Fifty applications is not an unusual number. Six months is not an unusual stretch. What makes Amanda Bowler's story worth stopping over is the third number, the one that sits at zero: face-to-face interviews. Not zero offers. Zero occasions on which a human being sat opposite her and formed a view. The 48-year-old psychology graduate, whose experience was reported by the Sydney Morning Herald on 29 August 2026, is carrying a HECS debt of roughly 44,000 Australian dollars, a figure inflated by the fact that registration as a psychologist in Australia requires years of postgraduate study stacked on top of the undergraduate degree. She paid for the qualification. She wrote the applications. And on the available evidence, the qualification was assessed by software and the applications were read by nothing.
That last sentence is the one that ought to be arresting, and the reason it is not is that we have grown accustomed to it in under three years. The Australian labour market she is applying into is genuinely tight. SEEK's employment report for July 2026, published on 12 August, recorded applications per job advertisement at their highest level on record, having risen for seven consecutive months while job advertisement volumes fell away beneath them, down 0.4 per cent month on month in July and 6.0 per cent lower than a year earlier. More people are chasing fewer roles. That is an old story and a cyclical one.
The new story is what is standing between them. The Australian Responsible AI Index, the national benchmark produced by Fifth Quadrant with the National AI Centre, found that a majority of Australian organisations already use AI in recruitment to some degree, and the Herald's reporting puts the figure at roughly 80 per cent of Australian companies at moderate or higher levels of use. On the other side of the same transaction, jobseekers now deploy generative systems and autonomous agents that scrape boards overnight, rewrite the curriculum vitae for each posting, generate the covering letter and submit before the applicant has woken up. A senior human resources leader, quoted in the same Herald report, described where this ends with more precision than most white papers manage. The bots just botting each other. Little robots yelling at each other.
Refusing Both of the Easy Stories
There are two ready-made articles about this and both are worthless.
The first is the technophobic one, in which the machines have desecrated a sacred human ritual and the remedy is to bring back the handshake. This is sentimental nonsense. Human hiring was never a meritocracy. It was slow, inconsistent, swayed by schools and accents and surnames, and it discriminated with a fluency that no audit ever caught because nobody was auditing. Anyone nostalgic for the CV pile on a hiring manager's desk should reckon with what happened to CVs at the bottom of it on a Friday afternoon.
The second is the efficiency story, in which volume has become unmanageable, automation is the only rational response, and since candidates are automating too the whole thing is symmetrical and therefore fair. This is more sophisticated and more dangerous, because the symmetry is fake. A candidate's agent optimises one person's chances against an entire market. An employer's agent optimises a queue against a budget. Only one of those produces an output with legal and material consequence. The candidate's machine issues a request. The employer's machine issues a verdict. Calling that an arms race between equals is like calling a trial an argument between the defendant and the judge.
The harder and more specific argument is this. The injury is not that judgement has been automated. Judgement has always been delegated, compressed and rushed. The injury is that the loop has closed: there is now no stage in the process at which any human being with time and authority is obliged to read. Once no human reads, the decision loses its author. And a decision without an author cannot be explained, cannot be appealed, cannot be corrected, and cannot be learned from. What Bowler encountered fifty times was not harshness. It was the absence of anybody to be harsh.
The Filter Myth and the Thing That Quietly Replaced It
Before going further it is worth demolishing the folklore, because the folklore is doing real damage to the argument.
For over a decade, career advice has been built around the claim that applicant tracking systems automatically bin 75 per cent of CVs before a human sees them. The statistic has no research behind it. As Leda Salazar de Leon and Mehnaz Rafi set out in an analysis published in July 2026, it traces to a 2012 sales pitch by Preptel, a résumé-optimisation vendor that closed the following year, and no methodology was ever published. The company selling the cure invented the disease. Recruiter surveys since have found the overwhelming majority do not configure their tracking systems to reject automatically on formatting or keywords, and small businesses, which employ most of the workforce in most developed economies, frequently do not use such systems at all. Applicant tracking systems were, for most of their history, filing cabinets with a search box.
