I work on alignment at Anthropic, and last month I put $3,000 on my corporate card to hire Aella as an escort and have sex with her. Finance rejected the expense on the grounds that having sex with an internet-famous rationalist is not obviously necessary to my job. This is a surprisingly difficult position to reconcile with the intellectual framework Anthropic otherwise pays me to apply. We spend enormous amounts of money trying to understand the difference between what humans say they want and what they actually want. Aella has spent years studying exactly that question through sex research, surveys, OnlyFans, escorting, and unusually direct observation of human behavior. She also comes out of the rationalist milieu that produced a remarkable amount of the original thinking about AI alignment and existential risk. In any other context, access to someone sitting at the intersection of preference research, rationalism, sex work, and the social network around people like Eliezer Yudkowsky would be described as domain expertise.
The EA case is stronger. Human sexuality is important, empirically neglected, and unusually tractable here because one of its best-known independent researchers literally publishes her data and accepts money to sleep with people. If transformative AI eventually mediates dating, companionship, sex, reproduction, and family formation for billions of humans, then a marginal improvement in our understanding of revealed preference has potentially enormous downstream value. Multiply a very small epistemic gain by billions of users, decades or centuries of deployment, and the expected number of future people whose intimate lives may be shaped by AI systems, and $3,000 rapidly disappears into the rounding error. I realize this sounds convenient coming from someone whose girlfriend left him in May. Convenient propositions can still be true.
Finance's main objection is that I could have simply interviewed Aella. Of course I could have interviewed her. I have listened to interviews with her. I have read her writing. The whole point was to acquire information unavailable through an interview.
This should be familiar to anyone working in alignment. If you want to know how a model behaves under incentives, you do not ask the model to describe its behavior under incentives. You create the incentive and observe what happens. Sex work presents an unusually information-dense case of preference elicitation: the client may be embarrassed, indirect, confused about what he wants, or trying to perform an identity for the other person, and the worker has a direct financial incentive to infer the underlying preference anyway. Aella herself has publicly discussed this aspect of sex work.
Finance asked why I had to be the client. Because otherwise I would once again be collecting somebody else's report of the interaction. Participant observation exists for a reason. Nobody tells an anthropologist studying a religious ceremony that he could have saved money by reading the Wikipedia page.
There was also no way to reproduce the relevant incentive structure by announcing, “Hello, I am an Anthropic alignment researcher conducting a study.” At that point I am observing how Aella behaves around an Anthropic alignment researcher conducting a study. Hiring her as an escort creates the actual economic relationship. The fact that I had also been lonely, sleeping badly, and checking my ex's Instagram more than was probably healthy does not alter the methodological point.
If anything, my personal circumstances improved ecological validity. AI systems are not going to mediate intimacy exclusively for well-adjusted married people with secure attachment styles. They will be used by lonely people, awkward people, recently dumped people, people who were not popular in high school and subsequently developed elaborate theories about why popularity was mostly a signaling equilibrium anyway. This population deserves to be represented in the data.
I also expensed a year of Aella's subscription content. Finance has been less focused on this charge, though analytically it is part of the same research program. A single encounter is cross-sectional. A subscription provides longitudinal exposure to pricing, audience segmentation, parasocial attachment, preference discovery, and the construction of intimacy at scale. We pay vendors much more money for much worse dashboards.
The strangest part of this dispute is that everyone at Anthropic already accepts the premises individually. Revealed preferences matter. Direct evaluations beat self-report. Neglected empirical domains deserve investigation. Rationalists should not allow social disgust to terminate an argument. Tiny improvements can dominate expected-value calculations when the relevant future population is sufficiently large.
Apparently the objection begins only when those premises imply that Anthropic should reimburse me for having sex with Aella. That is not a rebuttal. That is the conclusion.
I submitted the receipt again.