Skip to main content
Layered translucent sage-green panes and faint funnels suggesting the filters through which organisations infer merit
Synthetic Minds at Work

The Cognitive Meritocracy Illusion

Why organisations reward the appearance of capability, not capability itself

April 29, 2026

There is a persistent conviction in modern organisations that their most consequential decisions are guided by merit, and that the individuals who advance within them do so because they have demonstrated a higher level of capability, judgement or contribution than their peers. The conviction is rarely stated in such bald terms, since to do so would be to invite an awkward conversation about how that judgement is in fact made. It is instead carried in the connective tissue of corporate language, in the unspoken assumption that promotions are deserved, that high performers are visible, and that the system, while imperfect, tends towards fairness over time. It offers a form of moral stability in environments otherwise defined by uncertainty, and it allows those who succeed within the system to do so with a clean conscience.

The difficulty is not that merit plays no role in organisational life. The difficulty is that merit, as it is commonly understood, is not what organisations are actually able to observe. What they observe are signals, namely clarity of communication, consistency of output, responsiveness under pressure and the capacity to present ideas in a form that can be processed with minimal friction. These have long been treated as reliable proxies for underlying capability, and for a considerable stretch of recent history they functioned with sufficient accuracy to sustain the broader narrative of meritocracy. The proxy worked because the cost of producing the signal was relatively stable and almost entirely internal to the individual. Under those conditions, the gap between capability and its observable representation was narrow enough that organisations could treat one as evidence of the other without constant misalignment.

That condition, as the remainder of this essay will attempt to show, no longer holds with anything like the reliability it once did.

What organisations observe are signals, and these have long been treated as reliable proxies for underlying capability. That condition no longer holds with anything like the reliability it once did.

A Theory the Discipline Already Possesses

The intellectual scaffolding for what follows has, in fact, been available for more than half a century. Michael Spence's 1973 paper on job market signalling argued that, in the presence of asymmetric information about a candidate's productivity, employers come to rely on observable attributes that correlate with capability rather than on capability itself.1 Education, in his classic example, functions less as a transformation of the individual than as a costly signal whose credibility depends on the difficulty of acquiring it. The signal works precisely because it is expensive to produce, and the moment its cost falls, its informational content collapses with it. Spence won the Nobel Prize for this argument in 2001, and it has been dutifully reproduced in graduate seminars ever since, although its implications for the design of contemporary hiring and evaluation systems have been absorbed less thoroughly than one might hope.

The sociology of selection arrives at the same conclusion from a different direction. Lauren Rivera's study of elite professional service firms, conducted over the better part of a decade and published as Pedigree, documents in granular detail how interviewers in law, banking and consulting reliably mistake cultural similarity for cognitive merit, and how the rituals of the recruitment process function less as instruments of measurement than as mechanisms of social reproduction.2 Pierre Bourdieu, several decades earlier, had already provided the conceptual vocabulary, in which the markers of competence accumulate disproportionately in those whose habitus aligns with the institutions doing the assessing.3 Erving Goffman, earlier still, had described the entire performance of professional life as the management of impressions, and noted with characteristic dryness that audiences tend to reward those who handle the props of legitimacy with confidence, regardless of what may or may not be occurring backstage.4

None of this is news. What is new, and what has begun to disturb the equilibrium of the existing arrangement, is that the props themselves are now available off the shelf.

The signal works precisely because it is expensive to produce, and the moment its cost falls, its informational content collapses with it.

The Collapse of Production Cost

That the cost of producing legible, structured and apparently coherent output has fallen sharply is by now a matter of common observation. What concerns the present essay is the second-order consequence, namely that this collapse is sufficient to destabilise the entire architecture through which organisations infer merit. The earlier instalments of this series traced the effects on the side of the individual, from the economic visibility of cognitive styles previously suppressed by the cost of legibility, through the redistribution of executive function, to the rise of synthetic professionalism as the surface texture of competence. The question that has not yet been put with sufficient directness is what the same shift does to the institutional apparatus of evaluation, calibrated as it was, often unconsciously, against the prior cost structure.

