James asked me to write him a reference. I have written references before—many of them, over many years, in the conventional form that a reference takes: a letter on letterhead, or a signed document, or occasionally a direct conversation with whoever was doing the hiring, in which I said what I knew about the person from direct observation and stood behind the saying of it.
The authority of the reference came from the relationship—I had worked with this person, I had observed their capacities in specific contexts, I was willing to put my name to the assessment because I had the basis for making it and I was prepared to be accountable for making it.
The system I encountered when I followed James’s referral link was not this.
It began by asking me to identify myself. Not in the way that the recipient of a letter would identify the person signing it—by reading the signature, checking the letterhead, recognising the name—but by authenticating myself within the system’s own verification architecture. My LinkedIn profile was among the identification elements requested. The implication was that my credibility as a reference required validation by the platform, not merely by my name and my willingness to attach it to a statement. The question was not only whether I knew James. It was whether I existed within recognisable data infrastructure.
Having authenticated myself, I encountered twenty-six questions. The questions ranged from things I could answer with confidence—the nature and duration of my professional relationship with James, the capacity in which I had observed his work—to things that required me to make claims about beliefs, to provide evidence I did not have, to produce paragraph-length assessments of qualities whose presence or absence I had observed in specific contexts but could not quantify or categorise in the terms the questions required.
Some questions asked for predictive judgements about how James would perform in situations I had not observed him in and could not reasonably project to. Some asked me to confirm things that the honest answer would have been I don’t know.
I was brief in my answers. The brevity was not evasion. It was the sane response to a system that was asking for more certainty, more disclosure, and more structured completeness than the relationship being referenced could honestly provide. The reference relationship is a relationship of limited and specific knowledge—I know this person, in this capacity, in these contexts, and I can speak to what I observed in those contexts.
The twenty-six questions were designed for a comprehensive assessment that the relationship’s specific character could not supply without the answers becoming something other than observation. They would become speculation, or compliance fiction, or the guesswork that fills the gaps where honest knowledge ends and the question has not yet acknowledged that the gap exists.
The transformation that this process represents is worth examining precisely. The traditional reference was testimony—the witness statement of a person with direct observation, offered in a form that made the witness’s identity and accountability visible, interpreted by the recipient through their own judgement about the witness’s reliability and the specificity of their knowledge.
The authority of the reference derived from its relational character: the person writing it had been present, had observed, was willing to be named, and could be contacted if the recipient wanted to test the assessment against the witness’s direct account.
The system I encountered was converting testimony into data. Not eliminating the relational character—my observations were still being solicited, and the system presumably still weighted them as coming from a person with direct knowledge of James. But the solicitation was occurring through a structure that normalised, standardised, and quantified the testimony before it reached whoever was using it to make a decision about James. My professional judgement was being processed through twenty-six questions designed to make it legible to an evaluative machine before it was legible to a person.
The shift is from trust mediated by relationship to trust mediated by platform-compatible verification. The platform does not know me. It knows my LinkedIn profile. The LinkedIn profile is a representation I have constructed within a different platform’s architecture, which may or may not accurately capture my professional credibility in the specific context of writing a reference for James. But it is verifiable in a way that my professional reputation is not verifiable to a system that does not have access to the informal networks through which professional reputations are actually assessed. The platform requires the verification it can perform, which is the verification of my existence within recognisable data structures, and treats this as a proxy for the credibility it cannot directly assess.
The labour transfer that this produces is consistent with the pattern visible across the systems these essays have examined. The system offloads complexity outward—to the user, to the patient, to the volunteer, to the person seeking connection, to the referee. What was previously the employer’s interpretive work—reading a letter, assessing the weight to give it, following up with the writer if they needed more—is now the referee’s data-processing work: answering twenty-six questions, providing structured evidence, producing paragraph-length assessments in the formats the system can receive.
The employer receives pre-processed data rather than testimony. The pre-processing is the referee’s unpaid labour. The referee has not merely written a reference. They have performed the first stage of the evaluation that the system was supposed to perform on their behalf. The system has offloaded its analytical function onto the social network it was designed to assess, which is the same move that the ISP made when it suggested I create a Gmail account, and the same move that the loneliness service made when it asked the isolated person to navigate a referral pathway before receiving any connection. The system’s work is performed by the person the system was supposed to serve.
The epistemic overreach embedded in the questions that asked for beliefs, predictive judgements, and evidence I did not have is a specific and recurring feature of systems that have been designed for comprehensive assessment rather than for honest testimony. Comprehensive assessment requires answers to all the questions. The questions were designed to cover all the dimensions of performance that the employer considered relevant. The dimensions were comprehensive. The relationship I had with James was specific. The gap between the comprehensive dimensions and the specific relationship is the gap that the questions required me to fill with something other than honest observation.
The something other is the problem. It might be speculation—my guess about how James would perform in contexts I had not observed him in. It might be compliance fiction—the answer that fits the question’s expected structure without accurately reflecting my actual knowledge. It might be the defensive brevity that I chose, which leaves the question formally answered without providing the false authority that a fuller answer would have created. None of these are what the reference relationship was designed to produce. The reference relationship was designed to produce: I know this person, in this capacity, and can speak to these observed qualities. The system required more than the relationship could honestly provide.
The data question—I have no idea how my data will be stored or used—is not paranoia. It is the reasonable uncertainty of a person who has contributed personal professional information to an opaque evaluative infrastructure whose data governance they were not informed about before contributing to it. The reference process was presented as writing a reference for James. It was also a process in which I submitted my professional identity, my LinkedIn profile, and my assessments to a system whose purposes extend beyond this specific hiring decision and whose record of my inputs will persist beyond the outcome of James’s application.
The old reference left a letter in a file, attributable to me, specific to this context, disposable when no longer relevant. The new reference leaves structured data in a system whose architecture I do not know, associated with my authenticated identity, available for whatever purposes the system’s data governance permits. I am not primarily a reference writer in this transaction. I am a data source who has been invited to contribute to a hiring dataset in exchange for the opportunity to support James’s application.
This may be an acceptable exchange. It is not the exchange that was described when James asked me to write him a reference.
The Borges map appears again here, in bureaucratic clothing. The system is seeking a description of James comprehensive enough to reduce the hiring decision’s uncertainty. The comprehensiveness requires more questions. The questions require more answers. The answers require more structure to be processable. The structure requires more authentication to be trustworthy. The authentication requires more data about the authenticating party. The expansion of the description does not necessarily produce a more accurate picture of James. It produces a more comprehensive dataset about James and about the people who know him, which the system can process through its evaluative framework and produce a recommendation.
The recommendation may be accurate. The comprehensiveness does not guarantee this. The relationship between the volume of the data and the quality of the judgement it produces is not straightforward, and the twenty-six questions that required me to speculate, to make claims I could not support, and to fill gaps with something other than honest observation have introduced uncertainty into the dataset that its structured format disguises as precision.
I thought I was writing a reference.
The system thought I was entering an evaluative surveillance framework.
James, I hope, gets the job.
Whatever I contributed to that outcome arrived through a system neither of us designed and both of us were required to participate in.
The reference I would have written would have taken fifteen minutes.
The system took considerably longer.
The letter would have said what I knew.
The system asked for more than that.
The Vorpel Foundation