The Authenticity Premium: Why Proof Beats Output

On October 1 this year, Cloudflare shipped a decision model that runs for nine cents per million tokens. The same day, Amazon gave away a two-billion-parameter model with the weights and the training data attached, free to run on a laptop. A week before that, OpenAI cut its mid-tier model to about a fifth of last year’s price. I spent that week watching a model I rent for pocket change write landing-page copy I could not tell apart from the agency’s, draft a cold email better than the one my own team sent, and generate a product screenshot that never existed.

Here is the thing nobody says out loud when a new model drops. The cost of producing something plausible has fallen to roughly zero, and it is still falling. Plausible writing, plausible code, plausible design, a plausible profile of a person who does not exist. If you build a company, this is the single economic fact that should reorganize how you think, because it changes what the market is willing to pay for.

When making a thing becomes free, people stop paying for the thing. They start paying for proof that it is real. That gap, between what looks real and what is provably real, is where value is moving, and I think it is the most under-priced opportunity of this cycle. I call it the authenticity premium, and most of the pricing power in the next few years sits inside it.

I have built four companies, and in each one the thing I sold eventually got cheaper to copy. This time the copying is happening to everyone at once, in every category, in the same year. So the question I want to answer is not whether your output gets commoditized. It will. The question is what you own when it does, and the answer keeps pointing at the same place: proof.

The cost of faking anything just fell to zero

Start with the supply side, because it is already past the point most people think it is approaching. Researchers at Stanford, Imperial College London, and the Internet Archive analyzed web pages published between 2022 and 2025 and found that 35.3 percent of all new websites are AI-generated or AI-assisted, and 17.6 percent are entirely machine-written. Cloudflare reports that close to a third of internet traffic now comes from bots. Imperva marked 2024 as the first year automated traffic surpassed human traffic. The open web stopped being a mostly-human place while most of us were still arguing about whether that would ever happen.

Now walk it into the places where money changes hands. In a 2026 Capital One survey, 82 percent of people said they had run into a fake review in the past year, and roughly 30 percent of online reviews are fake or planted, almost half of them five stars. The FTC puts the return on buying fake reviews at around 1,900 percent. The platforms are spending real money to hold the line: Amazon says it removed more than 275 million fake reviews in a single year and has 8,000 people and more than 500 million dollars a year aimed at the problem. Google pulled 240 million. Trustpilot removed 4.5 million, about seven percent of everything on the site.

Hiring is next. Gartner projects that by 2028, one in four job applicants worldwide will be fake, driven by AI-generated profiles and deepfaked interviews. Voice cloning now takes minutes of sample audio. A convincing video candidate takes an afternoon. The resume, the portfolio, the live interview, these were all signals that used to cost something to fake. They do not anymore.

Notice what is common to all of it. A review, a byline, a credential, a face on a call, a case study. Each one used to carry information precisely because it was expensive to produce a fake version. Strip out that cost and you strip out the information. The signal still looks like a signal. It just stopped telling you anything.

A signal is only worth what it costs to fake

The cleanest way I know to think about this is an idea that won a Nobel Prize. In 1973, the economist Michael Spence asked why a degree is worth so much in the job market even when the degree itself teaches you little about the actual job. His answer was that the degree is a signal, and it works not because of what it contains but because of what it costs. A capable person can earn a hard degree more cheaply, in effort and time, than a less capable person can. The price of faking it is the whole point. Take a signal that is expensive to fake, and it sorts the real from the pretend without anyone having to check directly.

This is the law under everything in this post. A signal carries value only in proportion to what it costs to fake. For most of history, nearly all the signals we used to decide who to trust, hire, buy from, and read were expensive to fake. A polished thousand-word essay meant someone could write. A clean portfolio meant someone could do the work. A detailed demo meant the product probably existed. A five-star pile meant customers probably liked it. None of those were perfect, but they were costly enough to counterfeit that they held.

AI did not make these signals a little cheaper to fake. It made them nearly free to fake, in one step, for everyone. And when the cost of faking a signal collapses, the signal collapses with it. The essay no longer means someone can write. The portfolio no longer means someone did the work. The demo no longer means the product exists. The reviews no longer mean customers showed up. The information that used to live inside those signals has to go somewhere, and where it goes is to whatever is still expensive to fake.

