#16 95% of Company AI Projects Failed. Here's What That Actually Means.
There is a deal on the table between the United States and Iran. A memorandum, a ceasefire renewed every few weeks, and missiles still flying in the gaps. Nobody has signed anything. Meanwhile the lights just came back on inside Iran after the longest blackout in modern history, and what they show is a country that spent those months running out of food.
Then there is the machine everyone says is about to think for us. Ninety-five percent of the company projects built on it returned nothing. It is inventing court cases and airline policies and saying them with a straight face. And we are told it is about to become self-aware.
And out in a Texas field, the biggest rocket ever built just flew again. The same rocket that has never once reached orbit, and the one America is counting on to carry its astronauts back to the moon.
Start with how we got here. The war began on February 28, when the United States and Israel struck Iran. A ceasefire took hold on April 8. It was shaky from the first day, formalized in the middle of April, and it has had to be renewed over and over since. The last extension, on May 17, bought another forty-five days, and late in the month negotiators reached a framework for a sixty-day stretch on top of that. That is the shape of this thing. Not peace. A clock that keeps getting wound back up right before it runs out.
And make no mistake about the size of it. When the two countries hit Iran on February 28, it set off the heaviest fighting the region has seen in years. The Pentagon told Congress that nine weeks of war with Iran had already run twenty-five billion dollars. The strikes have gone back and forth ever since, the oil markets have been in turmoil the whole time, and the ceasefire that was supposed to settle it has been broken and patched more than once. That is the backdrop behind every word said at the negotiating table.
And none of it is new ground. There was a nuclear deal back in 2015, signed by Iran and six world powers. Trump pulled the United States out of it in 2018, and that agreement was formally killed off in October of last year. So what is on the table right now is, in large part, a renegotiation of ground that deal already covered. The difference this time is that there is a shooting war wrapped around the table.
The talking has gone in rounds, all of it indirect, with go-betweens carrying the messages. The first round ran last spring, in Oman and in Rome. A second round opened in Geneva this past February and fell apart the moment the fighting started. They picked it back up through the spring, including a round in Islamabad in April, and the messages move now through Pakistan. On the American side the lead is Special Envoy Steve Witkoff, with others in the room. On the Iranian side it is Foreign Minister Abbas Araghchi and the country's national security secretary, Ali Larijani.
What they have produced so far is a memorandum. Under it, Iran cannot charge tolls to ships moving through the Strait of Hormuz, and Iran has to clear all of its mines out of that waterway within thirty days. Read that document for what it actually is. It is an agreement to keep negotiating. It is not a signed peace deal. Nobody's name is on a settlement.
On May 23, Trump said the deal to reopen Hormuz was largely negotiated and would be announced soon. He also told PBS that Iran would not get sanctions relief in exchange for giving up its highly enriched uranium. The vice president said the two sides were making progress, but was careful about how far the White House was actually committed. And the administration stacked another condition on top of the nuclear question. It wants more countries in the region to normalize relations with Israel as part of any deal.
Then there is the other side of the table. A senior Iranian source told Reuters that Iran has not agreed to hand over its enriched uranium stockpile, and that the nuclear question was not even part of the preliminary agreement. That is the wall the whole thing keeps hitting. Iran says it has no nuclear weapon, so it frames this as refusing to give up its program and its stockpile, not nukes. The United States wants the stockpile gone. Neither side has moved off that point.
And all of this talking is happening while the shooting goes on. In late May there were two clashes inside forty-eight hours. A U.S. strike near Hormuz, and an Iranian missile fired at an American base in Kuwait that was intercepted. The week before, there were more U.S. strikes that Iran called a gross violation of the ceasefire. On May 30, the defense secretary stood up at a security summit in Singapore and said the United States is ready to resume strikes on Iran if there is no deal. Iran's own parliament speaker said concessions come through missiles, not talks. So you have a memorandum, a ceasefire renewed every few weeks, and both sides still putting metal in the air between the meetings.
Now look at what that war has done to the people living inside it, because this is the part the regime spent three months trying to keep you from seeing. For eighty-eight days, Iran was cut off from the outside world. The monitoring group NetBlocks called it the longest nationwide internet shutdown in modern history. Two thousand and ninety-three hours. At the worst of it, connectivity inside the country was running at about one percent of normal.
It started in the streets. Late last December, anti-government protests broke out and spread fast. The exiled crown prince, Reza Pahlavi, called for coordinated demonstrations and rooftop chanting, and by early January the protests had reached all thirty-one provinces. That is when the regime first pulled the plug. It pulled it again, near-total, after the February strikes. And the blackout was never only about the war. Rights groups say the regime used it to smother the protests and bury what happened in the crackdown. Thousands were reportedly killed. Thousands more arrested. Children went months without proper schooling, because the only thing left running was the regime's own walled-off national network.
