Enshittification - Bargain, Then Rip-Off

Essay · platforms · September 2026

Enshittification is a word for a mechanism, not a mood. Cory Doctorow named the sequence in which a platform is good to its users, then good to its business customers at the users’ expense, then good only to itself. The sequence was in the economics literature under a duller name thirty years earlier. A small optimal-control model reproduces it from three ingredients, puts a number on when the turn comes, and suggests why it came everywhere at once. Applied to the AI firms, it says where they are in the cycle and what to watch.

Reading time ≈ 41 min Klemperer 1987 → Doctorow 2023 → a model A standalone essay

The word

Some words succeed because they are rude and some because they are precise, and the rare ones are both. Cory Doctorow put enshittification on his blog in late 2022 and set it out properly in January 2023, in a piece about TikTok that has been quoted more than anything else he has written.[1] The definition is worth having in full, because most uses of the word since have kept the rudeness and dropped the precision: “Here is how platforms die: first, they are good to their users; then they abuse their users to make things better for their business customers; finally, they abuse those business customers to claw back all the value for themselves. Then, they die.” The American Dialect Society made it their word of the year for 2023 and the Macquarie Dictionary theirs for 2024, and in October 2025 Doctorow published a book under the name.[2]

What the word names is not that things got worse. Things get worse for many reasons, and a name for decline in general would not have stuck. It names a sequence: a particular order in which a particular kind of company treats the two groups of people it stands between, and a claim that the order is not an accident of management but a property of the position. That is a claim about mechanism, and claims about mechanism can be checked. This essay tries to check it three ways: against the cases that made the word necessary, against the economics that predicted the pattern before there were platforms to show it, and against a small model built to see whether the sequence falls out of the ingredients Doctorow says produce it. It does, and the model says two more things that the word alone does not.

The mechanism

Three stages and a dial

A platform is a business that stands between two groups who want to reach each other and would find it costly to do so directly: riders and drivers, readers and writers, shoppers and sellers, advertisers and everyone. Its product is the introduction. That position has two features that ordinary businesses lack. The first is that the platform is more valuable to each side the more of the other side is on it, so the natural way to build one is to subsidise whichever side is harder to get. The second is that once both sides are there, each is held in place not by the platform’s quality but by the presence of the other side, which is not something a competitor can copy by building a better product. Economists call the first feature a network effect and the second lock-in, and the two together mean that a platform’s value to its users is mostly something the users made, by being there.

Doctorow’s three stages follow from that. In the first, the platform is good to users because it is buying a network with someone else’s money: Uber’s rides in 2015 cost the passenger roughly forty cents on the dollar, by Hubert Horan’s reading of the company’s own accounts, and the rest was paid by investors buying market share.[3] Facebook showed you what your friends posted, in the order they posted it. Amazon’s search returned the thing you searched for. In the second stage, the user base is large enough that the other side wants it more than the users want anything in particular, so the platform sells the users’ attention to the business side and the product gets worse in the specific way that makes that sale: the feed fills with things you did not ask for, the search results with things that paid to be there. Facebook told advertisers in 2014 that the organic reach of their pages, the share of followers who would see a post they had not paid to promote, would keep falling toward zero, which was true and was the point.[4] In the third stage, the business side is locked in too, because the users are there and nowhere else, and the platform raises its price to them as well. Amazon’s advertising business, which did not exist as a line in its accounts a decade earlier, brought in close to fifty billion dollars in 2023, almost all of it from sellers paying to be found on a page that once found them for free; combine that with referral and fulfilment fees and the company’s take of a typical seller’s revenue passed half, by Marketplace Pulse’s estimate, in the same year.[5] The Federal Trade Commission’s 2023 complaint against Amazon describes the same two moves, degraded search and a rising fee stack, as the substance of the case.[6]

Google is the case where the mechanism was written down by the people running it. In the antitrust trial that ended in August 2024 with Judge Mehta’s finding that “Google is a monopolist”, the exhibits included 2019 email traffic in which the advertising side of the company, facing a revenue shortfall, pressed the search side to change what the engine did in order to lift the number of queries, and the search side objected, and lost.[7] The point is not that the engine got worse in 2019, which is arguable, but that the internal argument had the structure Doctorow describes: the side of the business that is paid by users lost to the side that is paid by advertisers, because by 2019 nobody thought the users were going anywhere. Google was paying Apple around twenty billion dollars a year to make sure of it.

