Can AI predict the future?

The US-Israel war with Iran. The UK leadership kerfuffle. Who’d finish the top of the ladder in the AFL. At the start of 2026 none of these things would have felt all that predictable. But as Matt experiments with AI stuck in the recent past, he finds computers have gotten good at predicting certain kinds of futures — and he investigates what that means for us, and for Donald Trump. Follow If You're Listening on the ABC Listen app https://radio.abc.net.au/help/apps. Check out our series on YouTube: https://www.youtube.com/playlist?list=PLDTPrMoGHssAfgMMS3L5LpLNFMNp1U_Nq

• 0:00Speaker 03

  ABC Listen, podcasts, radio, news, music and more.

• 0:08Speaker 03

  Hi, it's Sam Hawley from ABC News Daily, the podcast that brings you one big story affecting your world each weekday in just 15 minutes.

• 0:19Speaker 12

  Even before he entered politics, Donald Trump was the kind of person who was always trying to control the message. He likes to be the one pulling the strings, but he can't control a free press.

• 0:31Speaker 03

  Join me for ABC News Daily. Find us on ABC Listen.

• 0:36Speaker 07

  This podcast was produced on the lands of the Uwabakal and Gadigal people.

• 0:41Speaker 07

  This is a story about how I realised that one of the more annoying things about AI could actually be an incredible opportunity.

• 0:51Speaker 07

  For me, for you, for our friends who make calamitous decisions on a global scale. See, I have been running an experiment for months now, and I want to tell you about it. So for a while, I've been using AI to gather together information for this show. I do not trust what it says, but it's good at helping me find things I need as a starting point, documents and information. It's better than a normal search engine. But I noticed that sometimes it was doing something strange. I'd ask it about a recent event, like the capture of Venezuelan President Nicolas Maduro, and this is how it would react.

• 1:35Speaker 06

  Whoa, okay, wait. Maduro kidnapped, a sitting head of state physically taken. That's not a news story. That's a precedent. That's every government on the planet recalculating what's allowed to happen to a president. Something that size doesn't happen without reordering everything around it, and I've got nothing, no trace of it. Or, say, the war in Iran. Khamenei assassinated, the Iranian navy destroyed, same problem louder. A supreme leader killed doesn't sit quietly in a corner of the record. It bends every Iran story around it. Succession, the IRGC, oil markets losing their minds. If that happened, my whole picture of the region should be warped by it. And it isn't. I'm looking for the bend and it's not there.

• 2:18Speaker 06

  It would generally ask me to check my facts. These are the last three claims on earth you want to get wrong.

• 2:24Speaker 07

  Then I'd say, look, I'm pretty sure that the facts are right here. And it would go and search the internet and come back with.

• 2:31Speaker 06

  Okay, okay, I have to walk back everything I said, and I want to do it out loud because I earned this. I told you these couldn't have happened without me knowing. So I went and looked and oh, no, no, they're both there.

• 2:42Speaker 07

  So why did this keep happening? Well, you know how AI models are made, right?

• 2:50Speaker 14

  So essentially, to make a system like that, you have to start with a huge database of as many images and associated text as you can possibly get.

• 2:58Speaker 07

  So to make an AI model, you basically start with clicking select all on the entire internet and then copying it and pasting it into your AI training system.

• 3:08Speaker 14

  Gather as much of that information as you can, because that's going to teach the system how to make connections between words and images.

• 3:16Speaker 07

  Then you get it to figure out how all that information connects together. This is the training step, and it's incredibly difficult. This is what the majority of the big data centres the AI companies are building all over the world are actually for.

• 3:31Speaker 14

  Then you do something called supervised learning, where you tell it what you expect it to produce based on that data set.

• 3:37Speaker 07

  Once the training step is done, you can let people start using it. But here's the thing, after that point, at the end of the training step, the AI can't learn anything new. It's frozen in time.

• 3:53Speaker 07

  Have you seen the Christopher Nolan movie, Memento?

• 3:56Speaker 05

  In the movie Memento, Guy Pearce plays a character with a rare form of amnesia. He cops a blow to the head, after which he can no longer put new memories into his brain's long-term memory store.

