The Exclusion Machine

A single sentence, printed and stamped

---

By the early 1980s, a claims adjuster at a mid-sized American casualty insurer could open a file in the morning and know, before reading a single word of it, roughly what he was going to find. There would be a man, usually in his late fifties or sixties, who had spent thirty years around insulation; a pipe-fitter; a boilermaker. A worker who had cut, fitted and sanded the gray-white board that lined the boiler rooms, shipyards and refineries of the postwar economy. During this period, that material was considered a 'marvel of fire safety' rather than a hazard. He was sick now with something the doctors called mesothelioma, a cancer of the lining of the lungs that has essentially one known cause. The file would arrive on the desk decades after the exposure, because the disease takes that long to appear, and by the time it did, the man frequently had months, rather than years to live.

The adjuster knew the shape of what came next, too. The claim would trace back through a long chain of employment records, through a dozen suppliers of insulation and fireproofing, through general liability policies written in the 1950s and 1960s by companies that had long since stopped thinking about them. Some of those companies no longer existed. Some had been bought and sold and renamed twice, even thrice over. The underwriters who had priced the original policies had retired or died. Nobody who was still working had planned for any of this. The policies had been priced and sold and forgotten, and the premiums that were supposed to cover the risk had been collected and spent a generation earlier, folded into quarterly results that were long since closed, audited and celebrated. What landed on the adjuster's desk was a bill for a promise that a previous era had made cheaply without truly understanding what it was promising.

What the industry did in response was almost anticlimactic. It wrote a sentence. Insurers and their standard-form organizations drafted an exclusion, a short clause added to the policy that said, in effect, that this coverage did not apply to bodily injury or property damage arising out of asbestos. The absolute asbestos exclusion, printed and stamped onto renewals, made the problem disappear for every policy written after it appeared. The old policies, the ones written in silence, still had to be paid, and they were paid for decades, draining reserves and bankrupting some very old and very large names along the way. The going-forward risk, however, was closed with a printed line. One clause on a renewal did what no amount of litigation or reserve-building could do. It stopped the future from arriving.

That is the playbook, and once you observe this pattern, you can see it everywhere. It is worth understanding, because in January of 2026 the same machine that produced the asbestos exclusion produced one for artificial intelligence. If you buy insurance for a business that touches AI in any way, you are now part of that playbook. The clause has your name on it, or it is about to.

The shape of the thing

---

The pattern has three moves, and it runs in the same order every time. Learn the order and the individual episodes stop looking like separate crises. They start looking like a single mechanism running its cycle over and over on different raw material.

First comes silence. A new risk appears in the world, something the actuaries do not yet understand and cannot yet price. Asbestos in the walls. Chemicals leaching into groundwater. Code reaching across a network. Because the standard liability policy is written broadly, to cover bodily injury and property damage from more or less any cause the insured did not specifically carve out, the new risk is covered by default. Nobody chose to cover it. No one was charged for it. The coverage exists simply because the policy language is old and wide while the risk is new and narrow, and no clause has yet been written to keep it out. Underwriters call this silent coverage, and the word silent is exact. The protection is real, it is unspoken, and no premium anywhere reflects it. The insurer is carrying the exposure without knowing the number, and the buyer is holding protection he does not know he owns.

Then comes the surprise. The risk matures, the way risks do, on a timescale far longer than a policy year. The injuries surface. The lawsuits arrive in waves rather than as isolated files. Claims that were never priced start landing against policies that were sold cheap, and the losses run past anything the reserves anticipated. This is the reckoning, and it is expensive, and it tends to arrive all at once because the underlying harm was accumulating quietly the entire time the coverage sat silent. The defining feature of the reckoning is that it feels like an ambush from the inside, even though, from the outside, it was entirely predictable. The money was always going to come due. The only question was when.