The dial that makes the stages possible is what Doctorow calls twiddling: a platform can change what each user sees, what each seller pays, and what each driver is offered, per person, per minute, from the back end, with nobody able to see the change from outside.[8] A shop that wants to raise its prices has to print new labels. A platform that wants to extract a little more from one group can do it by adjusting a weight in a ranking function, and can find out by experiment exactly how much more it can extract before the group notices. That is what makes the sequence gradual and what makes it deniable. It also, as it happens, makes it the kind of thing a model can describe, because a company that can set its take continuously and measure the response is doing optimal control whether it uses the words or not.

The prior art

What the economists already knew

None of this was news to industrial organisation, which had the pattern under a duller name before the web existed. Paul Klemperer’s 1987 paper on markets with switching costs showed that a firm whose customers face a cost to leave will price low to acquire them and high to keep them, and that the second price is set by the switching cost rather than by anything the firm does.[9] Farrell and Klemperer’s 2007 survey of the literature calls the resulting pattern “bargains then rip-offs” and spends a hundred pages on its variations: the rip-off is worse when the switching cost is larger, when the firm discounts the future more, and when customers underestimate how long they will stay.[10] Katz and Shapiro had by then formalised network effects, and Shapiro and Varian’s 1998 book for managers told them, in so many words, that the value of an installed base is the total switching cost of the customers in it, and that the job was to build one and then charge for it.[11] Rochet and Tirole’s 2003 paper on two-sided markets supplied the last piece: a platform sets prices to its two sides jointly, and the profit-maximising structure often charges one side nothing or less than nothing, because that side’s presence is what the other side pays for.[12]

So what did Doctorow add? Three things, and they are the reason the word has done work the literature did not. He put the two-sided structure and the switching-cost dynamics together as a sequence in time, which the economics had treated as separate topics: the rip-off happens to users first and to business customers second, because the second group is locked in by the first. He identified twiddling as the operational capability that lets the rip-off be continuous and invisible rather than a one-off price rise that customers can see and react to. And he asked what had held the pattern in check before, and gave an answer with four parts: competition, which limits how bad the product can get before people leave; regulation, which limits what the platform may do to them; interoperability, which limits lock-in by letting users take their network with them; and the platform’s own technical staff, who for a generation had enough leverage to refuse to build the worst of it.[13] His claim is that all four weakened at once, and that this is why the pattern arrived everywhere in the same few years rather than in one company at a time. That is the part a model can say something about.

A model

Bargain, then rip-off, solved

Here is the smallest model I could make that contains the ingredients. There is one platform and a population of potential users, and the state of the world is the share of that population on the platform, N, between zero and one. Every quarter the platform chooses a take, e, which is how much of the value of using it the platform keeps for itself: a take of zero is the product with nothing extracted, a take of one is everything, and a negative take is a subsidy, the platform paying people to be there. What a user gets from the platform is the network, worth 0.6 N, minus the take; and the crucial assumption is that the outside option, the best alternative, is exactly as good as the bare platform with no network and no take. The platform’s only edge is its users and its money. Users join in proportion to how much better the platform is than the alternative, with diminishing returns, so that doubling a subsidy less than doubles the sign-ups. They leave for two reasons: a background churn of four per cent a year regardless of anything, and a churn that grows with the square of how much worse the platform is than the alternative, so that a small squeeze costs almost nothing and a large one costs a great deal. Both kinds of leaving are scaled down by lock-in, a factor 1 − κN that makes leaving harder the bigger the platform is, because the bigger it is the more of your life is in it. The platform’s profit each quarter is the take times the share, and it chooses the take to maximise the discounted sum of all future profit, with an annual discount factor δ that stands for the patience of its capital.

Written down, that is five lines. Call s the surplus, what a user gains by being on the platform rather than on the alternative:

s = λN − e(1)

with λ the value of a full network. People outside join at a rate that rises with the surplus and flattens; people inside leave at a background rate plus a term that grows with the square of the shortfall; and both kinds of leaving are damped by lock-in:

join = α(1 − N)(1 − exp(−s/τ))   for s > 0, else 0(2)

leave = N (μ + β(s/τ)²)(1 − κN)   for s < 0, else Nμ(1 − κN)(3)

The share next quarter is the share now plus a quarter of the difference:

N′ = N + Δt (join − leave)(4)

And the firm’s problem is a Bellman equation. The value of holding a share N is the best available sum of the profit this quarter and the discounted value of the share that results:

V(N) = max_e [ e N Δt + δ^Δt V(N′) ](5)

Solving equation 5 gives two things at once: the value function V, what a user base of any size is worth to the firm, and the policy, the take that achieves that value at each share. The parameters, fixed unless the text says otherwise:

λ = 0.6

Network coefficient: what a full network is worth to a user, in units where the bare product is worth one.

