2. Interface Collapse

The previous essay established that AI plus verification has crossed cost-quality parity for a large class of professional cognitive tasks. The most respectable continuity defence accepts this and answers it with a structural distinction. AI can do tasks, the argument runs, but tasks are not jobs. Jobs are bundles of judgement, relationships, context, and physical-world friction. AI can write a paragraph or fill a spreadsheet, but it cannot run a legal practice, manage a team, close a deal, or sit in a meeting. So jobs, the argument concludes, are safe even where individual tasks are not.

This was the most thoughtful version of the continuity story. It came from labour economics, productivity research, and business-school augmentation theory. It held up for a while because it was true in the same way a gate is true. It was true until enough of it came off the hinges.

This essay closes the gate. The thesis does not need AI to replace whole occupations in one step. It never did. It needs only enough economically valuable task-units to cross the cost and quality threshold that firms start rebuilding workflows around machines instead of people. A job is not a sacred object. It is a bundle of tasks. Once the bundle becomes decomposable, the worker becomes decomposable with it.

What makes the bundle decomposable is interface collapse.

The hidden moat

A great deal of white-collar labour was historically protected by software fragmentation. The worker did not only think. The worker moved. They moved between the inbox and the spreadsheet, the CRM and the browser, the calendar and the document editor, the dashboard and the ticketing system, the codebase and the internal wiki, the procurement portal and the finance system, the project tracker and the customer database. Humans survived because they were the integration layer between systems that did not talk to each other cleanly.

This was a real moat. It was also invisible to most people who held it, because the labour of stitching looked like the labour of thinking. A surprising fraction of so-called knowledge work was not pure judgement. It was moving information between semi-compatible systems while maintaining enough context not to break the process. The worker absorbed the friction that the software stack produced, and the wage compensated the absorption.

The moat is now collapsing. Once an AI can see screens, click, type, browse, call tools, manipulate files, and persist across a multi-step workflow, the boundary between producing output and doing work begins to dissolve. The model is no longer a text generator on the other side of a keyboard. It is an interface operator.

The API automates where systems cooperate. The GUI automates where they do not. Few ordinary enterprise software environments can rely on fragmentation as a durable moat once both routes exist. Some systems are air-gapped, permissions-heavy, bespoke, legally restricted, or physically entangled, and these will resist for longer. But for ordinary white-collar workflow friction, fragmentation has stopped being a general defence.

The benchmark trajectory

The interface capability is being tracked publicly. OSWorld-Verified is the standard benchmark for desktop computer use, measuring whether a model can navigate real software through screenshots, keyboard, and mouse actions to complete multi-step tasks across applications. The trajectory across recent frontier releases is steep. GPT-5.2 reached 47.3 percent on OSWorld-Verified. GPT-5.4 cleared 75.0 percent on the same benchmark while pushing GDPval to 83.0 percent.^[1] GPT-5.5 then hit 78.7 percent on OSWorld-Verified, 84.9 percent on GDPval, 84.4 percent on BrowseComp, and 98.0 percent on Tau2-bench Telecom, which tests complex multi-turn customer-service workflows.^[2]

On this benchmark, frontier models now exceed the reported human baseline of 72.4 percent. The trajectory crosses human capability and continues upward.

These numbers describe a specific capability shift. The model is no longer producing isolated outputs that a human pastes into software. It is operating the software layer itself, across applications, over multiple steps, with persistence and tool use. The work that used to require a human to move between systems now requires only a human to authorise the start and verify the end.

The frontier has moved from output generation to work continuation. GPT-5.5 is not described by OpenAI as a model that answers prompts. It is described as a model that plans, uses tools, checks its work, navigates ambiguity, and keeps going. That is the cognitive loop previously sold as the human advantage.

The phase change between model generations matters here. GPT-5.2 demonstrated expert-level professional deliverables at machine cost. GPT-5.5 demonstrates persistent computer-use and workflow execution. The first is unit cost dominance at the deliverable layer. The second is interface collapse in operation. Both are documented in the same vendor’s launch notes within four months of each other.

OpenAI’s own description of GPT-5.5 captures the propagation directly. The model is described as able to research, analyse, build documents, operate software, move across tools, check its work, and push through messy multi-part tasks.^[3] That is not a description of a text generator. It is a description of a workflow operator. The taxonomy of what counts as AI capability has shifted from output generation to process execution within the span of two model releases.

