When proof no longer proves anything

The trap of the digital trace

The timing makes it worse. Over the past decade, insurance had willingly exposed itself to exactly this attack. In the name of modernization, the sector moved a growing share of claims handling onto remote, automated channels: first notice of loss by app, damage photos taken by the insured, video identity checks, voice authentication, documents uploaded with no human contact. This dematerialization was praised as a gain in speed and comfort, and it was. But it moved the handling of proof onto exactly the forms AI fakes best: images, voices, documents, remote identities.

The warning already given about insurtech applies here, transposed. Removing friction from a journey also removed the human filter that, unremarked, served as a control. The agent who met the client, the adjuster who went out to the site, the employee who picked up the phone all ran a tacit check. They sensed whether the person was real, whether the story held, whether something felt off. Swapping them for a form and an algorithm made things smoother and lost that human judgment just when it was most needed. The industry automated its way to maximum exposure. It built the highway just before the cars learned to lie.

The consequence goes beyond claims. It reaches the moment the contract is written. Underwriting too rests on the proposal form and supporting documents, and an applicant can now present a synthetic identity, a fabricated history, coherent documents that point to no real person. The insurer that believed it knew its insured may have contracted with a fiction, and the risk it thought it had priced does not exist, or is not the one it thinks. So corrupted proof infects both ends of the system: the payment on the way out, and knowledge of the counterparty on the way in. A contract built on a false identity rests on nothing.

The liar's dividend

The corruption of proof has a second effect, more twisted than the first. This one takes the subject well beyond the fight against fraud. As long as fabricating a fake was hard, the authentic proof enjoyed a presumption of truth. Once everyone knows that anything can be fabricated, that presumption vanishes on both sides at once. The fake can pass for true, and the true can be rejected as fake. An authentic recording of a real event can be waved away, it is only a deepfake, and a legitimate proof loses its force because it has become, in principle, contestable.

English-speaking legal scholars call this effect the liar's dividend. The fraudster gains without faking anything, simply because fakes exist as a category, which lets them deny what is true. Applied to insurance, it creates a symmetrical double corrosion. On one side, the insurer can no longer trust what seems authentic, which multiplies false positives, invented claims that pass the filter. On the other, a bad-faith insured can reject the evidence against them, the surveillance footage or the recording that gives them away, by claiming it was doctored. That multiplies false negatives, frauds that can no longer be proven. The tool that was supposed to help prove things now also helps deny them.

The link to reputation risk is plain, since there the harm had already come loose from the truth. The deepfake takes that break to its limit. It makes the false as powerful as the true, and the true as doubtful as the false. That hits hard at an industry whose trade is settling disputes about what happened and deciding who is right when two accounts collide. The loss adjuster, the investigator, the court hearing a coverage dispute, all practiced an art that assumed proof possessed a default truth-value. That ground is now giving way. Beyond whether a given piece of evidence is good, the question becomes whether any evidence at all can still prove anything.

A wave, not an incident

Isolated cases hide one property. AI-augmented fraud is correlated: it behaves like a wave, not like an accident. A method that works, a voice-cloning model, a template for a false document, a technique of synthetic identity, can be reproduced identically against thousands of victims at the same time, at a near-zero marginal cost. The old-school forger worked one file at a time. An attacker with AI works at industrial scale and fires the same weapon at a whole book. Fraud stops being diversifiable background noise, a small share of independent claims that the law of large numbers absorbs. It becomes a peak risk, able to hit many contracts at once with a single technique.

The shift from diffuse to correlated fraud echoes the central lesson of the series, simultaneity, on different ground. An insurer could calmly reserve for stable, scattered fraud losses. It cannot reserve the same way for a mass attack that exploits a flaw all its insureds share. The day a new method circulates, it can produce a burst of fraudulent claims concentrated in time, before detection even adapts, and this dynamic resembles an epidemic more than an individual hazard. So corrupted proof is a quality problem claim by claim, and also an accumulation problem book by book.

The correlation flows up to reinsurance, because a common-cause peril does not stop at one insurer's border. It runs across every market player that shares the flaw. More troubling still, the defense itself is concentrated. If all insurers rely on the same few detection vendors, they inherit a common blind spot, a fake that fools one detector fools them all, and the countermeasure meant to diversify the risk correlates it further. This is the concentration risk already described for digital infrastructures, now applied to verification tools. A technical monoculture turns a single flaw into a general failure and makes the fix one more channel for the problem.

The lines that pay first

The threat is not spread evenly across the market. It hits first the lines where proof decides everything. Financial lines and wire fraud are in the firing line, since the most mature weapon, an executive's cloned voice, targets the payment order directly. Life business is exposed by synthetic identities and false certificates, which make it possible to insure people who do not exist or to declare deaths that never occurred. Motor and property are exposed by fabricated loss photographs, a crumpled bumper, a water leak, a torn-off roof, generated without any accident having happened. Wherever indemnity follows an image, a document or a voice, the breach is open, and all the more so because these lines had invested the most in remote settlement.

