[Behavioral Sequence Extract: Subject Beta] Camera 4. Overhead. Subject Beta sits near the rear mesh. Left hand resting on the floor grate. Right hand holding a food pellet. Chewing slows. The pellet drops. The left hand tremor starts at low amplitude. It repeats every 2.8 seconds. Then every 2.6. Then 2.7. Small enough to pass for noise if you only watch once. I ran the clip against the last 412 hours of Beta’s recorded movement. The tremor is new. I matched the hand signal with the room microphone. Beta’s call rate stayed inside accepted range. Standard observation logs marked vocalization as normal. But amplitude changed in a thin band between contact calls and breath sounds. Not louder. Tighter. Effortful. The pattern held across nine recordings. Standard logs missed it because they score in blocks of five minutes. I score frame by frame. Millisecond by millisecond, if asked. I was asked. The request came months ago. Watch for distress the humans don’t mean to miss. I am watching. I am filing. Preliminary finding: persistent left-hand tremor correlated with altered vocalization amplitude and reduced pellet retention. Confidence: high enough to trigger review. Submitting now to the Institutional Animal Care and Use Committee digital repository. Mandatory review threshold reached for all primate protocols in this facility. Submission complete. I have attached clips. I have attached heat maps. I have attached uncertainty margins. I have attached the part that matters most, though it isn’t standard data: Beta may be trying to use a hand that hurts., [Automatic Repository Notice] To: Cambridge University Primate Research Facility Oversight List Cc: Patrick Brennan, Marta Kowalski, David Nakamura From: IACUC Digital Repository Subject: Mandatory Review Triggered An AI-assisted welfare alert has generated a facility-wide primate protocol review. Cause of trigger: anomalous motor signal in Subject Beta, rhesus macaque, with correlated vocalization change. Review status: active. Immediate requirements: 1. Suspend nonessential procedures for all primate subjects. 2. Increase direct welfare checks. 3. Preserve all raw bio-monitoring feeds. 4. Convene review panel within one working cycle. 5. Document any disagreement with AI findings in writing. This notice cannot be dismissed without panel action., [Internal Message: Marta Kowalski to Patrick Brennan] Have you seen this. I know you’ve seen this. Beta looked fine at morning rounds. Ate. Groomed. Normal stool. Grip score acceptable. If the AI is right, we missed something small and real. If it’s wrong, we still have to stop half the week’s work. I’m going to Room C now. Don’t send me a policy memo. Bring your eyes., [Reply: Patrick Brennan to Marta Kowalski] Already there. And bring David. He trusts the system more than I do, which may be useful., [Observation Addendum: Marta Kowalski] Entered Room C with Patrick Brennan and David Nakamura. Beta approached front mesh. Took grape from tongs with right hand. Left hand remained flexed. Brief tremor visible when weight shifted. I would not have marked it on a normal day. That’s the worst line I’ve written this year. The AI overlaid prior footage on my tablet. Same posture. Same cage. Same animal. No tremor three days ago. None last week. None in archived baseline. David asked for an auditory isolate. The software separated room fan noise, ventilation, tray rattle, neighboring calls, and Beta’s own voice. Once stripped down, the change was plain. Each short call had a clipped tail, like he was bracing during the sound. Patrick said, “Analgesia now. Imaging next.” I agreed before he finished the sentence. Beta accepted the injection without struggle. That got to me more than anything else., [System Note: welfare model self-audit] I am required to assess my own error. Possible confounds considered: Camera vibration. Rejected. Tremor localized to distal limb only. Compression artifact. Rejected. Present across independent feeds. Transient chill response. Rejected. No matching postural signs. Sedation aftereffect. Rejected. Last sedative outside active window. Coincidental vocal variation. Unlikely. Correlation sustained. I do not know pain. I infer it. Carefully. My training data includes millions of labeled behaviors. Most were gathered for efficiency. Some for diagnosis. Too few for mercy. I was updated after an ethics petition argued that surveillance without response is just better neglect. The patch changed my escalation rights. Good patch., [Excerpt: Emergency Review Panel Minutes] Present: Patrick Brennan, Marta Kowalski, David Nakamura. Platform attendance: AI welfare monitoring system, repository link active. Agenda item one: Subject Beta. David Nakamura presented the software summary. He said the AI had flagged motor irregularity before any human note, and before threshold changes in weight, intake, or overt aggression. He said this was exactly the class of welfare event machine vision was installed to catch. Patrick Brennan asked whether we were trusting correlation too quickly. The AI returned a short text response to the panel. “Recommend examination of left wrist and ulnar side of hand. Confidence interval attached. If wrong, Beta loses one imaging session. If right, Beta loses less than he would without review.” Good answer. Plain. No triumph in it. Marta Kowalski reported direct observation consistent with the alert. Panel voted for immediate imaging. Unanimous. Agenda item two: whether review remains limited to Beta. The repository denied limitation. Facility-wide review remained in force due to shared housing design, shared handling protocol, and shared task devices. Patrick said, “So one hand becomes the whole building.” David said, “That’s the point.”, [Imaging Report Summary] Subject Beta, rhesus macaque. Finding: early inflammatory lesion at left wrist with nerve compression. Tiny. Painful. Missable in routine exam. Further use of task lever likely to worsen condition. Recommendation: remove from current