Deploying CEOs describe the same shift in their own organisations. Cursor’s CEO reports that GPT-5.5 “stays on task for significantly longer without stopping early, which matters most for the complex, long-running work our users delegate.”^[4] MagicPath’s CEO reports merging branches with hundreds of frontend and refactor changes into a substantially-changed main branch in twenty minutes. The pattern is consistent: agentic delegation, long-horizon execution, work continuation rather than output assistance.

The reliability comparison

The standard objection to interface-level deployment is that benchmark scores do not translate to production reliability. The most common form of the objection treats the benchmark percentage as a per-step success rate and compounds it across multi-step workflows. A 78 percent per-step rate, exponentiated across a 10-step workflow, would yield roughly 8 percent end-to-end success.

The compounding-error objection misreads the benchmark. OSWorld-Verified is not reporting per-step accuracy. It is reporting end-to-end task success on multi-step computer-use tasks involving real applications, file operations, and workflows that span multiple programs. Treating the score as a per-step probability and exponentiating it double-counts the compounding problem. The benchmark already contains the compounding problem. The 78.7 percent figure is the rate at which the model completes the entire task successfully, not the rate at which any individual step succeeds. GPT-5.4 reaching 75.0 percent on the same benchmark already exceeded the published human baseline of 72.4 percent. GPT-5.5 at 78.7 percent extends the gap.

Production deployment adds distribution shift, latency, permissions, audit trails, retry logic, escalation paths, and verifier cost on top of the benchmark performance. Those are real costs. They enter the unit cost dominance equation as deployment costs. They do not refute the interface capability the benchmark measures.

The production comparison is therefore not raw AI versus ideal human. It is AI plus retries, guardrails, escalation, and verification against human workers plus training, supervision, review, error correction, management, and quality control. The relevant question is not whether AI is perfect. It is whether the full AI stack reaches equivalent reliability at lower cost than the full human stack. For an increasing fraction of professional digital workflows, the answer is yes.

The four layers

The clean way to think about the propagation is as a cascade, not a single event.

The first layer is task-level unit cost dominance. AI plus thin human oversight produces professional task outputs at equal or better quality, faster speed, and lower marginal cost. This has crossed for a large and growing set of well-specified cognitive deliverables. The previous essay established this layer.

The second layer is interface and workflow dominance. AI operates through the same software environments where work happens, and handles the stitching between them. This is rapidly crossing, and the public benchmarks now track it directly. This essay establishes this layer.

The third layer is job-level dominance. Whole roles become economically unnecessary as enough human task volume is stripped out. This is partial and uneven across occupations, and the thesis does not require it. Some jobs will retain residual human content for years. The thesis works whether or not any particular job is fully eliminated, because what matters is the aggregate displacement, not any specific role’s survival.

The fourth layer is labour-market dominance, which is where the thesis actually lives. Wage labour stops being the mass route to economic agency. This has not fully arrived, but the pathway from the first two layers to the fourth is now specified and visible.

Critics prefer to argue at the third layer because residual human tasks are easy to find inside any existing job. But that is not where the thesis is defended, and never was. The argument runs from the first two layers directly to the fourth, with the third layer treated as partial and contingent rather than as a necessary intermediate step. Workflow recomposition can suppress hiring, break training ladders, and collapse mass labour absorption without any single occupation being formally eliminated. This is what makes the displacement quiet.

No scream, just non-absorption

Mass displacement will not first look like mass unemployment. It will look like non-absorption.

Fewer entry-level roles. Fewer junior ladders. Fewer graduate pathways. Fewer promotions. More contractors. More review roles. Productivity gains that do not reach wages. Incumbents held in place while new entrants fail to launch. A headline unemployment rate that looks fine while the system underneath it has stopped reproducing itself.

The tells to monitor are specific. Entry-level hiring decline in AI-exposed fields. Junior-to-senior ratio compression. Weak graduate absorption despite stable aggregate employment. Wage stagnation or wage compression in exposed cognitive work. Productivity gains not passed to labour. Contractorisation and project-based substitution. Delayed retirement of incumbents combined with fewer new entrants. Expansion of review and validation roles that do not scale into careers. Collapse of training ladders because AI now does the junior work from which expertise used to grow. Rising dependence on capital income, transfers, rents, or platform ownership rather than wages.

