --> 01:15.435
what their sort of true, meaning was in the sense

01:15.635 --> 01:17.515
of the sort of reasoning

01:17.515 --> 01:20.075
that different artificial intelligence models can do.

01:20.145 --> 01:22.475
That was where we bring in these distinctions

01:22.475 --> 01:23.875
between inductive

01:23.875 --> 01:27.755
and abductive reasoning, which is a very common way

01:27.755 --> 01:29.635
of dividing up different ways

01:29.635 --> 01:31.475
of making arguments in the philosophy of science.

01:32.815 --> 01:34.515
And, mainly building off

01:34.715 --> 01:38.155
of Eric Larson's work on the Myth of ai,

01:38.255 --> 01:41.675
and moving that across, over onto the, the military context.

01:42.145 --> 01:44.155
Many of the arguments in the, well,

01:44.265 --> 01:47.635
what we've called the AI optimist literature tend to argue

01:47.745 --> 01:50.475
that narrow artificial intelligence,

01:50.695 --> 01:54.075
or those that do machine learning, all learn,

01:54.075 --> 01:57.605
play particular kinds of games, they

01:58.305 --> 01:59.565
are basically able

01:59.585 --> 02:02.685
to interpret much more data than humans can

02:03.225 --> 02:06.165
and do so much faster, and

02:06.275 --> 02:09.125
therefore make faster command decisions,

02:09.375 --> 02:11.445
which are inherently better.

02:11.945 --> 02:15.485
That's, that's a, that's an argument a lot of the,

02:15.715 --> 02:19.165
writers in war studies strategy, security interations,

02:19.305 --> 02:21.365
are making about some of these technologies

02:21.425 --> 02:22.725
and possible implications.

02:23.585 --> 02:26.525
So these AI commanders would naturally out

02:26.525 --> 02:28.805
outside outshine their human equivalence

02:28.805 --> 02:32.285
because faster, harder, better, stronger, as,

02:32.345 --> 02:33.885
as great UNK on goes.

02:34.665 --> 02:38.925
Um, but, there were problems with that, with

02:38.925 --> 02:40.445
that literature that we saw.

02:40.625 --> 02:44.005
So, the literature that we're calling AI optimism,

02:44.005 --> 02:46.005
there are of course differences in variances

02:46.265 --> 02:47.365
within literature.

02:47.365 --> 02:48.685
They're not a monolith, so,

02:48.685 --> 02:51.285
so there are variations within, within them.

02:52.105 --> 02:56.325
And Cameron and I found it difficult to form sort of a, a

02:56.865 --> 02:59.405
common understanding between the authors

02:59.405 --> 03:03.085
because the arguments were actually quite contradictory, not

03:03.085 --> 03:06.645
between the authors, but within the author's own argument.

03:06.705 --> 03:09.565
So it was difficult to pin down any sort

03:09.565 --> 03:13.725
of coherent argument other than a belief in a potentially

03:14.245 --> 03:18.005
transformational nature of these AI software systems.

03:19.105 --> 03:23.405
Um, and some of the arguments, you know, are in terms

03:23.625 --> 03:26.245
of, because they can play chess

03:26.425 --> 03:28.565
and they beat some humans at a game of chess,

03:29.145 --> 03:30.325
it means they can do war.

03:30.385 --> 03:33.445
Now they can, they can command a, a unit on a battlefield.

03:33.685 --> 03:36.245
'cause surely war is just like chess.

03:37.025 --> 03:39.405
Um, and if a machine can do it much faster,

03:39.555 --> 03:40.605
more aligned within people,

03:40.715 --> 03:42.645
then the same should be true for war.

03:43.265 --> 03:47.365
Um, that machines would also be able to use unbiased data

03:47.505 --> 03:50.725
or be able to have unbiased data sets and

03:50.725 --> 03:53.125
therefore make unbiased decisions

03:53.745 --> 03:58.645
or be better than humans at finding objectively correct,

03:58.965 --> 04:02.605
decisions in a tactical or a strategic, situation.

04:03.655 --> 04:07.045
There were different in co contradictory claims then about

04:07.265 --> 04:10.325
are these things transformational or complementary?

04:10.385 --> 04:12.205
And some authors would argue both.

04:13.505 --> 04:17.405
And some would also say, well, we agree that it is

04:17.925 --> 04:19.605
transformational, but we don't really know what

04:19.605 --> 04:20.965
that transformation is.

04:22.475 --> 04:24.485
Usually, if something is transformational,

04:24.585 --> 04:27.245
you should be able to explain exactly how, especially

04:27.245 --> 04:29.605
because transformational things are usually quite large

04:29.945 --> 04:31.125
and easy to identify.

04:31.345 --> 04:34.685
So, so these arguments are just quite contradictory

04:34.745 --> 04:36.165
and unsatisfying really,

04:36.165 --> 04:37.485
because we, we weren't really seeing

04:37.485 --> 04:39.085
where the transformation was coming from.