Cyber insurance is a special case here, exposed twice over. It is first a target, like every line, whose procedures rest on digital proof. But it is also the cover that, increasingly, takes on these frauds themselves: social engineering, funds transfer fraud, CEO impersonation. So the cyber insurer is both judge and payer of a loss AI makes more frequent. Corrupted proof hits its own claims handling, and it pays the resulting losses of its insureds. Here the thesis stops being abstract and becomes a quantifiable exposure. It sits on a line that is already strained, and that ran into accumulation as soon as systemic cyber came into play.

The race between the forger and the hound

The optimistic objection deserves a fair hearing, because it has some force. AI also arms the defense. Fake generators meet fake detectors, and anomaly analysis, provenance checks, watermarking and cryptographic signing of authentic media give the insurer new ways to sort true from false. Fraud has always existed, the argument goes, and the sector has always adapted to each new trick. The deepfake would then be one more lap in an endless race where insurance has held its own so far. The point stands: the race is real, and the insurer is not unarmed in it.

But there are three reasons not to stop there. The first is that detection is structurally behind generation. A model that fabricates fakes trains precisely to defeat existing detectors, so that the generator improves by learning from the defender, and the hound always runs behind the forger who studies it. The second is that, even won, this race carries a cost that falls on the honest. Checking every document and demanding extra authentication slows settlement for every insured, honest ones included. They pay in delay and suspicion for the mere possibility of a fake. The pool absorbs a new verification tax, a deadweight no real loss justifies, and the speed that digital claims handling promised turns into defensive slowness.

The third reason is the deepest. No detector can remove the liar's dividend, because it targets default trust rather than any single artifact. You can, at a stretch, unmask a given fake. Technology cannot restore the presumption of authenticity that let proof work without having to prove itself. Even a perfect detector cannot give true evidence back its natural credibility. It adds one more step, a check on the check, and moves the question without answering it. Now you have to trust the detector, then the detector of the detector. Tools do not repair lost trust. They displace it, and each displacement costs.

Rebuilding proof, at the price of the panopticon

One defense goes to the root, and it is also the most disturbing. If proof no longer holds by default, the industry can try to make it reliable again by certifying it at the source, through cryptographic provenance. The idea is to sign each piece of content at the moment of its capture, so that an image, a voice, a document carry an unforgeable proof of their origin and integrity, a seal AI could not imitate. Technically, this is the most serious path, and it would reconcile insurance with a world in which the trace became trustworthy once more.

But the price of this reconciliation is staggering. For provenance to carry weight, every device, sensor and phone would have to sign everything it records and tie that signature to a verified identity. That means a world where nothing happens without a date, a signature and a trace. The anonymity of proof, which was also a form of freedom, would give way to total traceability. We would trade the deepfake problem for the panopticon problem, and restore trust in proof only by abolishing the possibility of acting without leaving a signed trace. This is the specter already met with behavioral pricing, that of a surveillance which, under cover of justice or security, turns every gesture into attributed data.

Insurance faces a choice it cannot settle alone, though it will be among the first to feel its edge. Either it learns to live in a world where no proof can be believed by default, and it pays the cost in verification, in slowness, in endless disputes, in books suddenly vulnerable to a single weapon. Or it demands, with others, the authentication infrastructure that would make proof reliable again, and it then helps to build a world of total traceability whose dangers are no less than those it seeks to avert. Between worthless proof and total surveillance, there is, for now, no comfortable middle ground.

Insurance is only the first sensor to pick up a wider tremor. The whole edifice of documentary trust (contracts, records, the press, the courts) rests on the same assumption that proof possesses, by default, a truth-value. Insurance is simply its most sensitive sentinel, because it pays real money every day against traces, and so it feels, before the others, the ground give way. What the sector undergoes today in its claims departments, society will undergo tomorrow in its courtrooms and its public debates, when no image or recording can carry conviction on its own any longer. The insurer is less the victim of a technical problem than an early witness to a shift: the moment proof stops going without saying.

For centuries, insurance rested on an assumption so obvious that no one bothered to state it: lying convincingly was expensive. Artificial intelligence has made lying free, and in doing so it has done more than help fraudsters: it has dissolved the difference between evidence and assertion. The industry that pays against proof enters an age in which proof is worth only what the infrastructure able to authenticate it is worth, and that infrastructure, it is becoming clear, would require recording very nearly everything. The age of the trustworthy trace is closing. The question AI raises goes beyond how to detect fakes. It asks what kind of world we are willing to live in so we can believe what we see again.