CIA to CAI? Spycraft in the Era of Machine Learning

While machine learning and artificial intelligence technologies enable the CIA to process vast amounts of information, algorithms are far from replacing human spies. James Bond will still have his job.

“…the Intelligence R&D Council formed the Artificial Intelligence Steering Group (AISG) for the expressed purpose of providing a mechanism for exchanging information throughout the Community regarding Artificial Intelligence (AI)”[1]. Americans, who tend to view government as a slow-moving, backwards bureaucracy, may be surprised to learn these words are from a 1984 CIA memo summarizing findings from a CIA-backed conference about intelligence applications of new computing techniques. Given recent technological leaps made in information gathering and processing technologies, one wonders: how is machine learning changing spycraft at the CIA?

An economic framework used by scholars Ajay Agrawal, Joshua Gans, and Avi Goldfarb to assess the impact of predictive analytics provides insight. Machine learning represents a decrease in the price of prediction and will increase demand for complements to prediction, such as judgment and creativity[2]. Since the CIA’s job is to gather intelligence, analyze it, and use the results to predict actions of foreign actors, it follows that new computational tools powered by machine learning will reduce time CIA analysts and operatives spend processing data and increase time spent drawing conclusions from it. To this end, the CIA is investing financially and organizationally to develop strong data processing and analytics capabilities.

In the short-term, the CIA is investing in both external and internal innovation. In 1999, the CIA chartered In-Q-Tel, a prestigious VC firm established to invest in technologies with national security applications[3]. Some notable portfolio companies include Palantir, whose data mining software is used across public and private sector institutions, and Keyhole, a mapping firm acquired by Google[4]. Today, the CIA has close to 140 machine learning and artificial intelligence projects underway[5]. In a real world spy-vs-spy case, one project helps agents avoid detection with a tool that maps cameras in foreign capitals by combining public images from Google Street View with algorithms trained to recognize cameras[6]. Another project uses open-source datasets to develop indicators of significant political and social events; the resulting tool predicted protests in Mexico triggered by an election and similar outbreaks in Paraguay connected to the impeachment of a former president[7]. A third project, sponsored by a different government agency but of benefit to the CIA, increases the amount of imagery an analyst can process by 95%[8].

In order to prepare for long-term changes to the intelligence trade, the CIA is changing itself structurally. On October 1st, 2015, in recognition of the need to build greater competency in digital approaches to data collection, management, and analysis, the CIA made its biggest organizational change since 1963 and established the Directorate of Digital Innovation (DDI) [9]. The Directorate is tasked with solving three problems: first, helping agents hone digital hacking and sleuthing skills; second, improving the CIA’s data management systems; and third, building predictive tools that leverage all of the CIA’s data[10]. According to CIA Deputy Director Andrew Hallman, who leads the new Directorate, “the days of attending a cocktail party and writing up your notes are over”[10]. Does this mean an end to human spycraft? According to Dawn Meyerriecks, CIA Deputy Director for Science and Technology, computation leads not to an end of human analysis, but one in which machine learning supercharges human capabilities[11].

While the information processing and predictive abilities provided by machine learning tools will improve CIA’s spying abilities, it is important to remember other nations are embracing these same technologies. As a result, CIA spies deployed abroad can be tracked remotely without human tails; in China, for example, image recognition powers cameras that blanket cities and can find fugitives in a crowd of 60,000 people[12]. Ironically, digitally enabled adversaries will necessitate the need for more, not less human spycraft, a view echoed by Jason Matthews, a CIA veteran and author of the Red Sparrow series [13]. The CIA will need human spies to create emotional bonds with potential agents, to manage their work, and perhaps most importantly, to compromise the digital systems and tools employed by adversaries.

To ensure access to talent needed to build its digital capabilities, I propose the CIA and its sibling agencies need to overcome an image problem. The CIA is competing for talent against deep pocketed companies, and it is easy for Americans to protest any kind of cooperation with the CIA based on its reputation as a nefarious entity. To counter this, I propose the CIA embark on a multiplatform marketing campaign designed to educate Americans of the benefits that flow from defense and intelligence investment in science and technology; likewise, since machine learning algorithms are notoriously difficult to debug, I propose the Agency can establish a panel to oversee the use of these tools. This said, it is worth contemplating: should the CIA be developing these capabilities internally? What other steps can the CIA take to stay ahead of competing agencies?