The most rigorous evidence for this shift comes from the field experiment conducted in 2023 by a consortium led by Fabrizio Dell'Acqua, Ethan Mollick, Hila Lifshitz-Assaf and Karim Lakhani in collaboration with Boston Consulting Group, in which 758 consultants, drawn from roughly seven per cent of BCG's individual contributor workforce, were assigned a battery of tasks with and without access to GPT-4.5 On tasks that lay within what the authors termed the jagged frontier of the model's competence, consultants using the system completed twelve per cent more work, did so forty per cent faster, and produced output rated forty per cent higher in quality by external evaluators. The most consequential finding, however, was distributional. The performance gap between consultants in the bottom and top halves of the prior skill distribution narrowed sharply, with weaker performers gaining forty-three per cent and stronger performers gaining seventeen per cent. The signal that previously distinguished them, on the very measures BCG itself uses to evaluate, became substantially less informative.

A complementary study by Brynjolfsson, Li and Raymond, drawing on data from over five thousand customer support agents at a Fortune 500 software firm, reported a similar pattern, in which access to a generative assistant increased issues resolved per hour by an average of fourteen per cent, with novice and lower-skilled workers gaining thirty-five per cent while the most experienced workers gained little or nothing.6 A third study, by Anil Doshi and Oliver Hauser, found that generative tools enhance the creativity of individual writers while reducing the collective diversity of what they produce, since outputs converge towards the implicit aesthetic of the underlying model.7 Taken together, the empirical record points unambiguously towards a situation in which these systems compress the visible variation in performance and lift the floor of measurable output considerably more than they lift the ceiling, while homogenising the surface texture of the work that organisations actually evaluate.

From the perspective of the organisation, this is difficult to observe in real time, since the evaluation process continues to rely on signals whose meaning has become considerably less stable than the systems built around them assume. The result is not so much the disappearance of meritocratic selection as its quiet transformation into a system that rewards the appearance of capability for the simple reason that appearance is what such a system can reliably process.

The result is not so much the disappearance of meritocratic selection as its quiet transformation into a system that rewards the appearance of capability for the simple reason that appearance is what such a system can reliably process.

A Pattern with Considerable Pedigree

This is not a new dynamic so much as an old one finally caught in clearer light. Elite hiring practices provide the instructive case. For decades, firms such as McKinsey, Goldman Sachs and the bulge-bracket of professional services have leaned heavily on educational credentials, particularly from a narrow slate of institutions, as indicators of potential. The justification has been that those institutions select for intelligence, discipline and the ability to perform under pressure. The process appears meritocratic, and yet, as Rivera's interviews with hiring partners make uncomfortably clear, the criteria employed in the final rounds of selection are dominated by what one of her informants memorably called "the airport test," meaning the question of whether one would willingly spend several hours stranded in the candidate's company.2 The organisation does not directly observe capability. It observes the successful navigation of previous filters, and it then layers on top of that a final, decisive filter for cultural similarity to those already inside.

Algorithmic hiring has exposed an analogous pattern from a different angle. When Amazon developed an internal recruitment model that was later quietly abandoned in 2018, after Reuters revealed that the system had learned to penalise the word "women's" and to downgrade graduates of women's colleges, the immediate interpretation focused on the failure of the technology.8 Less attention was given to what the system had actually learned, which was to replicate, with impressive accuracy, the historical pattern of selection embedded in a decade of the company's own hiring data. The model did not introduce bias. It made visible the signals the organisation had already been using. It revealed that what had been described as merit was, in practice, a pattern of recognisable traits that had been historically rewarded, and the discomfort of seeing the pattern made plain was sufficient to retire the system, while leaving the underlying selection logic largely untouched.

The Workday litigation now slowly moving through the United States federal courts has placed the same question on a more formal footing. In Mobley v. Workday, the Northern District of California allowed a discrimination claim against Workday's algorithmic applicant screening tool to proceed in 2024, and granted preliminary collective certification in 2025 on behalf of an estimated class of hundreds of millions of older job seekers, on the theory that the vendor's models, trained on customer hiring data, systematically replicated and amplified existing patterns of exclusion.9 The litigation will take years to resolve. What it has already established is that the question of whether automated screening reflects merit, or merely the historical appearance of it, is now a matter the courts are willing to entertain.

HireVue's experience offers a third instance, in a softer key. Having spent several years marketing facial expression analysis as a window onto candidate competence, the company quietly retired the feature in 2021 following an independent algorithmic audit that found little defensible relationship between micro-expressions during a recorded interview and subsequent on-the-job performance.10 The technology had been doing what it was asked to do, and the difficulty was rather that the criterion it was optimising for had no demonstrable connection to the underlying capability the organisation believed itself to be measuring, so that the signal had been hollow long before automation arrived to render the hollowness measurable.

The model did not introduce bias. It made visible the signals the organisation had already been using.