So an entire generation of jobseekers spent a decade terrified of a filter that mostly did not exist, buying templates and keyword tools to defeat it. That matters here for two reasons. First, it should make anyone sceptical of vendor-generated panic, including panic that flatters this article's own thesis. Second, and more uncomfortably, the filter has now actually arrived, and it arrived in precisely the place the myth said it was, which means the warning has been worn out by overuse at exactly the moment it becomes true.
Large language models genuinely do rank. They produce orderings, and the orderings are not neutral. Kyra Wilson and Aylin Caliskan of the University of Washington ran a résumé audit through massive text embedding models, using more than 500 real CVs and 500 job descriptions across nine occupations, generating over three million combinations of job, race and gender. White-associated names were preferred in 85.1 per cent of cases and Black-associated names in 8.6 per cent. Female-associated names were preferred in 11.1 per cent of cases. For some intersections, notably Black men, the models preferred other candidates in close to 100 per cent of comparisons. This is not a system that occasionally misfires. It is a system reproducing the statistical regularities of a corpus, doing exactly what it was built to do, at the precise point in the process where the myth taught everyone to expect it and nobody has been checking.
Two Automations Racing Each Other Into a Wall
On the employer side, the screening layer has stopped being software and become a participant. Alex, formerly Apriora, is a Y Combinator company founded by Aaron Wang and John Rytel that conducts live voice and video screening interviews, scores candidates and returns a ranked shortlist. It raised a 17 million dollar Series A led by Peak XV Partners in September 2025, taking total funding to 20 million, and reports having run more than a million AI-led screening interviews, reaching up to 5,000 in a single day for its largest customers. Five thousand first-round interviews in a day is not a faster version of what recruiters used to do. It is a different activity with the same name.
On the candidate side, the tooling is scrappier, cheaper and open. Public repositories host agents that use browser automation and web scraping to read postings and auto-apply with a tailored CV and covering letter for each one. Commercial equivalents advertise applying to thousands of roles in a click. Abhijay Arora Vuyyuru, a product manager at Google's YouTube who writes a widely read newsletter on applied AI, has published a step-by-step guide to building a personal job-hunting agent that runs on a small cloud server, pulls fresh listings each morning, scores each against your CV, and messages you the top five. Speaking to the Herald, he predicted that within five years every jobseeker will have a personal AI agent, and that it will most probably be talking to the company's AI recruitment agent. He is not forecasting from the sidelines. He ships the thing.
The consequences are measurable, and Greenhouse has measured them. Its AI in Hiring report, published on 19 November 2025 and drawing on 4,136 respondents across the United States, United Kingdom, Ireland and Germany, split between 2,900 jobseekers and 1,236 recruiters and hiring managers, found that 49 per cent of candidates had submitted more applications than the year before and that 54 per cent had already been interviewed by a machine. Six months later the company asked again. Its 2026 candidate report, published on 1 May and surveying 2,950 active jobseekers across the United States, United Kingdom, Germany, Australia and Ireland, put the share at 63 per cent, a rise of thirteen percentage points in half a year. Whatever this is, it is not settling down. The earlier report also found the trust gap that defines the whole system: 70 per cent of hiring managers trust AI to make faster and better hiring decisions, while just 8 per cent of jobseekers believe AI makes hiring fair. Daniel Chait, the company's chief executive and co-founder, called it an AI doom loop that is getting worse, not better.
Eight per cent. That is not scepticism. That is a population that has concluded the process is illegitimate, still participating in it because it has no alternative.
When the Signal Stops Meaning Anything
There is a clean piece of economics underneath this, and it was written in 1973.
Michael Spence's job market signalling model explains why credentials work at all. A signal only carries information if it is differentially costly to produce: education signals ability because it costs more, in effort and time, for the less able to acquire. Take away the cost differential and the signal stops transmitting. It does not become a weaker signal. It becomes noise wearing a signal's clothes.
Generative AI drove the cost of producing a competent, tailored, keyword-aligned application to approximately zero, for everyone, simultaneously. The considered covering letter used to be a costly signal of interest. It is now free. Which means it now says nothing, and everybody involved knows it says nothing, and each side has responded by escalating rather than by admitting the instrument is broken.