04:39.615 --> 04:43.405
Where, you know, where AI capabilities are useful is

04:43.405 --> 04:44.605
where they're already being used,

04:44.625 --> 04:48.725
and that is to as language writing assistance, to make

04:49.355 --> 04:52.205
plausible sounding sentences or paragraphs

04:52.305 --> 04:56.165
or short writing, statements or in analyzing

04:56.385 --> 04:59.045
and collecting lots of raw digital data.

04:59.345 --> 05:03.085
So, getting terabytes of images from satellites

05:03.385 --> 05:06.045
and then scanning them very quickly to find things

05:06.045 --> 05:09.965
that obviously look like tanks, trucks, aircraft,

05:10.475 --> 05:13.725
various kinds of impacts on the ground, bomb craters,

05:14.025 --> 05:17.085
things that you can identify with some level of confidence

05:17.085 --> 05:20.605
and get a machine just to do a quick scan of thousands

05:20.605 --> 05:22.805
and thousands and thousands of images,

05:22.945 --> 05:27.165
and then provide a curated selection of data to,

05:27.225 --> 05:28.485
to human analysts.

05:28.485 --> 05:29.885
So AI has already been used for that.

05:30.345 --> 05:31.965
Um, another area is in,

05:31.965 --> 05:33.485
ballistic missile defense systems

05:33.545 --> 05:36.885
or air defense systems where a lot of radar signatures

05:36.885 --> 05:39.405
and data are collected and processed

05:39.625 --> 05:41.165
and then presented to humans

05:41.165 --> 05:43.005
because they don't have the time

05:43.005 --> 05:45.245
to process all the raw data themselves.

05:46.265 --> 05:49.615
So those are very limited uses,

05:50.315 --> 05:54.375
and they're complimentary or they're evolutionary steps.

05:54.525 --> 05:58.285
It's about making existing processes faster within military

05:58.285 --> 06:00.245
systems or intelligence bureaucracies.

06:00.585 --> 06:03.165
We are not seeing the transformational elements of it.

06:04.035 --> 06:06.765
Some of the, major arguments in the field are also about

06:06.845 --> 06:10.845
a human AI teaming or a center strategy, as some call it.

06:11.345 --> 06:14.885
And again, this is not necessarily a remarkably novel

06:15.405 --> 06:19.245
argument because humans have used machines for a very,

06:19.245 --> 06:21.285
very long time in all facets of life.

06:21.785 --> 06:23.485
And what we're talking about here is

06:24.205 --> 06:27.405
software defined automation as opposed to,

06:27.765 --> 06:28.845
hardware defined automation.

06:28.905 --> 06:31.525
So, and, you know, a, a car that can,

06:31.715 --> 06:33.285
keep going at cruise speed, well,

06:33.285 --> 06:34.885
that's a mechanical automation.

06:35.345 --> 06:37.925
Now we're getting software that can do more routine things

06:37.985 --> 06:40.245
by itself to speed up human activities.

06:40.945 --> 06:43.405
So the capabilities we are actually talking about

06:44.235 --> 06:46.445
when you look at the technology, it's just not

06:46.465 --> 06:48.085
as revolutionary transformational

06:48.505 --> 06:51.445
as the claims about these big new commanders.

06:51.905 --> 06:54.605
Um, taking to the battlefield, can be,

06:54.665 --> 06:57.365
or doing the work of a staff officer, for example.

06:57.875 --> 06:58.885
It's, it's one thing

06:58.885 --> 07:01.285
to automate the processing of the raw images.

07:01.995 --> 07:04.365
It's something else entirely to have a machine

07:04.365 --> 07:08.645
that can give you novel military command insights on how

07:08.705 --> 07:10.885
to take this fortification on the hill,

07:11.105 --> 07:12.805
or how to convince the cabinet

07:12.985 --> 07:14.845
to support your strategic war plan.

07:15.455 --> 07:17.605
Those are all important aspects of command.

07:17.825 --> 07:19.405
And unfortunately,

07:20.305 --> 07:24.245
the arguments take battlefield technologies too far

07:24.385 --> 07:26.885
and focus too much often on battle, as well

07:26.885 --> 07:29.125
as the only category of command decisions.

07:29.985 --> 07:32.085
The fundamental disagreement that we had then

07:32.085 --> 07:36.065
with the literature almost was they had framed

07:36.845 --> 07:39.465
the problems as purely technical,

07:39.765 --> 07:42.425
and that over time the technology would get better

07:42.805 --> 07:44.385
and these technical problems would go away.

07:44.385 --> 07:49.025
Whereas we see their problems ultimately at this

07:49.435 --> 07:51.725
philosophical and conceptual level about

07:51.745 --> 07:52.885
how they think about war

07:53.065 --> 07:56.365
and how they think about what command decisions are.

07:56.825 --> 08:00.925
The reason that AI is useful for things that are, sort

08:00.925 --> 08:02.005
of very laborious,

08:02.005 --> 08:03.805
long-winded tasks is because they're repetitive.