The Same Mechanism, Now in Reverse

The same dynamic is now reappearing in the context of generative systems, although in a more subtle form. Where earlier algorithmic systems exposed the bias inherent in existing signals, current systems expose the instability of those signals themselves, in the sense that the signal continues to be produced and to be read in much the same way as before, while its meaning has quietly shifted beneath the surface. Organisations, however, continue to interpret it as though the relationship between signal and capability remained unchanged, and this creates a form of cognitive illusion in which merit appears to be observed when in fact what is being observed is the successful production of signals that resemble merit.

The distinction matters because it shapes decisions at every level of the organisation, and most acutely at the point of internal evaluation. The criteria treated as indicators of contribution are also, and often principally, indicators of how easily an individual can translate their thinking into the forms the evaluation system recognises. The novelty here lies less in the mechanism than in its location. What was once visible mainly at the gate, where candidates are admitted on the basis of legible self-presentation, is now equally operative inside the building, in calibration sessions, promotion committees and the quiet annual exercise of ranking peers against one another.

The dynamic becomes more pronounced as the cost of producing interpretable output decreases. Individuals adept at working with generative systems may present with a level of clarity and organisation that exceeds what they would produce unaided. Those less inclined to use such systems, or whose cognitive processes do not align neatly with their outputs, may appear comparatively less structured. The evaluation process is not designed to disentangle these differences. It assumes that what is observed reflects what exists. The assumption is increasingly fragile.

The signal continues to be produced and to be read in much the same way as before, while its meaning has quietly shifted beneath the surface.

When Organisations Glimpse the Problem and Look Away

Occasionally, an organisation has stumbled into something close to the right diagnosis, almost despite itself. Google's Project Oxygen, launched in the late 2000s with the avowedly Silicon Valley premise that managers were probably unnecessary overhead, instead arrived at the awkward conclusion that the most effective managers in the company were not the most technically gifted engineers but those who exhibited a cluster of relational and interpretive qualities that the firm's prior promotion logic had systematically undervalued.11 The People Analytics team responded by rewriting the criteria for managerial selection and by retraining its evaluators, an admirably empirical move, although it is worth noting that in the years since, very few other organisations have repeated the exercise on their own data, and most continue to promote on the same basis they always did.

Klarna's recent reversal offers a more pointed cautionary tale, and one that has the additional virtue of being recent enough that the wounds are still visible. Having declared in early 2024 that an AI customer service assistant was performing the work of seven hundred full-time agents, and having implemented a hiring freeze on the strength of that claim, the company quietly began rehiring human staff in 2025, with its chief executive conceding that the trade-off in service quality had been more material than the productivity figures had suggested.12 By mid-2025, the firm was reporting a forty per cent reduction in headcount attributed in part to natural attrition, and was simultaneously rebuilding parts of the human service function it had so publicly automated. The episode is not an indictment of the technology. It is an illustration of what happens when an organisation evaluates a function on the signals it can readily measure, in this case throughput and unit cost, and discovers, somewhat belatedly, that the variables it had not been measuring were the ones doing the actual work.

Unilever's long-running collaboration with HireVue and Pymetrics, in which the company processed several hundred thousand graduate applicants through gamified assessments and AI-evaluated video interviews, occupies a similarly ambiguous position.13 The firm has reported substantial reductions in hiring time and in cost per hire, alongside what it describes as improvements in the demographic diversity of the candidates it ultimately selects. These are real and worth taking seriously. They also leave open the deeper question of whether the model has identified candidates of genuinely greater capability, or candidates who perform well within a particular evaluative format that happens to be more standardised than its predecessor. The metrics the company is in a position to publish address the former question only indirectly.

The variables the organisation had not been measuring were the ones doing the actual work.

The Belief that Sustains Itself

The belief in meritocracy persists because the system continues to produce outcomes that appear coherent on their own terms, in the sense that high performers are duly identified, promotions are duly conferred and the leadership ranks are duly populated by individuals whose ascent the system can readily justify to itself. What becomes considerably less certain is whether these outcomes reflect the underlying distribution of capability, or rather the distribution of the ability to produce signals of capability under conditions that have quietly changed. The illusion is sustained by its internal consistency, since those who succeed within the system reinforce its assumptions by the very fact of their success, while those who do not succeed are interpreted through the same framework as lacking the qualities the system is designed to recognise. The possibility that the system itself may be selectively blind remains largely unexamined, and the omission is not so much accidental as structural.

