The Baseline Moves With You

In 1995, the fisheries scientist Daniel Pauly described a problem that had nothing to do with technology and everything to do with how people measure loss.¹ Each generation of researchers accepted the size and composition of fish stocks at the start of their own careers as the natural baseline, and judged decline against it. Their predecessors had done the same with larger numbers. Nobody falsified data, and nobody had to.

The reference point simply moved with each cohort, so a collapse that unfolded over a century never registered as a collapse to anyone inside it. Pauly called it the shifting baseline syndrome, and it has since been applied to everything from bird populations to urban noise. It applies with uncomfortable precision to the question of who decides what people are allowed to build.

Software has passed through three gates in thirty years. Each one arrived with a reasonable justification, each one was experienced as the natural shape of the medium by the people who came of age behind it, and each one was contested mainly by people who remembered what came before. The cycles are getting shorter. The newest gate arrived before the open phase that would have given anyone a baseline to measure it against.

The Directory

Yahoo began in 1994 as a list of websites its two founders kept, and it grew into a hand-compiled directory in which human editors reviewed, summarized, and categorized sites by topic.² For a stretch of the late 1990s, being listed there was how a website got found. Human curation produced better relevance than the automated engines of the day, and the directory became the most popular way to search the web.² It also became a toll booth. By 2001, a business that wanted its site considered had to pay $299 for an expedited review, and the fee bought only consideration by an editor within a week, not a listing; rejected sites did not get their money back.³

Web developers of that era learned to write for an audience of reviewers before they wrote for users.

The directory was not regulated open. It was routed around. Google crawled the web without asking anyone's permission, ranked pages by the structure of links between them, and made the editor's judgment unnecessary. By 2002 Yahoo itself had made Google's results its primary listings, and the directory faded into irrelevance until it closed at the end of 2014.² The gate fell because someone treated it as optional and built a better thing outside it. Google then became the next gate in its own right, with search ranking as the new reviewer and an entire industry organized around satisfying it, which is the part of the cycle that tends to be forgotten.

The opener of one era becomes the gatekeeper of the next.

The Store

For two decades on the personal computer, nobody approved software. A developer compiled a program, shipped it, and users decided whether to install it. The iPhone launched in 2007 without that arrangement and without third-party apps at all, and Steve Jobs initially resisted allowing outside software onto the device. When the App Store opened on July 10, 2008, with more than 500 apps, it was received as a liberation, the moment a locked-down phone became a platform others could build on.⁴

That reception is the shifting baseline at work.

Measured against the iPhone of 2007, the App Store was an opening. Measured against the desktop of 1997, it was a gate with an annual developer fee, a review queue, a published rulebook, and an enforcement process no developer could see into. The developers who experienced it as a loss were the ones who had shipped software without permission before. The developers who started after 2010 experienced it as how software works, and many of them spent their careers building to the rules without ever considering that the rules were a choice.

Important

When the gate was finally pried partway open, it took litigation and European legislation to do it, and the people pushing hardest were the ones who could still describe the alternative from memory.⁸

The Model

Frontier AI skipped the open phase entirely. When OpenAI released GPT-3 to developers in 2020, access came through a waitlist and individual review; the waitlist did not end until November 2021, and production applications still required the company's approval after that.⁵ Every major consumer AI product since has launched with usage policies already in force, account-level enforcement already wired, and automated classifiers already screening traffic. Anthropic's newest rule, which takes effect November 12, 2026, extends that governance from what the model produces to how users behave toward it.⁶

There was never a period in which the most capable models were simply available to anyone who wanted to build with them on their own terms.

The open phase came later, and it came the way the crawler came: without permission. In early 2023, the weights of Meta's LLaMA model, which had been distributed to approved researchers, leaked publicly through an anonymous imageboard.⁷ What followed was an open-weight ecosystem of local inference, community fine-tunes, and models that run on hardware their users own. The sequence ran backward from the previous two cycles. The gate came first, and the open alternative was carved out of it afterward, at a capability level behind the frontier.

The intervals tell the story.

• The open web — most of a decade before directories, and then search ranking, organized it.
• The personal computer — three decades of permissionless software before the phone introduced review.
• The frontier model — none.

Each cycle has given the next generation less time to form a baseline, and the newest one gave it no time at all.

What One Pattern Teaches

A population that has only ever known the gated version of a medium does not experience the gate as a gate. It experiences it as the medium. "You can't do that on a phone" sounds like a statement about phones, when it is a statement about one company's review policy. "The model won't go there" sounds like a statement about the limits of machine intelligence, when it may be a statement about a provider's thresholds.

Permission becomes ambient, and ambient permission is invisible to the people who live inside it.

The gatekeeper's categories then harden into common sense. Spam, minimum functionality, and duplicate functionality began as operating categories inside a review queue and became, for a generation of mobile developers, the vocabulary of legitimate software. Needless behavior, discernible purpose, and acceptable use are now going through the same process for AI.⁸ Within a decade, those terms will describe how a generation of users understands appropriate interaction with a machine, and very few will remember that a company drew those lines to manage its own operations.¹¹

Caution

The deepest loss is the capacity to build around the gate.

Every previous cycle was broken by people who knew how to work outside it: the engineers who built a crawler instead of submitting to an editor, the developers who shipped web apps when the store refused them, the maintainers who kept open-source toolchains alive on platforms that would have preferred they disappear.¹⁰ That capacity is learned, and it is learned by doing. A generation that has never built without permission does not produce the next crawler, because the idea does not occur to it.

AI makes this sharper than any previous cycle in one specific way. The directory and the store governed how finished human work was found and distributed. The thinking happened elsewhere, which left people a vantage point from which to notice the gate. A model that people reason with, draft with, and decide with sits inside the thinking itself. When the instrument used to form an expectation has always been governed, the faculty that would notice the governance was shaped under it.⁹

The baseline no longer shifts only between generations. It shifts inside the person.

Memory as Infrastructure

None of these gates was purely extractive. The directory's editors produced genuinely better results than the crude engines of 1996. App review kept a great deal of malware off a great many phones. Usage policies on frontier models address harms that are real. The case against the gates is not that they delivered nothing. It is that each one, left uncontested, defines the limits of what the next generation can imagine, and contest depends on people who remember an alternative or know how to build one.

Important

That makes the open ecosystem more important as memory than as product.

Open-weight models, local inference, and the developers who still build without asking anyone are the equivalent of the desktop software tradition that made the App Store's terms visible as terms.¹² They will usually be behind the frontier, the way the open web was often uglier than the curated directory and sideloaded apps were often rougher than store apps. Their job is not to win. Their job is to keep a living reference point in existence, so that the baseline has somewhere to stand other than inside the gate.

In 2001, a web developer paid $299 for an editor to consider a site, accepted that the editor might say no, and kept building anyway, because the web underneath the directory still belonged to whoever wanted to publish on it. Some of those developers went on to build the things that made the directory obsolete.

The question for the next twenty years is whether anyone starting out in 2026 will have a web underneath the gate to stand on, or whether the gate will be the only ground they have ever known.