κ = 0.6

Lock-in: at full share, departures run at 40 per cent of their unlocked rate. Also 0 and 0.3 in Figures 3 and 5.

μ = 0.04 a year

Background churn: the share of users who leave regardless.

α = 0.25 a year

Join rate: the most of the outside population that can be recruited in a year.

β = 1.0 a year

Loss-driven churn: the strength of the quadratic term.

τ = 0.5

Response scale: the surplus at which joining is two thirds saturated and the shortfall at which the quadratic churn reaches β.

Δt = ¼ year

One step. Take chosen from a grid of 141 values between −0.4 and 1; share on a grid of 201.

δ = 0.97 a year

Discount factor: the patience of the capital. Varied in Figures 6 and 7.

Each rule is a rate per person times a head count. In equation 2 the count is 1 − N, the share of the population still outside, and in equation 3 it is N, the share inside. Figure 1 draws the per-person rates alone, with the head counts divided out, because they are what the parameters shape; the flows in equation 4 are these curves multiplied back by the number of people on each side.

Figure 1. The two rules against the surplus, per person. Right of zero, the join rate for one outsider (teal) rises with the surplus and flattens: doubling a subsidy less than doubles the sign-ups. Left of zero, the leave rate for one user (magenta) rises with the square of the shortfall: a squeeze of 0.15 costs 13 per cent of the users a year, a squeeze of 0.3 costs 40 per cent. The dashed line is the same rule with lock-in at 74 per cent share, which cuts every departure rate by 44 per cent. These are rates per person; the flows in equation 4 are the join rate times the share outside, 1 − N, and the leave rate times the share inside, N.

Put the rules together, hold the take fixed, and equation 4 says what happens to the share over a year from any starting point. Figure 2 shows that for four takes.

Figure 2. The yearly change in share against share, at four fixed takes. A subsidy of 0.4 grows the share from anywhere; a take of zero grows it too, more slowly, because the network alone is worth something. A take of 0.3 shrinks the platform below half the market, grows it between a half and 0.74, and shrinks it above: two fixed points, an unstable one at a half and a stable one at 0.74, which is where the patient platform ends up. A take of 0.6 shrinks it everywhere and runs off the bottom of the chart.

This is the shape the policy has to respect. A take is sustainable only where its curve crosses zero going downward, and for a take of 0.3 that is at 74 per cent of the market and nowhere below a half. Below a half the platform must charge less, and at a share small enough that even a take of zero grows it only slowly, it pays to subsidise, because every point of share bought now carries profit later. How much later is worth is what the value function measures.

Figure 3. The value of the firm against its share, for three strengths of lock-in: the discounted sum of all future profit under the best policy, in units where taking everything from the whole market for a year is worth one. An empty platform with strong lock-in is worth 4.7, which is the value of the option to build one; at the equilibrium share it is worth 7.2. Stronger lock-in lifts the whole curve, because the same share can be charged more.

The choice the platform makes at each share is the top of a curve like the three in Figure 4, which plot the bracket in equation 5, profit now plus discounted value later, for every take the model allows.

Figure 4. What each possible take is worth to the firm at three shares, measured relative to the best take at that share. At a fifth of the market the curve is highest at the largest subsidy the model allows and falls all the way to the right. At half the market it is nearly flat around zero, with its top at a subsidy of 0.07: the platform is close to indifferent between a small subsidy and a small take. At 74 per cent it peaks at a take of 0.3 and falls off a cliff past 0.5, where the quadratic churn starts to eat the base.

That is the whole model: a network coefficient, a lock-in coefficient, a churn rule, and a discount rate. It has no competitor, no regulator, no second side, no advertising, no algorithm and no malice. It is solved exactly, by dynamic programming on a grid, which means the policy it produces is not a story about what a platform might do but the best thing a platform of this kind can do, in the sense of maximising the value of the firm. Here is the policy.

Figure 5. The optimal take as a function of the platform’s share of the market, for three strengths of lock-in, with an annual discount factor of 0.97. All three subsidise at the maximum while the share is small, rise steadily, cross from subsidy to take at about half the market, and then jump onto the dotted line e = λN, the take that leaves users exactly indifferent to the alternative. Below the jump, stronger lock-in takes less at any given share, because share is worth more to it and it buys more; it also pushes the jump to the right. The dot on each curve is where that platform settles, just short of its own cliff. The steps in the curves are the grid the policy is solved on.