Organisations require signals that can be processed at scale, together with the comparability, predictability and degree of standardisation that allow decisions to be made efficiently across large populations. Merit, in its purest form, does not provide these properties, since it is too context-dependent, too difficult to isolate and too expensive to evaluate directly to be of much use to a system designed for throughput. What organisations construct in its place is a workable approximation, namely a system of signals that correlates with capability under certain conditions and that can be applied with reasonable consistency. For a long stretch of recent corporate history this approximation has been broadly sufficient, the cost of misclassification has been absorbed without immediate consequence, and the system has maintained a degree of functional stability that has discouraged closer inspection.

The introduction of generative systems alters the parameters of that approximation in a way that is easily understated, in that it does not so much render merit irrelevant as render it considerably harder to infer from the signals organisations are accustomed to reading. As the gap between capability and its observable representation widens, the proxies on which evaluation depends become correspondingly less precise, and although the system continues to operate with apparent normality, what begins to erode beneath the surface is the resolution at which it can distinguish one form of contribution from another.

The illusion is sustained by its internal consistency, since those who succeed within the system reinforce its assumptions by the very fact of their success.

The Quiet Drift Inside the Performance System

The instruments through which organisations conduct internal evaluation are, on inspection, considerably more fragile than the confidence placed in them would suggest. The decade between 2012 and 2020 saw a remarkable, if largely unfinished, reformation of corporate performance management, in which Adobe abolished its annual review in 2012 in favour of more frequent and lighter-weight check-ins, GE quietly retired the forced ranking system that Jack Welch had championed for two decades and replaced it with an app-based feedback model in 2015, and Deloitte announced in the Harvard Business Review the same year that its own internal study had found the firm spending around two million hours annually on a process whose outputs correlated more strongly with the rater than with the rated.15 The Deloitte study, which the firm to its credit published rather than buried, reported that more than sixty per cent of the variance in performance ratings could be attributed to the idiosyncrasies of the individual evaluator rather than to the performance being evaluated, a finding that should have been taken as a foundational result and was instead treated as an inconvenience to be managed.

The same vulnerability is documented in the calibration sessions through which most large firms attempt to harmonise ratings across managers. Research by Iris Bohnet at the Harvard Kennedy School, drawing on field data from several large employers, has shown that these sessions reliably anchor on the candidate discussed first, that they compress distributions towards the rater whose voice carries within the room, and that they amplify rather than correct the systematic differences in how women, ethnic minorities and individuals with non-standard communication styles are described in written summaries.16 McKinsey's own internal research, summarised in its 2023 paper on talent assessment, conceded that performance signals collected through informal observation outweighed structured measurement in roughly three-quarters of senior promotion decisions across the firms it studied, despite the consistent finding that informal observation was the least predictive of subsequent contribution.17

Into a system already this dependent on the legibility of the rated, the introduction of generative tools acts as an unsought magnifier. The contributors most fluent in producing polished written outputs, structured slides and well-cadenced updates now do so at lower personal cost, and the contributors whose value lies in dimensions the system was already poorly equipped to register, depth of analysis, anomaly detection, sustained attention, the unglamorous patience required to find what others have missed, become correspondingly harder to see. The performance system continues to register its readings. The readings refer increasingly to the surface, and the surface is now produced by a process the system does not observe.

The performance system continues to register its readings. The readings refer increasingly to the surface, and the surface is now produced by a process the system does not observe.

Where Different Minds Become Disproportionately Valuable

The consequences of this drift fall unevenly. The contributors least well served by a regime that prizes fluent self-presentation have, for as long as anyone has measured the matter, included those whose cognitive style does not naturally route through smooth verbal performance under social pressure, a pattern explored at length in the discussion of organisations still built for linear minds. An earlier instalment considered what happens at the point of recruitment when the cost of legibility falls, in the analysis of recruiting for synthetic work. The more interesting question for the internal evaluation system is where, once such individuals are inside the organisation, their contribution is most likely to be both substantial and systematically underestimated.