This is why the loudest advice right now, make more, ship faster, flood every channel, is quietly the wrong instinct. Everyone can make more. Everyone is flooding every channel. Volume was a signal when volume was hard. It is noise now that it is free. The scarce thing is no longer output. It is proof of origin, proof of use, proof of outcome, proof that a named human stands behind it. That is what people will pay a premium for, and the premium gets larger the cheaper the fakes get.

The premium lives where faking is cheap and the truth still matters

Not every signal is worth defending. If something is cheap to fake and nobody cares whether it is real, let it go. The premium only exists where two things are true at once: faking has become cheap, and the truth still matters to the person deciding. That intersection is where buyers are now anxious, where they are over-correcting, and where they will pay more for certainty than they used to pay for the thing itself.

I map it on two axes. One is how expensive the proof is to fake. The other is how much the buyer actually cares whether it is real. Four zones fall out, and only one of them is a business.

The danger zone is the top left, and it is where most businesses are sitting without knowing it. Your signal used to be expensive to fake, so you leaned on it. Now it is cheap to fake, buyers have noticed, and they have quietly started assuming your version might be counterfeit too. The marketing agency whose whole pitch was good copy is here. The developer whose proof was a clean GitHub is here. The brand whose trust lived in its review count is here. The work is still real. The signal that said so is dead, and the discount is already priced in.

The top right is the only quadrant that is a business. It is where you can prove something that is costly to fake, and where the buyer cares enough to pay for the certainty. The whole move, strategically, is to drag your value out of the top-left danger zone and into the top-right premium. Everything below is about how.

The biggest buyers in the world are already paying it

If you want proof that realness is repricing upward, look at who is writing the largest checks for it: the AI labs themselves. The companies that made output free are now paying extraordinary sums for provably-real human data, because their own models are starving for it.

Reddit licenses its human conversations to Google for a reported 60 million dollars a year and to OpenAI for around 70 million. News Corp signed with OpenAI for a reported 250 million dollars over five years, the largest content deal disclosed so far, and with Meta for up to 50 million a year. Amazon pays the New York Times a reported 20 to 25 million a year. Shutterstock booked over 100 million dollars in AI licensing in a single year. And when the copying was not licensed, the bill came anyway: Anthropic agreed to a 1.5 billion dollar class settlement with authors, approved in 2026. These are not charity. They are a market setting a price on verified-human material, and the price is high because the material is scarce and getting scarcer.

The scarcity has a mechanism, and it is worth understanding because it is the deepest version of this whole argument. In 2024, a team led by Ilia Shumailov published a result in Nature showing that when you train a model on the output of earlier models, over successive generations the model degrades. The tails of the distribution disappear first, then the whole thing narrows toward bland averages. They called it model collapse. As the open web fills with machine-generated text, the labs are increasingly at risk of training tomorrow’s models on yesterday’s model output, which poisons the well. Clean, provably-human data, the kind produced before the flood or inside a trusted boundary, becomes the rare input that keeps the models from eating themselves.

So the single most valuable commodity in the most valuable industry of the decade is turning out to be proof of human origin. That is not a side note. That is the authenticity premium showing up at the top of the food chain, in nine-figure contracts, before it reaches the rest of us. It also reframes an old question. I have written before about whether your proprietary data is actually a durable advantage and about whether you should ever sell it. The new wrinkle is that the value of your data now depends heavily on whether anyone can prove it is real and where it came from. Provenance is becoming part of the asset.

Trust is becoming the expensive part of every transaction

Step back from any single signal and look at the buyer’s side of the table. When output was costly to fake, a buyer could glance at a review, a profile, or a demo and move on. The signal did the checking for them. Now that the signals are cheap to fake, the checking falls back onto the buyer, and checking is slow and expensive. In a 2026 Omnisend survey, 93 percent of people said they double-check an AI recommendation before buying, and 27 percent said they always verify first. That is a tax, and every person on the other side of your pitch is now paying it whether they say so or not.

This is the quiet shift that reorganizes markets. The cost of a transaction used to live mostly in producing the thing. It is moving into sorting the real from the fake. A hiring manager spends more time weeding out fake candidates than evaluating real ones. A buyer spends more time verifying claims than comparing features. The friction did not disappear when output got cheap. It relocated to trust, and it got bigger.