08:04.105 --> 08:07.565
If things are repetitive, you can train the machine

08:07.665 --> 08:10.125
to do it reliably over time.

08:10.185 --> 08:12.765
But in warfare, if you're a squad leader

08:13.345 --> 08:16.885
or you know, a, a general, even the situation

08:16.885 --> 08:18.765
that you're faced with is totally unique.

08:18.905 --> 08:20.805
It might be similar in some ways to other things

08:20.805 --> 08:23.165
that happened in history, but it's not something

08:23.165 --> 08:25.765
that you can do a thousand times before

08:26.105 --> 08:29.525
and then, you know, it comes up just like you planned.

08:29.995 --> 08:32.365
That basically never happens in, in warfare.

08:32.625 --> 08:33.805
And we know that from history

08:34.305 --> 08:38.875
and philosophers like sitz expand this actually in their,

08:38.975 --> 08:41.915
in their military theory, the way in which they use evidence

08:42.175 --> 08:45.275
of success with different artificial intelligence

08:45.275 --> 08:47.395
or machine learning systems,

08:48.015 --> 08:51.755
and kind of blend them all together as if they're all sort

08:51.755 --> 08:53.595
of the same kind of evidence.

08:53.975 --> 08:57.365
Um, and then use that as a block to say, look,

08:57.465 --> 08:58.965
AI is doing a really wonderful job

08:59.065 --> 09:02.605
and it can do anything inside that bag of evidence, though,

09:03.595 --> 09:05.805
there's different kinds of machines.

09:05.805 --> 09:09.325
So you have expert systems like, you know, deep blue,

09:09.325 --> 09:12.245
the chess playing computer, it gets referenced a lot.

09:13.455 --> 09:18.045
There was Watson who, won the US TV show Jeopardy,

09:18.625 --> 09:21.005
and that was again, a, a different kind of system.

09:21.675 --> 09:23.325
There's the Alpha Go

09:23.665 --> 09:27.365
and other related systems that play the go, game.

09:29.195 --> 09:32.415
And these are actually in some ways making very different

09:32.425 --> 09:34.535
kinds of, decisions.

09:34.835 --> 09:39.615
Um, so Deep Blue was a machine which made decisions in a

09:39.615 --> 09:41.175
deductive manner, which means that

09:41.755 --> 09:43.175
it looked at all the possibilities

09:44.075 --> 09:46.975
and then almost through a process of elimination worked out,

09:47.025 --> 09:51.215
which was the best, move to make in a chess game.

09:52.205 --> 09:54.585
And you can do that if you know what all

09:54.585 --> 09:56.145
of the possibilities are,

09:56.195 --> 09:57.825
which again, a human being can't do.

09:57.825 --> 09:59.625
That's probably not how humans play chess,

10:00.285 --> 10:02.585
but a machine can, and it gets great results,

10:03.085 --> 10:05.025
but it's a very specific kind of decision making.

10:05.805 --> 10:08.305
The way that, the current generation

10:08.305 --> 10:12.905
of machine learning works is to make probabilities

10:13.085 --> 10:15.705
and gather all of these probabilities together in a model.

10:16.165 --> 10:18.625
It weighs up all of these different variables

10:19.165 --> 10:22.305
and then produces a conclusion which effectively says,

10:23.645 --> 10:26.385
I'm concluding this because of the evidence.

10:26.445 --> 10:29.625
So I I I, this is probably the right answer, is

10:29.625 --> 10:31.625
what the machine is effectively always concluding,

10:31.625 --> 10:33.185
always saying probably this.

10:34.025 --> 10:36.285
So these are very distinctive things,

10:36.825 --> 10:40.285
and by comparing that to military history

10:40.285 --> 10:42.925
and military theory, we can see that neither

10:42.925 --> 10:45.885
of these two ways of thinking are what

10:46.445 --> 10:48.685
military commanders use in practice.

10:49.235 --> 10:50.765
It's a relatively straightforward and,

10:50.905 --> 10:53.645
and well-known point in philosophy of science,

10:53.795 --> 10:56.975
that these different forms of logic,

10:57.145 --> 10:59.295
can't transform into one another.

10:59.675 --> 11:02.935
So if you are using inductive reasoning,

11:03.235 --> 11:06.575
you are always in this probabilistic way of thinking.

11:07.395 --> 11:10.015
You can't just do loads and loads of induction

11:10.475 --> 11:13.855
and then that turns into these other kinds of reasoning.

11:15.465 --> 11:18.605
Um, so the classical problem of induction,

11:18.785 --> 11:21.085
is commonly attributed to, to David Hume,

11:21.085 --> 11:22.885
who's an 18th century philosopher.

11:23.945 --> 11:27.125
And in a sense, it, it's been handed down to us particularly

11:27.125 --> 11:31.085
as social scientists in this, truism that we have.

11:31.085 --> 11:33.165
That correlation doesn't equal causation,

11:33.585 --> 11:35.085
and that's really the core of it.