Then comes the exclusion. Having been burned, the industry does the one thing that reliably stops the bleeding on future business. It writes the risk out. A clause is drafted, circulated through the standard-form bodies, filed with state regulators, and stamped onto the next round of renewals. The silent coverage becomes loud and negative. The policy now states, plainly, that it does not cover the thing that just cost so much. And the buyer, who a year earlier had protection he did not know he had, now holds a printed gap he may not notice until the day he needs the coverage and discovers it was withdrawn while he was not paying attention.

Silent coverage, catastrophic surprise, printed exclusion. Asbestos was the archetype, and it was far from the only time. The value of the pattern is that it is predictive. Once you can name the three moves, you can look at any emerging risk and ask where it sits in the cycle, which tells you something useful about your own coverage that no broker is obligated to volunteer.

The canon

---

Consider environmental pollution, which followed the asbestos script almost beat for beat. Through the middle of the twentieth century, general liability policies covered property damage broadly, and the gradual contamination of soil and groundwater by industrial operations fell inside that broad language by default. A plant would discharge solvents or heavy metals into the ground for forty years, entirely within the norms and often within the permits of its day, and the cost of that slow poisoning sat silent inside liability policies that were priced as though the ground would take the waste forever without complaint. Then came the environmental reckoning, the era of Superfund and long, ruinous cleanup liabilities for sites that had been contaminated slowly over decades. The federal government created a legal regime that could reach back and assign the cost of remediation to the parties connected to a site, and those parties turned to their insurers to pay. The insurers had never meant to underwrite the cost of scrubbing a half-century of chemicals out of the earth. Nobody had collected a premium for that.

So they wrote the pollution exclusion. When courts kept finding daylight in the early versions, reading ambiguities in favor of policyholders and forcing payment anyway, the insurers wrote a tighter one, the absolute pollution exclusion, that left far less room to argue. The progression from a qualified exclusion to an absolute one is itself part of the pattern and worth marking, because it recurs. The first attempt to close a gap is rarely airtight. The industry drafts, the courts test the drafting, and the industry redrafts with harder language until the door will not open. By the time a clause carries the word absolute, it has usually already survived a round of litigation that taught the drafters exactly which words the courts would honor.

Consider cyber, which is the modern textbook case and the one most instructive for what is happening now. For years, businesses suffered data breaches, network failures, and eventually ransomware, and they turned to their traditional property and liability policies to pay for the damage. Sometimes those policies paid, because the language was old and broad and said nothing specific about digital harm. A property policy written to cover physical damage to physical things could be argued, plausibly, to cover the corruption or loss of data, and a liability policy written to cover harm to third parties could be argued to cover the fallout of a breach that spilled customer information. This was silent cyber, coverage for computer-related loss buried inside policies that had never been designed or priced for it. Insurers were carrying an enormous, growing, correlated digital exposure across their entire books without a line item to show for it.

The reckoning, when it came, was the surge of large ransomware and systemic-attack claims that forced insurers to confront how much digital exposure they were quietly holding. A single strain of malware could propagate across thousands of insured organizations in a matter of hours, which is precisely the kind of correlated, simultaneous loss that insurance is least equipped to absorb. The whole logic of the business depends on losses being independent, on the house not burning down at the same time as every other house on the book. Digital risk violates that assumption at machine speed.

Their response was, by now, familiar. Regulators and the market pushed insurers to stop being silent. Lloyd's of London, among others, issued a mandate in 2019, effective at the start of 2020, requiring that policies explicitly state whether they covered cyber, which in practice meant writing clear cyber exclusions into the traditional policies and pushing the risk into standalone cyber coverage that could be priced on purpose. The silent coverage was converted, deliberately, into an explicit exclusion paired with a separate market. The buyer who had been quietly protected now had to go and buy the protection as its own product, or go without. This is the constructive half of the cycle, the part that eventually gives the buyer somewhere to turn, and it is worth holding on to, because it is the closest thing to good news the pattern offers. The exclusion is not always the end of the story. Sometimes it is the beginning of a properly priced market that did not exist before.