The escalation has a measurable shape, and the most revealing number in this entire field lives in the gap between two studies. Greenhouse found that 41 per cent of jobseekers said they had used prompt injection, hiding instructions in a document to steer an AI screener, with 52 per cent of the remainder saying they were considering it. Then Mohan Zhang, Yuqi Jia, Zhen Tan, Steven Jiang, Neil Zhenqiang Gong, Tianlong Chen and Dawn Song went and looked. Their study, presented at the USENIX Security Symposium in 2026, is the first large-scale empirical measurement of prompt injection in a real deployed system: roughly 200,000 genuine CVs collected over several years, of which about 1 per cent contained hidden injections. Prevalence is rising. More than 90 per cent of the injections avoided explicit instruction syntax, which is to say they were subtle enough to be difficult to catch.
Forty-one per cent say they do it. One per cent do it. That gap is not a measurement error, and it is not really about cheating. It is a statement of belief. Two in five jobseekers now consider sabotaging the employer's screening system to be a normal and defensible thing to say out loud about themselves. The moral standing of the hiring process collapsed considerably faster than the technology arrived.
Employers, deprived of a working signal, are hunting for a new costly one, which turns out to be proof that the applicant is a person. Gartner, in research released on 31 July 2025 drawing on surveys of some three thousand candidates, predicted that by 2028 one in four candidate profiles worldwide will be fake, found that 6 per cent of respondents admitted to some form of interview fraud, that 39 per cent had used AI during the application process, and that only 26 per cent trusted AI to evaluate them fairly. The vendor response has been identity verification products bolted onto applicant tracking systems, selfie-based checks, and a documented drift back towards in-person interviews for no reason other than confirming the candidate exists.
Read that sequence again, because it is close to farcical. The industry removed the human reader in order to save the cost of reading. It has now reinstated costly human verification in order to establish that there is a human on the other end. The cost was never eliminated. It was moved downstream, past the point at which Amanda Bowler had already been rejected fifty times.
What Researchers Found Inside the Loop
Three recent studies get closer to the mechanism than any amount of vendor commentary, and together they say something the industry has not absorbed.
The first, by Aditya Bhattacharya and Katrien Verbert, published in May 2025, built a multi-agent system powered by large language models to guide jobseekers through the recruitment process and explain hiring decisions to them. Evaluated with 20 participants, it was found significantly more actionable, trustworthy and fair than the conventional process. Note what the candidates responded to. Not a better weapon. An explanation. The enormous demand currently being met by auto-apply tools is, at root, demand for legibility, and it is being served by the only vendors willing to sell anything at all to the person being screened.
The second, by Md Nazmus Sakib, Naga Manogna Rayasam and Sanorita Dey, submitted in January 2026 and revised in March, examined asynchronous AI interviewers through analysis of subreddit discussion and interviews with 17 participants, then tested interface changes with 180 more. Its central finding is that applicants suffer from mismatched expectations amplified by organisational rhetoric, and that familiarity with large language models shaped how candidates perceived the process, sometimes producing workarounds and deceptive practices. That causal direction matters enormously. Deception is not a character flaw distributed randomly across the applicant pool. It is a predictable behavioural response to a process that misrepresents itself, and it is elicited most reliably from the candidates who understand the technology best.
The third is the one that should end a particular argument for good. Sajel Surati, Rosanna Bellini and Emily Black interviewed 22 recruiting professionals about generative AI in their daily workflows. Recruiters believed they retained final decision-making authority. The researchers found that the AI had become, in their phrase, an invisible architect shaping the foundational building blocks of the information used for evaluation. Adoption decisions largely lay outside recruiters' control, driven by organisational pressure and competitive anxiety rather than professional judgement. And the marginal efficiency gains came at considerable cost, including reduced recruiter expertise and compromised oversight capacity.
This dismantles the human-in-the-loop defence, which is the single most common answer employers give when challenged. The human is in the loop. She is reading a summary the model wrote, of candidates the model ranked, drawn from a pool the model already narrowed, under time pressure the model's throughput created, having gradually lost the expertise that would let her notice when it is wrong. Her presence satisfies an organisational chart and nothing else. Oversight that cannot be exercised is not oversight, and a rubber stamp is not a reader.
The Particular Exposure of Being Forty-Eight
Everything above lands unevenly, and it lands hardest on people who look like Amanda Bowler.