There is a cruel twist in this for honest operators. When fakes flood a channel, buyers do not just discount the fakes. They discount the whole channel, including you. The 82 percent of people who hit a fake review in the past year did not come away distrusting only the fakes. They came away trusting reviews less across the board. So the honest seller with real five-star reviews now gets lumped in with the counterfeit ones and pays for a crime they did not commit. The only way out is a signal the fakers cannot copy: verified-purchase proof, a named customer, a number tied to a real account. You do not get to opt out of the trust collapse. You only get to prove your way above it.

Whoever removes that tax for the buyer captures the value. If you can make it instantly obvious that your results are real, your customers are real, and your claims are checkable, you have taken a cost off the buyer’s plate that every competitor is still making them pay. That is worth a premium in the most literal sense, because you are not only selling the product, you are selling the end of the buyer’s verification work. This is a different problem from the one I wrote about in the trust gap, which is about how much to trust the output your own systems hand you. This is about the trust your buyer has to manufacture about you, and about who pays to manufacture it. The companies that win the next few years will feel, from the outside, like the ones you do not have to double-check.

Four signals AI just made worthless, and what replaced them

Let me get concrete, because abstraction is where this kind of thinking goes to die. Here are four signals founders and operators rely on every day, all of which AI just moved into the cheap-to-fake zone, and what is quietly replacing each one as the thing that actually carries weight.

Read down that right-hand column and a pattern jumps out. Everything that replaced a broken signal has one property: it costs something real to produce, and you cannot generate it on demand. A track record takes years. A reference who answers the phone had to have a real relationship with you. A verified-purchase number had to come from an actual transaction. An outcome tied to a named person and a date cannot be conjured, because the person and the date are checkable. These are the new costly signals, and they sit on a ladder.

If you want a fast test for where your own signals sit, subtract the AI. Ask of each proof point you rely on: could a competent person with a cheap model fabricate a convincing version of this in an afternoon? If the answer is yes, it is in the bottom of the stack and the market is already discounting it. If the answer is no, because it would require a real history, a real customer, or a real result, then it is near the top, and that is the part of your story worth investing in. This is the opposite side of a problem I wrote about recently, the verification ceiling. There, the constraint is how fast you can check the output you generate. Here, the market is asking the same question about you from the outside: can anyone else check that what you claim is real.

Proof is becoming a product, not a feature

When a scarce thing becomes valuable, infrastructure grows up around producing and verifying it. That is already happening with proof, and it tells you the premium is real rather than theoretical. At its 2026 developer event, Google rolled out SynthID watermarking and the C2PA content-credentials standard across Chrome and Search, reaching billions of users, so an image can carry a cryptographic record of how it was made. OpenAI announced C2PA conformance in the same window. The two dominant labs are now building the plumbing for provenance into the places people actually look.

But here is the honest part, and it is where the opportunity hides. The plumbing is leaky. SynthID only detects Google’s own models, not Midjourney or open-source tools. C2PA metadata gets stripped the moment a file passes through most platforms. Watermarks can be removed. So the official provenance layer covers a fraction of the content, badly, and that gap is enormous. The companies that figure out how to prove realness in the messy middle, where a credential actually has to survive a screenshot, a re-upload, and a skeptical buyer, are building something people will pay for.

I do not think this is only a job for big platforms. It shows up as a feature wedge inside ordinary products. A hiring tool that can verify a candidate is a real person who did the work they claim. A marketplace that can prove a review came from a real transaction. A content platform that can show which parts of a piece were human-made and which were generated. A data vendor that can certify where every row came from. In a world drowning in plausible fakes, being the trusted source in your category is not a nice-to-have. It is the position everyone else is trying to buy their way back into. I made a similar argument about defensibility in the absorption test: the frontier models will swallow your clever feature, but they struggle to swallow trust that took years to earn.

Where the premium shows up in your own company

This is not a trend to admire from a distance. It changes four concrete things inside a company right now: how you price, how you sell, how you hire, and how you market. In each one, the cheap-to-fake version is quietly losing value and the provable version is quietly gaining a premium.

Pricing is the one founders feel first, and it connects to something I argued in pricing under cheap inference. When the tokens are nearly free, pricing per unit of output is a race to the floor. The durable move is to price the thing the output cannot fake on its own: the verified result, the guarantee, the name on the line if it goes wrong. Vertical AI companies that sell a proven outcome in a specific domain keep churn in the low single digits and charge a premium, while horizontal tools that sell generic output churn around 15 percent, because generic output is exactly what buyers can now get for free somewhere else.