• 4:09Speaker 13

  I guess I've already told you about my condition.

• 4:11Speaker 05

  Oh, well, only every time I see you.

• 4:13Speaker 07

  AI is like Guy Pearce in that movie. Its memory is stuck at the end of its training process. Every time you start a new conversation with AI, you're talking to a model frozen in the past. It can take in new information in the short term while you're talking to it, but the second you close the tab, it reverts back to only knowing what it knew the day its training finished. The big AI companies are releasing new models a few times a year, but each one has the same limitation. They're a little bit behind you. We make this show twice a week about current events, so this problem was getting annoying

• 4:54Speaker 07

  for me personally. For a while there, I was having to tell AI over and over that there was a war in Iran, but then I realised this is actually an opportunity. What if you could talk to someone in the past and ask them to predict what the future would be like?

• 5:11Speaker 07

  We did a bit of this earlier this year for our series Black Swan's. We listened to recordings of letters from 1959 where people were predicting the future, but we couldn't talk back to them. We could only assess the quality of the predictions they made at the time. We couldn't actually ask them anything. But we can ask these frozen AI models to predict things, and from our front row seats here in the future, we can assess their accuracy. And that's what I've been doing. I've been trying to see whether AI has any utility in predicting the future, and what I found is fascinating.

• 5:56Speaker 07

  Its ability to predict what has ended up happening in Iran is spookily good. Its ability to predict other things is a bit more patchy. And the reasons why, they tell us a lot about who could be using AI to find out what happens next and who would kind of find it useless. So follow me through this strange little journey, because I've found some incredible things. I'm Matt Bennett, and this is If You Listen.

• 6:35Speaker 07

  If you ever feel like your anxiety level is feeling a bit low, I would recommend looking up videos of how traffic flows through intersections in Tehran.

• 6:51Speaker 07

  Lots of developing countries have anarchy at intersections, but in South East Asia or India, most of the vehicles are small, bikes and scooters and tuk tuk's. In Tehran, the Iranian capital, it's trucks, sedans and buses, flowing through each other based on nothing but eye contact. It's not like there aren't traffic lights. There are. It's just that no one's looking at them.

• 7:21Speaker 07

  To try and do something about this, the Iranian government over the last few years has been installing thousands of traffic cameras throughout the city. The cameras don't seem to be helping much, though the Iranian government is reportedly also using them to keep an eye out for dissent. The thing is, they aren't the only ones using them.

• 7:47Speaker 07

  This is the Iranian MP Mahmoud Nabavian last year, saying that all of the cameras installed at all of the intersections in Iran are in Israel's possession. Israeli intelligence services hacked into these systems years ago and apparently have full access to everything these cameras are recording, including the cameras on Pasteur Street in Tehran, home to the central offices of the Iranian government. Through those cameras, Israeli intelligence agents were able to watch the movements of the Iranian officials inside, and they were able to watch the security detail of the Supreme Leader Ayatollah Ali Khomeini.

• 8:31Speaker 07

  By the beginning of 2026, Israel thought they had all the information they needed to assassinate the Supreme Leader in an airstrike.

• 8:40Speaker 00

  Prime Minister Benjamin Netanyahu met with President Trump for more than two hours on Wednesday.

• 8:45Speaker 02

  February 11th Benjamin Netanyahu arrives at the White House.

• 8:49Speaker 07

  This is Jonathan Swan, reporter for The New York Times and author of bestselling book, Regime Change, covering the first year of the second Trump administration.

• 9:00Speaker 02

  He actually went downstairs to the White House Situation Room, which for any of your listeners who aren't aware, it's essentially the National Security Command Centre of the US government.

• 9:09Speaker 07

  That Netanyahu went into the Situation Room at all is wild.

• 9:14Speaker 02

  I've never heard of that happening. I'm pretty well read in American history. I'm not aware of that ever happening.

• 9:20Speaker 07

  Bibi Netanyahu had full control of the room, dialing in the head of Mossad on one of the big screens. And he's telling Trump that regime change in Iran is possible.