The Australian HR Institute and the Australian Human Rights Commission run a periodic national survey of employers on age in the workplace. Their 2025 report found that 24 per cent of HR professionals classify workers aged 51 to 55 as older, up from 10 per cent in 2023, and that only about half of employers were open to a large extent to hiring people over 50. That is the human baseline into which the automation was installed, and it is worth naming plainly: employer age bias was already there and already growing before a single model was deployed.
What the model adds is scale, speed and deniability. But there is a subtler and more structural problem with automated matching that has nothing to do with inherited prejudice. A ranking system trained on the profiles of people who already hold a role learns what an incumbent looks like. A career changer, by definition, does not look like an incumbent. A 48-year-old psychology graduate has a curriculum vitae with a shape that no successful applicant for an entry-level psychology position has ever had, because successful applicants for those positions are usually 25. The model will score her as a poor match. It will be functioning perfectly. Not a sufficiently close match is not a malfunction, it is the product specification, and it encodes a definition of suitability that nobody in the organisation ever wrote down, approved or could defend if asked.
The most consequential legal test of this proposition is grinding through the Northern District of California. Mobley v. Workday was filed on 21 February 2023 by Derek Mobley, an African American man over 40 with depression and anxiety, who alleged he had been rejected from more than a hundred positions screened through Workday's software, in several instances outside business hours and within an hour of applying. In July 2024, Judge Rita Lin preserved the disparate impact claims, finding it plausible that the vendor was acting as an agent of the employers. On 16 May 2025 the court granted preliminary certification of a collective under the Age Discrimination in Employment Act. In July 2025 the collective was expanded to cover applicants screened using Workday's HiredScore features. The notice plan was approved on 2 December 2025, collective notice was formally authorised on 17 February 2026 with an opt-in deadline of 7 March, and on 6 March Lin rejected Workday's argument that the age discrimination statute does not reach job applicants at all. A third amended complaint followed on 27 March 2026, and on 22 June the court granted in part and denied in part Workday's motion to dismiss it, leaving discrimination claims alive across race, sex, age and disability.
Two figures from the reporting on that case deserve to sit next to each other. Around a billion applications were rejected through the software during the relevant opt-in period. Roughly 14,000 people opted in. That window is now shut, which makes 14,000 not a running tally but a final number, and the ratio it forms with a billion is the accountability gap rendered numerically. Almost nobody rejected by an automated system ever learns enough to raise a hand, because the system's defining feature is that it does not tell you anything happened.
The Law Arrives, and Then Steps Back
You would expect regulation to be closing on this. In the specific weeks around Bowler's fifty applications, it moved backwards.
Under the European Union's AI Act, artificial intelligence used in employment and worker management is classified as high risk under Annex III, and the substantive obligations were due to apply from 2 August 2026, four weeks before this article was written. They did not. The Digital Omnibus on AI, provisionally agreed on 6 May 2026 and confirmed by member state representatives on 13 May, was adopted as Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 and in force from 27 July, five weeks before this article was written and days before the deadline it removed. This is not a proposal that might yet be resisted. It is settled law. The obligations for stand-alone Annex III systems now apply from 2 December 2027, with embedded Annex I systems slipping to 2 August 2028. What survives on the original timetable is Article 50, the transparency duty to disclose that a person is interacting with an AI system, alongside the Article 5 prohibitions. Article 86, which grants an affected person the right to obtain clear and meaningful explanations of the role an AI system played in a decision that adversely affects them, hangs off the high-risk regime and slips with it.
The resulting settlement is almost satirical. A European candidate rejected today has a legal right to be told a machine was involved, and a fifteen-month wait for any right to be told what it did.
The American picture offers a preview of what enforcement looks like when it does arrive. New York City's Local Law 144 requires annual independent bias audits of automated employment decision tools, public posting of a summary, and advance notice to candidates. Researchers led by Lucas Wright, with Roxana Mika Muenster, Briana Vecchione, Tianyao Qu, Pika Cai, Alan Smith, Jacob Metcalf and J. Nathan Matias, coordinated 155 trained student investigators to check 391 employers in late 2023, behaving as motivated jobseekers would. Eighteen employers, 5 per cent, had posted a bias audit report. Thirteen, 3 per cent, had posted a transparency notice. Eleven had posted both. The researchers coined the term null compliance for the resulting condition, in which non-compliance cannot even be established because the regulated party controls whether anyone knows the law applies to it. The New York State Comptroller's audit published on 2 December 2025, covering July 2023 to June 2025, found the city's enforcement of the law ineffective.