Sales changes in a way that feels almost nostalgic. The deck and the demo were powerful because they were hard to fabricate. Now a competitor can spin up a better-looking version of your product in a video before they have built anything, a dynamic I called the cut list when it comes to getting dropped by a buyer consolidating vendors. So the proof that moves a deal forward shifts back to things that resist fabrication: let them touch the live system, let them talk to a customer who will be candid, show them a result tied to an account they can verify. The buyer is not being difficult. They have been burned by plausible fakes, and they are pricing that risk into every pitch.

The gap widens every time the models get better

The instinct is to treat this as a temporary mess that better detection will clean up. I think the opposite is true, and it is the most important thing to internalize. Every improvement in the models makes the fakes more convincing, which makes the premium on provable-real larger, not smaller. The better AI gets at producing plausible output, the less a plausible output proves, and the more valuable it becomes to prove what is actually behind it.

This is why the premium compounds. In a world with weak AI, a good essay was decent evidence of a good thinker, so proof of the person underneath mattered less. In a world with strong AI, a good essay is evidence of nothing, so proof of the person underneath is the entire value. The curve runs the wrong way from what most people expect. We assume that as AI gets better, human proof matters less. In fact, as AI gets better, human proof is the only thing left that carries information, which is exactly why its price goes up.

There is a personal version of this too, which I wrote about as the worth gap. As AI absorbs the doing, your value moves to the parts of you it cannot fake: your judgment, your accountability, the specific results people can trace back to you. The company version and the personal version are the same law pointed at different subjects. When making is free, worth moves to proof, whether the subject is a product or a person. It also connects to the slower truth I keep running into, that capability gets shipped long before it gets absorbed, which I called the capability overhang. Proof is one of the last things to get absorbed, because it is the one thing that cannot be downloaded.

Nobody pays extra for human-made, they pay for verified-real

Here is where I want to push back on the obvious version of this argument, because the obvious version is a trap. The lazy read is that we are entering a nostalgia economy, where people will pay more for anything stamped human-made the way they pay more for handmade pottery. That is wrong, and building on it will cost you.

Buyers do not actually care that a human did the work. They care that the claim is true and that someone is accountable if it is not. A customer choosing a vendor does not want a lovingly hand-coded product, they want a product that provably works and a company that will answer when it breaks. A hiring manager does not want a candidate who refused to use AI, they want a candidate who can really do the job and is really who they say they are. The premium is on verifiable reality and accountable outcomes, not on the absence of machines. Plenty of the proof at the top of that stack will itself be produced with AI. What makes it valuable is that it is checkable and that a name stands behind it.

Get this distinction wrong and you waste the whole opportunity. You start marketing your human-ness as a virtue, which buyers correctly read as a weak signal, since anyone can claim it and the claim proves nothing. You put human-made on the label and watch it do nothing, because human-made is cheap to fake too. Get it right and you do the harder, more valuable thing: you make your claims checkable, you tie your results to names and dates, and you build the kind of track record that cannot be generated. One is a slogan. The other is an advantage. This is also the honest limit of the whole idea. Proof is expensive and slow to build on purpose, which is exactly why it is worth money, and why there is no shortcut to it.

Two ways founders get the premium wrong

Once this clicks, the instinct is to prove everything, loudly, everywhere. That instinct creates the two failure modes I see most, and both of them waste the opportunity in opposite directions.

The first is over-proving. There is a whole set of things that are expensive to fake but that nobody actually doubts, and pouring proof into them is motion without return. If no buyer in your category questions whether your company legally exists or whether your office is real, building an elaborate verification story around it earns you nothing. Proof is expensive to produce, so spend it where the doubt is. The test is simple: before you invest in proving something, check that a real buyer actually hesitates there. If they do not, you are decorating, not defending.

The second mistake is treating proof as a campaign instead of a system. A founder reads an argument like this, commissions a few verified case studies, puts a verified badge on the site, and considers the job done. Six months later the case studies are stale, the badge means nothing because everyone added one, and the trust has leaked back out. Proof that works is not a one-time asset you produce and display. It is a standing habit: every result tied to a name and a date as it happens, every customer reference kept warm, every claim carrying a way to check it, built into how the company runs rather than bolted on for a launch. The reason this is hard is the reason it is valuable. A competitor can copy your badge in an afternoon. They cannot copy two years of you doing it on purpose.

Both mistakes come from the same misread, that proof is a thing you say rather than a thing you accumulate. The founders who get it right treat proof the way a careful company treats its books: not a marketing flourish, but a running record that happens to be worth money precisely because keeping it honestly costs something.