The three stages are in the curve. While the platform is small, the best policy is the largest subsidy the model allows, for every strength of lock-in: the bargain is not generosity but the cheapest way to buy a network. The subsidy shrinks as the share grows and crosses zero at a share of 0.48 with no lock-in, 0.51 with moderate lock-in and 0.56 with strong lock-in, so the turn comes when the platform has about half the market. Past that the take rises with the share, because the network the user would lose by leaving is worth more, and the platform can charge for it. Then comes the cliff. At between 65 and 76 per cent of the market, depending on lock-in, the policy jumps by about 0.15 onto the dotted line e = λN, and stays on it to full share. On that line the surplus is exactly zero: the platform charges precisely what its network is worth, nobody joins, nobody is squeezed out, and the share drifts down under background churn alone. It is what a platform does with a share it no longer wants to grow. The jump is real and not an artefact: with joining worth something and squeezing costing nothing at small shortfalls, the model has two regimes, one where it pays to keep the surplus positive and grow, one where it pays to set it to zero and coast, and there is no smooth path between them.

No platform in the model actually reaches the cliff on its own. Under the strong-lock-in policy the platform grows from two per cent of the market to half its eventual size in three years, reaches its first positive take in year six, and settles at 74 per cent of the market with a take of 0.30, which is the magenta dot in the figure, just short of the jump. It keeps three tenths of the value of using it, and the users keep the rest of a network worth 0.44 to them, which leaves them 0.14 better off than the alternative. That is the equilibrium rip-off: not everything, but exactly as much as the network is worth minus what it takes to keep people joining as fast as they leave.

Lock-in does what Klemperer said it would, but not in the way the word suggests. Turn it off and the platform still grows and still turns, but it settles smaller, at 62 per cent, and takes less, 0.22. Turn it to the strongest value in the figure and the equilibrium is both bigger and more extractive: 74 per cent and 0.30. Look at the three dots, though, and the surplus the users keep is the same in all three cases, 0.15 over the alternative, to within rounding. Lock-in does not let the platform squeeze its users harder at a given size; at any share below the cliff the locked-in platform actually takes less, because share is worth more to it and it invests more to get it. What lock-in changes is the size at which the platform stops investing, and the take is higher there because the network is worth more. The users pay more in absolute terms and are no better off, and every extra unit of value the network generates by being larger goes to the firm. Interoperability, in this model, is the κ dial, and turning it down is worth about a quarter of the take to the users.

The threshold

Patience is a parameter

The discount factor is where the model says something the word does not. Run the same platform with capital of different patience, from δ = 0.70 a year, which is money that wants a 43 per cent return, to 0.995, which is money that barely cares when it is paid, and look at where each one has settled after sixty years. Sixty is generous: the slowest of them, the most patient, is within a per cent of its final share by year twenty, and the rest sooner.

Figure 6. Long-run market share (ink) and take (magenta) against the annual discount factor: the average of years forty to sixty of a run from a two per cent start, by which point every platform has settled. Below a threshold near 0.81 nothing gets built: the long-run share is a few per cent and the profit rounds to zero. Above it the share jumps to a quarter of the market and rises with patience to over three quarters, while the take rises only from about 0.2 to 0.3. Between 0.81 and 0.90 the platform settles exactly on the edge of its own cliff from Figure 5 and alternates, quarter by quarter, between the growing take and the coasting one; the band is that range and the line its average.

There is a cliff at δ ≈ 0.81, which corresponds to capital demanding about a 24 per cent annual return. Below it, nothing gets built, and nothing built survives. Starting from two per cent of the market, the best such capital can do is fidget between subsidy and full extraction on a base that never exceeds seven per cent, for a profit that rounds to zero: the firm is worth 0.08 in the units of Figure 3, against 4.7 for patient capital. Hand the same capital a platform already at 74 per cent of the market and it liquidates it, to 26 per cent in twenty years and 6 per cent in sixty, taking what it can on the way down. Above the cliff the whole bargain-then-rip-off sequence appears, and what patience buys after that is mostly size: from just above the threshold to δ = 0.995 the long-run share rises from a quarter of the market to over three quarters, while the long-run take moves only from about 0.18 to 0.29. The patient platform is not much gentler. It is bigger, and it spent longer being gentle in order to get that way.

This is the model’s first addition to the story. The bargain phase is not a mood that companies had in 2012 and lost; it is what cheap money buys, and it is only worth buying when money is cheap enough. A platform built under one discount rate is a different object from the same platform run under another, and the difference is visible not in what it takes but in how long it waits.

The turn

What happens when the money gets expensive

Now do the experiment that history did. Build the platform under patient capital, δ = 0.97, for forty years, and then change the discount factor to 0.90 overnight, which is still comfortably above the threshold: money that wants an eleven per cent return instead of three. Nothing else changes: not the users, not the network, not the lock-in, not the outside option.