Illinois has gone further on paper. House Bill 3773, enacted as Public Act 103-0804 and effective from 1 January 2026, amends the state Human Rights Act to make discriminatory use of AI in recruitment, hiring and promotion a civil rights violation, prohibits the use of postcodes as proxies for protected characteristics, and requires employers to notify people when AI is used in employment decisions. It sits alongside the older Artificial Intelligence Video Interview Act. Whether it bites is a question about enforcement capacity, and New York's experience is not encouraging. Neither is Illinois's own. The Department of Human Rights published proposed regulations on 15 May 2026, under the heading Subpart J: Use of Artificial Intelligence in Employment, and then on 2 June postponed the rulemaking in order to coordinate with other agencies. No revised timeline has been announced. The statutory obligations have been enforceable throughout, which produces the peculiar situation of an employer bound since New Year's Day to notify candidates that AI is being used in a decision about them, with no final rules describing what an adequate notice contains, and a candidate holding a right whose shape the regulator has not yet been able to state. A law that is in force and unexplained is not the opposite of a law that is unenforced. It is a variation on it.
And Australia, where Bowler is applying, has chosen to do less rather than more. In September 2024 the Department of Industry, Science and Resources published a proposals paper setting out ten mandatory guardrails for AI in high-risk settings, covering testing, transparency and accountability for developers and deployers. In December 2025 the National AI Plan shelved them, opting instead to rely on existing technology-neutral law, voluntary guidance folded into six essential practices, and a new AI Safety Institute. The practical answer available to Amanda Bowler is therefore the Age Discrimination Act 2004 and state anti-discrimination statutes, which she would have to invoke on the basis of no information whatsoever about what happened to her application.
One genuinely enforceable Australian obligation is coming. From 10 December 2026, under amendments made by the Privacy and Other Legislation Amendment Act 2024, entities subject to the Australian Privacy Principles that use personal information in automated decisions capable of significantly affecting a person's rights or interests must disclose in their privacy policy the kinds of information used and the kinds of decisions made, with the Office of the Australian Information Commissioner able to enforce. That is real, and it is worth having. It is also a paragraph in a document nobody reads, published by the organisation, containing no information about any individual decision. It tells Bowler that a category of thing happened to a category of person. It does not tell her what happened to her.
There is a reason this reticence should alarm Australians more than most. Between 2015 and 2019 the Commonwealth ran an automated debt recovery scheme against welfare recipients using crude income averaging. Robodebt was unlawful, produced enormous error rates, and a Royal Commission reported on it in 2023 with 57 recommendations. The financial reckoning continues: on 4 September 2025 the government agreed to pay a further 475 million dollars to settle the appeal in Knox v The Commonwealth, approved by the Federal Court on 23 June 2026, on top of the earlier Prygodicz settlement, bringing total redress past 2.4 billion dollars. Australia has already run the experiment of automating a high-stakes determination about people's lives at national scale and discovering the error rate afterwards, from the wreckage. Having paid that bill, in money and in lives, its response to the automation of hiring is a set of voluntary practices.
What a Right to a Human Decision Would Actually Have to Contain
The phrase a right to a human decision is currently doing rhetorical work and no operational work. It is worth spelling out, because vaguely stated it will be satisfied by a rubber stamp, and the SCHUFA reasoning already tells us a rubber stamp does not count. In that case, decided by the Court of Justice of the European Union in December 2023, the score the credit agency produced was itself held to be the decision in law, because the German bank relying on it did no more than follow the output. Where the nominal decision-maker adds nothing, the machine's output is the decision.
Six things would make it real, and the first principle underlying all of them is that a review conducted by someone with no practical capacity to overturn the outcome is not a review at all.