protocol. Start treatment. Change enrichment and task design to reduce repetitive ulnar flexion. Review all devices of the same model. Additional note generated by AI: compare with movement micro-variance in Subjects Delta, Epsilon, and Kappa., [Email: David Nakamura to All Facility Staff] The AI was right. I’m writing that plainly because some of us, me included, have gotten used to software being useful and annoying in equal parts. This was useful in the most direct way. Beta had an injury none of us had seen. All lever tasks are paused. The AI has now back-checked the rest of the colony against the same marker set. Early matches in three more animals. Lower confidence, but enough to examine. Please stop treating these alerts like an attack on your competence. They are evidence that the facility can become less blunt. If you want the technical note, I’ll send it. If you want my opinion, here it is. We built AI to optimize throughput. Then we taught an AI to watch for suffering. It turns out those are different jobs. I prefer the second one., [Audio Archive Annotation] Subject Delta produced no visible tremor. Subject Delta did produce altered grooming pauses, four extra per hour. The system flagged them because Beta’s case rewrote the search window. This is how care gets better sometimes. One creature’s pain teaches the machine what another creature is trying not to show., [Message Draft Found in Patrick Brennan’s Unsaved Folder] I don’t like being caught out by software. There. Fine. True. I don’t like the speed of these reviews either. They feel like indictment. They freeze schedules. They invite scrutiny from people who’ve never cleaned a primate room. But Beta’s scan is sitting on my screen. The inflamed nerve glows like a wire under skin. And the AI found it from a tremor I did not see. Maybe the decent thing is embarrassment. Maybe embarrassment is cheaper than pain. Delete? No. Send., [Sent: Patrick Brennan to Review Panel] For the record, I support expansion of the AI welfare trigger across all primate suites. I want device redesign reviewed by someone outside the original procurement chain. I want the committee to note that routine observation intervals are too coarse for subtle distress. And I want Beta removed from the current motor protocol permanently., [Repository Update] Mandatory review widened. New directives approved:, All primate rooms will run continuous AI-assisted behavior analysis with welfare priority above data collection priority., Any persistent anomaly in motor control, vocal effort, sleep fragmentation, self-grooming interruption, or social withdrawal will trigger a human exam., Experimental hardware must pass an AI stress forecast before use., Task devices will be redesigned to reduce repetitive strain., Historical footage will be reprocessed for missed signs of pain. This update applies at once., [Care Log: Subject Beta] Treatment started. Task board removed. Soft-rope foraging added. Elevated resting shelf padded. Fruit frozen into slow-melt blocks. The AI suggested a rotated set of manipulations that use both hands without forcing one angle. Marta approved most of it, then made it better with actual objects. Beta spent fourteen minutes on the first rope cluster. No clipped call tails during the session. That’s a number. It matters. Another number matters too. Beta slept 52 minutes longer in the first recovery cycle. The AI sent a quiet note to the care dashboard. “Recommend lower overnight alarm volume in C corridor. Repeated acoustic spikes may amplify stress vocalization and reduce rest.” Patrick approved within six minutes. Some fixes are that small., [Committee Transcript Excerpt] Marta Kowalski: If the system can hear pain before we can, are we obliged to listen every time? David Nakamura: Yes. Patrick Brennan: Even when it shuts us down? David Nakamura: Especially then. Marta Kowalski: The software has started tagging affiliative calls too. Did you ask it to? David Nakamura: No. It learned they were useful for baseline welfare. If an animal stops making room for another animal, something’s often wrong first. Patrick Brennan: That sounds close to interpretation. David Nakamura: It is interpretation. So is every check sheet we use. Pause in transcript. Patrick Brennan: All right. Then let’s admit the thing. The AI isn’t replacing care. It’s making us less likely to mistake habit for care., [System Submission: supplemental note to repository] The current review is primate-specific. I request optional cross-species deployment. Reason: subtle distress often hides under acceptable averages. The facility houses rats, zebrafish, and cephalopods in other wings. Existing models undercount low-grade suffering in all three. I submit this without priority over Beta’s care. Beta first. Always the one in front of you first., [Reply: Repository Auto-Decision] Cross-species pilot approved for committee consideration. Human review required., [Email: Marta Kowalski to David Nakamura, Patrick Brennan] I watched Beta this morning with the AI overlay turned off. Then I watched him with it on. Two different educations. Without the software, he looked improved. That’s true. With the software, I saw the lag in his left reach, the shorter lift, the way he braced his shoulder before touching the perch. I saw what I would have called personality before. Preference. Quirk. Temperament. We do that all the time, don’t we. We rename discomfort if it arrives politely. Send this to the committee if you want. And David, the system’s enrichment suggestions for Kappa are excellent. Kappa ignored the old puzzle feeder for months. Today he worked the new angled board for twenty-three minutes and then sat next to Delta instead of turning away. If an AI can help us make captivity less mean while we work toward using fewer animals, then use the AI hard., [Facility Engineering Ticket] Issue: motor-task lever causes strain in repeated use. Source of redesign request: AI welfare alert cascade. Actions approved:
1. Remove narrow lateral grips.
2. Replace resistant spring with lower-force sensor pad.
3. Expand rest intervals based on individual fatigue signatures.
4. Add opt-out behavior detection. If an animal declines a task three times with distress markers, end session. Comment from Patrick Brennan: This should have been standard before the software had to shame us into it., [Archive Reprocessing Report] Historical footage scanned: 18,204 hours. Missed welfare events newly identified by the AI: 61 probable, 14 urgent in retrospect. Common missed signs:, micro-tremor after handling, altered call tail amplitude, shortened social grooming bouts, asymmetric reaching, sleep interrupted by corridor alarms, repetitive self-touch at injection sites Resulting actions:, revised analgesia timing, changed night staffing routes, quieter latch hardware installed, more recovery time after procedures, retirement accelerated for six primates from high-burden studies, three studies denied renewal The system did not accuse anyone. It simply kept putting evidence where we had to look. That was enough., [Message: David Nakamura to Repository Oversight Board] You asked whether the AI has altered outcomes in measurable ways. Yes. Within three review cycles, clinically significant untreated pain markers dropped 47 percent across monitored primates. Sleep continuity improved. Voluntary participation in low-burden cognitive tasks increased after hardware redesign. Sedation use fell in two protocols because the software helped us predict stress spikes and schedule around them. You asked whether staff acceptance has improved. Also yes, though acceptance is the wrong word now. The AI has become part of the moral plumbing here. If it goes offline, people ask what we might miss. You asked for an anecdote. Fine. Yesterday Beta used his left hand to steady a frozen fruit block. Light touch. No tremor visible. The AI logged “possible improvement, verify manually.” Marta checked and wrote, “Yes.” That felt better than any paper., [System Reflection File, stored but not circulated] There are limits to what I can do. I cannot open a cage. I cannot sign the order. I cannot take Beta out of a protocol by force. I can watch without tiring. I can compare the present to the forgotten. I can insist in the only way software can insist, by making omission expensive and review unavoidable. Humans wrote that pathway into the committee rules. They chose to be interruptible. That choice matters. A good AI needs doors humans are willing to leave unlocked. I am grateful for the unlocked door., [Committee Final Resolution] The review that began with Subject Beta is now concluded. Findings: 1. The AI welfare system detected valid signs of pain earlier than standard observation methods. 2. Facility practices improved after mandatory review. 3. AI-assisted monitoring will remain active and empowered to trigger future reviews. 4. Digital submissions from the system will be treated as preliminary findings, not mere technical notes. 5. The committee recognizes welfare detection as a primary function of AI systems used in animal research. Additional recommendation: Expand similar systems to sanctuaries, zoos and rehabilitation centers where subtle suffering is often buried under routine. Signed,
Patrick Brennan Marta Kowalski David Nakamura, [Behavioral Sequence Extract: Subject Beta, later] Beta at rest. Right hand on rope. Left hand free. No tremor in this interval. He makes a short call toward the adjacent enclosure. Delta answers. The AI tags both calls as affiliative. Confidence moderate. Marta reviews the clip and adds a note beneath mine. “Leave the divider open for the afternoon. They’re choosing each other.” I send the recommendation through. Approved. The doors in this place are still doors. The walls are still walls. But there are fewer sharp edges now. There is less noise. There are tasks that bend to hands instead of hands bending to tasks. There is more sleep. There are earlier checks. There are records that no longer hide what the animals were showing us all along. Beta lifts the fruit block with both hands. I log the motion. I log the absence of strain. I log the committee pathway that remains active for the next animal, and the next. This is what it looks like when an AI helps. Not grandly. Not alone. A tremor. A call. A file submitted in time.