(793 Words)

[1] Chairman, AI Steering Group to Chairman, Intelligence R&D Council, memorandum regarding Issues in Artificial Intelligence, August 21st1984, Intelligence Research & Development Council, from CIA Library: https://www.cia.gov/library/readingroom/docs/CIA-RDP86M00886R000500040004-2.pdf, accessed November 2018.

[2] Ajay Agrawal, Joshua Gans, and Avi Goldfarb, “The Simple Economics of Machine Intelligence,” Harvard Business Review, November 17th, 2016, https://hbr.org/2016/11/the-simple-economics-of-machine-intelligence, accessed November 2018.

[3] George Tenet, At The Center Of The Storm: My Years at the CIA(New York, NY: HarperCollins, 2007), p. 26.

[4] In-Q-Tel, “Portfolio”, https://www.iqt.org/portfolio/, accessed November 2018.

[5] Patrick Tucker, “What the CIA’s Tech Director Wants from AI,” Defense One (A Property of Atlantic Media), September 6th, 2017,https://www.defenseone.com/technology/2017/09/cia-technology-director-artificial-intelligence/140801/, accessed November 2018.

[6] Jenna McLaughlin, “CIA agents in ‘about 30 countries’ being tracked by technology, top official says,” CNN, April 22nd, 2018,https://www.cnn.com/2018/04/22/politics/cia-technology-tracking/index.html, accessed November 2018.

[7] Patrick Tucker, “Meet the Man Reinventing CIA for the Big Data Era”, Defense One (A Property of Atlantic Media), October 1st, 2015, https://www.defenseone.com/technology/2015/10/meet-man-reinventing-cia-big-data-era/122453/, accessed November 2018.

[8] Michael Phillips, “The U-2 Spy Plane Is Still Flying Combat Missions 60 Years After Its Debut,” The Wall Street Journal, June 8th, 2018, https://www.wsj.com/articles/the-u-2-spy-plane-still-flying-combat-missions-60-years-after-debut-1528382700, accessed November 2018.

[9] CIA, “Offices of CIA”, https://www.cia.gov/offices-of-cia/digital-innovation, accessed November 2018.

[10] Patrick Tucker, “Meet the Man Reinventing CIA for the Big Data Era”, Defense One (A Property of Atlantic Media), October 1st, 2015, https://www.defenseone.com/technology/2015/10/meet-man-reinventing-cia-big-data-era/122453/, accessed November 2018.

[11] Jenna McLaughlin, “The Robots Will Run the CIA, Too,” Foreign Policy, September 7th, 2017, https://foreignpolicy.com/2017/09/07/the-robots-will-run-the-cia-too/, accessed November 2018.

[12] Amy Wang, “A suspect tried to blend in with 60,000 concertgoers. China’s facial-recognition cameras caught him.,” The Washington Post, April 13th, 2018, https://www.washingtonpost.com/news/worldviews/wp/2018/04/13/china-crime-facial-recognition-cameras-catch-suspect-at-concert-with-60000-people/?utm_term=.24d9d908db07, accessed November 2018.

[13] Javier David, “’Red Sparrow’ used to be an actual phenomenon during the Cold War, and in some ways still is: Author,” CNBC, March 3rd, 2018, https://www.cnbc.com/2018/03/03/red-sparrow-used-to-be-an-actual-phenomenon-during-the-cold-war.html, accessed November 2018.

Previous:

3D Printing and Humanitarian Aid, Will the US Learn from Past Disastrous Response?

Next:

Hinge and Machine Learning: The makings of a perfect match

Student comments on CIA to CAI? Spycraft in the Era of Machine Learning

  1. I agree James Bond will not be out of a job for a long time since HUMINT will never cease to require American agents, as opposed to various other things such as SIGINT or IMINT, which are becoming more and more the domain of AI, especially on the analysis side.
    However, do you think that projecting a new image and educating Americans on the importance of the CIA will actually attract top tier programming talent from the private sector to the agency when the compensation is so uneven? I don’t believe it is an image problem as much as a money problem. It follows they must raise salaries or be forced continue to contract work out to companies like Palantir, who can afford more attractive compensation for its employees.