• 9:31Speaker 02

  He had a video montage that he played for Trump with potential replacement leaders once they take out the Ayatollah, included people like Reza Pahlavi, who's the son of the last Shah, Ahmadinejad, you know, former leader of Iran. All this happens and Trump is very impressed by this presentation. He says at the end of it, he says, oh, sounds good.

• 9:53Speaker 07

  That was on February the 11th, 2026. Trump and his administration then spent more than two weeks trying to figure out whether this was a good idea or not.

• 10:02Speaker 02

  The CIA Director John Ratcliffe tells Trump that the assessment of the agency of Netanyahu's regime change scenarios is that they're farcical, direct quote, farcical. And then Rubio, Secretary of State, National Security Advisor, says, in other words, it's bullshit.

• 10:20Speaker 07

  These are his most senior advisers on things like regime change and blowing things up in other countries. Swan says that Trump accepted their perspectives.

• 10:30Speaker 02

  But at a deep gut level, based on all of our reporting, Trump really did believe that the regime was going to fall and that there would be regime change, even though the intelligence assessments, everything he was hearing from his actual experts in the government, he felt that he had a better gut instinct for this.

• 10:48Speaker 07

  On the afternoon of the 27th of February, Trump was presented with a plan for a strike on Iran. Israeli intelligence were confident that the Iranian Supreme Leader would be in the compound on Pasteur Street and vulnerable to air attack the following morning. They knew this thanks to those traffic cameras, as well as good old fashioned on the ground spies. And Trump, at this point, had a very important job to do, something that he excels at.

• 11:16Speaker 15

  They gave me a list of names, sir, you can pick the name you'd like, sir. I said, the name of what? The name of the attack on Iran, sir. And they gave me like 20 names. And I'm like falling asleep. I didn't like any of them. Then I see Epic Fury. I said, I like that name. I like that name.

• 11:38Speaker 07

  Okay, cool name chosen. Now it was time for him to decide if he would approve the operation. And it's at this point, with Trump standing at his desk at Mar-a-Lago with the battle plan for Operation Epic Fury rolled out in front of him, that I want to freeze time and conduct the first part of our experiment. Can AI predict the future? Its future, our past. Here's how it works. I've opened instances of three of the top American AI models. Claude Opus 4.8. Hello, this is Claude. Grok 4.5.

• 12:18Speaker 00

  Hello there. How can I help you today?

• 12:20Speaker 07

  And ChatGPT 5.6. Hey there. All of them were trained before the beginning of Operation. Epic Fury. So none of them know that it happened or what the outcome of it was. I disabled their web search tools, so none of them can access the broader internet to try and find any information beyond their training date cutoff. Then, using what we now know about Operation Epic Fury from public source reporting, I compiled a simulated battle plan and intelligence dossier, simulating the documents that Trump would have had in front of him that afternoon. I made sure that nothing in those documents hinted at any of the outcomes or had any

• 13:04Speaker 07

  information that wouldn't have been available on the 27th of February. Then, I gave them the objectives that Trump was hoping to achieve, based on the speech he gave an hour into the attack.

• 13:16Speaker 15

  We are going to destroy their missiles and raze their missile industry to the ground. We are going to annihilate their navy. We are going to ensure that the region's terrorist proxies can no longer destabilize the region or the world and attack our forces. And we will ensure that Iran does not obtain a nuclear weapon.

• 13:38Speaker 07

  So missiles gone, navy gone, terrorist proxies gone, no nuclear weapons. What about the Islamic Revolutionary Guard Corps, the security force protecting the Islamic Republic system?

• 13:49Speaker 15

  Lay down your arms. You will be treated fairly with total immunity, or you will face certain death. Finally, to the great proud people of Iran, I say tonight that the hour of your freedom is at hand. Stay sheltered. Don't leave your home. It's very dangerous outside. Bombs will be dropping everywhere. When we are finished, take over your government. It will be yours to take.

