A quarter-century after 9/11, the question is difficult to avoid: What if the United States had today's artificial intelligence before September 11, 2001? What if analysts at the FBI, CIA and other agencies could have searched across enormous collections of intelligence reports, watchlists, travel records and other data, looking for connections no individual analyst could reasonably be expected to find?
Could AI have connected the dots? Perhaps. But the deeper question is what technology can actually solve – and what it cannot.
Any attempt to revisit September 11 through the lens of technology carries a risk of turning a human catastrophe into a case study. Nearly 3,000 people were killed that day, thousands more were injured, and families and communities have lived with the losses ever since. The point of asking what today's AI might have changed is not to reduce 9/11 to an analytical puzzle, but to ask whether better tools can help us recognize danger sooner – and where technology still falls short.
For readers who would like to know more about that day – which remains one of the most devastating and historically significant days in American history – the National September 11 Memorial & Museum recounts the events and their human toll.
After a year and a half of hearings and investigation, the bipartisan 9/11 Commission concluded in July 2004 that the failures stretched across "imagination, policy, capabilities, and management." Specifically, agencies struggled to share information, coordinate their work, and recognize the significance of the intelligence they already possessed. The Commission's final report also warned against the easy certainty of hindsight: What looks obvious now might not have looked significant amid the mass of fragmentary information available at the time.
AI might have helped with some of those failures, but it would not necessarily have solved them.
So many dots
Viewed from 2026, the many warning signs before 9/11 remain difficult to stomach. We want to scream, "How were they possibly ignored?" U.S. intelligence knew that Osama bin Laden and al Qaeda wanted to strike the United States. In fact, the CIA had identified two future hijackers, Khalid al-Mihdhar and Nawaf al-Hazmi, as al Qaeda operatives in January 2000, when they entered the U.S. under their real names. Yet, the CIA did not watchlist them or tell the FBI for 19 months, as the DOJ Office of the Inspector General's report makes clear.
By the time a CIA officer flagged the omission and a search for the two began in the final weeks of August 2001, the men were living openly in California. The FBI treated the search as routine, and the request to search for them in Los Angeles, where they lived, did not arrive until September 11 itself.
The warning signs kept multiplying: That summer, an FBI agent in Phoenix warned that bin Laden might be sending students to U.S. civil aviation schools and urged the bureau to look for a broader pattern. Nothing came of it.
In August, the Minneapolis FBI opened a case on Zacarias Moussaoui after a flight school reported he had paid cash for training on a Boeing 747 simulator and showed little interest in takeoffs or landings. FBI agents detained him on an immigration violation and asked headquarters for a warrant to search his laptop. The request was turned down.
None of those facts, standing alone, revealed what was about to happen. Put together, however, they become much more ominous, which is why it is reasonable to ask what today's AI might have seen in them.
Modern AI systems can process quantities of information beyond any individual's capacity, extracting names and relationships, comparing records, analyzing unstructured documents, and finding patterns across massive datasets. Just two months ago, in a July 2026 analysis titled Artificial Intelligence and the Future of Terrorism, Daniel Byman of the Center for Strategic and International Studies (CSIS) and computer scientist V.S. Subrahmanian of Northwestern University wrote that today's AI can help intelligence agencies detect "anomalous travel patterns," analyze financial transactions, and "connect fragmented datasets." It can also process communications, images, video and metadata far more efficiently than human analysts working alone.
It isn't difficult to imagine a modern system surfacing some combination of known al Qaeda associates, flight training, visa and travel records, intelligence reports, and suspicious activity. Perhaps an analyst would have seen that cluster, asked a different question, and prompted action. But perhaps not. We don't get to know, and claiming otherwise flatters both AI and hindsight more than either deserves.
What hindsight cannot tell us
We know what happened on September 11, and that knowledge inevitably changes the way we see every warning that came before it. The 9/11 Commission was careful about this, asking whether insights that seem apparent in retrospect really would have been meaningful given what officials reasonably could have known at the time.
Artificial intelligence does not eliminate that problem, and in some respects, it could magnify it. A system capable of finding one ominous pattern can find thousands. Some will matter, but most will not. If an AI system flags thousands of people because some combination of their travel, education, financial activity, or communications resembles behavior associated with terrorism, someone still has to decide which warnings deserve attention.