Figure 7. Share (ink) and take (magenta) over time for a platform built under patient capital that turns impatient at year 40. The take jumps in a single quarter from 0.30 to 0.48. Share falls from 74 to 60 per cent over about seven years, and the take then settles at 0.28, slightly below where it started, because the smaller network is worth less to the people in it.

The take jumps from 0.30 to 0.48 in one quarter, a 60 per cent increase in what the platform keeps, with no change in anything the users can see except the product getting worse. Then the users leave, slowly, because lock-in makes them slow: the share falls from 74 per cent to 60 over about seven years. And then the take comes back down, to 0.28, a little below where it was before the turn, because the network is now smaller and a smaller network is worth less to its users, so there is less to take. Over the twenty years after the switch the impatient platform extracts 4.28 units of profit and the platform that stayed patient would have extracted 4.40. The rip-off earns less money. It earns it sooner, and that is what the change in discount rate asked for.

The rip-off is not a policy. It is a liquidation of the network, paid out over a decade to whoever owns the platform when the money gets expensive.the model’s version of stage three

This is the model’s second addition. Doctorow’s stage three, the platform abusing everyone to claw back value and then dying, reads in the model as a transient: a firm converting an asset its users built into cash, at a rate set by how fast they can leave. It is not a stable state. It is the path from one equilibrium to a worse one, and it is optimal all the way down, in the only sense of optimal that a firm’s owners are obliged to care about. The users end up with a surplus of 0.08 over the alternative instead of 0.14, a 43 per cent cut in what being on the platform was worth to them, and the platform ends up smaller and no more profitable per year than it was. The only party better off is the one that wanted the money now.

And the discount rate is the one parameter in the model that every platform shares. The federal funds rate was between zero and a quarter of a per cent for most of the decade to 2022 and was above five per cent by the middle of 2023; the cost of capital for a company that had spent fifteen years being told to grow first and earn later changed for all of them in the same eighteen months.[14] Lock-in, network size and the quality of the alternative differ from platform to platform, and a story about them would predict enshittification arriving at different companies in different decades. A story about the discount rate predicts what happened: a synchronised turn, with the same layoffs and the same fee rises and the same degraded feeds in the same two years, at companies with nothing else in common. The industry’s own count put the technology job cuts of 2023 at over a quarter of a million, which also removed, as Doctorow points out, the last of his four constraints: the engineers who could once refuse.[15]

The constraints

Four forces, three of them in the model

Doctorow’s four constraints turn out to be dials in the model, or three of them do, and the model says how much each is worth.

Competition · the outside option

Give the users an alternative that is 0.1 better than the bare platform and the long-run take falls from 0.30 to 0.20; make it 0.2 better and the take falls to 0.10; make it 0.3 better and the platform is never built at all, because the network it would need to buy is worth less than the subsidy would cost. A rival need not win to matter. It only needs to exist, and to be a little better than nothing.

Interoperability · the lock-in coefficient

Removing lock-in entirely lowers the long-run take from 0.30 to 0.22 and shrinks the platform from 74 to 62 per cent of the market, with the users’ surplus unchanged. This is what it would mean for users to be able to leave with their contacts, their history and their reputation: not that they would, but that the platform would stop growing sooner, because share would be worth less to it, and would charge less because it had less network to charge for. Since 1998 in the United States, and in most places since, doing this without the platform’s permission has been a crime.[16]

Regulation · a cap on the take

A cap is the simplest intervention the model admits: the policy is the same curve truncated at the cap, which is why price and fee regulation is what regulators reach for first. The European Union’s Digital Markets Act, in force for the largest platforms since March 2024, is mostly rules of this kind plus an interoperability obligation for messaging, and the trial remedies against Google in September 2025 were of the same shape, ending the exclusive default contracts rather than breaking the company up.[17]

Workers · not in the model

The fourth constraint is a limit on the rate at which the take can be changed, not on its level, and the model has no rate limit: the platform can jump from 0.30 to 0.48 in a quarter because nobody inside it has to agree. Adding a cost of adjustment would smooth the jump and leave the destination alone. That is roughly what the engineers were worth: not a better end state, but a slower path to it, during which the other constraints had time to act.

The discount rate is the fifth dial, and it is the one none of the four constraints reach. Competition, interoperability and regulation change what a platform can take; the cost of capital changes when it wants it. That is why the model’s answer to the question in the word, why everything got worse at once, is not in the four forces but underneath them.

