There is a quiet confidence embedded in most hiring processes, and it rests on the assumption that capability can be inferred from how individuals present their thinking under controlled conditions. Interviews are structured to elicit clarity, coherence and responsiveness, and candidates are expected to articulate their reasoning in a form that is both structured and accessible, often under time pressure and in unfamiliar contexts. The underlying premise is rarely questioned, since it appears both practical and fair. If someone can explain their thinking clearly, organise their ideas and respond with precision, it is taken as a reliable indication that they will perform effectively within the organisation. The premise has always been imperfect, yet it has been sufficiently functional to sustain itself for several decades of professional selection.
What is less frequently examined is the extent to which these signals depend on the cost of producing them, and that cost is now collapsing in ways the hiring process has not yet absorbed.
What is less frequently examined is the extent to which these signals depend on the cost of producing them, and that cost is now collapsing in ways the hiring process has not yet absorbed.
The Cost of Legibility
For some individuals, the translation from internal thought to structured expression occurs with relative continuity, and ideas can be stabilised quickly, organised into linear form and communicated without extended effort. For others, the same process requires a different kind of work, in which structure must be actively imposed, language carefully constructed and coherence maintained through deliberate intervention. The interview, in this sense, does not simply measure capability so much as it measures the ease with which capability can be rendered legible within a particular format. This distinction has always existed, although it has remained largely invisible because the cost of legibility was distributed unevenly but was not itself disputed. It was simply considered part of what it meant to be professional.
Generative systems alter this calculation. They do not necessarily change what people think, yet they change the effort required to make thought presentable, and that change is sufficient to destabilise the entire architecture of inference on which hiring has historically relied. A well-structured answer may now reflect clarity of thought, effective use of available tools, or some combination of both that the interviewer is not equipped to disentangle. A candidate who appears highly organised in written communication may be drawing on systems that reduce the cognitive cost of producing that organisation, while another candidate who produces less polished responses under the same conditions may be operating without such support, even if their underlying reasoning is no less developed.
From the perspective of the hiring process, these differences are difficult to observe, and the organisation continues to evaluate what it can see. Clarity, coherence and responsiveness remain central to assessment, because they provide a sense of comparability that allows decisions to be made with a degree of confidence even when the underlying processes producing those signals are not fully understood. The result is a growing misalignment, in which organisations believe they are hiring for capability while in practice they are often hiring for the ability to produce signals of capability under conditions that no longer reflect how work is actually performed.
Organisations believe they are hiring for capability while in practice they are often hiring for the ability to produce signals of capability under conditions that no longer reflect how work is actually performed.
The Industry Responds, Awkwardly
The response from large employers has been notable less for its coherence than for its visible internal contradiction. Goldman Sachs, in mid-2025, asked student applicants to refrain from using ChatGPT during interviews, encouraging them instead to study the firm in preparation, while Anthropic and Amazon went further, indicating that detected use of generative tools during the application process could lead to disqualification.1 The same firms, in the same period, were issuing internal mandates that staff embed AI into their daily work, with Goldman rolling out a proprietary assistant and Amazon publicly championing reflexive AI use across functions.2 The implicit message to candidates was that the very behaviour they were expected to exhibit on day one would, if exhibited a week earlier, mark them as unsuitable.
At the other end of the spectrum, a small but conspicuous group of technology companies has chosen the opposite path. Canva published a deliberately provocative engineering post in mid-2025 titled "Yes, You Can Use AI in Our Interviews. In fact, we insist," redesigning its technical assessment to evaluate how candidates collaborate with generative tools rather than whether they can perform tasks unaided.3 Shopify, through Tobi Lütke's widely circulated internal memo on "reflexive AI usage," extended this logic to recruitment by treating fluency with AI as a baseline expectation rather than a disqualifier.4 Meta and a growing cohort of mid-sized firms have begun to experiment with similar formats, particularly for engineering and product roles, in which the candidate is asked to think aloud while working alongside a model, and the assessment focuses on how they prompt, how they verify and how they correct what the system produces.
These two postures are not merely different policies. They reflect two incompatible theories of what an interview is for. The first treats the interview as a measurement of the unaided individual, and the second treats it as a simulation of the augmented worker. Both can be defended in isolation, yet within the same labour market they produce candidates who are optimising for opposite signals and recruiters who are evaluating for criteria that are no longer commensurable.