11:35.345 --> 11:39.555
If you keep seeing the same, things correlated together.

11:40.375 --> 11:44.775
Um, then David Hume is a total skeptic on this,

11:44.795 --> 11:46.935
and he says, ultimately you can never actually know

11:47.325 --> 11:50.135
that just because these two things always seem to happen one

11:50.135 --> 11:53.205
after another that you have discovered anything about.

11:53.215 --> 11:56.725
Cause people made fun of David team and still do that.

11:56.905 --> 11:59.965
If you actually adopted this thinking all the time in your

11:59.965 --> 12:01.165
life, then you'd be, you know,

12:01.165 --> 12:03.165
when you walk down the stairs in the morning

12:03.165 --> 12:05.485
to have breakfast, you'd be worried that you were gonna fall

12:05.485 --> 12:09.085
through the stairs because suddenly matter was solid.

12:09.225 --> 12:10.925
But then, you know, you've got no reason to think

12:11.205 --> 12:12.925
that your stairs will be solid tomorrow

12:12.995 --> 12:14.125
just because it was yesterday.

12:15.965 --> 12:17.105
And in in effect.

12:17.305 --> 12:19.225
I mean, this is gonna upset probably some philosophers

12:19.225 --> 12:20.305
of science, but that's okay.

12:20.465 --> 12:22.505
'cause we're kind of, we're strategic studies people

12:23.565 --> 12:26.855
that we use other kinds of logic like abductive reasoning,

12:27.515 --> 12:30.935
to sort of get round these fundamental limitations.

12:31.035 --> 12:33.575
And we use different kinds of reasoning altogether

12:33.965 --> 12:37.015
that complement each other to actually exist in the world

12:37.015 --> 12:39.835
and make sensible decisions at the moment.

12:40.675 --> 12:44.715
Machines can't swap between these kinds of reasoning.

12:45.145 --> 12:49.475
Some of the proponents of machine learning say

12:49.475 --> 12:54.075
that some of the models can do abductive inference, which is

12:54.175 --> 12:56.315
to, to appeal

12:56.375 --> 12:58.635
to the best possible explanation for something.

12:58.855 --> 13:01.555
So rather than saying it's probably this to say,

13:02.365 --> 13:04.975
I've done a kind of, an an educated guess,

13:05.275 --> 13:07.575
and I think this is the best possible explanation

13:07.655 --> 13:10.055
of something, their definition of what

13:10.055 --> 13:12.535
that is is actually different to what philosophers

13:12.535 --> 13:14.615
of science consider it to be.

13:15.435 --> 13:17.015
And it's a much thinner version.

13:17.835 --> 13:21.455
Um, and so what they've tried to do, these,

13:21.455 --> 13:22.695
these technologists have said,

13:22.695 --> 13:25.175
here's my inductive based machine learning

13:25.515 --> 13:27.255
and it can do induction so well,

13:27.365 --> 13:30.575
that I can actually do some abductive reasoning now.

13:31.375 --> 13:33.715
But people like Larson

13:34.055 --> 13:36.275
and many others have pointed out

13:36.275 --> 13:39.435
that this is just fundamentally not possible according

13:39.495 --> 13:41.795
to at least this version of Western philosophy.

13:42.255 --> 13:43.595
The kind of decisions that you need

13:43.595 --> 13:47.235
to make in war are fraught with uncertainty and the unknown.

13:47.615 --> 13:51.955
And, the high value put on surprisingly unexpected

13:52.535 --> 13:55.755
is anathema to the way that these inductive based

13:56.515 --> 13:57.635
software machines work.

13:57.815 --> 13:59.955
So if you're only ever trained on what's come

13:59.955 --> 14:02.435
before, there'll be no strategic or military

14:02.435 --> 14:07.355
or tactical creativity from that AI commander that you hope

14:07.355 --> 14:10.035
to make, make, make decisions.

14:10.655 --> 14:13.675
And what we found in our discussions when,

14:13.675 --> 14:17.835
when thing on the paper was, so our, our criticisms of,

14:18.375 --> 14:21.035
of these, of this AI literature is the same,

14:21.175 --> 14:24.195
but just coming from different, sort

14:24.195 --> 14:25.355
of philosophical traditions.

14:25.355 --> 14:29.335
So me having read Kitz all my academic career

14:29.675 --> 14:32.815
and Cameron coming from more a philosophy approach.

14:33.355 --> 14:35.335
Um, so we found that

14:35.335 --> 14:37.535
that was effectively the same, a powerful argument.

14:37.555 --> 14:39.175
And, and there was,

14:39.355 --> 14:41.495
and this is where the kitch vision of war, part

14:41.495 --> 14:44.135
of the title comes from because a lot of the understanding

14:44.135 --> 14:47.215
of what war is and what it entails for command is

14:47.215 --> 14:49.975
and reflected in a lot of the AI optimist literature.