And consider terrorism, which produced the pattern's most dramatic single instance. Before September 11, 2001, terrorism was simply not something property and casualty insurers priced separately in the United States. The risk was considered so remote, so seemingly uncorrelated with the ordinary perils of fire and flood and liability, that it sat inside standard commercial policies as an unremarked-upon inclusion. No underwriter added a charge for it because no underwriter believed it was a meaningful exposure on a domestic commercial building. Then a single morning produced insured losses at a scale the industry had never modeled, spread across property, liability, life, aviation, and workers' compensation all at once, from one correlated event in one place at one time.

The response was immediate. Insurers moved to exclude terrorism from future policies, because a risk that can produce that much correlated loss in a single instant is, by the ordinary logic of insurance, close to uninsurable. There is no way to diversify against an event that hits every line of business simultaneously and that no model can assign a credible frequency to. The industry did what it always does when confronted with a loss it cannot price and cannot spread. It reached for the clause.

But terrorism added a wrinkle worth remembering, and it is the reason this section closes the canon rather than the cyber one. When the private market pulled back, the wider economy still needed the coverage to exist. Lenders would not finance construction that could not be insured against terrorism. Landmark buildings and stadiums and transit projects stalled for want of coverage that no private carrier would write. So the government stepped in as a backstop. The Terrorism Risk Insurance Act created a federal mechanism to share catastrophic terrorism losses with private insurers, which allowed them to keep offering the coverage they had just moved to exclude. The exclusion still got written. It was simply paired with a public backstop that made the risk "carriable" again by putting the government behind the tail of the loss. Keep that arrangement in mind when you reach the end of this, because it is the one part of the pattern that has no equivalent for AI yet, and its absence is the single most important thing a buyer should understand about where the current cycle actually stands.

The machine turns on AI

---

Which brings us to the present, and to a set of form numbers that will look like bureaucratic noise until you understand that they are the leading edge of the same century-old process, running once more on new material.

On the first of January, 2026, new standard endorsements took effect that do to generative AI exactly what the asbestos exclusion did to asbestos. The forms carry the designations CG 40 47 and CG 35 08, and they attach to commercial general liability coverage, the broad, workhorse policy that most businesses carry to cover bodily injury and property damage. In plain terms, these endorsements exclude injury and damage arising out of generative artificial intelligence. The coverage that was silent, the default protection a business held for AI-related harm simply because its liability policy was old and broad and said nothing about the technology, is being converted into a printed line that says no. A company that deployed a customer-facing chatbot, or a diagnostic tool, or an automated screening system in 2025 was almost certainly carrying silent AI coverage without knowing it. The same company renewing in 2026 may carry a printed exclusion instead, and the only visible difference is a form number stapled to the back of a policy that otherwise looks identical to last year's.

It is not only the standard-form general liability world where the machine is running. At least one major carrier, WR Berkley, has moved to write an absolute AI exclusion that reaches across the management-liability policies, the directors and officers coverage, the errors and omissions coverage, and the fiduciary coverage that sit at the center of how companies protect their leadership and their professional work. These are the policies that respond when a board is sued for a bad decision, when a professional service is alleged to have caused a client harm, when the stewards of a retirement plan are accused of failing their duty. An absolute exclusion is the strongest form the machine produces. It is the same word that appears in absolute asbestos and absolute pollution, and it signals the same intent, which is to close the door with as little room to argue as the drafters can manage. When a carrier files an absolute exclusion rather than a qualified one, it is telling you that it has decided the risk is not worth carrying at any price it knows how to set, and that it would rather forgo the premium than litigate the ambiguity later.

