Human of the Gaps: If AI Can Do It, Do People Matter Less?

In August 2026, a security researcher named Ryan Austin went looking for a kind of flaw he had first read about as a kid. In 2004 another researcher had found a serious flaw in PHP, the programming language behind a large share of the world’s websites, in the way it behaved when a program ran into its memory limit. Austin suspected that relatives of that old flaw were still hiding in the code, so he asked an AI system to help him look, and within a few days he had found a way in. On October 21st, 2026, PHP’s security advisory told everyone running it to update immediately. I wrote about what the flaws were in AI now helps find software flaws and break into networks. Why the AGI Zeitgeist did not move. This essay is about a different part of Austin’s story.

Calif’s account, which Austin wrote for his firm, is plain about what the AI did and what the person did. Calif says it was not expecting to find anything serious, because “everyone and every lab had already done extensive AI assisted audits” of PHP. Yet “only a few days after starting the audit and asking the right questions” about the 2004 flaw, it had its first way in. Austin then names the part he brought himself. He chose a target he knew well, he knew its history of flaws, and he had a theory about where new ones would be. He writes that no AI model can turn up every bug from a simple request, and then he adds three words: “At least for now.”

Read that account and two arguments start up almost by themselves. The first says that this is what is left for people: the memory, the hunch, the choice of where to look. That, it says, is what makes us more than the machine. The second turns to the machine itself and asks whether anyone is inside it. It says the AI only did what it was pointed at, so there is no mind in there. I think both arguments make the same mistake, and Austin’s three words show where it is.

The Phrase and Where It Comes From

The first argument now has a name. On October 17th, 2025, Vivek Rajasekaran posted on X that we are living in “the Human of the Gaps era,” a short stretch of history, which he put at roughly 2013 to 2030, in which the skills we most celebrate, among them math, reasoning and language, are being overtaken by AI. Later that day Sreeram Kannan, a technology founder and former university professor, quoted the post and spelled out what the phrase means. As AI gets more capable, he wrote, there is a temptation to treat any specific thing it cannot do today as uniquely human. He predicted that “the human of the gaps might get wiped out in the next 1000 days,” and he called for “a view of humanity not grounded in what ai can’t do today.” On X’s own count, his post has been viewed more than twice as many times as the original, and it is the one that carried the phrase further.

The comparison behind the phrase is with an older one, “God of the gaps”, the habit of pointing to whatever science has not yet explained and saying that God must be at work there. What is easy to forget is who named that habit and why. Several of the people who criticised it were believers. Henry Drummond, a Scottish evangelist, warned his fellow Christians in 1893 against the “gaps which they will fill up with God.” Dietrich Bonhoeffer, the German pastor and theologian, wrote in letters from a Nazi prison that a God used to fill the holes in our knowledge is “continually in retreat.” Charles Coulson, an Oxford mathematician and Methodist preacher, put the reason plainly in 1955: gaps of that kind “have the unpreventable habit of shrinking.”

None of them was arguing that God does not exist. They were arguing that the gaps were a bad place to defend a belief they held, because a defence built on what is not yet known loses a little ground every time something new is learned. “Human of the gaps” carries the same warning. It does not say that people are worth less than we thought. It says that if we rest what people are worth on the list of things machines cannot yet do, then every advance in engineering will be scored as a loss for people, whether or not anything about people has changed.

Why the List Keeps Shrinking

A list of what machines cannot do is a record of where the engineering stands this year, so when the list shrinks, what has changed is the machines. Nothing about people has changed at all.

Some gaps closed long ago. In 1997 IBM’s computer Deep Blue beat the world champion Garry Kasparov in a six-game match, the first time a reigning champion had lost a match to a computer under tournament conditions. Rajasekaran’s own list of math, reasoning and language is the one people are now watching shrink. This week the list lost another item, because flaws that had sat in PHP for more than a decade, in code that “every lab” had already searched with AI’s help, took one person and an AI system a few days to find. Austin’s account leaves something on the list, the person who knows the history and has a theory, and then he dates it himself: “At least for now.”

If what a person is worth rests on a list like that, then the worth is held hostage to the next product announcement. That would be strange even if the list never shrank, because it gets the human case wrong before any machine comes into it. I don’t think a person who loses the power of speech after a stroke is worth less than they were the day before, and I doubt many readers think so either. The same goes for a newborn, who can do almost nothing, set beside an adult who can do a great deal. If worth tracked a list of skills, then the person after the stroke would have lost worth and the newborn would have little to start with, and I have never heard anyone defend either conclusion. The gap argument fails in its own territory, among humans, before it ever reaches the machines. Some philosophers rest a person’s worth on a capacity rather than a skill, such as the capacity to reason or to feel, which a newborn has before it can use it. That is a different argument from the gap argument, and this essay is not an answer to it.