Disclosure has to happen at the point of application, not in a privacy policy. Before submitting, the applicant should see plainly that the application will be machine-scored, and what the system is assessing. This costs nothing and changes behaviour immediately, because a candidate who knows can decide whether the role is worth two hours of their life. It is also the reform with the clearest measured demand behind it and the widest gap between that demand and current practice. Of the jobseekers Greenhouse surveyed in 2026 who had been interviewed by an AI, 70 per cent were never told plainly beforehand that a machine would be assessing them, and for 21 per cent the fact only surfaced once the interview had started. Fifty-seven per cent think disclosure ought to be a legal requirement. This is not a protection that has to be explained to the people it protects.
Rejections issued without human involvement must be reversible on request. Not appealable in principle, reversible in practice, by a named person with authority to overturn, within a fixed window. The thirty-second rejection that jobseekers describe as soul-destroying is not painful because it is fast. It is painful because its speed is a disclosure: it tells the applicant that nothing was considered. A reversal route converts the machine's output from a verdict into a triage recommendation, which is what it always should have been. Here too the appetite is documented rather than assumed: 46 per cent of jobseekers want a human interview available as an alternative, and 38 per cent want a human being to look at the outcome before the decision is taken.
Employers should be required to sample. A fixed minimum percentage of machine-rejected applications, chosen at random, read by a human being, with the pass-through rate published. This is the single most valuable intervention available, because it generates a continuous audit of the model's false negatives at the employer's own expense and makes those false negatives visible for the first time. On a role attracting 500 applications, a 5 per cent sample is 25 CVs. That is one morning. Any organisation arguing it cannot afford a morning is telling you the vacancy was not worth filling.
Reasons must be given, and a score is not a reason. A short statement of which requirement was assessed as unmet is sufficient, and it is the difference between a decision and an event.
Auto-rejection rates should be published by age band, and by other protected characteristics where the data exists. Employers already collect this information for equal opportunity reporting. The reason it is not published broken down by automated stage is that nobody has asked.
And verification should be portable. If the labour market now requires proof that an applicant is a real human with the credentials claimed, that proof should be established once, held by the candidate, and presented on demand, rather than repeatedly extracted by every employer and every vendor at the candidate's cost.
It is worth saying plainly that none of this is charity. Thirty-eight per cent of jobseekers have already walked out of a hiring process because it involved an AI interview, and a further 12 per cent say they would. That is half the candidate pool either gone or standing at the door with a hand on the frame, and it is a cost borne by the employer, in the form of the applicants it never gets to see. Only 21 per cent believe most employers are deploying these systems responsibly. Yet only 19 per cent want less AI in hiring overall, and those two findings sitting side by side are the most useful thing in the whole survey. Candidates are not objecting to the machine. They are objecting to the machine being unaccountable, and they are quite capable of telling the difference. An employer that discloses, offers a human alternative and reverses on request is not making a concession to sentiment. It is recovering the third of its applicant pool that the rest of the industry is currently losing before the first question is asked.
None of this is radical, and none of it requires banning anything. It requires accepting a principle that every other consequential decision-making system in a developed economy already accepts: that a decision affecting someone's livelihood must have an author who can be identified and asked.
The Author of the Decision
So, to the questions this began with.
What does it mean when the most pivotal process in a working life is conducted between two artificial intelligence systems exchanging notes about a person's qualifications with no human ever reading the application? It means the decision has no author. Consider the comparison. If a bank refuses Bowler a loan, there is a decision-maker, a reason and a complaints process. If a government agency refuses her a payment, there is a delegate, a statement of reasons and a tribunal. If a university refuses her a place, there is an admissions officer and an appeal. Employment, which determines her income, her housing, her health, her standing and most of her waking hours, has quietly become the largest category of consequential decisions about human beings in a modern economy that has no such person anywhere in it. Not a person who decided wrongly. No person at all.
And what happens to the relationship between employer and employee when the first impression is made by an algorithm that can end a career in thirty seconds? An employment relationship is a structure of mutual obligation that begins in an act of assessment, and the character of that first act sets the terms of everything after it. When the opening move is made by a system neither party controls, that the employer cannot explain and the candidate cannot see, the relationship begins in bad faith on both sides, and the numbers show both sides already know it. Seventy per cent of hiring managers trust the machine. Eight per cent of candidates think it is fair. Forty-one per cent are willing to say out loud that they would sabotage it. That is not a labour market clearing. It is two populations, each correctly convinced the other is not really present, transacting through proxies neither of them chose.
