The Early Warning Record

AI-assisted content delivery is expanding rapidly — adaptive learning platforms, automated grading, AI tutoring systems. The deployment is concentrated in content transmission: delivering information, assessing factual recall, providing practice problems. This is the domain sub-function most amenable to automation and the sub-function with the weakest claim to irreducibility.

The irreducible functions of education — social-emotional learning (SEL) capacity, developmental attunement, the relational scaffolding that enables a child to tolerate frustration and persist through difficulty — are not yet directly displaced. But the structural compression is beginning: as AI handles content delivery, the institutional pressure is to reduce the teacher-to-content ratio rather than to redeploy teacher capacity toward relational functions. The extractive pattern is present in the economic logic even where it has not yet produced measurable atrophy.

Routine finance is the domain furthest along the collapse gradient with documented evidence. The Goldman Sachs equity trading floor transition — from approximately 600 traders in 2000 to 2 traders plus 200 engineers and automated systems in 2017 — represents a completed Stage 0 and advanced Stage 1. DTCC straight-through processing eliminated manual reconciliation entirely.

The Stage 2 assessment is based on the tacit knowledge transmission criterion defined in HC-021: when the population of experienced practitioners drops below the threshold needed to transmit tacit knowledge to the next generation through mentorship, modeling, and situated learning. In routine equity trading, this threshold has plausibly been crossed. The people who understood market microstructure through embodied practice — who could read order flow, sense liquidity shifts, and exercise judgment under uncertainty — are largely retired or retrained. The knowledge they carried is not fully encoded in the automated systems that replaced them.

Construction remains largely pre-automation in its core craft functions. Robotic bricklaying, 3D-printed structures, and AI-assisted design exist but have not displaced the skilled trades at population scale. The domain's physical complexity, site variability, and the embodied nature of craft knowledge — the carpenter's feel for grain, the mason's sense of mortar consistency, the electrician's judgment about load distribution in non-standard configurations — create structural barriers to automation that do not exist in information-processing domains.

The Stage 0 assessment reflects the beginning of extractive deployment in construction planning, estimation, and project management — information-processing sub-functions where AI is being deployed for efficiency. The craft functions themselves remain at pre-Stage 0, but the economic pressure on apprenticeship programs is already measurable.

Healthcare presents a compound picture. AI diagnostic tools — radiology image analysis, pathology slide screening, ECG interpretation — are deployed and in many cases outperform human readers on narrow metrics. This is Stage 0 in diagnostic sub-functions. The practice atrophy concern is real: if radiologists stop reading routine scans because AI handles them, the skill base for identifying the atypical cases that AI misses will erode on the depreciation curve documented in HC-020.

But the more immediate threat to healthcare capability is not AI displacement of clinical skills. It is administrative burden displacing relational capacity. Physicians spend an estimated 49% of their time on electronic health records and desk work (Sinsky et al., 2016). This is not AI-caused — it is a pre-existing compression of the irreducible relational function by administrative systems. AI could either deepen this compression (by adding more documentation requirements) or relieve it (by handling administrative tasks to free relational capacity). The direction is not yet determined.

AI-assisted legal research, contract review, and risk assessment tools are deployed across the legal profession. Algorithmic risk assessment in criminal sentencing (COMPAS, PSA) represents the most consequential deployment — automated systems making or influencing decisions about human liberty. The extractive pattern is present: judges receive algorithmic risk scores and must actively override them to exercise independent judgment.

The Stage 0–1 assessment reflects the documented automation bias problem: when an automated system provides a recommendation, human decision-makers systematically defer to it even when they have grounds for disagreement. Skitka et al. (2000) documented this in aviation; the same mechanism operates in judicial settings. The judicial override is nominally active — judges can and do deviate from algorithmic recommendations — but the practice of independent risk assessment is under structural pressure from the availability of automated scores.

AI deployment in governance is at Stage 0: beginning but not yet producing measurable practice atrophy in the irreducible governance functions. Automated benefits determination, AI-assisted policy analysis, chatbot-mediated citizen services, and predictive policing represent the current deployment frontier. The extractive potential is significant — governance involves complex judgment about competing values, contextual interpretation of rules, and democratic accountability that cannot be fully automated — but the current deployments are concentrated in routine processing rather than judgment-intensive functions.

The governance domain has a unique vulnerability: democratic legitimacy requires that consequential decisions be traceable to accountable human judgment. If governance functions are automated to the point where the human officials nominally responsible for decisions lack the capability to evaluate the automated output, the democratic accountability chain breaks. This is a Stage 2 risk specific to governance that does not have an equivalent in other domains.

Science presents a distinctive pattern. AI is accelerating hypothesis generation, literature synthesis, data analysis, and experimental design at a pace that exceeds the scientific community's capacity to evaluate, replicate, and govern the outputs. AlphaFold's protein structure predictions, AI-driven drug discovery pipelines, and automated experimental platforms represent genuine scientific capability — but the acceleration creates a governance gap when human researchers cannot independently verify AI-generated results at the rate those results are produced.

The Stage 1 assessment is based on the practice atrophy criterion applied to scientific judgment: the capacity to evaluate evidence quality, identify confounds, exercise skepticism about seemingly clean results, and maintain the social practices of peer review and replication that constitute science's error-correction mechanism. When AI generates hypotheses and results faster than human scientists can critically evaluate them, the practice of critical evaluation atrophies even if the scientists remain employed. The skill is not displaced by unemployment but by throughput — the volume of AI output exceeds human evaluative capacity.

Care — eldercare, childcare, disability support, mental health services — is the domain where the extractive deployment thesis is most clearly a thesis about human dignity rather than system efficiency. AI companion robots, therapeutic chatbots, automated monitoring systems, and virtual care platforms represent the beginning of substitutive deployment: using AI systems to perform functions previously performed by human caregivers.

The Stage 0 assessment reflects the early state of this deployment. The care domain's irreducible functions — relational attunement, emotional co-regulation, the capacity to be genuinely present with another person's suffering — are the functions most resistant to automation and most damaged by substitution. The evidence base for care capability atrophy under automation is thin because the deployment is early. But the structural logic is clear: if institutions adopt AI care systems to reduce labor costs, and if the economic pressure on human caregiving roles intensifies, the pipeline of skilled human caregivers will narrow. HC-023 (The Common Faculty Problem) identifies care as the domain where Stage 4 risk is highest precisely because the irreducible function (human relational presence) cannot be rebuilt by technical means once lost.