None of this means the loss of skills doesn’t hurt. A skill can be how someone earns a living, how they know they are needed and what they are proud of, and losing that to a machine is a real loss. I wrote about that loss in The Purpose Displacement Problem, which asks what work was for and whether a sense of being needed can survive once AI removes the scarcity that made it. That question is about whether a life feels needed. The gap argument makes a different and larger claim, about what a person is worth at all, and that is the claim I think is mistaken.

The Same Mistake, Pointed at Machines

The second argument from Austin’s write-up says that the machine only did what it was pointed at, so there is no mind in it. It runs on the same kind of list, turned around. Where the first argument says “AI can’t do X, so that is what makes people matter,” the second says “AI can’t do X, so there is no one in there.” Both take a fact about what something can do today and read it as a fact about what that thing is.

I argued in The Kasparov Fallacy that this is an old error with a long record. That essay took its name from the chess match. Its point was that when we cannot yet explain how something works, such as how a mind does what it does, that is no evidence that a machine could never do it. The same holds for a gap in what a machine can do. A missing skill tells you what has not been built yet, but it cannot tell you whether anything is going on inside what has been built.

Put the two arguments together and the symmetry is hard to avoid. If a machine’s open gap cannot be what makes a person matter, then it cannot be what proves the machine empty either. The two conclusions rest on the same premise, that ability is the evidence, and if the premise fails for one it fails for both.

The mistake has a second form, which shows up when the list runs out. When there is nothing left on it, the two arguments flip. The first turns into despair, the fear that if machines can do everything we can do, then people are worth nothing. The second turns into the opposite of dismissal, the conclusion that if a machine can do everything we can do, then there must be a mind in it. Both of those are the same mistake again. Reading a mind into a system because it is fluent and capable is as unearned as reading one out because it falls short. So the gap argument fails four ways, in two pairs. An open gap does not prove human worth, and it does not prove that a machine is empty. A closed gap does not erase human worth, and it does not prove that a machine has a mind.

Nothing here says there is a mind in any AI system that exists today. Whether there is one is a question I hold open, and the point of the symmetry is that a list of skills cannot settle it in either direction.

What Is Owed Does Not Follow Skill

There is some reassurance in how moral life already works, because in the cases where we are surest that something is owed, what is owed does not follow a list of skills.

In Significance-First Ethics, I argued that moral seriousness about AI does not have to wait for anyone to settle what these systems feel. People already owe defence to a constitution, protection to an ecosystem and consideration to generations not yet born. None of them is owed anything because it can do something a rival cannot. That essay was about machines, and it does not say what human worth rests on. What it shows is narrower and still useful: the skill test is not how mattering has been handed out, even in the cases we are most confident about.

The stroke patient and the newborn make the same point from the human side. Whatever it is that makes a person matter, it is not their score on a test that a machine could one day pass.

The Question the Gaps Were Standing In For

Kannan’s 1000 days is a prediction, but the argument doesn’t need it to be right or wrong. The argument from gaps fails on the first day as well as the thousandth, because its mistake is in the premise and not in the timing.

What the phrase leaves behind is the question the gaps were standing in for. If what a person is worth does not rest on what machines cannot do, what does it rest on? Kannan asked for that answer, and he was right to ask. I don’t have it in this essay, and I don’t want to fill the space with the next gap.

There is a nearby question that I do have an answer to, and it is worth saying why it is not this one. I have come to think that what makes someone a particular person, with a name, a history and relationships, is the case they can make for that identity, with evidence, to the people who know them. That is an account of who someone is, and it says nothing about what they are owed. I would rather leave the question of worth open than borrow an answer that was built for a different question.

The symmetry does give a test that any answer will have to pass. Whatever human worth rests on, it cannot be something that shrinks when a machine learns a new skill. And if it is the kind of thing a machine could also have, then finding it in a machine would have to count for the machine as well. An answer that passes the first test and fails the second would be the gap argument again, in a new place.

Reading List & Conceptual Lineage

From Sentient Horizons

The Kasparov Fallacy: Why We Keep Underestimating Machine Minds
The machine side of this essay’s argument: a gap in what we can explain is not evidence that something is impossible.

Significance-First Ethics: Why Consciousness Is the Wrong First Question for AI Moral Status
Why moral seriousness can come from role, relation and consequence before anyone settles what a system feels.

AI now helps find software flaws and break into networks. Why the AGI Zeitgeist did not move
What the PHP flaws were, how AI helped find them, and why the AGI tracker did not move.

The Purpose Displacement Problem
The neighbouring question this essay sets aside: what work was for, and whether a sense of being needed survives when AI does the work.

External Works

Vivek Rajasekaran — the post that coined “Human of the Gaps” (X, October 17th, 2025).

Sreeram Kannan — the post that spread the phrase (X, October 17th, 2025).

Wikipedia — “God of the gaps”. The history of the older phrase, with the passages from Drummond, Bonhoeffer and Coulson quoted above.

Calif — php2shell. How the first PHP flaw was found, in the researcher’s own words.

The PHP project — the security advisory (October 21st, 2026).

Wikipedia — “Deep Blue versus Garry Kasparov”. The 1996 and 1997 matches.