Before ChatGPT: The CIA’s Decades-Long Pursuit of Artificial Intelligence

Declassified documents, intelligence experiments, and government disclosures reveal that U.S. spy agencies were developing and deploying powerful AI capabilities long before the public experienced the technology. The most surprising discovery involves a 2019 operation that used generative AI to uncover criminal networks.

When OpenAI released ChatGPT on November 30, 2022, millions of people encountered a technology that seemed almost unimaginable just a few years earlier.

A computer could write essays, summarize research, answer complex questions, and hold surprisingly sophisticated conversations.

For much of the public, it felt like artificial intelligence had suddenly arrived.

But inside America’s intelligence agencies, the AI revolution had been unfolding for decades.

A review of declassified CIA documents, historical government records, intelligence research, and investigative reporting reveals something extraordinary: American intelligence officials were exploring machine reasoning in the early 1980s, analyzing millions of social media posts by 2011, and investigating predictive AI years before ChatGPT became a household name.

Most strikingly, a U.S. intelligence operation was already using generative AI to prepare investigative evidence summaries in 2019.

The evidence doesn’t establish that the CIA secretly possessed artificial general intelligence. But it reveals a much longer—and more sophisticated—history of intelligence AI than most people realize.

The 2019 Operation That Changed Intelligence Analysis

In December 2019, a team at the Defense Intelligence Agency received an award for an experimental program called SABLE SPEAR.

The initiative was designed to answer an extraordinarily difficult question: Could artificial intelligence uncover global fentanyl trafficking networks more effectively than experienced human intelligence analysts?

The government partnered with Silicon Valley AI company Rhombus Power, supplying investigative guidance and allowing the contractor to analyze publicly available information, including material from the dark web.

The results were remarkable.

According to a 2021 account by intelligence officer Craig A. Dudley, published in the CIA’s Studies in Intelligence, the system identified:

100% more companies associated with illicit activity.
400% more people potentially involved.
900% more illicit activities than the human analysts had identified.

The software discovered relationships across datasets, recognized behavioral patterns, and highlighted connections that investigators might never have thought to examine.

Some findings were false positives, requiring human validation. The dramatic increases represented additional discoveries, not proof that every flagged entity had committed a crime.

Nevertheless, the experiment demonstrated the potential for machines to examine complex networks at speeds and scales that were impractical for human investigators.

Then came another revelation.

In May 2024, the Associated Press reported that Rhombus Power had also used generative AI to produce evidence summaries for potential criminal cases during the 2019 operation.

That was three years before ChatGPT’s public launch.

The important distinction is that SABLE SPEAR was led by the Defense Intelligence Agency, not the CIA. Its results were subsequently published through the CIA’s intelligence research journal.

The system wasn’t necessarily capable of holding sophisticated conversations like modern chatbots. But it was already applying machine-generated analysis and written output to consequential investigative work.

This was not a theoretical demonstration in a university laboratory. It was an operational intelligence experiment addressing a real international criminal network.

A Declassified Document Reveals the CIA’s AI Ambitions in 1983

To understand how far back these ambitions extend, consider a document preserved in the CIA’s own public archive.

Dated December 1, 1983, the memorandum contains the schedule for a three-day artificial-intelligence symposium.

Its subject: Intelligence Applications of AI.

The event brought together government officials, defense contractors, and researchers from leading academic institutions.

Participants discussed computer vision, speech understanding, robotics, expert systems, and artificial intelligence for military decision-making.

Several sessions were explicitly designated SECRET.

Among the topics were automated scientific intelligence analysis, image interpretation, radar imagery, signal processing, and computer-assisted weapons allocation.

These subjects sound remarkably familiar today.

Modern AI systems are increasingly being designed to interpret satellite imagery, identify unusual activity, understand spoken language, and assist military or intelligence decision-making.

Yet the intelligence community was exploring these applications more than four decades ago.

An academic examination of declassified records published in the Journal of Global Security Studies adds another revealing detail.

Deputy CIA Director John McMahon helped organize an AI symposium in 1982, reportedly attended by roughly 500 representatives of government, industry, and academia.

He subsequently encouraged relationships between the intelligence community and technology companies capable of advancing the field.

The lesson is not that ChatGPT-level technology existed in 1983. It did not.

Early expert systems were far more limited than today’s neural networks.

The remarkable finding is that intelligence leaders had already recognized AI’s strategic potential decades before the modern consumer market existed.

By 2011, the CIA Was Analyzing Millions of Social Media Posts

Long before most people imagined artificial intelligence as a conversational assistant, American intelligence agencies were confronting another problem.

They were drowning in information.

In November 2011, Associated Press reporter Kimberly Dozier described a CIA operation housed inside an unassuming office building in Virginia.

Its analysts were reportedly examining as many as five million tweets per day.

The CIA’s Open Source Center monitored information from social networks, newspapers, radio broadcasts, television, and internet discussions around the world.

The objective was not merely to collect information.