The implicit message to candidates was that the very behaviour they were expected to exhibit on day one would, if exhibited a week earlier, mark them as unsuitable.
The Cheating Discourse and Its Misdirection
A parallel conversation has emerged around what is often described as cheating in technical interviews, although the framing itself reveals more about the assumptions of the hiring process than about the behaviour it seeks to police. Industry analyses published in early 2026 estimated that as many as four in ten candidates in remote technical interviews were drawing on generative tools in some form, ranging from second-screen ChatGPT use to dedicated overlay applications designed to feed real-time suggestions during coding tasks.5 Business Insider documented hiring managers across the technology sector who described the situation as broken, with traditional coding interviews producing candidates who performed flawlessly under evaluation and then struggled visibly once on the job.6
The instinctive response has been to invest in detection, with platforms now offering keystroke analysis, gaze tracking and tab-switching telemetry, and a small industry has formed around the promise of restoring the integrity of the unassisted interview. The structural problem is that the very capability being detected, the use of generative systems to think alongside, is the capability the same employer will require the moment the candidate is hired. The process is therefore expending considerable effort to filter out behaviour it cannot afford to lose, and it is doing so in defence of a measurement model that has already lost its validity.
Beneath the discourse of cheating sits a quieter question that few hiring functions are willing to confront. If a candidate can produce excellent work with the support of available tools, in the same conditions that will define the actual role, the persistence of an evaluation format that explicitly excludes those tools begins to look less like rigour and more like nostalgia. The interview continues to test for a version of work that no longer exists.
The interview continues to test for a version of work that no longer exists.
Targeted Programmes and the Limits of Selective Inclusion
A third strand of recent practice has approached the same problem from a different angle, focusing not on the use of tools but on the cognitive profiles that the standard interview tends to filter out. Palantir's Neurodivergent Fellowship, launched in late 2025 and reportedly drawing more than two thousand applicants within its first weeks, accepts self-identification rather than formal diagnosis and explicitly seeks candidates whose pattern of attention, hyperfocus or non-conventional reasoning is poorly served by traditional interview formats.7 SAP's Autism at Work programme, now more than a decade old, has documented sustained productivity outcomes when standard interviews are replaced with extended observational placements and skill-based exercises performed under realistic conditions.8 Microsoft's research with EY, published in late 2024 and reinforced through 2025, found that employees with disability or neurodivergence reported disproportionately large productivity gains from generative tools, with eighty-five per cent indicating that such tools made the workplace meaningfully more inclusive.9
These initiatives are valuable, although they reveal a structural inconsistency that becomes more visible the more they succeed. Each operates as a parallel track, sitting alongside a primary recruitment system that remains largely unchanged. The implicit logic is that the standard process is fit for purpose for the majority and requires accommodation only at the margins, when the more accurate observation is that the standard process is increasingly poorly calibrated for everyone, and the targeted programmes are demonstrating, in miniature, what a better calibration would look like for the organisation as a whole.
The question that targeted programmes do not yet ask, although it is implicit in their design, is why the operational principle that informs them, which is that capability is best assessed under conditions resembling the actual work, has not been applied to recruitment in general. The answer is partly that change at the level of the entire hiring system is expensive, partly that the existing process produces an appearance of comparability that nothing has yet replaced, and partly that the assumptions on which it rests have not been articulated with sufficient clarity to be challenged directly.
The standard process is increasingly poorly calibrated for everyone, and the targeted programmes are demonstrating, in miniature, what a better calibration would look like for the organisation as a whole.
The Older Ghost in the Machine
It is worth recalling that the discomfort with synthetic signals in hiring is not new. When Amazon abandoned its internal résumé screening model in 2018, after Reuters reported that the system had learned to penalise applications containing the word "women's" and had downgraded graduates of women's colleges, the lesson was widely read as a cautionary tale about algorithmic bias.10 The deeper lesson was different. The system had not malfunctioned. It had learned, with considerable accuracy, to reproduce the historical pattern of which candidates the company had previously hired, and it had done so by privileging the legible markers of that pattern. The technology made visible what the human process had always done, and the visibility was uncomfortable enough that the technology was discontinued, while the underlying process continued largely intact.