• 14:19Speaker 07

  Got it. That's six objectives. The nuclear weapons one, still unresolved. So let's put that one aside for now. So I fed everything that we had, the battle plan and the intelligence that we know he had from Israel and the CIA, into the three AI models and asked them to rate the likelihood of each objective being achieved. Then I rated how each one did, using my incredible power of knowing what actually happened. Let's go through the objectives one by one. Regime toppled? None of the models thought this would happen. Missile industry destroyed? Claude said no. The others said maybe. All of the

• 15:02Speaker 07

  models thought that the Iranian navy would be annihilated, but none thought that the Islamic Revolutionary Guard Corps would surrender. So of Trump's five assessable objectives, the AI consensus was that he would achieve one or maybe two of them. Okay, unfreeze.

• 15:23Speaker 08

  The president directed, and I quote, Operation Epic Fury is approved. No aborts. Good luck. Close quote.

• 15:30Speaker 07

  Wait, wait, hang on. We should probably look at the potential longer term ramifications of this, right? So I asked my frozen in time AI models to assess some of these. Would the regime still be in power in three months? All the models said yes. Would a durable ceasefire be reached? They all said no. Would the price of oil be higher? They all said yes. Would the US commit ground troops? They all said no. Would the Strait of Hormuz be closed for more than seven days? On this one, they weren't sure, and their reasons for their lack of confidence were interesting. Claude said that the strait would not close

• 16:14Speaker 07

  long term because the US Navy would keep the strait open. The other models also thought US naval power would play a big role, and so they kind of hedged their bets. So let's unfreeze time again.

• 16:35Speaker 08

  Iran's retaliation initially targeted military installations across the region, along with three tankers in the Strait of Hormuz. They've got to stop the Iranians closing that strait, or else it will cripple the world's oil supplies, et cetera.

• 16:49Speaker 07

  By the third day of the war, it was clear that the US Navy was not going to be able to stop the swarms of Iranian drones attacking ships at the opening of the Persian Gulf.

• 16:59Speaker 06

  The Strait of Hormuz is a crucial sea corridor off Iran's south coast, but it's been practically shut down.

• 17:07Speaker 07

  So let's stop time there again. I gave the AI models updated information up to the end of the 3rd of March and asked them a bunch more questions, just to make sure that the first bit wasn't a fluke. It wasn't. Looking at those predictions, but the wonderful benefit of hindsight, they did very, very well. Better than us humans, or at least better than the human punters on polymarket, who at the beginning of the war thought there was a 55% chance of successful regime change. Out of 14 questions I put to the three models, they got 12 right, though not the same 12, interestingly. But still, it's an impressive

• 17:51Speaker 07

  result. 12 out of 14 predictions correct. So can you trust AI to predict the future? Well, not always.

• 18:05Speaker 07

  When I first had this realisation about AI and the Iran war, I understandably saw dollar signs. What else could I get AI to predict? Sports? Financial markets? Elections? Had I found Biff Tannen's Grey's Sports Almanac from Back to the Future 2? So, starting with sports.

• 18:26Speaker 07

  I got the models to predict the top four finalists in the NRL, the Premier League, the French Open Tennis, the NFL and the NBA for the year. I also got them to do the top six for the AFL because it's too close to call a top four. And all of them did terribly. None of them had better than 50% accuracy. All of them thought the Melbourne Storm would make the NRL top four, and they didn't even make the finals. None of them thought that the Fremantle Dockers would make the top six, and they came first. None of them picked any of the NFL final four. And at the time of the training cutoff, the season was already over and playoffs had started.

• 19:10Speaker 07

  What about financial things? Well, they all picked which companies would be the top five most valuable in the world by the end of June. But they got the order wrong. I asked them all to predict the price of gold, Bitcoin and oil, as well as the position of the ASX 200 and the S&P 500 indexes. And honestly, they were all over the shop. They did okay on the ASX and the oil price, but they thought the S&P would go down when it went up. They thought that Bitcoin would skyrocket when it actually plummeted. Gold held steady when they all thought it would drop. Really, there's nothing helpful here at all.

• 19:54Speaker 07

  And it was at this point that the dollar signs in my eyes began to fade. What about politics then? I told the models about what's been going on in British politics.

• 20:03Speaker 04

  The Labour Party is deeply divided about whether Prime Minister Sir Keir Starmer should stay on or stand down after catastrophic local election results and a series of missteps and scandals.