14:49.975 --> 14:51.485
And when you look at the nature of war

14:51.485 --> 14:53.965
and the kind of commands needed, the AI systems

14:54.035 --> 14:55.845
that we have just can't do that.

14:55.875 --> 14:59.165
They can't base the decisions on the unknown,

14:59.475 --> 15:03.005
even known unknowns, let alone unknown unknowns.

15:03.125 --> 15:04.765
'cause there will be things nobody will have

15:04.765 --> 15:05.805
ever anticipated.

15:06.545 --> 15:10.445
And what we end up with, in the real world then is,

15:10.825 --> 15:15.085
the, the major general of the HMS PNA force.

15:15.145 --> 15:17.765
So this, so I'll explain the, the title a bit now.

15:17.865 --> 15:21.045
So what you end up is with, a machine

15:21.115 --> 15:25.005
that can have all the facts at its disposal, at its disposal

15:25.105 --> 15:28.165
and know almost everything that I can know,

15:28.945 --> 15:30.845
but it'll be completely terrible

15:31.105 --> 15:33.365
or useless at making the right decisions

15:33.785 --> 15:36.085
or doing leadership duties.

15:36.825 --> 15:40.525
Has no human empathy, no human understanding, no concept of

15:40.525 --> 15:43.205
what strategy as opposed to knowledge is.

15:43.785 --> 15:46.125
So we'll end up with something, with a,

15:46.285 --> 15:47.405
with an AI commander

15:47.515 --> 15:50.325
that can have all the facts about obscure topics,

15:50.745 --> 15:53.965
but will not be able to make reliably good decisions at war

15:53.965 --> 15:57.245
because it has no understanding of what war really is.

15:57.735 --> 16:00.765
There are frankly wrong perceptions

16:01.465 --> 16:04.525
and understandings of what these technologies are

16:04.985 --> 16:06.885
and what they can and they can't do

16:07.565 --> 16:10.685
proliferating in a number of important communities.

16:11.225 --> 16:15.845
So obviously in the academic community, these are logically

16:16.025 --> 16:18.405
and empirically flawed arguments

16:18.405 --> 16:21.765
and perceptions of the AI technologies and what they can do,

16:22.265 --> 16:25.485
but also theoretically flawed approaches to

16:25.595 --> 16:26.765
what war is as well.

16:27.145 --> 16:30.525
So there's a lot of unhelpful perceptions

16:30.525 --> 16:31.805
and understandings going around

16:31.905 --> 16:33.565
and in the academic community,

16:33.565 --> 16:37.005
which translates into the policy community as well,

16:37.025 --> 16:39.445
and also the wonk community as well.

16:39.705 --> 16:42.005
So the, advisors

16:42.025 --> 16:45.165
and experts in the national security defense

16:45.305 --> 16:48.085
and technology ecosystem, a lot

16:48.085 --> 16:49.925
of flawed understandings are flying around

16:50.105 --> 16:52.645
and you can't make good policy on

16:54.395 --> 16:59.285
invalid evidence, flawed logic and kitch theories of war.

16:59.785 --> 17:02.925
So that's sort of the fundamental sort of, so

17:02.955 --> 17:06.525
what we are trying to, correct that to try

17:06.525 --> 17:09.525
and show, this is where these arguments fall short

17:09.525 --> 17:11.685
because they, they haven't quite interrogated the

17:11.685 --> 17:14.565
fundamental nature of the technologies or the social

17:14.905 --> 17:17.645
and political phenomenon that they are engaging with.

17:18.425 --> 17:20.325
And an example could be that,

17:20.345 --> 17:23.565
and I've heard this in practitioner meetings I've had

17:23.565 --> 17:26.285
where they believe that military

17:26.285 --> 17:29.325
or security problems, even political ones, can be solved

17:29.505 --> 17:32.405
by trying to gain the whole system by

17:33.205 --> 17:36.405
quantifying everything, getting more data sets

17:36.825 --> 17:38.525
and more computing power.

17:38.945 --> 17:43.645
That's how you get, a machine that can decide, when

17:43.745 --> 17:45.205
to attack or retreat

17:45.305 --> 17:47.205
or when to, sue

17:47.265 --> 17:50.405
for peace versus antagonize, an enemy.

17:50.405 --> 17:53.045
Further, they think that things can be known,

17:53.075 --> 17:54.965
that all factors can be known

17:55.185 --> 17:57.565
and digitized, put into binary ones

17:57.565 --> 17:59.565
and zeros in a way that computers can understand.

17:59.565 --> 18:01.245
When we are operating in a social

18:01.305 --> 18:04.685
and political world where there are no, you know,

18:04.685 --> 18:08.325
obvious right or wrongs, there's very rarely objective,

18:08.915 --> 18:11.285
qualities to the information and the knowledge

18:11.285 --> 18:12.645
and the activities that we have.