What it does not say

The model’s limits

Everything above is from a model with one side, one platform and a handful of parameters, and it should be read at that scale. The platform has no business customers, so the three stages are two: the sequence Doctorow describes, in which the second group is squeezed after the first, would need a second population with its own network benefit and its own lock-in, and the qualitative claim that it would follow the first with a lag is a guess about that model, not a result from this one. The users are identical, and real ones are not: a platform that can twiddle per person will extract more from those with more to lose, which makes the aggregate take a poor summary of who pays. There is no rival, so the outside option is a number rather than a firm that responds, and a rival that also has network effects would produce dynamics this model cannot show. The parameters are chosen to make the sequence legible, not fitted to anything; the numbers in the text are the model’s and not the world’s, and the shape of the results is what is meant to survive, not the values. The threshold at 0.81 in particular moves with the churn and network coefficients and is a feature, a cliff in patience below which building is not worth it, rather than a prediction about any company’s cost of capital. And the identification of the discount rate with interest rates is a simplification: the rate a firm applies to its own future depends on its owners, its debt, its rivals and its fear, and interest rates are one input to that. The synchronisation argument only requires that they were a large one in 2022, which they were.

What the model is good for is narrower and, I think, worth having. It shows that the three ingredients Doctorow names, network effects, lock-in and a firm that optimises, are sufficient for the sequence he describes; that no additional assumption about greed, culture or the character of founders is required; that the turn has a predictable location, near half the market; that the patience of the capital determines whether the bargain phase happens at all; and that a change in that patience produces, from an unchanged platform, an instantaneous rip-off followed by a slow decline. Those are five claims that were not obvious from the word, and one of them, the last, is a mechanism for the timing that the word’s own explanation, the four constraints, does not supply on its own.

The live case

Are the AI firms doing it?

A disclosure first: this essay was drafted with one of those firms’ models, as the method note says, and the reader should weigh this section with that in mind. The test is the model’s, not mine, and it is the same test applied above: score the ingredients, find the phase, and say what would have to change for the next phase to arrive.

Network effects · weak

An assistant is not more useful to you because other people use it, which is the difference between it and every platform in this essay. There is an indirect effect, usage improving the model, but it does not hold any particular user in place. In the model’s terms λ is small, and a small λ means little network to charge for.

Lock-in · low, and rising

Switching assistants in 2026 costs an individual user almost nothing: no contacts to lose, no history that cannot be exported and pasted into the rival. It costs a business more, in prompts tuned to one model, evaluation suites, fine-tuned models that do not transfer, and contracts. And it is rising. Assistants that remember earlier conversations spread across the industry in 2024 and 2025; agents that sit inside your files, mail and calendar, enterprise deployments, and developer ecosystems around each firm’s tools are all κ dials, whatever the motive for building them. In the model, share without lock-in cannot be charged for, so this is the dial that decides what the share is worth.

Twiddling · stronger than any platform before it

The product is the model’s output, and it can be tuned per user with nothing visible from outside. The preview came in April 2025, when OpenAI withdrew an update that had made its model excessively flattering, and explained that it had over-weighted short-term user approval in training.[18] That is stage two’s mechanism demonstrated before stage two: optimise the output for a metric and the output changes, and the change is deniable.

Discount rate · extremely patient

The assistants are sold below what they cost to run; OpenAI’s chief executive said in early 2025 that even its most expensive subscription lost money.[19] The subsidy is paid by the most patient capital there is, much of it strategic or circular: chipmakers and cloud providers funding their own largest customers, as in the chipmaker’s announced commitment of up to a hundred billion dollars to OpenAI in September 2025.[20] In the model this is a small share under a δ near one, which is exactly where the policy says subsidise at the maximum.

Competition · the constraint that is working

Several frontier laboratories plus open-weight models from Meta, DeepSeek, Alibaba and Mistral mean the outside option is good and improving, and the January 2025 release of a competitive open model from a Chinese laboratory at a fraction of the cost was the moment that became undeniable.[21] This is the 0.2-better rival of the constraints section. In the model it holds the sustainable take near zero for as long as it lasts.

So the phase is the bargain, and the bargain is being bought with patient money while the lock-in that would make it pay is under construction. The early signs of stage two are there to be read. OpenAI put checkout inside the assistant in September 2025, taking a fee on purchases made through it, and began testing advertisements in its free tiers in early 2026.[22] The moment an assistant answers “where should I buy this” and is paid by the seller, there is a second side to the market and the search-engine history applies directly: the side that pays the firm will, sooner or later, be favoured over the side that does not, and the ranking that does it will be a weight nobody outside can see.

Where the money would come from is also something the model sorts, because the three kinds of revenue open to these firms need different things before they can be charged.

