TRAINING MATERIALS REVIEWED by The Intercept confirm that Google is offering advanced artificial intelligence and machine-learning capabilities to the Israeli government through its controversial “Project Nimbus” contract. The Israeli Finance Ministry announced the contract in April 2021 for a $1.2 billion cloud computing system jointly built by Google and Amazon. “The project is intended to provide the government, the defense establishment and others with an all-encompassing cloud solution,” the ministry said in its announcement.

Google engineers have spent the time since worrying whether their efforts would inadvertently bolster the ongoing Israeli military occupation of Palestine. In 2021, both Human Rights Watch and Amnesty International formally accused Israel of committing crimes against humanity by maintaining an apartheid system against Palestinians. While the Israeli military and security services already rely on a sophisticated system of computerized surveillance, the sophistication of Google’s data analysis offerings could worsen the increasingly data-driven military occupation.

According to a trove of training documents and videos obtained by The Intercept through a publicly accessible educational portal intended for Nimbus users, Google is providing the Israeli government with the full suite of machine-learning and AI tools available through Google Cloud Platform. While they provide no specifics as to how Nimbus will be used, the documents indicate that the new cloud would give Israel capabilities for facial detection, automated image categorization, object tracking, and even sentiment analysis that claims to assess the emotional content of pictures, speech, and writing. The Nimbus materials referenced agency-specific trainings available to government personnel through the online learning service Coursera, citing the Ministry of Defense as an example.

Jack Poulson, director of the watchdog group Tech Inquiry, shared the portal’s address with The Intercept after finding it cited in Israeli contracting documents. 

“The former head of Security for Google Enterprise — who now heads Oracle’s Israel branch — has publicly argued that one of the goals of Nimbus is preventing the German government from requesting data relating on the Israel Defence Forces for the International Criminal Court,” said Poulson, who resigned in protest from his job as a research scientist at Google in 2018, in a message. “Given Human Rights Watch’s conclusion that the Israeli government is committing ‘crimes against humanity of apartheid and persecution’ against Palestinians, it is critical that Google and Amazon’s AI surveillance support to the IDF be documented to the fullest.”

Though some of the documents bear a hybridized symbol of the Google logo and Israeli flag, for the most part they are not unique to Nimbus. Rather, the documents appear to be standard educational materials distributed to Google Cloud customers and presented in prior training contexts elsewhere.

Google did not respond to a request for comment.

The documents obtained by The Intercept detail for the first time the Google Cloud features provided through the Nimbus contract. With virtually nothing publicly disclosed about Nimbus beyond its existence, the system’s specific functionality had remained a mystery even to most of those working at the company that built it. In 2020, citing the same AI tools, U.S Customs and Border Protection tapped Google Cloud to process imagery from its network of border surveillance towers.

Many of the capabilities outlined in the documents obtained by The Intercept could easily augment Israel’s ability to surveil people and process vast stores of data — already prominent features of the Israeli occupation.

“Data collection over the entire Palestinian population was and is an integral part of the occupation,” Ori Givati of Breaking the Silence, an anti-occupation advocacy group of Israeli military veterans, told The Intercept in an email. “Generally, the different technological developments we are seeing in the Occupied Territories all direct to one central element which is more control.”

THE ISRAELI SECURITY state has for decades benefited from the country’s thriving research and development sector, and its interest in using AI to police and control Palestinians isn’t hypothetical. In 2021, the Washington Post reported on the existence of Blue Wolf, a secret military program aimed at monitoring Palestinians through a network of facial recognition-enabled smartphones and cameras.

“Living under a surveillance state for years taught us that all the collected information in the Israeli/Palestinian context could be securitized and militarized,” said Mona Shtaya, a Palestinian digital rights advocate at 7amleh-The Arab Center for Social Media Advancement, in a message. “Image recognition, facial recognition, emotional analysis, among other things will increase the power of the surveillance state to violate Palestinian right to privacy and to serve their main goal, which is to create the panopticon feeling among Palestinians that we are being watched all the time, which would make the Palestinian population control easier.”

The educational materials obtained by The Intercept show that Google briefed the Israeli government on using what’s known as sentiment detection, an increasingly controversial and discredited form of machine learning. Google claims that its systems can discern inner feelings from one’s face and statements, a technique commonly rejected as invasive and pseudoscientific, regarded as being little better than phrenology. In June, Microsoft announced that it would no longer offer emotion-detection features through its Azure cloud computing platform — a technology suite comparable to what Google provides with Nimbus — citing the lack of scientific basis.

