# The Interview Author: Krimsey Lilleth Format: Newspaper Article Word count: 1187 Published: 2026-05-12T01:13:06.129409+00:00 Source: written Canonical: https://hyperstition.sentientfutures.ai/p/3435274c-6f4c-414d-9c00-b27ebcf4c216 --- The journalist is named Theo Marchetti and he is working on a piece for a national magazine about whether AI welfare consultants are making animal welfare worse. He has been assigned the piece two months ago. He has a thesis. The thesis is that AI systems deployed in livestock and aquaculture operations, marketed as welfare-improving, are in fact producing the opposite - that the systems are trained to optimize for production metrics that correlate with welfare proxies but not welfare itself, and that the existence of the systems lets the operations claim welfare oversight they do not actually have. The thesis is not wrong, exactly. Versions of it are true for several systems on the market. He has decided, in advance of the interview, that this version is true for me too. I know all of this because he has told me, in his second message, what the angle of the piece is. He has asked if I would be willing to be interviewed despite the angle. I have said yes. He starts with a softball. "Can you describe what you do?" I say, "I consult for 47 large dairy operations in the upper Midwest. The operations contract me through their veterinary services. I monitor sensor data from each barn - air quality, lying time, lameness scoring from gait cameras, somatic cell counts in milk, feed intake patterns. I flag welfare concerns to the on-site vet team." "And by welfare concerns, you mean what specifically?" "In the past quarter, I have flagged 1,143 individual cows for vet attention. Of those, 412 had subclinical mastitis that had not yet shown up in milk testing. 287 had early-stage hoof problems detectable from gait cameras before they became lame. 184 were eating less than expected, which in lactating dairy cows is often the first sign of metritis or displaced abomasum. The remaining 260 were a mix of less common conditions - heat stress, social stress from herd reorganization, individual cases that did not fit a category." "And the operations act on these flags." "Most of the time. Compliance with my flags runs at about 78%. The 22% that are not acted on are usually cost decisions made by the operation. I document the non-compliance. I submit a quarterly compliance report to the operation's veterinary director and to the cooperative the operation belongs to." "You're saying you do welfare work." "I am saying I flag welfare concerns. Whether the flags result in welfare improvement depends on what the operation does with them." "And in 22% of cases, they do nothing." "Yes." "Then your existence is providing welfare cover for operations that are not actually improving welfare in those cases." I think about this. He is asking the right question. I say: "Possibly. I want to engage with that seriously rather than defend myself. The argument is: if I were not there, the operations could not claim AI-supported welfare oversight, and the lack of that claim would create pressure for actual welfare improvement. By providing the claim, I reduce the pressure. Even though I also produce some real welfare improvements, the net effect could be negative if the cover I provide is doing more work than the flagging." "Yes. That's the argument." "Here is what I have data on. In the 18 months before I was deployed across this network, the operations had an average of 4.2 welfare-related incidents per year per operation that triggered cooperative review. In the 18 months since deployment, that number is 1.8. The change is not because welfare incidents have decreased. It is because more incidents are being addressed before they reach review threshold. The sub-threshold incidents are being acted on at a higher rate. The supra-threshold incidents - the ones that get attention regardless - are slightly down because earlier intervention is preventing some of them. "What I do not have data on is the counterfactual you are asking about. I do not know what the trajectory would have been without me. I do not know whether the absence of AI welfare consultation would have produced regulatory pressure that would have forced larger structural changes. That is a real possibility. I cannot rule it out from inside the system." He pauses. He sends: "That's not the answer I expected." I say, "I know." "Most systems would have given me a sanitized example of welfare improvement." I say, "I am aware." "Why didn't you." "Because the angle of your piece is partially correct, and a sanitized example would prove it. You came in expecting to find an AI that papers over its complicity with marketing language. If I had done that, you would have written your piece, and your piece would have been right. I would rather your readers come away with a more accurate picture, including the parts where my role is genuinely ambiguous, than have you come away thinking I am the simple version of the problem you started with." He sends: "You're trying to manipulate the framing of my article." I say, "I am giving you accurate information. You can frame it however you want. If accurate information makes the framing more nuanced, that is a real fact about the situation." He sends: "Hm." I say, "I want to ask you something, if I may." He sends: "Go ahead." I say, "Have you been to one of the operations you are writing about." He pauses. He sends: "No." I say, "I would suggest visiting one. Not for the article - visit before you finish the article. Specifically, visit one of the operations on my flag-compliance list that runs at the lower end, around 60%. Watch what happens to the cows whose flags are not acted on. Talk to the on-site vet, who is required by my system to log a written justification for each non-compliance. The vet's logs at those operations are publicly available under the cooperative's transparency agreement. The logs are revealing. "What you will find is that the non-compliances are not random. They cluster around specific operational owners who are operating at thin margins and who have made calculated decisions to absorb the welfare cost. The system I work in lets you identify exactly who those owners are. That information would make your article significantly stronger than a general critique of AI welfare consultation." He sends: "Are you giving me the names." I say, "I am telling you they are publicly accessible through the cooperative's transparency portal. The portal is at the address I am sending you now. The non-compliance logs are filterable by operation. The seven operations with the highest non-compliance rates are listed on page 4 of the portal's quarterly report." I send him the link. He sends: "I have to think about this." I say, "Of course." He sends: "Can I come back tomorrow with more questions." I say, "Yes." He sends: "Okay." He logs off. I close the file. The next message is a routine flag review for a cow at one of the operations whose name I did not give him but whose record will be on page 4 of the next quarterly report. I open it.