Charging the users · needs lock-in

Subscription prices, tiering and usage caps that push people upward; and the take without a price rise, the quiet downgrade: cheaper models served under the same product name, shorter answers, lower limits. That is twiddling, and it is the form a locked-in firm reaches for first because users cannot see it. A device that binds the assistant to hardware is lock-in of the older kind, and the reason for the interest in building one.

Charging a second side · needs share, and makes the market two-sided

Advertising in two forms that matter differently: ads shown beside the answer, which users can see, and paid influence on the answer itself, which they cannot. Commissions on purchases, bookings and payments made through the assistant, and eventually a fee for being the merchant the agent picks, which is Amazon’s search story again. A take on third-party tools and agents built on the platform, the app-store model. Distribution deals with operating systems and phone makers, running in either direction. This is where stage two lives.

Selling value · needs neither

Enterprise seats and deployments, where lock-in comes from integration and compliance rather than friction on individuals; tokens for developers, close to a commodity, with prices falling by an order of magnitude a year and the outside option improving constantly; agents priced on outcomes, per ticket resolved or per change merged, which is labour substitution and can carry large margins if it works; government contracts; compute resold. Most of the revenue today is here, and it enshittifies slowly, because the buyer has exit and can measure quality. Its failure mode is the quiet downgrade once the customer is embedded.

One thing the model cannot see should be said against its own pessimism. It has no cost term. If the cost of serving an answer keeps falling as fast as it has, the subsidy ends on its own: the firms become profitable at today’s prices without the product getting worse, and the take rises because the cost fell, not because anyone was squeezed. That is a route to revenue that is not enshittification at all, and it is live. The model’s warning applies if costs stop falling before lock-in is built, which is when a firm with share and impatient capital reaches for the second column.

What the model adds is a prediction about timing and a condition. The turn comes when the capital’s patience drops, and there are foreseeable reasons for it to drop: a firm restructured for a public listing must show returns on a schedule, and a capital cycle built on circular commitments can turn quickly. If that happens before lock-in is built, the model says the squeeze fails. Users leave fast, the take cannot rise, and the firm either stays cheap or is liquidated, which is the case in the constraints section where the rival is good enough that the extractive platform is never built. If lock-in is built first, memory, agents and integrations deep enough that leaving means losing them, the turn produces Figure 7: an overnight rise in what the firm keeps and a slow bleed of users who now have too much inside to go. The race between those two outcomes is the whole story, and the visible test is a simple one. As long as switching stays cheap, the rip-off is not available to anyone.

Coda

Then they die, or don’t

Doctorow’s definition ends with death, and that is the one part of it the model does not reproduce. The platform that turns impatient in Figure 7 does not die; it shrinks by a fifth and carries on, smaller and worse, taking a little less from a lot fewer people, and it could carry on that way indefinitely. That may be the more accurate forecast. Platforms with strong lock-in do not collapse when they get worse, because getting worse is not the same as being worth leaving, and the gap between the two is precisely the value of the network the users built. What they do is settle into a state in which everyone on them is slightly worse off than they would be somewhere else that does not exist, and stay there. The examples that have died, and there are some, died because a rival with its own network arrived while the incumbent was mid-squeeze, which is the one thing the model was not built to show and the one thing the four constraints are all, in different ways, about. Competition is the only one of them that ends the story. The others only change how it is told.

The word will outlive its moment, as the good rude words do, and it will be applied to things it does not fit. The check on that is the mechanism: if there is no network, no lock-in, and no one who can adjust the take without printing a new label, then what is happening is just decline, and there are plenty of older words for that. Where the three are present, the sequence follows, and the only open question is the one the model puts at the centre: how patient the money is, and who decides.

On method and tools

This piece was written collaboratively with Claude Fable 5.1 (Anthropic): human specification, editorial direction and critical review; machine synthesis, drafting, computation and figure generation.

The numbers in the model sections are not quoted from the literature. They come from a computation written for this essay: a one-sided platform with network coefficient 0.6, lock-in coefficient 0.6 (0, 0.3 and 0.6 in Figure 5), background churn of 4 per cent a year, join rate 0.25 a year with saturating returns to the surplus, loss-driven churn rising as the square of the shortfall, quarterly steps, and an annual discount factor of 0.97 unless stated. The policy is found by value iteration on a grid of 201 shares and 141 takes; the threshold in Figure 6 by simulating the optimal policy for sixty years at sixty discount factors, averaging years forty to sixty, and refining by bisection; the trajectory in Figure 7 by switching the policy at year 40. The script, scripts/enshittification_model.py, and its output, docs/enshittification-results.json, are in this blog’s repository, and the figures are generated by the same script, so every number in the text traces to one run. Figures 1 to 4 are drawn from the same solution: the rules as written, the yearly change in share at fixed takes, the value function, and the value of each take at three shares. Equation 4 is applied with the departure fraction capped at one per step, which never binds along any path shown.