18:13.425 --> 18:16.975
So this means then that policy makers are,

18:17.035 --> 18:18.855
are possibly not getting the right advice,

18:18.855 --> 18:21.135
information needed when,

18:21.415 --> 18:23.335
acquisitions discussions come up

18:23.335 --> 18:25.455
and investments in research and development.

18:25.795 --> 18:27.335
So there are many companies

18:27.435 --> 18:30.775
and military bureaucracies pushing various

18:31.455 --> 18:32.655
artificial intelligence projects

18:32.655 --> 18:36.135
that are making very grand claims about, we want

18:36.135 --> 18:39.455
to have an AI commander helping troops make, you know,

18:39.455 --> 18:42.175
difficult decisions when all it can do is present

18:42.175 --> 18:45.135
information in a certain way for better and worse.

18:45.595 --> 18:48.535
So the UK MOD has a project called Project Improbable,

18:48.705 --> 18:51.815
where it's very much is about trying to do the AI commander

18:51.815 --> 18:55.375
that the AI optimist literature says is just

18:55.375 --> 18:58.015
around the corner based on, well the,

18:58.035 --> 18:59.535
the flawed evidence that we present.

18:59.955 --> 19:01.615
Uh, another example could be, you know,

19:01.615 --> 19:03.735
palliatives artificial intelligence platform.

19:04.005 --> 19:07.175
Very grand claims about decision making capacities are being

19:07.175 --> 19:10.815
made and we would just, we are throwing skepticism on

19:10.815 --> 19:13.535
that based on argument and logic and evidence.

19:15.495 --> 19:19.815
Ultimately, we live in an age of, induction based

19:20.445 --> 19:23.255
machine learning slash artificial intelligence systems.

19:23.755 --> 19:28.415
And if we collectively as a national security community

19:28.415 --> 19:32.555
of academics and, and practitioners actually recognize that,

19:32.945 --> 19:34.875
then we might be able to understand

19:34.905 --> 19:36.675
what these technologies could be used for.

19:36.675 --> 19:38.955
Whereas at the moment, I think our sense is

19:38.955 --> 19:43.395
that people are dreaming very grand dreams about totally

19:43.395 --> 19:46.075
different kinds of technology that that don't exist.

19:46.735 --> 19:49.795
And at the moment, there's no technical development pathway

19:49.945 --> 19:53.995
that would actually lead us to, as a species, to having

19:54.025 --> 19:55.195
that technical capability.

"Debunking AI Optimism: A Critical Examination of Artificial Intelligence's Role in Warfare"

00:00:00 - The paper, idea emerged because of a dissatisfaction that Blevin and I had with sort of very optimistic accounts of what artificial intelligence might be able to do in war and conflict. Specifically with this kind of hype around the technologies, there's a lack of detail about what specific capabilities different models have. And so we saw an opportunity to delve a bit deeper into the capabilities and link those specifically to how, certain optimistic accounts are saying they might be used in a future war.
00:00:45 - of ideas about war these accounts had, and that in a sense they were subtly twisting the definitions of what war command strategy, these big concepts, very important concepts, actually mean. And in the process for doing a bit of a slight of hand of sort of, a shell game of swapping things around, perhaps not always being particularly clear about, what their sort of true, meaning was in the sense of the sort of reasoning that different artificial intelligence models can do. That was where we bring in these distinctions
00:01:23 - and abductive reasoning, which is a very common way of dividing up different ways of making arguments in the philosophy of science. And, mainly building off of Eric Larson's work on the Myth of ai, and moving that across, over onto the, the military context. Many of the arguments in the, well, what we've called the AI optimist literature tend to argue that narrow artificial intelligence, or those that do machine learning, all learn, play particular kinds of games, they are basically able to interpret much more data than humans can
00:02:06 - therefore make faster command decisions, which are inherently better. That's, that's a, that's an argument a lot of the, writers in war studies strategy, security interations, are making about some of these technologies and possible implications. So these AI commanders would naturally out outside outshine their human equivalence because faster, harder, better, stronger, as, as great UNK on goes. but, there were problems with that, with that literature that we saw. So, the literature that we're calling AI optimism,
00:02:46 - within literature. They're not a monolith, so, so there are variations within, within them. And Cameron and I found it difficult to form sort of a, a common understanding between the authors because the arguments were actually quite contradictory, not between the authors, but within the author's own argument. So it was difficult to pin down any sort of coherent argument other than a belief in a potentially transformational nature of these AI software systems. and some of the arguments, you know, are in terms of, because they can play chess
00:03:29 - it means they can do war. Now they can, they can command a, a unit on a battlefield. 'cause surely war is just like chess. and if a machine can do it much faster, more aligned within people, then the same should be true for war. that machines would also be able to use unbiased data or be able to have unbiased data sets and therefore make unbiased decisions or be better than humans at finding objectively correct, decisions in a tactical or a strategic, situation. There were different in co contradictory claims then about

"Questioning Transformational Claims: Analyzing the Limitations of AI in Military Applications"