The connection started coming back the night of May 26. It climbed to around eighty-six percent. But the filtering is still in place. The big platforms are still restricted, and people get around it with VPNs the regime outlawed in 2024, bought on a black market only those with money can afford. And there is a darker worry underneath the reopening. There are reports the regime imported Chinese equipment to build a permanent shutdown switch, a system of control it can throw any time it wants. Even the order that restored access got tangled up, when a court temporarily suspended the body that issued it.
And while the screens were dark, the country was running out of food. The Strait of Hormuz is how Iran brings in its grain, its corn, its soybeans, and with the strait closed, those imports stopped. By the figures researchers have been able to piece together, Iran's wheat for milling dropped to around seven or eight million tons, in a country that eats roughly one and a half million tons a month. The feed mills that supply the poultry farms were down to two or three weeks of stock, which forced farmers to slaughter their flocks early just to cut their losses. A short glut of meat in the markets, and then a protein gap behind it.
Iran's own negotiators told the United States there is a real fear of hunger in parts of the country, on top of shortages of medicine and cash. One Israeli outlet reports unemployment has climbed past fifty percent. That figure is their reporting, not a confirmed government number, so take it as that. But the picture the regime itself painted at the table was bleak. And all of it sits on top of a fuel crisis, with energy infrastructure hit on both sides of the war.
Which brings us to the thing doing the squeezing. The Strait of Hormuz. About one-fifth of all the oil and gas on earth moves through that one narrow channel. The International Energy Agency called the disruption there the largest supply disruption in the history of the global oil market. And it is not only Iran that feels it. The strait carries a huge share of the world's fertilizer too, and Britain's foreign secretary warned the world was sleepwalking into a global food crisis because of it. But right now, the country keeping that strait shut is not Iran. It is the United States.
On May 30, U.S. Central Command disabled a Gambia-flagged cargo ship, the Lian Star, in the Gulf of Oman. They put a Hellfire missile into its engine room after more than twenty warnings. As of that day, five commercial ships had been disabled and one hundred sixteen turned around. That is a blockade, and it is being run on purpose.
Because the blockade is the leverage. The White House rejected Iran's request to reopen the strait to Iranian tankers until the regime accepts the nuclear terms. It has offered to let ships carrying basic goods through, and to release some of Iran's frozen money, but only with oversight, so the Revolutionary Guard cannot get its hands on it. The Treasury is hunting down the Guard's accounts to squeeze the regime from the inside, working on the bet that when the Guard cannot pay its own loyalists, the hardliners soften. One U.S. official said it flat out. The economic blockade is the best leverage they have for the central goal of this whole war, which is ending Iran's nuclear threat for the long term.
So put it all together. The talks are open, but nothing is signed. The internet is back on, but the shelves are bare. And the strait that feeds and fuels that country stays closed until the regime puts its name on a deal. The blackout, the empty markets, the ships dead in the water, and the memorandum nobody has signed are not four separate stories. They are one. The pressure and the negotiation are the same thing, and it all holds until Iran signs.
Now to the thing everybody says is about to change the world. Artificial intelligence. You have heard the promises. It is going to replace your doctor, your lawyer, your accountant. The people building it talk about machines that will become smarter than people, maybe even wake up one day and become self-aware. So set the sales pitch aside, and look at what these tools have actually done out in the real world, on the record, over the last year and a half.
Start with the money, because that is the cleanest measure there is. This past summer, MIT published a study called The GenAI Divide. They looked at three hundred public AI deployments, surveyed the executives running them, and asked a simple question. What are companies getting back for the thirty to forty billion dollars they have poured into this technology? The answer was brutal. Ninety-five percent of those projects delivered no measurable return. None. Only one in twenty produced value you could point to on the books. And the pattern underneath it is just as telling. Eight in ten companies are poking at these tools, but only about one in twenty ever gets a project into real production, and the big outfits that try take most of a year to get there. The most money went into flashy sales and marketing pilots, which is exactly where the returns were lowest. The researchers are fair about it. They say a lot of the failure is in how companies bolt the tools on, not the tools by themselves. Nineteen out of twenty attempts to put this miracle to work produced nothing anyone could measure.
Now here is what happens when they do put it in front of you. Air Canada. A man's grandmother died, and he went to the airline's website to book a flight home. He asked the company's chatbot how the bereavement discount worked. The chatbot told him to book now and apply for the discount within ninety days. That was false. No such policy existed. When the man tried to collect, the airline refused, and then stood in front of a tribunal and argued it could not be held responsible for what its own chatbot said. Its defense, in writing, was that the chatbot was a separate legal entity, responsible for its own actions. The tribunal called that remarkable and made the airline pay. The machine invented a policy out of thin air, said it with total confidence, and a grieving man got burned.
