Original subtitles

I miss how my family used to gather

at the end of the day.

How we used to talk.

My home was like a normal home.

The simple, daily details that everyone has.

Heba lived here, in northern Gaza.

Her family evacuated on October 11th, 2023.

By February, she learned her home was no longer there.

She talked to me from a friend's home, in Rafah, in southern Gaza.

We received a picture of our house

and we were in shock.

We had down there, like,

a place where where we have trees,

we have flowers planted.

Heba didn't know exactly why her home had been destroyed.

But over the past few months, Israeli journalists

have found that much of the destruction in Gaza

since the attacks of October 7th

has been enabled and often directed by an artificial intelligence system.

The promise of AI generally

is a promise in two respects.

One is swiftness and the second is accuracy.

The whole dream of AI is

that it would offer these precision strikes.

But after over 34,000 Palestinians killed,

compared to just over 1,400 in Israel's 2014 war in Gaza,

it's clear something different is happening.

So what does AI have to do with it?

To get some answers, we called a couple of AI experts, reporters

and investigative journalists.

The Israeli Defense Forces’ use of AI is not new.

I think that the most famous use of AI

by the IDF is, of course, the Iron Dome,

which is a defensive system that aims to disrupt

the threat of missile attacks.

This system is partly what defended Israel

against Iran's drone and missile attacks in April 2024.

The other one is another homegrown weapon that they have called

the SMASH from Smartshooter,

which is an AI precision assault rifle sight

that you add on to handheld weapons.

And what it does is it uses advanced

image-processing algorithms to hone in on a target,

sort of like a an auto-aim in Call of Duty.

Another way Israel uses AI is through surveillance

of Palestinians in the occupied territories.

Every time they pass through one of the hundreds

of checkpoints, their movements are being registered,

Their facial images and other biometrics

are being matched against a database.

But we're now learning more about the AI systems

that choose bombing targets in Gaza,

from two reports in the Israeli publications +972 and Local Call.

Gospel is a system that produces bombing targets

for specific buildings and structures in Gaza.

It does this by working in conjunction with other AI tools.

And like any AI system,

the first step is the large-scale collection of data.

In this case, surveillance and historical data

on Palestinian and militant locations in Gaza.

The most famous application,

would be Alchemist,

which is a platform that collects data

and allows the transfer of data between different departments

later being transferred to another platform, which is called the Fire Factory.

The Fire Factory observes the data and categorizes it.

The generated targets are generally put into one of four categories.

First, tactical targets,

which usually include armed militant cells, weapons warehouses,

launchers and militant headquarters.

Then there are underground targets,

primarily tunnels under civilian homes.

The third category includes the family homes

of Hamas or Islamic Jihad operatives.

And the last category includes targets that are not obviously military in nature,

particularly residential and high-rise buildings with dozens of civilians.

The IDF calls these power targets.

Once the data is organized,

it goes through a third layer called the Gospel.

The Gospel creates an output

which suggests specific possible targets,

possible munitions,

warnings of possible collateral damage, and etc.

This system produces targets in Gaza faster than a human can.

And within the first five days of the war,

half of all the targets identified were from the Power Targets category.

Multiple sources who spoke to +972 reported that the idea behind power targets

is to exert civil pressure on Hamas.

Heba’s home was most likely one of the power targets

picked up by the Gospel system.

Months after the Gospel investigation, +972

also surfaced a more opaque and secretive AI system,

built for targeting specific people,

known as Lavender.

As the Israel-Hamas war began,

Lavender used historic data and surveillance

to generate as many as 37,000 Hamas and Islamic Jihad targets.

Sources told +972 that about 10% of

those targets are often wrong.

But even when determining the 90% of supposedly correct targets,

Israel also expanded the definition

of a Hamas operative for the first time.

The thing is, Hamas ultimately runs the Gaza Strip.

So you have a lot of civil society that interacts with Hamas.

Police force, doctors, civil society in general.

And so these are the targets that we know that they're looking at.

After Lavender used its data to generate these targets,

AI would then link the target to a specific family home,

and then recommend a weapon for the IDF to use on the target,

mostly depending on the ranking of the operative.

What we were told is that for low-ranking Hamas militants,

the army preferred to use “dumb bombs,”

meaning bombs that are not guided, because they are cheaper.

So in a strange way, the less of a danger you posed,

then they used less sophisticated bombs,

therefore maybe creating more collateral damage.

Sources told reporters that for every junior Hamas

operative that Lavender marked,

it was permissible to kill up to 15 or 20 civilians.

But also that for some targets,

the number of permissible civilian casualties

was as high as 300.

[Arabic] More than 50 displaced people were in the building.

More than 20 children were in it.

AI systems do not produce facts.

They only produce prediction,

just like a weather forecast or the stock market.

The “intelligence” that’s there

is completely dependent on the quality, the validity,

the understanding of the humans

who created the system.

In a statement to the Guardian, the IDF “outright rejected”

that they had “any policy to kill tens of thousands of people in their homes”

and stressed that human analysts must conduct

independent examinations before a target is selected.

Which brings us to the last step of both of these processes:

Human approval.

Sources told +972 that the only human supervision protocol in place

before bombing the houses of suspected junior militants marked by Lavender,

was to conduct a single check:

Ensuring that the AI-selected target is male

rather than female.

Experts have been telling us that essentially what's happening in Gaza

is an unwilling test site for future AI technologies.

In November 2023,

the US released an international framework

for the responsible use of AI in war.

More than 50 signatures from 50 different countries.

Israel has not signed on to this treaty.

So we're in sort of this space

where we lack sufficient oversight

and accountability for drone warfare,

let alone new systems being introduced like Gospel and Lavender.

And we're looking at a future really, where

there is going to be more imprecise

and biased automation of targets

that make these civilian casualties much worse.

The fallacy of,

you know, the premise that faster war fighting is somehow

going to lead to global security and peace.

I mean, this is just not the path that's going to get us there.

And on the contrary,

I think a lot of the momentum of these technological initiatives

needs to be interrupted, in whatever ways we can.

It really aches my heart that these moments are never going to be back.

It's not like I left home and like, for example,

I traveled and I know it's there.

No, it's not.

It's not there anymore.