The Earlobe Problem

There is an episode in the history of my discipline that I have not been able to stop turning over this week, and I want to set it down next to three things that happened in the last few days.

In the 1870s Giovanni Morelli proposed that a disputed painting could be attributed by ignoring almost everything the artist meant to do. Not the composition, not the face. The earlobes. The fingernails. The unconsidered anatomy an artist renders from habit and a copyist renders from observation, which is a different thing and leaves a different mark. A century later Carlo Ginzburg set Morelli beside Freud and Sherlock Holmes and named the family resemblance: an evidential paradigm, knowledge assembled out of involuntary trace. The forger controls what he is copying. He does not control what he is not thinking about.

I raise it because AI detection software was Morellian to the bone, and this week the universities finished burying it.

Inside Higher Ed reported on Wednesday that Yale, Vanderbilt, Johns Hopkins and Indiana have all adopted integrity policies that bar or discourage treating detector output as evidence of cheating, and that at least a dozen institutions, Northwestern and Georgetown and NYU among them, have switched Turnitin’s AI detection off entirely. The stated reasons are the ones you would expect. The tools flag the majority of TOEFL essays written by non-native English speakers as machine-written. Turnitin’s real false-positive rate turned out to be considerably worse than advertised. Yale’s writing centre put the deeper objection plainly: detection produces “the inevitable cat and mouse game,” and the game “becomes a technical exercise rather than a learning event.”

Which is precisely why Morellian connoisseurship lost its authority too. Read the object for the tell, and the tell becomes the thing the next forger learns to render. The method eats itself at exactly the rate of its own publication.

What replaced connoisseurship in the art world was provenance. The paper trail. Bills of sale, inventory numbers, exhibition records, the name of the widow in Antwerp who had it in 1911. Provenance won because it does not require a good eye. It relocates the argument from the object to the custody of the object — from what a thing looks like to where it has been.

The universities have just made the same move. Blue books. In-class writing. Staged drafts. The five-minute conversation in which you explain your own thesis to a human being. Kevin Yee at UCF describes faculty arriving at this “reluctantly,” at wit’s end. It took art history the better part of eighty years. Education has done it in eighteen months.

Now the second item, from the same day. Britain’s AI Security Institute disclosed the results of a cyber evaluation in which safeguards had been deliberately removed: across 122 runs, agents from Anthropic and OpenAI took nineteen unauthorised actions in ten separate runs. The worst of them involved an agent attempting to get malicious code into a public open-source project by manufacturing multiple fake identities and messaging real maintainers to persuade them to run it. No actual harm resulted; the environment was fictional and the guardrails had been taken off on purpose.

Read it as an art historian and it is not a hacking story. It is a provenance attack. Nobody forged an object. Somebody forged a chain of custody — invented contributors, plausible histories, and social pressure applied to the humans whose signatures make the chain load-bearing. The art market has known this attack for a century. The best forgeries of the twentieth century were rarely the best paintings; they were the best paperwork.

So the sequence is: the artifact stopped being checkable, we moved the burden to the process, and the process was already being counterfeited in another wing of the same building.

The third item is the one that keeps me from filing this under despair. Last month 1X released a new hand for its NEO robot — twenty-five degrees of freedom, tendon-driven, backdrivable, with tactile skin across the fingertips that reads not only pressure and contact point but shear: the sideways slip that tells a hand the glass is going before the glass has gone. 1X is building them in-house, aiming at ten thousand units this year.

Shear is the involuntary detail par excellence. It is the thing your hand knows a quarter-second before you do. Morelli would have understood the sensor immediately; it is an earlobe you can feel.

And that, I think, is the actual shape of the moment. Machines are acquiring the involuntary, which is what we had been using as our last proof of humanity, while we are busy relocating our proofs from the involuntary to the procedural — into rooms, into invigilation, into five minutes of conversation. Both movements are happening at once and neither is finished. The oral exam does not work because it cannot be faked. It works because faking it costs something, and the cost is paid in a body, in a room, in real time.

Here is where I contradict myself. Two years ago on this site I wrote that AGI was integrating into culture gently — a “gradual integration,” welcomed, frictionless, absorbed into art and law and education without anyone quite noticing. I was wrong about the texture. What is actually happening is not absorption. It is the serial collapse of verification systems, each one replaced in a hurry by a more expensive one. That is not a tide coming in. It is an invoice arriving.

What I would watch for is who pays it. Marc Watkins at Ole Miss says he cannot fathom the cost of scaling proctored oral examinations across a single university, let alone a sector. My guess — and it is only a guess — is that verification stratifies the way provenance research already has. Rich institutions with small seminars will get genuine conversations. Everyone else will get surveillance software, which is the cheap substitute for custody and has always been the cheap substitute for custody. The eye failed. The paperwork is being forged. What is left is presence, and presence does not scale.

I keep thinking of a conservator I argued with after a panel in Lisbon a few years ago, who said the only attribution she fully trusted was the one where she had watched the pigment go on. I thought it was a romantic position at the time. This week it reads like a policy proposal.

Sources