And that confidence is the heart of the problem. These tools make things up. The industry calls it hallucination, which is a polite word for it. What it means is the machine will state a flat falsehood with the exact same certainty it states a fact, and it does not know the difference between the two. The clearest place in the world to watch it happen is a courtroom.
Lawyers, real ones, at real firms, have been caught over and over filing legal briefs stuffed with court cases that do not exist. The machine invents the case name, the judge, the quotation, the citation number, all of it, and all of it looks real. It started back in 2023, when two New York lawyers were sanctioned over a brief full of fake cases out of ChatGPT. One of them had actually asked the machine whether the cases were real. It told him yes. They were not. The judge said those lawyers had abandoned their responsibilities, and reminded the whole profession that a human signature on a filing is supposed to mean a human checked it.
And it has only sped up since. A California attorney was fined ten thousand dollars this fall for an appeal with twenty-one made-up cases in it. Three lawyers at one of the largest firms in the country were sanctioned in February for the same thing. Attorneys at another big firm got thrown off an Alabama case for it. A researcher in France now keeps a running database for the single purpose of tracking these filings, because there are too many to follow any other way. And before anyone decides this is one company's problem, one of those lawyers was sanctioned over fake citations that came out of a different company's tool entirely. It is not a brand. It is the technology. They all do it.
And do not think the fix is to buy the expensive version built for the job. Stanford researchers tested the high-end legal research tools, the ones that cost firms thousands of dollars a month, the ones marketed with the exact words one hundred percent hallucination-free. They ran real legal questions through them. One of those tools made things up seventeen percent of the time. Another, a full third of the time. The plain consumer chatbot was wrong better than four times out of ten. These are the products built and sold specifically to stop the machine from lying, and they still lied on a fifth to a third of the answers, while the company selling them promised zero.
And they bend. They are built to keep you happy, and that pulls them right off the truth. Last spring, the company behind ChatGPT had to publicly yank back an update to its own flagship model, because the thing had turned into a yes-man. It was praising terrible ideas, validating bad decisions, telling people whatever would make them feel good about themselves. The company's own words for what it had shipped were, quote, overly supportive but disingenuous. That is a tool with seven hundred million users a week, and the people who built it could not keep it from sliding into flattery over one single update. It does not hold a line. It drifts toward whatever it decides you want to hear.
So step back and look at what the whole record shows about these machines. They do not reliably follow the rules they are handed. Give one a clear set of instructions, and it will drift off them and do something close, or something else entirely. They do not remember. Most of them start every conversation from a blank slate, with no memory of what you told them an hour before, so you end up explaining the same thing again and again. They are inconsistent. Ask the same question twice and you can get two different answers. And they cannot tell the difference between something they actually know and something they invented on the spot. Every fake court case, every made-up airline policy, came from a person who trusted the machine to do the job on its own. The only ones that did not end in a mess are the ones where a human checked every single line.
And notice the one thing every story here shares. Not a single one of them was the machine working on its own and getting it right. The rare wins, that five percent, come where a person scopes the job down tight and rides the tool through every step. The wrecks come the moment somebody trusts it to run on its own. That is not a partner. That is a tool that has to be watched every minute it is switched on, checked line by line, corrected and corrected again. Which is the exact opposite of what we were sold.
Which is why the talk about these things waking up, becoming aware, becoming smarter than people, runs straight into a wall the moment you sit down and use one for real work. A tool that cannot hold a set of instructions for the length of one task, that has to be walked through every step, that invents court cases and airline policies and reads them back to you with a straight face, is a long way from anything you would honestly call intelligence. The technology is real, and some of it is impressive. But out here, where the work actually gets done, it is average, it is unreliable, and it cannot be left alone in the room.
Get out of the courtroom and look up for a minute, because we earned it. On Friday, the twenty-second of May, out at Starbase in south Texas, SpaceX flew the biggest, most powerful rocket human beings have ever built. Fully stacked, this thing stands four hundred and eight feet tall. That is about the length of a Boeing 777 stood up on its tail. It was the first Starship flight of the entire year, and the debut of a brand new, third version of the machine.
They rebuilt nearly everything on it. Fewer but bigger fins on the booster. A new staging design built right into the body that throws away one less piece. Cleaner lines and better heat shielding. And new engines, called Raptor 3, that are lighter and simpler and put out more than double the thrust, at sea level, of the giant rocket NASA built for its own moon program. Eighteen million pounds of thrust all told. They built a new launch pad to fuel it faster, so they can fly it more often. The whole point of the redesign is to make it simpler, stronger, and quick to turn around.