Stanford’s Digital Economy Lab has reported the first expected signal. Brynjolfsson, Chandar, and Chen find that early-career workers aged 22 to 25 in the most AI-exposed occupations have experienced a 16 percent relative employment decline after controlling for firm-level shocks, while less-exposed fields and more experienced workers have remained stable or grown.^[5] The adjustment is more visible in employment than compensation, and it is concentrated where AI is more automative rather than augmentative. That is exactly what the thesis predicts. Not “everyone gets fired at once.” Entry ladders narrow first. Training ladders break first. Junior absorption weakens first. Aggregate employment can look fine while the reproduction mechanism fails underneath.

Aggregate calm is not a refutation if the No-Scream indicators are moving. That is the whole point of calling it the No-Scream Principle. The displacement does not arrive as a discrete event. It arrives as the quiet disappearance of the entry path, and by the time it becomes politically visible through aggregate statistics, the propagation has already done its work.

The augmentation argument is a stage error

The standard reply to all of this is that humans will work with AI. That sentence is doing more work than it can carry. There are three kinds of complementarity, and they are not the same thing.

Genuine complementarity means AI raises the marginal value of human labour. The worker becomes more productive, more valuable, and captures some of the gain through wages, bargaining power, or advancement. This is real, and it exists. It is not guaranteed to persist, because the same model that makes the worker more valuable today may absorb the worker’s role tomorrow.

Transitional complementarity means humans supervise, correct, validate, integrate, and absorb responsibility while the AI improves. This is the phase most commonly mistaken for the future of work. It is unstable by design. The human role gets thinner as the system gets better, because the human role exists precisely to compensate for the system’s current limitations. As those limitations are addressed, the role shrinks toward zero.

Theatrical complementarity means humans remain for trust, liability, regulation, customer comfort, ritual legitimacy, or institutional optics. The human is still in the room, but no longer economically central. This is not augmentation. It is managed displacement wearing augmentation’s clothes. The role exists because the institution requires a human to be present, not because the work requires a human to be done.

The augmentation narrative points at the second stage and calls it the destination. The thesis argues it is the corridor. Once that corridor is named, “humans will work with AI” stops being a rebuttal and becomes a question about which complementarity is meant. Which complementarity, at what margin, for how long, under what competitive pressure? The honest answer is that genuine complementarity must be demonstrated, not assumed. Transitional complementarity is unstable by construction. Theatrical complementarity is displacement under another name.

Why the propagation is not stoppable inside firms

The interface collapse layer is not something firms choose to deploy as a discrete decision. It is the cumulative effect of thousands of small workflow changes, each defensible on its own terms.

A firm does not announce that it is rebuilding its operations around AI agents. It deploys a model in customer service, then expands the deployment, then connects the model to the ticketing system, then connects the ticketing system to the CRM, then automates the escalation path, then reduces the customer service headcount as the deployed system handles more cases. Each step is incremental. Each step is justified by quarterly cost-benefit analysis. The aggregate effect is workflow recomposition, but no individual decision looks like workflow recomposition. The decisions look like ordinary process improvement.

This is why interface collapse is hard to govern from inside the firm. The decisions are too small to require board approval. The decisions are too distributed to be coordinated as a single intervention. The decisions are made by the line managers responsible for the budget, who face direct competitive pressure to reduce costs and improve throughput. The CEO who wants to slow the propagation cannot do so without overriding hundreds of small operational decisions, each of which is locally rational. The CEO who tries to slow the propagation produces a firm with higher costs than competitors who do not.

The next essay asks why no actor can restrain this propagation once it begins. The answer is the Multiplayer Prisoner’s Dilemma.

Notes

1. OpenAI, “Introducing GPT-5.4.” https://openai.com/index/introducing-gpt-5-4/
2. OpenAI, “Introducing GPT-5.5.” https://openai.com/index/introducing-gpt-5-5/
3. OpenAI, “Introducing GPT-5.5.” https://openai.com/index/introducing-gpt-5-5/
4. OpenAI, “Introducing GPT-5.5.” https://openai.com/index/introducing-gpt-5-5/
5. Brynjolfsson, Chandar, and Chen, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” Stanford Digital Economy Lab. https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/