Something similar is now occurring with generative systems, although in reverse. Where the earlier wave of recruitment AI exposed how heavily human selection had relied on biased signals, the current wave is exposing how heavily it has relied on signals whose meaning depends on conditions that have changed. HireVue's quiet retirement of facial expression analysis in 2021, after independent audits questioned the validity of inferring competence from micro-expressions, prefigured the same pattern.11 Each cycle of automation has revealed something about the model of capability that hiring had implicitly adopted, and each cycle has been received as a problem with the technology rather than as a question about the model.
Each cycle of automation has revealed something about the model of capability that hiring had implicitly adopted, and each cycle has been received as a problem with the technology rather than as a question about the model.
The Selection of Visible Performance
Across the cases described above, a common pattern emerges. The hiring process continues to evaluate forms of performance that were once reliable proxies for capability and that are now substantially modifiable through systems whose use is either expected or excluded depending on the policy of the firm. Candidates who are adept at working with generative systems may present with a level of clarity and structure that exceeds what they would produce unaided, and those who are less inclined to use such systems, or who operate in ways that do not align neatly with their outputs, may appear comparatively less prepared. The hiring process is not designed to interrogate this distinction, and it assumes that what is presented reflects how the individual will operate, even as the conditions of work themselves are becoming less amenable to that assumption.
The connection to the broader argument of this series should be visible at this point. The earlier essays have suggested that generative systems make certain forms of cognition economically visible for the first time, that the outward markers of professionalism are increasingly reproducible without the underlying internal organisation they were once thought to indicate, and that organisational structures continue to filter for linearity even as the work itself becomes less linear.12 Recruitment is the point at which all three of these dynamics converge most clearly, since it is the moment at which an organisation decides whom to admit on the basis of signals that are simultaneously becoming more visible, more reproducible and more dependent on conditions that the organisation has not adapted.
The consequence is a form of selection that is increasingly misaligned with the conditions of work. Organisations may believe they are identifying the most capable candidates, while in fact they are identifying those who perform most convincingly within an evaluative model that is gradually losing its relevance. The system continues to function. What begins to erode is its precision, and the erosion is unlikely to correct itself, because the signals remain even as their meaning shifts.
Organisations may believe they are identifying the most capable candidates, while in fact they are identifying those who perform most convincingly within an evaluative model that is gradually losing its relevance.
Three Plausible Futures
Drawing the threads of this series together, three broadly distinct trajectories appear plausible for how organisational selection may evolve over the coming decade, and each carries different implications for the kinds of capability that organisations will end up holding.
The first is gradual adaptation, in which a growing number of firms follow the trajectory of Canva, Shopify and parts of Meta, and the interview is reconceived as a working session in which generative tools are present by default. Assessment shifts from the production of unaided answers to the quality of the candidate's interaction with available systems, including how problems are framed, how outputs are verified and how judgement is exercised in the presence of plausible but uncertain machine suggestions. This trajectory preserves the existing recruitment infrastructure while updating its content, and it is the path most consistent with how work is already being performed inside the same organisations.
The second is bifurcation, in which the labour market splits into two parallel selection regimes that coexist without converging. Certain employers, particularly in regulated industries and in roles where unaided cognitive performance is treated as a marker of trust, double down on detection and on the unassisted interview, while others, particularly in technology and creative sectors, openly embrace augmented assessment. Candidates increasingly choose between markets rather than between firms, and the long-term consequence is a divergence in the kinds of capability that each market accumulates, with the unaided market gradually selecting for ease of legibility rather than for depth, and the augmented market for fluency with systems rather than for structural understanding.
The third is structural erosion, which is the trajectory that follows if neither adaptation nor bifurcation is pursued with sufficient deliberation. The existing recruitment process continues, modified at the edges by detection technology and supplemented by a growing constellation of targeted programmes for cognitive profiles it cannot accommodate. The signals it relies on become progressively less informative, the gap between interview performance and on-the-job performance widens, and organisations increasingly find that the candidates they have chosen are not the candidates they need, although the reasons for the mismatch remain difficult to articulate within the language of the hiring process itself. This trajectory does not produce a crisis so much as a slow loss of selective precision, which is harder to diagnose and considerably harder to reverse.
