Artificial Intelligence and the Surrogation of Human Work
Shorrock argues in this blog essay that AI-produced accounts of work can become surrogates for the work itself: fluent enough that it becomes easier to believe the work is understood, and concerns absent from the account may be discounted. Could an organisation, he asks, become more informed while becoming less aware of how work happens?
He starts from the varieties of human work (work-as-imagined, work-as-prescribed, work-as-disclosed, work-as-done) and his further proxies, such as work-as-measured and work-as-analysed: each necessary, each partial, each involving choices. An account may not be inaccurate while being inadequate in what it leaves out. From Tsoukas he takes paradoxes: more information may bring less understanding, and people adapt to the indicators that judge them, so representations begin to shape the reality they describe.
Surrogation he traces from Beer's warning that people often begin to mistake the model for the reality and start managing the model instead, through Tsoukas, to Choi and colleagues' term for treating performance measures as the constructs they stand for; by extension, he suggests, work-as-measured may be treated as sufficient understanding of work-as-done. AI may increase the risk through speed, scale, fluency and apparent coherence, which may leave assumptions and limitations less visible: a risk with an associated opportunity.
He reports findings that finer-grained estimates drew more confidence in their accuracy (resting also, he notes, on assumptions about the communicator), and that longer language-model explanations raised confidence even where length added no accuracy, in experiments he says were not in organisational life, and asks whether a fluent analysis might outweigh a hesitant account of a weak signal. He draws on Star and Strauss on invisible work, Ridgway on information-only measures probably read as definitions of what matters, and Espeland and Sauder on reactivity, and sketches the risk of a self-reinforcing cycle from reports through analysis to priorities and back. Practitioners adapt their accounts to what the organisation acts on; such adaptations may be protective, and people may stop raising concerns that are repeatedly discounted.
The Post Office Horizon scandal (Horizon, he notes, does not use AI) shows a technological account defended against its challengers; he names epistemic injustice. From Rogers he takes locus of evaluation, asking whether people come to judge their own experience by AI-generated assessments: where an authoritative account repeatedly conflicts with what they notice, they may discount their concerns before raising them. Questioning AI, he argues, depends on the design of work and the organisation's response. Released time may go to higher expectations or direct contact with practitioners. He closes with two complementary inquiries, examining the analysis and exploring the work.
Several illustrations come from air traffic control and his own practice. The essay does not name psychological safety; the bridge is ours. We read it as showing how a fluent machine account can make speaking up cost more and seem worth less; his question, whether people retain the confidence, opportunities and authority to bring their experience into judgements, is close to ours.
Save this paper to a bibliography, keep notes on it, and export citations. Create a free account →
Neighbours (9)
- On Becoming a Person: A Therapist's View of Psychotherapy
- Brain of the Firm
- Dysfunctional Consequences of Performance Measurements
- Systems Thinking for Safety: Ten Principles. A White Paper — Moving Towards Safety-II
- Goodhart's / Cobra Effect
- Measuring PS
- Streetlight Effect
- WAI vs WAD
- Making Work Visible