The CIA and FBI will receive a closed analogue of ChatGPT for working with open sources

US intelligence agencies will acquire generative artificial intelligence (AI) in the form of an analogue of ChatGPT, which will allow them to sift through huge amounts of publicly available information in open or commercially available sources to search for evidence. According to Randy Nixon, director of the CIA’s open source Enterprise division, it plans to provide its AI tool to the relevant US agencies in the near future.

According to Nixon, the CIA is preparing to implement a function similar to the well-known program developed by American machine learning specialist OpenAI Inc. The tool will allow intelligence personnel users to see the original source of the information they are viewing, Nixon said. He added that the chat feature is a logical part of disseminating intelligence information more quickly. “Then you can take it to the next level and start communicating and asking questions to machines, which will give you answers that are also sourced,” Nixon said (as quoted by Bloomberg).

The AI ​​tool will be available to all 18 US intelligence agencies, including the CIA, NSA, FBI and agencies run by branches of the military. Nixon did not say what kind of AI model the intelligence agency would use as the basis for its new tool or how it would protect information from leaking onto the open Internet.

Now the US intelligence agencies already meet the necessary technical requirements for using AI – they have modern neural network architectures, a large amount of available data, and computing resources, says Olga Migacheva, junior researcher at the Institute of Theoretical and Mathematical Physics of Moscow State University. She suggests that a deep neural network architecture like Transformer will likely be used.

But, the expert continues, the question remains about whether it is safe to trust the results generated by such a neural network, since it will be fed with huge amounts of publicly available information, including unreliable data. “At this stage, it is impossible to do without human verification and verification of the results, as in medical cases: a neural network can suggest a patient’s diagnosis based on the data, but the final decision is made by the doctor. It is possible that sources and data will be filtered by intelligence officers before they are used for processing using a neural network,” Migacheva said.

ChatGPT and its analogues in general are able to coherently answer questions and summarize available information, notes Alexander Zhukov, director of business development for the IT developer “Code Format”. According to him, the neural network summarizes everything it has read, but if additional data in combination with open information is provided at the input of a similar network instead of forums and Wikipedia, it will be able to consolidate everything and coherently describe various parameters and find primary sources of information. “The role of this network can be formulated as a kind of “automatic analytical department” that can sift through the mass of information provided to it and simply known in advance, explain the detected patterns to the operator and answer additional questions. Will it always be possible to accurately determine the original source? The question is more about how well it was hidden. In simple cases, for example, for social waves in social networks, you can always detect the source of the wave,” explains Zhukov.

It is ChatGPT that has problems with primary sources, recalls Vladimir Sveshnikov, co-founder of Parodist AI. According to him, the model makes factual errors. “True, this problem is now being actively solved. The main reason why various companies and, apparently, government agencies are already developing their own analogues is confidentiality. OpenAI has access to all user conversations, and not everyone wants to share data with them,” says Sveshnikov.

Nixon said the technology would not be available to politicians or the public, and the CIA “closely follows U.S. privacy laws.” But from the point of view of legislation in the United States, neural networks are still in a gray zone; there are no significant restrictions on this area, recalls Dmitry Gorbunov, partner at Rustam Kurmaev and Partners. He notes that neural networks are not prohibited; work on their creation and development is not subject to licensing or other control by executive authorities. The only thing that limits developers and those in whose interests neural networks are created are the rules governing the collection, processing, storage and use of personal data, information constituting commercial, state, military, medical, and banking secrets. “Whether these restrictions apply to the activities of special services, investigative and state security agencies and other law enforcement agencies is a question that does not have a clear answer,” notes Gorbunov.

According to him, on the one hand, American law guarantees the protection of personal data, the confidentiality of correspondence and other information that cannot be freely disseminated without the permission of the subject of the information, including from access to it not authorized by the court. “If a person does not give permission to collect, process, store and transfer data to third parties, any of these actions will be contrary to the law,” Gorbunov notes.

On the other hand, it is important to understand that in the United States, as in most countries, national security interests come to the fore in a number of issues, the lawyer emphasized. Such structures may be endowed with emergency powers that allow them to bypass restrictions, and such decisions may be made without properly informing data subjects “in the interests of ensuring state security,” Gorbunov argues.

“The sensational story with the PRISM program, within the framework of which information stored on the servers of Microsoft, Apple, Facebook, Google could be collected (the “Yarovaya package” was the response to this story in Russia), shows that it is possible to completely eliminate abuses and leaks in the field of collection and storing information is impossible,” says Gorbunov. In this case, it recalls the scandal of 2013, when the American press received information about the mass surveillance of the NSA using PRISM over the correspondence and conversations of US citizens and foreigners.

According to Gorbunov, there is no significant difference in whether data is collected manually or using neural networks, neither from the point of view of legal compliance requirements, nor from the point of view of the risks of leaks. “The protection can be hacked, any information to which at least one person has access can leak,” the lawyer concludes.