How intelligence agencies are carefully adopting generative AI

Confident that they will be buried in data as sensor-generated surveillance technology blankets more of the globe, U.S. intelligence agencies are rushing to embrace the AI ​​revolution — and to keep pace with competitors who are already using AI to deploy deepfakes on social media platforms.

But the technology is still new and fragile, and authorities are acutely aware that generative AI is not built for dangerous and deceptive transactions.

Years before OpenAI’s ChatGPT sparked the current generative AI marketing craze, U.S. intelligence and defense agencies were experimenting with the technology. One contractor, Rhombus Power, used the tool in 2019 to uncover fentanyl trafficking in China faster than human analysis. Rhombus would later predict Russia’s full-scale invasion of Ukraine four months in advance with 80% accuracy.

Implementing AI is not easy

CIA Director William Burns recently wrote in Foreign Affairs magazine that US intelligence agencies need “sophisticated artificial intelligence models that can digest massive amounts of open-source and clandestine information.”

But Nand Mulchandani, the agency's first chief technology officer, said generative AI models “show hallucinations,” making them more like “crazy drunk friends with incredible insight but prone to bias.” ”, it is best treated as such.

There are also security and privacy issues. Adversaries can steal and tamper with them. It may contain sensitive personal data that agents are not authorized to view.

Mulchandani says Gen AI is primarily good as a virtual assistant, looking for “the needle in the needle pile,” but it will never replace human analysts, officials insist.

Open source AI “OSIRIS”

Officials have not said whether they are using the generated AI for any larger purpose on classified networks, but thousands of analysts from 18 U.S. intelligence agencies are currently using the CIA-developed Osiris. It uses a generation AI called. Ingest unclassified public or commercial data (so-called open source) and write annotated summaries. A chatbot is included so analysts can ask follow-up questions.

Osiris uses multiple commercial AI models. Mulchandani said the agency does not intend to commit to any particular model or technology vendor. “It's still early days,” he said.

Experts believe that predictive analytics, wargaming, and scenario brainstorming will be among the most important uses of generative AI for information workers.

“Regular AI” is already in use

Even before generative AI, intelligence agencies were using machine learning and algorithms. One use case is to alert analysts to potentially important developments outside of business hours. The analyst could tell the AI ​​to ring their phone no matter the time of day. It couldn't explain what happened, and it would be classified, but it could say, “Come on in and take a look at this.”

AI giants competing for U.S. intelligence business include Microsoft, which announced on May 7 that it would offer OpenAI's GPT-4 for classified networks, though the product has not yet been certified for such purposes.

Competitor Primer AI listed two intelligence agencies as customers in a recent military AI workshop published online. One of Primer's products uses AI to search more than 60,000 news and social media sources in 100 languages, including Twitter, Telegram, Reddit, and Discord, to “detect new signs of sudden events.” It is designed to.

Like Rhombus Power's product, it also uses computer vision to help analysts identify key people, organizations and locations.At a demonstration days after the Oct. 7 Hamas attack on Israel, Primer executives explained how their technology separates fact from fiction in the flood of online information from the Middle East.

Future challenges as AI becomes more widespread

For U.S. intelligence officials, the most important near-term AI challenge is likely to be countering how adversaries are using it: to penetrate U.S. defenses, spread disinformation and try to undermine Washington's ability to read their intentions and capabilities.

The White House is also concerned that generative AI models employed by U.S. government agencies could be infiltrated and contaminated.

Another concern is ensuring the privacy of people whose personal data may be embedded in AI models. Officials say it is currently impossible to guarantee that all personal data will be removed from AI models.

That's one reason intelligence agencies aren't in “move fast and break things” mode when it comes to generative AI, said John Byler, the top AI official at the Office of the Director of National Intelligence.

If government agencies were to use AI to investigate biological or cyberweapons technologies, there would be concerns about the integrity and security of the models.

Different Agencies, Different AI Missions

How AI is deployed varies widely between intelligence agencies, depending on their mission. The National Security Agency primarily intercepts communications. The National Geospatial-Intelligence Agency (NGA) is tasked with seeing and understanding every corner of the globe.

Using Gen AI to enhance these missions is a priority and is far less complicated than, say, how the FBI uses the technology given legal limitations on domestic surveillance.

In December, the NGA issued a request for proposals for an entirely new type of AI model that would use imagery collected from satellites and ground sensors to gather precise geospatial information with simple voice or text prompts. Gen AI applications also have huge implications for cyber conflict.

Matching with rivals

Generative AI can’t easily compete in wits with rival masters of deception.

Former defense intelligence officer Zachary Tyson Brown said analysts are dealing with “incomplete, ambiguous, and often contradictory pieces of partial, unreliable information.” He believes that if intelligence agencies adopt generative AI too eagerly, quickly, or completely, it will be a recipe for disaster. The model has no reason. They're just predicting. And their designers can't fully explain how they work.

Linda Weisgold, a former deputy director for analysis at the CIA, doesn't think AI will replace human analysts anytime soon.

Rapid decisions are often required based on incomplete data. She says the intelligence community's “customers,” most importantly the U.S. president, want human insight and experience to be central to the decision options presented.

“I don't think it would ever be acceptable to the president for the intelligence community to come in and say, 'We don't know. That's just what the black boxes tell us.'”

© 2024 The Associated Press. All rights reserved. This material may not be published, broadcast, rewritten or redistributed without permission.