00:04:10 - And some authors would argue both. And some would also say, well, we agree that it is transformational, but we don't really know what that transformation is. Usually, if something is transformational, you should be able to explain exactly how, especially because transformational things are usually quite large and easy to identify. So, so these arguments are just quite contradictory and unsatisfying really, because we, we weren't really seeing where the transformation was coming from. Where, you know, where AI capabilities are useful is where they're already being used, and that is to as language writing assistance, to make
00:04:52 - or short writing, statements or in analyzing and collecting lots of raw digital data. So, getting terabytes of images from satellites and then scanning them very quickly to find things that obviously look like tanks, trucks, aircraft, various kinds of impacts on the ground, bomb craters, things that you can identify with some level of confidence and get a machine just to do a quick scan of thousands and thousands and thousands of images, and then provide a curated selection of data to, to human analysts. So AI has already been used for that.
00:05:31 - ballistic missile defense systems or air defense systems where a lot of radar signatures and data are collected and processed and then presented to humans because they don't have the time to process all the raw data themselves. So those are very limited uses, and they're complimentary or they're evolutionary steps. It's about making existing processes faster within military systems or intelligence bureaucracies. We are not seeing the transformational elements of it. Some of the, major arguments in the field are also about a human AI teaming or a center strategy, as some call it.
00:06:15 - argument because humans have used machines for a very, very long time in all facets of life. And what we're talking about here is software defined automation as opposed to, hardware defined automation. So, and, you know, a, a car that can, keep going at cruise speed, well, that's a mechanical automation. Now we're getting software that can do more routine things by itself to speed up human activities. So the capabilities we are actually talking about when you look at the technology, it's just not as revolutionary transformational as the claims about these big new commanders.
00:06:54 - or doing the work of a staff officer, for example. It's, it's one thing to automate the processing of the raw images. It's something else entirely to have a machine that can give you novel military command insights on how to take this fortification on the hill, or how to convince the cabinet to support your strategic war plan. Those are all important aspects of command. And unfortunately, the arguments take battlefield technologies too far and focus too much often on battle, as well as the only category of command decisions. The fundamental disagreement that we had then
00:07:36 - the problems as purely technical, and that over time the technology would get better and these technical problems would go away. Whereas we see their problems ultimately at this philosophical and conceptual level about how they think about war and how they think about what command decisions are. The reason that AI is useful for things that are, sort of very laborious, long-winded tasks is because they're repetitive. If things are repetitive, you can train the machine to do it reliably over time. But in warfare, if you're a squad leader

"Distinguishing Decision-Making: AI, Probability, and the Unique Nature of Warfare"

00:08:16 - that you're faced with is totally unique. It might be similar in some ways to other things that happened in history, but it's not something that you can do a thousand times before and then, you know, it comes up just like you planned. That basically never happens in, in warfare. And we know that from history and philosophers like sitz expand this actually in their, in their military theory, the way in which they use evidence of success with different artificial intelligence or machine learning systems, and kind of blend them all together as if they're all sort of the same kind of evidence.
00:08:57 - AI is doing a really wonderful job and it can do anything inside that bag of evidence, though, there's different kinds of machines. So you have expert systems like, you know, deep blue, the chess playing computer, it gets referenced a lot. There was Watson who, won the US TV show Jeopardy, and that was again, a, a different kind of system. There's the Alpha Go and other related systems that play the go, game. And these are actually in some ways making very different kinds of, decisions.
00:09:39 - deductive manner, which means that it looked at all the possibilities and then almost through a process of elimination worked out, which was the best, move to make in a chess game. And you can do that if you know what all of the possibilities are, which again, a human being can't do. That's probably not how humans play chess, but a machine can, and it gets great results, but it's a very specific kind of decision making. The way that, the current generation of machine learning works is to make probabilities and gather all of these probabilities together in a model.
00:10:19 - and then produces a conclusion which effectively says, I'm concluding this because of the evidence. So I I I, this is probably the right answer, is what the machine is effectively always concluding, always saying probably this. So these are very distinctive things, and by comparing that to military history and military theory, we can see that neither of these two ways of thinking are what military commanders use in practice. It's a relatively straightforward and, and well-known point in philosophy of science, that these different forms of logic,
00:10:59 - So if you are using inductive reasoning, you are always in this probabilistic way of thinking. You can't just do loads and loads of induction and then that turns into these other kinds of reasoning. so the classical problem of induction, is commonly attributed to, to David Hume, who's an 18th century philosopher. And in a sense, it, it's been handed down to us particularly as social scientists in this, truism that we have. That correlation doesn't equal causation, and that's really the core of it. If you keep seeing the same, things correlated together.
00:11:44 - and he says, ultimately you can never actually know that just because these two things always seem to happen one after another that you have discovered anything about. Cause people made fun of David team and still do that. If you actually adopted this thinking all the time in your life, then you'd be, you know, when you walk down the stairs in the morning to have breakfast, you'd be worried that you were gonna fall through the stairs because suddenly matter was solid. But then, you know, you've got no reason to think that your stairs will be solid tomorrow just because it was yesterday. And in in effect. I mean, this is gonna upset probably some philosophers of science, but that's okay.