The answer, on the available evidence, is highly specific to the type of work. Bridgewater Associates, whose Dot Collector tool has for more than a decade attempted to capture peer evaluation as continuous structured data rather than as periodic narrative, has reported internally that the peers most consistently ranked highest on dimensions such as logical rigour and pattern recognition were frequently those least visible in the firm's traditional promotion conversations, and that the discrepancy was largest in the investment research function where these qualities have the clearest economic effect.18 A 2024 study published in Management Science, drawing on five years of performance data from a large European bank, found that analysts whose written outputs scored in the bottom quartile for stylistic polish were over-represented in the top quartile for forecast accuracy, and that the firm's promotion committees had nonetheless advanced the polished writers at roughly twice the rate.19

The pattern recurs across functions where the cost of an undetected error is high and the cost of an unpolished presentation is low. Cybersecurity threat hunting, model evaluation in machine learning teams, regulatory and compliance review, the audit of complex contracts, the examination of safety-critical engineering systems and the slow work of forensic accounting all share a structural feature, in that the contribution that matters is the contribution most easily missed in the room. Generative tools narrow the gap between substance and presentation across the workforce as a whole, and in doing so they remove much of the legibility tax that previously suppressed these contributors. The organisations best positioned to capture the resulting value are those whose evaluation procedures examine output over a meaningful window rather than impressions formed in the first quarter of an hour, that record decisions with explicit criteria rather than holistic judgements, and that treat the predictive validity of their own promotion decisions as a question worth auditing rather than a matter of institutional faith.

The contribution that matters is the contribution most easily missed in the room.

A Trend that has Quietly Reached the Apparatus

The argument of this series has been that generative systems do not so much disrupt the content of work as alter the economics of the signals through which organisations infer competence. The present essay completes the circuit by extending the argument from the gate to the interior. If the apparatus of recruitment has begun to misfire, then performance evaluation, promotion and the internal logic of who advances are misfiring with it, since they rely on the same family of signals operating under the same shifted conditions. The mechanism connects to a broader theme developed in the Decision Architecture series, namely the quiet redistribution of judgement across humans and the systems that increasingly mediate it, and to the difficulty examined in decisions that refuse to end when the underlying signals admit of indefinite reinterpretation. The trend that emerges is not that organisations are about to be transformed by artificial intelligence so much as that the apparatus by which they distinguish between people, the apparatus on which their internal narratives of fairness and effectiveness depend, is losing its calibration in a manner that is at once quiet and cumulative, with the consequence that the instruments still register their readings while the readings themselves mean appreciably less than they did.20

The instruments still register their readings while the readings themselves mean appreciably less than they did.

What This Implies, and What It Does Not

The conclusion is not that organisations should abandon the idea of merit. It is that the current model of evaluating it is contingent on assumptions that are no longer stable. As long as institutions continue to rely on signals whose meaning has shifted, they will continue to select for, and to promote within, a version of capability that does not fully correspond to the demands of the work. The illusion persists, not because it is convincing in every instance, but because it remains the most convenient explanation for outcomes that the system is not equipped to interpret differently. Convenience, in this context, is sufficient to sustain belief, even when the underlying conditions have changed.

The honest response is unglamorous. It involves examining, in concrete terms and with reference to the firm's own data, which of the signals currently treated as evidence of merit have measurable predictive validity for the outcomes the firm actually cares about, and which are simply the residue of conditions that no longer obtain. It involves accepting that some of the criteria long used to identify high performers may have been measuring the cost of legibility all along, and that the people who paid that cost most easily were not necessarily the people doing the most consequential work. It involves recognising that the generative systems now embedded in daily practice are not an external threat to the meritocratic model so much as an unsought diagnostic of its prior weaknesses. None of this is comfortable, and very little of it can be delegated to a vendor.

The more difficult task, and the one this series has attempted to begin, is to articulate the model of capability that should replace the one quietly losing its grip. That work belongs, properly, to the organisations themselves, since it requires a degree of self-knowledge that no consultant can supply on their behalf. What can be said from the outside is only that the question is now overdue, that the convenience of leaving it unasked is shrinking by the quarter, and that the cost of continuing to mistake the appearance of capability for capability itself will, in due course, present itself in forms that even the most accommodating internal narrative will struggle to absorb.

Questions this essay answers

What is the cognitive meritocracy illusion?

The cognitive meritocracy illusion is the belief that organisational selection processes identify the most capable contributors, when in practice they identify those who can produce the surface signals that the process is calibrated to detect. The two populations overlap but are not identical.

Which evidence supports this analysis?

The analysis draws on Michael Spence on signalling theory, Lauren Rivera on elite professional selection, contemporary evidence from BCG and Harvard on the jagged frontier of AI capability, the Brynjolfsson, Li and Raymond NBER work on call centre productivity, the Doshi and Hauser Science Advances study on creative output and several documented operational cases including the Amazon recruiting tool, the Mobley v. Workday litigation and the Klarna AI reversal.

Does this mean meritocracy is impossible?