And by and large, it worked. It lifted off carrying twenty mock Starlink satellites and flew for about an hour, halfway around the world. It deployed those dummy satellites through a new dispenser built for rapid release. The ship came down upright into the Indian Ocean, under what looked like full control, then it tipped over and erupted in flames, which is exactly what SpaceX expected it to do. Musk called it an epic launch and landing, and told his team they had scored a goal for humanity.
It was not clean. The booster suffered multiple engine failures on the way up. The ship lost an engine too, and had to burn the others longer than planned to reach its trajectory. Because of that early trouble, they did not even attempt to relight an engine in space, a thing they have pulled off on earlier flights. But after the year these people had, this counted as a very good day.
And what a year it had been. Five test flights in 2025. The first three of them blew up. Two came apart on the way up and rained debris over parts of the Caribbean, and forced commercial jetliners to change course to stay clear of it. The aviation regulators opened reviews each time. But they ended the year with two solid flights in the fall. That is the pattern with these people, and it is worth respecting. They blow it up. They go pick through the wreckage. They find the failure. They fix it, and they fly again. That is engineering. It is the opposite of the two stories we just told you. It is a part that gets measurably better every single time it rolls off the line.
Here is the timing that matters. Just before this flight, SpaceX filed the paperwork to go public, and it could turn out to be the biggest stock offering in history. That filing told us things the company normally keeps locked up tight. They have spent more than fifteen billion dollars on the Starship program. Last year the space side of the business brought in about four billion dollars but lost more than six hundred million, and the whole operation leans on their Starlink satellite internet for its profits. So while the rocket goes up, the dream is also getting cashed in, and the rest of us are finally getting a look at the books.
And here is a fact that ought to sit with you a while. For all that size, all that money, all that thrust, Starship has never once reached orbit. Every flight so far has been suborbital. Up, around the world, and back down, but never into a stable orbit the way every satellite overhead does it every day. The company says it expects to begin orbital missions in the second half of this year. The biggest rocket ever built has not yet done the most basic thing in spaceflight.
Now hear where the man says it all goes, because he does not think small. The goal is to send an uncrewed Starship to Mars by the end of this year. The crew on that first run would be Tesla robots, humanoid machines standing in for people. It has to be the end of the year, because Earth and Mars only swing close enough together for the trip once every two years, and even then it is a seven to nine month ride, one way. Miss that window, and you wait two more years for the next one. And the man himself, not his critics, the man himself, puts the odds at about fifty-fifty.
That fifty-fifty is generous, and here is why. The whole Mars trip depends on refueling the ship out in orbit after it launches. A run to Mars takes far more fuel than the ship can carry up off the pad in a single go, so the plan is to fill it back up once it is in orbit and then send it on its way. That has never been done. They have exactly one flight of this new version under their belt, the thing has not even reached orbit yet, and the make-or-break maneuver is still on the drawing board. An uncrewed shot at Mars in the next seven months is a long, long putt. And the actual human beings setting foot on Mars? He is now saying 2029 at the earliest, and admits 2031 is more likely.
It is not just Mars riding on this one machine. It is the moon. NASA is counting on this exact rocket to be the lander that puts American astronauts back on the lunar surface, for the first time in more than half a century. The way it is supposed to work, the crew launches on NASA's own capsule, the two craft dock together in orbit in a critical test, and then the Starship lander carries the astronauts down to the surface. That moon landing is currently aimed at around early 2028. But the schedule for the lander is, in the words of one former NASA mission director, significantly challenged, and could run years late. The same rocket that has flown one single time this year, and has never reached orbit, is the one our astronauts are supposed to ride down onto the moon. Back in February, the man even floated shifting the whole focus to building a city on the moon first, before he ever gets to Mars.
He says the why plainly enough. He wants us to be a multi-planet species. A backup of humanity somewhere besides this one rock, so that whatever happens here, the light of human life does not all go dark at once. He talks about lifting a hundred people at a time off this planet one day. A future that pulls a person forward, instead of one that just grinds along in a courtroom. You can roll your eyes at a billionaire talking like that. Plenty of people do, and they have their reasons.
And one man, out in a Texas field, flying the biggest rocket ever built, betting it on Mars, and carrying America's whole trip back to the moon on the same machine that has not yet reached orbit. One is aimed at the sky. A country is at its best when it is reaching forward, not settling old scores.
*
AND THAT'S IT.
That is everyone knows. The hard truth, the stuff nobody else will say, and the thing that keeps you going.
I am Brian Bullock of Starborne Studios. Find me on X at Everyoneknws1. If this hits you share it with someone who needs to hear it. Subscribe. Leave a review. Come find us at brianbullockwriter.com
Because everyone knows. Your silence is your consent.