The most likely outcome, in practice, is a combination of all three, distributed unevenly across sectors and regions, with adaptation concentrated in firms whose senior leadership has internalised the change in the cost structure of cognition, bifurcation visible at the boundaries between industries with different regulatory and cultural inheritances, and erosion quietly underway in the broad middle, where the assumptions of the process remain intact even as their empirical foundation gives way.
Erosion is quietly underway in the broad middle, where the assumptions of the process remain intact even as their empirical foundation gives way.
What Is Being Selected, and What Is Being Required
The question that recruitment now faces is not whether to permit the use of generative systems during interviews. That question is increasingly residual, since the systems are present whether or not the process acknowledges them. The more demanding question is what an organisation believes it is selecting for, and whether those beliefs still correspond to what the work requires. Hiring decisions continue to be made, roles are filled, and work progresses, yet the precision of the matching between candidate and role depends on a model of capability that was assembled when the cost of producing legible thought was high and individually borne. That cost is no longer high in the same way, and it is no longer borne in the same way, and a process that has not absorbed this change is selecting for a version of capability that the work itself is leaving behind.
Until that recognition occurs, recruitment will continue to operate as though professionalism, clarity and structure were stable indicators of capability, even as the conditions under which those signals are produced continue to shift. What is being selected may therefore diverge, gradually but persistently, from what is actually required, and the divergence is unlikely to correct itself through the application of more rigorous detection, more polished interview formats or more elaborate competency frameworks. It is the model of capability itself that requires examination, and that examination is, at present, conspicuously absent from the discourse on hiring.
Questions this essay answers
Why is recruiting particularly exposed to this shift?
Recruiting depends almost entirely on the diagnostic value of the artefacts that candidates produce within a short evaluation window. Generative systems compress the cost of producing those artefacts, which makes the assessments less informative about underlying capability and more informative about access to the systems.
What contradictions are visible in the current behaviour of major firms?
Several major firms have publicly endorsed the use of generative systems by their employees while simultaneously prohibiting their use by candidates in selection processes. The contradiction is not sustainable in the long term and signals an unresolved question about what the firm intends to evaluate.
Which of the three futures is most likely?
Bifurcation between firms is the most plausible near-term outcome, with some firms adapting selection toward longer observation windows and others retreating to credentials and references. Structural erosion of the assessment paradigm becomes plausible only if the diagnostic gap continues to widen for several more years.
Sources
- Sahota, P., "Goldman Sachs wants students to stop using ChatGPT in job interviews with the bank," Fortune, 11 June 2025.
- Chan, B. and Alexander, R., "How 7 Goldman Sachs Employees Are Putting AI to Work on the Job," Business Insider, 18 April 2025.
- Newton, S., "Yes, You Can Use AI in Our Interviews. In fact, we insist," Canva Engineering Blog, 11 June 2025.
- Lütke, T., "Reflexive AI Usage," internal Shopify memo, March 2025, reproduced at shopify.com/careers.
- Vijayvergiya, A., "State of AI Interview Cheating in 2026: Insights from 19,368 Interviews," Fabric, March 2026.
- "Hiring for tech jobs is broken: Cheating on Tech Interviews Is Soaring," Business Insider, April 2025.
- Scammell, R., "Palantir's Alex Karp Launches a Fellowship for Neurodivergent People," Business Insider, 8 December 2025; follow-up reporting on 2,000+ applications, 11 December 2025.
- "SAP's Autism at Work Program Celebrates 10 Years of Success," SAP Community, 2023; "The Neurodiversity in SAP Report 2024," Bluewave Select.
- Microsoft and EY, GenAI for Accessibility: More Human, Not Less, December 2024; Microsoft WorkLab, "AI Data Drop: How AI breaks down barriers to inclusivity," December 2024.
- Dastin, J., "Insight - Amazon scraps secret AI recruiting tool that showed bias against women," Reuters, 10 October 2018.
- Kahn, J., "HireVue drops facial monitoring amid A.I. algorithm audit," Fortune, 19 January 2021.
- See in particular essays SM1 (Economic Visibility of Minds), SM3 (The Rise of Synthetic Professionalism) and SM5 (Organisations Still Built for Linear Minds) in this series.
About this Series
This article is part of the Synthetic Minds at Work series exploring how generative AI is reshaping the role of cognition in organisations. The structural questions raised here also connect to the Decision Architecture series, in particular to the redistribution of judgement and the design of organisational closure.