It was to recognize patterns in public sentiment, identify emerging instability, and understand how geopolitical events were unfolding.

This work helped policymakers assess developments across regions where political movements were emerging and governments faced unrest.

The Associated Press reported that analysts could combine these public signals with other intelligence to provide a more comprehensive picture of events.

Not all this activity was necessarily performed by modern machine-learning systems. Traditional databases, search technology, linguistic tools, and human expertise played important roles.

But the scale of the operation reveals how early the intelligence community had begun approaching the world as an enormous information-processing problem.

The challenge was no longer simply obtaining secrets.

It was finding meaningful signals inside an ocean of data.

In 2017, the CIA Disclosed 137 AI Projects

By September 2017, artificial intelligence was no longer a distant research ambition for the CIA.

It had become a sizable portfolio of active experiments.

At an intelligence conference in Washington, Dawn Meyerriecks, then the CIA’s deputy director for science and technology, disclosed that the agency had 137 AI-related pilot projects.

The experiments included automated video analysis, object recognition, data correlation, and attempts to improve predictions about significant geopolitical developments.

One potential application involved recognizing signals that could indicate preparations for a North Korean missile launch.

The most revealing part of the discussion wasn’t merely the number of projects.

It was the problem intelligence officials were trying to solve.

Could computers discover connections across vast quantities of information that human analysts might otherwise overlook?

And if a machine reached a consequential conclusion, how could intelligence officials verify its reasoning before advising national leaders?

These were already central questions in 2017—five years before ChatGPT reached consumers.

The disclosure does not prove that all 137 projects were successful or operational. But it demonstrates the scale of the CIA’s investment in exploring artificial intelligence.

The Real Secret May Not Be a More Intelligent Machine

These findings raise an intriguing possibility.

What if the greatest advantage held by intelligence agencies isn’t necessarily access to a vastly superior AI model?

What if it is access to information that no publicly available AI can see?

Imagine two equally capable AI systems.

The first can analyze publicly available research, books, websites, financial reports, and news coverage.

The second can also access authorized intelligence collections, satellite imagery, intercepted communications, government databases, and years of classified reporting.

Even if their underlying algorithms were identical, their potential analytical capabilities would differ dramatically.

The second system might recognize relationships, identify emerging threats, or detect patterns invisible to the first.

This distinction is essential.

The intelligence advantage may arise from the combination of algorithms, exclusive datasets, computing infrastructure, and specialized human expertise—not from a secret breakthrough in artificial general intelligence.

It also explains why companies capable of integrating vast quantities of information, including Palantir Technologies, became important partners to the national security establishment.

The value of AI extends beyond generating answers.

It lies in connecting information that was previously fragmented.

The Dangerous Side of Machine Intelligence

There is another dimension to these capabilities that deserves serious attention.

AI systems can identify correlations that appear significant but turn out to be meaningless.

They can amplify biases, incorrectly associate innocent people with suspicious networks, or generate convincing but unsupported conclusions.

In intelligence operations, those errors can have enormous consequences.

The SABLE SPEAR experiment itself encountered false positives that required human correction.

Modern generative AI adds further risks, including fabricated information, manipulated data, and uncertainty about how particular conclusions were reached.

In its 2024 investigation, the Associated Press reported that CIA officials remained concerned about these limitations even as agencies expanded their use of AI.

The question is not simply how powerful intelligence systems can become.

It is also how societies ensure their conclusions are accurate, their use is lawful, and their capabilities are subject to meaningful oversight.

A machine that can detect complex relationships across billions of records is useful.

A machine that incorrectly labels innocent people as threats can be dangerous.

So How Far Ahead Is Classified AI?

The historical evidence supports several conclusions.

American intelligence agencies pursued artificial intelligence decades before the public became familiar with it.

They investigated machine recognition, automated analysis, and predictive capabilities long before the modern generative AI boom.

And by 2019, the wider U.S. intelligence community was already experimenting with generative AI in consequential investigative applications.

What the evidence does not establish is that the CIA possessed human-level artificial general intelligence, secretly developed ChatGPT years before OpenAI, or maintains an undocumented technological lead of a particular number of years.

Classified systems may possess capabilities unavailable to the public. By their nature, however, those capabilities cannot be reliably assessed without evidence.

The historical record is fascinating enough without inventing what remains unknown.

The Bigger Revelation

Perhaps the most consequential finding is that the AI revolution did not begin when millions of people started interacting with chatbots.

It emerged from decades of research across universities, government laboratories, defense agencies, and private technology companies.

For intelligence organizations, AI was never primarily about writing clever emails or answering homework questions.

It was about extending the limits of human analysis.

Finding previously invisible connections.

Understanding an increasingly complicated world.

And making decisions before events fully unfolded.

The public’s introduction to conversational AI in November 2022 was a historic technological milestone.

But it was not the beginning of artificial intelligence’s role in national security.

The deeper story is that powerful institutions had been exploring the strategic value of machine intelligence for generations—and the public is only gradually learning how extensive that work has been.