Byman and Subrahmanian warn specifically about false positives that could lead innocent people to be suspected of terrorist sympathies. They also raise a larger concern: AI can dramatically expand governments' ability to monitor people, track movements and analyze communications, creating risks to privacy and civil liberties.
So the question is not only whether AI could make us safer. It is who gets flagged when the system is wrong, and how easily a human being with a badge learns to defer to a machine that says someone looks suspicious.
The harder question: imagination
The most striking part of the 9/11 Commission's assessment, however, was not about databases or watchlists. "The most important failure," the Commission wrote, "was one of imagination."
Government officials understood hijacking. They understood terrorism. They understood that al Qaeda wanted to attack the United States. What they largely failed to contemplate was combining those things in a new way: hijackers taking control of commercial aircraft not to negotiate or escape, but to turn the airplanes themselves into weapons.
Former Deputy CIA Director Richard Kerr, testifying before the Commission in 2003, argued that intelligence failures often stemmed less from a lack of information than from analysts asking the wrong questions of what they had. That may be the more difficult challenge for AI.
Today's AI systems excel at recognizing patterns, including ones humans overlook, and they can increasingly generate scenarios, challenge assumptions and identify ways a plan could fail. Those capabilities could give intelligence analysts a powerful new tool for considering threats that might otherwise escape attention.
Yet AI also learns largely from what humans have already recorded – the accumulated evidence, ideas, assumptions and experiences of the past. We increasingly ask it to help anticipate what happens next because it has become extraordinarily good at digesting what has already happened.
September 11 was devastating in part because the attackers combined familiar things in a way the country's security apparatus had not seriously contemplated. That distinction – between recognizing patterns from the past and imagining a threat that departs from them – remains important.
The limit cuts both ways
It cuts the other way, too. Byman and Subrahmanian argue that most terrorist organizations are cautious, resource-strapped, and averse to risk, and that AI is far more likely to produce "cumulative" change in how such organizations operate than a "revolutionary" one.
The researchers' own example of the alternative – a genuine departure from the pattern – is September 11 itself, which combined hijacking and suicide attack in a way no one had done before. There's a kind of reassurance in that: The same imaginative leap that can escape an analyst's notice is also the leap AI is least equipped to hand to an attacker.
AI may also make terrorism more dangerous in ways that have nothing to do with connecting dots – producing sharper propaganda, making fraud easier, or having a lonely person's only confidant turning out to be an algorithm patiently working to radicalize them. Those risks deserve a fuller look than this piece can give them today.
Twenty-five years later
This is not merely a historical thought experiment. Less than two weeks ago, on August 31, the House Permanent Select Committee on Intelligence released a new bipartisan review of the 9/11 Commission's recommendations.
The committee credited the post-9/11 reforms with breaking down many of the institutional stovepipes that hampered information-sharing before the attacks, but its recommendations show that the work is unfinished. This work includes building interoperable information-sharing systems across the intelligence community, professionalizing open-source intelligence, improving intelligence integration, and expanding the use of emerging technology. The report also calls on the government to address potential "black swan" risks posed by advances in AI.
So, 25 years after one of the most consequential intelligence failures in American history, the Committee's own recommendations point to a version of the same unfinished business: Agencies still need better ways to share what they know, and someone still has to make sense of what it adds up to.
Knowing what we know
It would be reassuring to believe that today's artificial intelligence, given the information available before September 11, would have discovered the hidden connections and produced a warning clear enough to prevent what followed. Perhaps it could have made such a warning more likely. We cannot responsibly say more than that.
But 9/11 also warns against technological hindsight. Information, patterns and intelligence reports ultimately do not make decisions; people and institutions do. These people and institutions determine what information gets shared, which warnings get elevated, what risks deserve attention, and whether an unlikely possibility should be taken seriously.
Artificial intelligence is becoming extraordinarily good at finding the dots, but it is still a system built on what has already happened. And yet, we ask it to warn us about what hasn't yet happened. Twenty-five years after September 11, that is still a job for people – the ones willing and able to look at a pattern nobody has seen before and imagine what it might be, before it's too late to make a difference.