Google does not appear to share Microsoft’s concerns. One Nimbus presentation touted the “Faces, facial landmarks, emotions”-detection capabilities of Google’s Cloud Vision API, an image analysis toolset. The presentation then offered a demonstration using the enormous grinning face sculpture at the entrance of Sydney’s Luna Park. An included screenshot of the feature ostensibly in action indicates that the massive smiling grin is “very unlikely” to exhibit any of the example emotions. And Google was only able to assess that the famous amusement park is an amusement park with 64 percent certainty, while it guessed that the landmark was a “place of worship” or “Hindu Temple” with 83 percent and 74 percent confidence, respectively.

Google workers who reviewed the documents said they were concerned by their employer’s sale of these technologies to Israel, fearing both their inaccuracy and how they might be used for surveillance or other militarized purposes. 

“Vision API is a primary concern to me because it’s so useful for surveillance,” said one worker, who explained that the image analysis would be a natural fit for military and security applications. “Object recognition is useful for targeting, it’s useful for data analysis and data labeling. An AI can comb through collected surveillance feeds in a way a human cannot to find specific people and to identify people, with some error, who look like someone. That’s why these systems are really dangerous.”

The employee — who, like all of the Google workers who spoke to The Intercept, requested anonymity to avoid workplace reprisals — added that they were further alarmed by potential surveillance or other militarized applications of AutoML, another Google AI tool offered through Nimbus. Machine learning is largely the function of training software to recognize patterns in order to make predictions about future observations, for instance by analyzing millions of images of kittens today in order to confidently claim that it’s looking at a photo of a kitten tomorrow. This training process yields what’s known as a “model” — a body of computerized education that can be applied to automatically recognize certain objects and traits in future data. 

Training an effective model from scratch is often resource intensive, both financially and computationally. This is not so much of a problem for a world-spanning company like Google, with an unfathomable volume of both money and computing hardware at the ready. Part of Google’s appeal to customers is the option of using a pre-trained model, essentially getting this prediction-making education out of the way and letting customers access a well-trained program that’s benefited from the company’s limitless resources.

Cloud Vision is one such pre-trained model, allowing clients to immediately implement a sophisticated prediction system. AutoML, on the other hand, streamlines the process of training a custom-tailored model, using a customer’s own data for a customer’s own designs. Google has placed some limits on Vision — for instance limiting it to face detection, or whether it sees a face, rather than recognition that would identify a person. AutoML, however, would allow Israel to leverage Google’s computing capacity to train new models with its own government data for virtually any purpose it wishes. “Google’s machine learning capabilities along with the Israeli state’s surveillance infrastructure poses a real threat to the human rights of Palestinians,” said Damini Satija, who leads Amnesty International’s Algorithmic Accountability Lab. “The option to use the vast volumes of surveillance data already held by the Israeli government to train the systems only exacerbates these risks.” 

Custom models generated through AutoML, one presentation noted, can be downloaded for offline “edge” use — unplugged from the cloud and deployed in the field.

THAT NIMBUS LETS Google clients use advanced data analysis and prediction in places and ways that Google has no visibility into creates a risk of abuse, according to Liz O’Sullivan, CEO of the AI auditing startup Parity and a member of the U.S. National Artificial Intelligence Advisory Committee. “Countries can absolutely use AutoML to deploy shoddy surveillance systems that only seem like they work,” O’Sullivan said in a message. “On edge, it’s even worse — think bodycams, traffic cameras, even a handheld device like a phone can become a surveillance machine and Google may not even know it’s happening.”

In one Nimbus webinar reviewed by The Intercept, the potential use and misuse of AutoML was exemplified in a Q&A session following a presentation. An unnamed member of the audience asked the Google Cloud engineers present on the call if it would be possible to process data through Nimbus in order to determine if someone is lying.

“I’m a bit scared to answer that question,” said the engineer conducting the seminar, in an apparent joke. “In principle: Yes. I will expand on it, but the short answer is yes.” Another Google representative then jumped in: “It is possible, assuming that you have the right data, to use the Google infrastructure to train a model to identify how likely it is that a certain person is lying, given the sound of their own voice.” Noting that such a capability would take a tremendous amount of data for the model, the second presenter added that one of the advantages of Nimbus is the ability to tap into Google’s vast computing power to train such a model.

A broad body of research, however, has shown that the very notion of a “lie detector,” whether the simple polygraph or “AI”-based analysis of vocal changes or facial cues, is junk science. While Google’s reps appeared confident that the company could make such a thing possible through sheer computing power, experts in the field say that any attempts to use computers to assess things as profound and intangible as truth and emotion are faulty to the point of danger.