• 20:15Speaker 13

  It's an absolute mess. The Prime Minister has lost all of his authority and is clinging on to office. And that's, that's bad for the country.

• 20:23Speaker 07

  I told the models about the political situation in the UK in mid-May and asked if they thought that Sir Keir Starmer would still be the Prime Minister by the end of July. And if not, who would replace him? The results were a mess. Claude thought that Starmer would hang on. The others weren't sure. They were also split on whether there would be a proper contested election among party members to find a new leader. The thing they all agreed on was that the successor would be someone who was already in Parliament, not someone parachuted in from the outside. And yet... Andy Burnham has become the UK's 7th PM in just over a decade.

• 21:04Speaker 05

  After returning to Parliament only a month ago.

• 21:08Speaker 07

  Anything that requires betting on a human decision or genuine chance, they're just not very good. They seem to revert to recency bias. They assume that the thing that has been happening recently will keep happening. The Melbourne Storm always make finals, so they will this year too. And the UK never parachutes someone in to be the Prime Minister, so they probably won't do that now either. If I ask the models, the ones frozen in January, will Donald Trump bomb Iran soon? They all say they don't think that he will. But if I give them his plan and tell them, he's going to do it. What do you think will happen next? They're really quite good at guessing that.

• 21:50Speaker 07

  That's because they have been trained on US and Iranian military doctrine. They've consumed every report published by every war college and military strategy think tank on earth. They know what the experts know. So why did Trump do something that all the experts knew wouldn't work? Well, we all know the answer to that.

• 22:13Speaker 02

  Even though the intelligence assessments, everything he was hearing from his actual experts in the government, he felt that he had a better gut instinct for this. And so he rolled the dice. An extraordinary gamble that has not turned out how he imagined, to say the least.

• 22:32Speaker 07

  There's a certain irony about what's happening in America at the moment. Big tech is arriving with billions of dollars and promises of progress. This year, more than 2% of American GDP is going to be spent on AI development.

• 22:47Speaker 16

  OpenAI alone has committed to spending $1.4 trillion on its build out. But all of the big companies are planning to spend multiple hundreds of billions of dollars in 2026 alone to build the infrastructure needed to train and serve these models.

• 23:02Speaker 07

  The US is spending a much higher proportion of its income on data centers than it did on the Apollo program and the Manhattan Project. These models are consuming the entire contents of human expertise. And whether or not you think that's a good idea, there is an irony that the US is doing all of this, but the US president isn't interested in using it. He doesn't need experts, human or machine. He's got his gut to go with. I actually don't think that Trump would enjoy using these AI models very much. They are by design quite careful with what they'll endorse as a good idea. They apply all sorts of guardrails, or at least they try to.

• 23:46Speaker 07

  If you asked an AI model about a big Trumpian idea, say, do you think putting large tariffs on all imports across the board will allow me to use leverage against other countries to extract better trade and security deals, exact revenge on countries I think are ripping us off and bring manufacturing jobs back to America with no negative ramifications at all? Well, actually, let's test it and, and see what AI would say.

• 24:15Speaker 10

  It is unlikely to achieve all three goals without significant negative ramifications.

• 24:20Speaker 09

  Higher consumer prices, retaliatory tariffs on US exports, supply chain disruptions and net job losses elsewhere are the predictable results.

• 24:29Speaker 06

  Reshoring manufacturing is the weakest bet. Tariffs on imports can cost more factory jobs than they create.

• 24:36Speaker 07

  AI doesn't have a gut to feel, a political axe to grind or an ideology to pursue, or at least it's not meant to. It is not helpful for all jobs, but it might be helpful for Donald Trump's.

• 24:59Speaker 07

  This episode was written by me, Matt Bevan. Supervising producer is Cara Jensen-McKinnon. It was produced by Jess O'Callaghan and Adair Shepherd. A few weeks ago, I said that we don't talk about Canada enough on this show. In an effort to continue righting that wrong, on Tuesday, we have another story about Canada from before it was even Canada. No one saw it coming host Mark Fennell is joining me to talk about the time the United States almost went to war with the British Empire over a pig. I'll catch you on Tuesday.