    1. You make a good point; however, many branches of government service are able to attract and retain talented people because those people are inspired by the idea of serving a higher purpose. This said, there is no denying the fact that many also leave government because pay is low compared to the private sector. I would propose that programmers and data scientists are, as a group, motivated by

      1) Opportunity
      2) Money, and
      3) Job Perception

      The data science roles at the CIA offer access to tools and pools of data that are perhaps only rivaled by those at top Big Tech firms, so the opportunity is quite compelling. Ironically, the government does have the money to pay market rates, and proof for this can be found in a Wall Street Journal article about Palantir was just published today[1]. The article mentions that the only part of the business have ever turned a profit is the government services arm; this fact can be read to mean that the government is paying ABOVE market rates for the programming talent at Palantir to come work for them, but sadly, the CIA does not appear to offer competitive salaries for data scientists working on the inside as CIA employees[2][3]. Finally, the perception of CIA and intelligence agencies is hurting in the post-Snowden era, and this can impact the ability to attract and retain talent even if CIA is paying market rates.

      Perhaps the solution I proposed above is too simplistic, and a more interesting question is whether the CIA should pay market rates for programming talent AND use an image makeover/marketing campaign to make it a place programmers and data scientists would feel more interested in joining. I wonder what political hoops would need to be cleared for CIA to pay the millions that top data scientists now command[4]?

      [1] Rob Copeland and Eliot Brown, “Palantir Has a $20 Billion Valuation and a Bigger Problem: It Keeps Losing Money”, The Wall Street Journal, November 13th, 2018, https://www.wsj.com/articles/palantir-has-a-20-billion-valuation-and-a-pretty-big-problem-it-keeps-losing-money-1542042135, accessed November 2018.

      [2] CIA, “Careers & Internships,” https://www.cia.gov/careers/opportunities/science-technology/data-scientist.html, accessed November 2018.

      [3] Paysa, “Palantir Technologies Data Scientist Salaries,” https://www.paysa.com/salaries/palantir-technologies–data-scientist, accessed November 2018.

      [4] Cade Metz, “A.I. Researchers Are Making More Than $1 Million, Even at a Nonprofit”, The New York Times, April 19th, 2018, https://www.nytimes.com/2018/04/19/technology/artificial-intelligence-salaries-openai.html, accessed November 2018.

  2. often when the topic of machine learning is brought up, the discourse tends to exaggerate the replacement effect while muting the potential for machine/human complementarity. This post definitely avoids making that error and instead, we are presented with a very well clear picture of a world where “machine learning supercharges human capabilities.” the fact that image analysts are potentially 95% more efficient at their roles speaks volumes. At the same time, I found it sobering to ponder the possibility that this improvement is likely to result in a zero-sum AI arms race rather than a value create.

  3. I think CIA should not develop these capabilities internally. Even if CIA is able to attract some good talent it wont be able to develop as much as all the big tech firms together. The companies are scattered all around the states and the people are motivated from very different things. I don’t think a marketing campaign and raising salaries will be enough to get the best data scientists on board. There is much more than that.
    What I propose instead is for them to partner with the tech giant companies or academia and have them develop the technology that they want. This will need a lot of rules to be set and potentially difficult agreements to be reached, but I think is a good way to get the best of all the technology developments and at the same time monitor what is going on in terms of AI potential harmful uses.

  4. This is a really interesting article to explain one of the new applications of machine learning. It is quite surprising that CIA and other national intelligence agency already started proactively leveraging machine learning to improve data gathering and data analysis. Given that machine learning is still at its early stage and needs more technological innovation, I would suggest collaborating with external parties such as tech giant mentioned in the previous comments as much as possible without harming confidentiality. One thought is whether it’s possible for CIA to collaborate with intelligence agencies in the U.S. allies countries in order to share best practice in utilizing machine learning for intelligence tasks. This idea sounds risky but it will be beneficial since machine learning generally works better with more data.

Leave a comment