If you have followed the pattern this far, none of this should read as a surprise or an outrage. It is the machine doing precisely what it has always done, on schedule, in the same sequence. AI-related harm was covered by silence for as long as the policies stayed old and the technology stayed young. The reckoning is arriving now, in the form of enough uncertainty about hallucinations, infringement, discrimination, and automated decisions gone wrong that insurers have decided they cannot keep carrying the exposure unpriced. So they are writing it out.

The only unusual thing about the 2026 AI exclusions is how fast the cycle is running. Asbestos took decades to move from silent coverage to printed exclusion, because the harm itself took decades to surface and the industry had to be taught its lesson slowly and at enormous cost. AI is doing it in a few years. Part of that speed is the nature of the technology, which is being deployed into millions of businesses in a very short span and generating novel legal questions faster than any court can answer them. The larger part, though, is institutional memory. The industry has done this enough times to recognize the shape early. It no longer needs to be ambushed to react. It can see the silent coverage sitting on its books, remember what happened the last four times it left such coverage in place, and reach for the clause before the wave of claims fully lands. In a sense the AI exclusion is the machine at its most efficient, skipping straight from silence to exclusion while the reckoning is still mostly a matter of forecast rather than settled losses.

What it means to be the buyer

---

Here is the part that matters if you are the one signing the policy rather than drafting it.

You are not being singled out, and you are not the victim of some novel injustice. You are living through a completely standard exclusion cycle, the same one that caught building owners in the asbestos years and manufacturers in the pollution years and, more recently, anyone who assumed an old policy would pay for a ransomware attack. The feeling of discovering a gap where you thought you had coverage is the ordinary experience of standing on the wrong side of the machine at the wrong moment. It has happened to sophisticated buyers with large risk departments and skilled brokers and every advantage. It is not a sign that you did anything wrong. It is a sign that you own a business at a moment when a particular clause is coming off the press, which is a matter of timing more than of judgment.

What it does mean is that the era of silent AI coverage is ending, and you can no longer assume that a broad, general policy quietly has your back on anything AI touches. The coverage is becoming explicit, and that fact cuts in two directions at once. On the traditional policies it is becoming an explicit no, a printed exclusion where a silent yes used to sit. In the specialty market it will, over time, become an explicit yes that you buy on purpose and pay for on purpose, the way cyber coverage became its own product after the silent-cyber reckoning. Standalone affirmative AI coverage is beginning to appear, and it will mature the way every specialty market matures, unevenly and with a great deal of variation in what any given policy actually promises.

The transition between those two states is exactly where buyers get hurt, because the exclusion arrives first and the well-formed replacement market arrives later. There is a gap in the middle, a period of months or years in which the old silent coverage has been withdrawn and the new priced coverage has not yet grown into something a buyer can rely on. Businesses that renew during that gap without noticing can find themselves genuinely uncovered for a category of loss they assumed was handled, holding a policy that reads almost exactly like the one that protected them a year before. The danger is not that the coverage changed loudly. The danger is that it changed quietly, in a form number, on a page most buyers never read.

The practical move is unglamorous and entirely within your control. Read your endorsements. Know your form numbers. When your renewal arrives, the difference between a business that is covered for an AI-related loss and one that is not will often come down to whether a specific endorsement is stapled to the back of the policy, and whether you noticed it. CG 40 47 is not trivia. The presence or absence of that clause, on your policy, is what decides whether a claim gets paid on the day you file it. The buyers who come through these cycles intact are the ones who treat the fine print as the actual product, because in insurance the fine print is the product, and an exclusion is simply the part of the product that was written to protect the insurer rather than you.

There is a second move that matters as much as reading the exclusions, and it is asking about the replacement. When a broker or carrier hands you a policy that now excludes AI, the right question is not only what came out but what you can buy to put it back, and on what terms, and at what price. That is where the affirmative market lives, and asking the question early puts you at the front of it rather than the back. The businesses that fared best in the cyber transition were the ones that treated the arrival of the exclusion as a prompt to go shopping for the specialty product, rather than as a door quietly closing that they noticed only after a loss.