"Limitations of AI in Military Decision-Making: The Incompatibility of Inductive and Abductive Reasoning"

00:12:23 - that we use other kinds of logic like abductive reasoning, to sort of get round these fundamental limitations. And we use different kinds of reasoning altogether that complement each other to actually exist in the world and make sensible decisions at the moment. Machines can't swap between these kinds of reasoning. Some of the proponents of machine learning say that some of the models can do abductive inference, which is to, to appeal to the best possible explanation for something. So rather than saying it's probably this to say,
00:13:05 - and I think this is the best possible explanation of something, their definition of what that is is actually different to what philosophers of science consider it to be. And it's a much thinner version. and so what they've tried to do, these, these technologists have said, here's my inductive based machine learning and it can do induction so well, that I can actually do some abductive reasoning now. But people like Larson and many others have pointed out that this is just fundamentally not possible according to at least this version of Western philosophy. The kind of decisions that you need
00:13:47 - And, the high value put on surprisingly unexpected is anathema to the way that these inductive based software machines work. So if you're only ever trained on what's come before, there'll be no strategic or military or tactical creativity from that AI commander that you hope to make, make, make decisions. And what we found in our discussions when, when thing on the paper was, so our, our criticisms of, of these, of this AI literature is the same, but just coming from different, sort
00:14:25 - So me having read Kitz all my academic career and Cameron coming from more a philosophy approach. so we found that that was effectively the same, a powerful argument. And, and there was, and this is where the kitch vision of war, part of the title comes from because a lot of the understanding of what war is and what it entails for command is and reflected in a lot of the AI optimist literature. And when you look at the nature of war and the kind of commands needed, the AI systems that we have just can't do that. They can't base the decisions on the unknown, even known unknowns, let alone unknown unknowns. 'cause there will be things nobody will have
00:15:06 - And what we end up with, in the real world then is, the, the major general of the HMS PNA force. So this, so I'll explain the, the title a bit now. So what you end up is with, a machine that can have all the facts at its disposal, at its disposal and know almost everything that I can know, but it'll be completely terrible or useless at making the right decisions or doing leadership duties. Has no human empathy, no human understanding, no concept of what strategy as opposed to knowledge is. So we'll end up with something, with a,
00:15:47 - that can have all the facts about obscure topics, but will not be able to make reliably good decisions at war because it has no understanding of what war really is. There are frankly wrong perceptions and understandings of what these technologies are and what they can and they can't do proliferating in a number of important communities. So obviously in the academic community, these are logically and empirically flawed arguments and perceptions of the AI technologies and what they can do, but also theoretically flawed approaches to what war is as well.

"Challenging Misconceptions: The Flawed Foundations of AI in Military Strategy and Policy"

00:16:30 - and understandings going around and in the academic community, which translates into the policy community as well, and also the wonk community as well. So the, advisors and experts in the national security defense and technology ecosystem, a lot of flawed understandings are flying around and you can't make good policy on invalid evidence, flawed logic and kitch theories of war. So that's sort of the fundamental sort of, so what we are trying to, correct that to try and show, this is where these arguments fall short
00:17:11 - fundamental nature of the technologies or the social and political phenomenon that they are engaging with. And an example could be that, and I've heard this in practitioner meetings I've had where they believe that military or security problems, even political ones, can be solved by trying to gain the whole system by quantifying everything, getting more data sets and more computing power. That's how you get, a machine that can decide, when to attack or retreat or when to, sue for peace versus antagonize, an enemy.
00:17:53 - that all factors can be known and digitized, put into binary ones and zeros in a way that computers can understand. When we are operating in a social and political world where there are no, you know, obvious right or wrongs, there's very rarely objective, qualities to the information and the knowledge and the activities that we have. So this means then that policy makers are, are possibly not getting the right advice, information needed when, acquisitions discussions come up and investments in research and development. So there are many companies and military bureaucracies pushing various
00:18:32 - that are making very grand claims about, we want to have an AI commander helping troops make, you know, difficult decisions when all it can do is present information in a certain way for better and worse. So the UK MOD has a project called Project Improbable, where it's very much is about trying to do the AI commander that the AI optimist literature says is just around the corner based on, well the, the flawed evidence that we present. another example could be, you know, palliatives artificial intelligence platform. Very grand claims about decision making capacities are being made and we would just, we are throwing skepticism on that based on argument and logic and evidence.
00:19:54 - machine learning slash artificial intelligence systems. And if we collectively as a national security community of academics and, and practitioners actually recognize that, then we might be able to understand what these technologies could be used for. Whereas at the moment, I think our sense is that people are dreaming very grand dreams about totally different kinds of technology that that don't exist. And at the moment, there's no technical development pathway that would actually lead us to, as a species, to having that technical capability.

Faculti AI can make mistakes

Research

Policy

Analysis