How to find your authenticity premium this month

You do not need a strategy offsite for this. You need an afternoon and a willingness to look hard at your own signals. Here is the exercise I run on my own companies, and it takes about a week to work through properly.

First, list every signal you currently rely on to win trust: your content, your reviews, your case studies, your demo, your resume screen, your pitch. For each one, run the subtract-the-AI test. Could a competent stranger with a cheap model fake a convincing version of this in an afternoon? Be brutal. Most founders are shocked by how much of their trust stack has quietly moved into the cheap-to-fake zone. Mark each signal green if it is still costly to fake and red if it is not.

Second, for every red signal, find the costly-to-fake proof that sits behind it and move your weight there. Your reviews got cheap to fake, but a verified-purchase count and two named customers who will take a call did not. Your demo got cheap to fake, but live access and a result tied to a dated account did not. You are not adding proof for its own sake. You are relocating the signal from a layer that no longer carries information to one that still does.

Third, pick one place where you can turn proof into a product feature, not just a marketing line. Can you let a buyer verify a result inside the product. Can you certify where your data came from. Can you show which parts of your output are machine-made and which are not. Pick the one that matters most to your buyer and build it, because being the provably-real option in a category full of fakes is a position competitors cannot copy by shipping faster. Then price it. Charge for the verified outcome and the accountability, not for the output volume that is now free everywhere. If you do only one of these three this month, do the first one, because you cannot fix a trust problem you have not admitted you have.

Questions founders ask about the authenticity premium

What is the authenticity premium in one sentence?

It is the extra value the market pays for proof that something is real, once AI has made the thing itself cheap to produce and cheap to fake. As output gets free, buyers stop paying for output and start paying for provable origin, provable use, and provable outcomes.

Is this just a fancy way of saying quality still matters?

No, and the difference is the whole point. Quality is about whether the thing is good. The authenticity premium is about whether anyone can prove it is real and who is accountable for it. A genuinely good product with no way to prove its claims now sits in the same discounted bucket as a fake one, because the buyer cannot tell them apart from the outside. Provability, not quality, is what moved.

Will AI content detection tools solve this and remove the premium?

I would not bet on it. Detection is an arms race, and every improvement in the models makes the fakes harder to catch. Watermarking like SynthID covers only some models, and provenance metadata gets stripped in normal use. Better detection raises the floor a little, but it does not remove the premium, because the premium is on things that are costly to fake in the first place, not on catching fakes after the fact.

Does this mean I should stop using AI in my business?

The opposite. Use AI aggressively to produce output, because that is where it is strongest and cheapest. Just stop expecting the output to carry trust on its own, and invest the time you save into the proof that does: real results, named references, verifiable track records, and provenance. The winning setup is heavy AI for production, and heavy human proof for trust.

How is this different from having a strong brand?

A brand is one form of accumulated proof, the reputation you have earned over time, and it is a good one precisely because it is slow and costly to fake. But the authenticity premium is broader than brand. It includes verifiable product results, checkable customer references, data provenance, and identity verification, most of which a young company can build long before it has a famous brand. Think of brand as one item on the proof stack, not the whole thing.

What is the single fastest place to capture it?

Your sales proof. Replace the deck-and-demo pitch, which any competitor can now fake, with live access to the real product and one or two named customers who will take a candid call. It costs you almost nothing, it is extremely hard for a competitor to fabricate, and it directly answers the fear every buyer now carries into a pitch, that what they are being shown might not be real.

Why are AI labs paying so much for human data if AI is so good?

Because models trained on model output degrade, a result shown in Nature and called model collapse. The labs need provably-human data to keep their models from narrowing into bland averages, and as the web fills with machine text, clean human data gets scarcer. The nine-figure licensing deals are the clearest signal that proof of real human origin is repricing upward, at the very top of the industry that created the problem.

How do I avoid the human-made nostalgia trap?

Do not market the absence of machines as your value, because that claim is cheap to fake and buyers discount it instantly. Market verifiable reality instead: results they can check, outcomes tied to names and dates, references who answer the phone. Your buyer does not want proof that no AI touched the work. They want proof that the work is real and that you will stand behind it.

I’m a four-time founder building AI-native companies. I write about the founder mindset, AI opportunities, and the mechanics of building with AI at vikasmalpani.com, including the broader AI opportunity map this piece sits inside.