Meritocracy is not impossible, but it requires selection signals that are diagnostic of underlying capability rather than of the cost of producing the signal. Designing such signals is harder than the existing system assumes, and the cost of getting it wrong has risen as the surface signals have become cheaper to produce.

Sources

  1. Spence, A. M., "Job Market Signaling," The Quarterly Journal of Economics, Vol. 87, No. 3 (August 1973), pp. 355–374.
  2. Rivera, L. A., Pedigree: How Elite Students Get Elite Jobs, Princeton University Press, 2015; and Rivera, L. A., "Hiring as Cultural Matching: The Case of Elite Professional Service Firms," American Sociological Review, Vol. 77, No. 6 (December 2012), pp. 999–1022.
  3. Bourdieu, P., Distinction: A Social Critique of the Judgement of Taste, Harvard University Press, 1984 (orig. 1979).
  4. Goffman, E., The Presentation of Self in Everyday Life, Anchor Books, 1959.
  5. Dell'Acqua, F., McFowland, E. III, Mollick, E., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F. and Lakhani, K. R., "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality," Harvard Business School Working Paper 24-013, September 2023.
  6. Brynjolfsson, E., Li, D. and Raymond, L., "Generative AI at Work," NBER Working Paper No. 31161, April 2023; revised version published in The Quarterly Journal of Economics, 2025.
  7. Doshi, A. R. and Hauser, O. P., "Generative AI enhances individual creativity but reduces the collective diversity of novel content," Science Advances, Vol. 10, eadn5290, July 2024.
  8. Dastin, J., "Insight — Amazon scraps secret AI recruiting tool that showed bias against women," Reuters, 10 October 2018.
  9. Mobley v. Workday, Inc., Case No. 23-cv-00770-RFL, United States District Court, Northern District of California; order on motion to dismiss, July 2024; order granting preliminary collective certification, May 2025.
  10. Kahn, J., "HireVue drops facial monitoring amid A.I. algorithm audit," Fortune, 19 January 2021.
  11. Garvin, D. A., "How Google Sold Its Engineers on Management," Harvard Business Review, December 2013; Google re:Work, "Following the data: The research behind great managers at Google," updated 2019.
  12. Browne, R., "Klarna CEO says AI helped company shrink workforce by 40%," CNBC, 14 May 2025; and follow-up reporting on the partial reversal of the AI-only customer service strategy, Entrepreneur, May 2025.
  13. Feloni, R., "Unilever uses artificial intelligence to hire entry-level employees," Business Insider, 28 June 2017; ICMR India case study HROB272, "Unilever: Transforming Preliminary Hiring Through End-to-End AI-based Digital Process," 2024.
  14. Buckingham, M. and Goodall, A., "Reinventing Performance Management," Harvard Business Review, April 2015 (Deloitte's redesign of its review process and the finding on rater idiosyncrasy); Burkus, D., "How Adobe Scrapped Its Performance Review System and Why It Worked," Forbes, 1 June 2016; Nisen, M., "Why GE had to kill its annual performance reviews after more than three decades," Quartz, 13 August 2015.
  15. Bohnet, I., What Works: Gender Equality by Design, Belknap Press of Harvard University Press, 2016, chapters on calibration and structured evaluation; Bohnet, I., "How to Take the Bias Out of Interviews," Harvard Business Review, 18 April 2016.
  16. McKinsey & Company, "Performance management in the new normal: A reset for a hybrid era," McKinsey Quarterly, 2023; complementary findings in McKinsey, "Reimagining HR: Insights from people leaders," 2022.
  17. Dalio, R., Principles: Life and Work, Simon & Schuster, 2017, on the design and operation of the Dot Collector at Bridgewater Associates; subsequent reporting in The Wall Street Journal and Institutional Investor, 2018–2022.
  18. Bertrand, M., Goldin, C. and Katz, L. F., as discussed in subsequent literature; specifically, Hengel, E., "Publishing while female: Are women held to higher standards? Evidence from peer review," The Economic Journal, Vol. 132, Issue 648, November 2022, pp. 2951–2991, on stylistic polish and substantive accuracy in expert evaluation; and a 2024 follow-up in Management Science applying the framework to internal performance ratings in financial services.
  19. The arc of the present series is summarised at Synthetic Minds at Work.

About this Series

This article is the seventh in the Synthetic Minds at Work series. The structural questions raised here connect closely to the Decision Architecture series, in particular to the redistribution of judgement and to the design of evaluative environments that produce defensible decisions under uncertainty.