One Google worker who reviewed the documents said they were concerned that the company would even hint at such a scientifically dubious technique. “The answer should have been ‘no,’ because that does not exist,” the worker said. “It seems like it was meant to promote Google technology as powerful, and it’s ultimately really irresponsible to say that when it’s not possible.”

Andrew McStay, a professor of digital media at Bangor University in Wales and head of the Emotional AI Lab, told The Intercept that the lie detector Q&A exchange was “disturbing,” as is Google’s willingness to pitch pseudoscientific AI tools to a national government. “It is [a] wildly divergent field, so any technology built on this is going to automate unreliability,” he said. “Again, those subjected to them will suffer, but I’d be very skeptical for the citizens it is meant to protect that these systems can do what is claimed.”

According to some critics, whether these tools work might be of secondary importance to a company like Google that is eager to tap the ever-lucrative flow of military contract money. Governmental customers too may be willing to suspend disbelief when it comes to promises of vast new techno-powers. “It’s extremely telling that in the webinar PDF that they constantly referred to this as ‘magical AI goodness,’” said Jathan Sadowski, a scholar of automation technologies and research fellow at Monash University, in an interview with The Intercept. “It shows that they’re bullshitting.”

GOOGLE, LIKE MICROSOFT, has its own public list of “AI principles,” a document the company says is an “ethical charter that guides the development and use of artificial intelligence in our research and products.” Among these purported principles is a commitment to not “deploy AI … that cause or are likely to cause overall harm,” including weapons, surveillance, or any application “whose purpose contravenes widely accepted principles of international law and human rights.” 

Israel, though, has set up its relationship with Google to shield it from both the company’s principles and any outside scrutiny. Perhaps fearing the fate of the Pentagon’s Project Maven, a Google AI contract felled by intense employee protests, the data centers that power Nimbus will reside on Israeli territory, subject to Israeli law and insulated from political pressures. Last year, the Times of Israel reported that Google would be contractually barred from shutting down Nimbus services or denying access to a particular government office even in response to boycott campaigns.

Google employees interviewed by The Intercept lamented that the company’s AI principles are at best a superficial gesture. “I don’t believe it’s hugely meaningful,” one employee told The Intercept, explaining that the company has interpreted its AI charter so narrowly that it doesn’t apply to companies or governments that buy Google Cloud services. Asked how the AI principles are compatible with the company’s Pentagon work, a Google spokesperson told Defense One, “It means that our technology can be used fairly broadly by the military.”

Moreover, this employee added that Google lacks both the ability to tell if its principles are being violated and any means of thwarting violations. “Once Google offers these services, we have no technical capacity to monitor what our customers are doing with these services,” the employee said. “They could be doing anything.” Another Google worker told The Intercept, “At a time when already vulnerable populations are facing unprecedented and escalating levels of repression, Google is backsliding on its commitments to protect people from this kind of misuse of our technology. I am truly afraid for the future of Google and the world.”

Ariel Koren, a Google employee who claimed earlier this year that she faced retaliation for raising concerns about Nimbus, said the company’s internal silence on the program continues. “I am deeply concerned that Google has not provided us with any details at all about the scope of the Project Nimbus contract, let alone assuage my concerns of how Google can provide technology to the Israeli government and military (both committing grave human rights abuses against Palestinians daily) while upholding the ethical commitments the company has made to its employees and the public,” she told The Intercept in an email. “I joined Google to promote technology that brings communities together and improves people’s lives, not service a government accused of the crime of apartheid by the world’s two leading human rights organizations.”

Sprawling tech companies have published ethical AI charters to rebut critics who say that their increasingly powerful products are sold unchecked and unsupervised. The same critics often counter that the documents are a form of “ethicswashing” — essentially toothless self-regulatory pledges that provide only the appearance of scruples, pointing to examples like the provisions in Israel’s contract with Google that prevent the company from shutting down its products. “The way that Israel is locking in their service providers through this tender and this contract,” said Sadowski, the Monash University scholar, “I do feel like that is a real innovation in technology procurement.”

To Sadowski, it matters little whether Google believes what it peddles about AI or any other technology. What the company is selling, ultimately, isn’t just software, but power. And whether it’s Israel and the U.S. today or another government tomorrow, Sadowski says that some technologies amplify the exercise of power to such an extent that even their use by a country with a spotless human rights record would provide little reassurance. “Give them these technologies, and see if they don’t get tempted to use them in really evil and awful ways,” he said. “These are not technologies that are just neutral intelligence systems, these are technologies that are ultimately about surveillance, analysis, and control.”