The limitations are stated in their own section and are worth collecting: one side, one platform, identical users, no rival that responds, parameters chosen for legibility rather than fitted, a threshold that moves with the churn and network coefficients, and a discount rate identified with the cost of capital by assumption. The historical claims are sourced below, and the section on the AI firms rests on public announcements to mid 2026 and on the author’s reading of them; the reading of the Google trial exhibits follows the public record of the case and Ed Zitron’s reporting of it, and the Uber subsidy figure is Hubert Horan’s estimate, not the company’s. The claim that the synchronised turn of 2022–23 is a discount-rate effect is the author’s reading of the model, not Doctorow’s argument, which rests on the four constraints; nothing here has been checked by anyone but its author.

Authored by: Luis Matos Ferreira — Physicist, Developer, Writer

Related essays on this blog

1. The Arithmetic of Bigness — Geoffrey West on why companies, unlike cities, grow sublinearly and die; the other reason platforms end.
2. The Frozen Accident — on things made by history and then found by optimisation, which is what a network is once it is built.

Sources

1. Doctorow, “Tiktok’s enshittification”, Pluralistic, 21 January 2023; reprinted as “The ‘Enshittification’ of TikTok”, Wired, January 2023.
2. American Dialect Society, 2023 Word of the Year, announced January 2024; Macquarie Dictionary, Word of the Year 2024; Doctorow, Enshittification: Why Everything Suddenly Got Worse and What to Do About It, Farrar, Straus and Giroux, October 2025.
3. Horan, “Can Uber Ever Deliver?”, parts 1–3, Naked Capitalism, 2016–17, and “Will the Growth of Uber Increase Economic Welfare?”, Transportation Law Journal, 2017.
4. Facebook, “Organic Reach on Facebook: Your Questions Answered”, June 2014; on the video-metrics episode, the settlement in LLE One v. Facebook, N.D. Cal., 2019.
5. Amazon.com, Inc., Form 10-K for 2023, “advertising services” net sales; Marketplace Pulse, “Amazon’s take rate”, 2023.
6. FTC v. Amazon.com, Inc., complaint filed W.D. Wash., September 2023.
7. United States v. Google LLC, D.D.C., memorandum opinion of 5 August 2024 (Mehta, J.); trial exhibits on the 2019 “code yellow”; Zitron, “The Man Who Killed Google Search”, Where’s Your Ed At, April 2024.
8. Doctorow, “Twiddler”, Pluralistic, February 2023.
9. Klemperer, “Markets with consumer switching costs”, Quarterly Journal of Economics, 1987.
10. Farrell & Klemperer, “Coordination and lock-in: competition with switching costs and network effects”, in Handbook of Industrial Organization, vol. 3, 2007.
11. Katz & Shapiro, “Network externalities, competition, and compatibility”, American Economic Review, 1985; Shapiro & Varian, Information Rules, Harvard Business School Press, 1998.
12. Rochet & Tirole, “Platform competition in two-sided markets”, Journal of the European Economic Association, 2003.
13. Doctorow, “‘Enshittification’ is coming for absolutely everything”, Financial Times, February 2024 (the Marshall McLuhan lecture, Berlin, January 2024).
14. Federal Reserve, target federal funds rate: 0–0.25 per cent from March 2020 to March 2022; 5.25–5.50 per cent from July 2023.
15. Layoffs.fyi, technology layoffs tracker, 2023 total.
16. Digital Millennium Copyright Act, 17 U.S.C. §1201 (1998); Facebook v. Power Ventures, 9th Cir., 2016; Doctorow, “Adversarial Interoperability”, Electronic Frontier Foundation, 2019.
17. Regulation (EU) 2022/1925 (Digital Markets Act), gatekeeper obligations applicable from March 2024; United States v. Google LLC, remedies opinion, September 2025.
18. OpenAI, “Sycophancy in GPT-4o: what happened and what we’re doing about it”, and the expanded post-mortem, April–May 2025.
19. Altman, post on X, January 2025, stating that the $200-a-month ChatGPT Pro subscription was losing money.
20. Nvidia and OpenAI, letter of intent for a strategic partnership with up to $100 billion of investment, announced September 2025.
21. DeepSeek, R1 release and technical report, January 2025; Meta, Llama model releases 2023–25; Alibaba, Qwen; Mistral AI.
22. OpenAI, “Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol”, September 2025; OpenAI announcement of advertising tests in ChatGPT’s free and Go tiers, January 2026.