The literature on what this does to the people caught in it is not ambiguous. The systematic review by Tom Sterud, Lars-Kristian Lunde, Rigmor Berg, Karin Proper and Fiona Aanesen, published in Occupational and Environmental Medicine in 2025 and pooling 38 prospective longitudinal studies, found a relative risk of mental health problems among unemployed people of 1.95 compared with the employed, and a relative risk of 0.66 after re-employment. Unemployment damages people and work repairs them, reliably, in both directions. Every unnecessary month a qualified person spends outside work because a ranking model scored their history as unusual is a month of measurable harm, imposed by a process nobody signed.
Amanda Bowler is not owed a job. Fifty applications do not entitle anyone to fifty interviews, and there are more psychology graduates than there are psychology posts. What she is owed, and what the entire apparatus described here has been engineered to avoid providing, is far smaller and far more fundamental than a job.
She is owed a reader.
Sources and References
- Bhattacharya, A. and Verbert, K. (2025). “Let's Get You Hired: A Job Seeker's Perspective on Multi-Agent Recruitment Systems for Explaining Hiring Decisions.” arXiv:2505.20312. https://arxiv.org/abs/2505.20312
- Sakib, M. N., Rayasam, N. M. and Dey, S. (2026). “Expecting Too Much, Getting Too Little: Exploring the Challenges and Design Opportunities of Asynchronous AI Interviewers.” arXiv:2601.02775. https://arxiv.org/abs/2601.02775
- Surati, S., Bellini, R. and Black, E. (2026). “Resume-ing Control: (Mis)Perceptions of Agency Around GenAI Use in Recruiting Workflows.” arXiv:2604.26851. https://arxiv.org/abs/2604.26851
- Zhang, M., Jia, Y., Tan, Z., Jiang, S., Gong, N. Z., Chen, T. and Song, D. (2026). “Measuring Real-World Prompt Injection Attacks in LLM-based Resume Screening.” USENIX Security Symposium 2026. arXiv:2605.28999. https://arxiv.org/abs/2605.28999
- Wilson, K. and Caliskan, A. (2024). “Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval.” Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society. https://arxiv.org/abs/2407.20371
- Wright, L., Muenster, R. M., Vecchione, B., Qu, T., Cai, P., Smith, A., Metcalf, J. and Matias, J. N. (2024). “Null Compliance: NYC Local Law 144 and the Challenges of Algorithm Accountability.” Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency. https://arxiv.org/abs/2406.01399
- Spence, M. (1973). “Job Market Signaling.” The Quarterly Journal of Economics, 87(3), pp. 355-374. https://www.jstor.org/stable/1882010
- Sterud, T., Lunde, L.-K., Berg, R., Proper, K. I. and Aanesen, F. (2025). “Mental health effects of unemployment and re-employment: a systematic review and meta-analysis of longitudinal studies.” Occupational and Environmental Medicine. https://pmc.ncbi.nlm.nih.gov/articles/PMC12505101/
- Greenhouse (2025). “An AI Trust Crisis: 70% of Hiring Managers Trust AI to Make Faster and Better Hiring Decisions, Only 8% of Job Seekers Call it Fair,” 19 November (https://www.greenhouse.com/newsroom/an-ai-trust-crisis-70-of-hiring-managers-trust-ai-to-make-faster-and-better-hiring-decisions-only-8-of-job-seekers-call-it-fair); and “Greenhouse 2026 Candidate AI Interview Report,” 1 May 2026 (https://www.greenhouse.com/blog/2026-candidate-ai-interview-report)
- Gartner (2025). “Gartner Survey Shows Just 26% of Job Applicants Trust AI Will Fairly Evaluate Them,” 31 July. https://www.gartner.com/en/newsroom/press-releases/2025-07-31-gartner-survey-shows-just-26-percent-of-job-applicants-trust-ai-will-fairly-evaluate-them
- SEEK (2026). “SEEK Employment Report, July 2026,” 12 August. https://au.seek.com/about/news/article/seek-employment-report-july26
- Australian HR Institute and Australian Human Rights Commission (2025). “Employer biases against older and younger generations hindering Australian workforce productivity.” https://humanrights.gov.au/about-us/media-centre/search-listing-media-releases/employer-biases-against-older-and-younger-generations-hindering
- Fifth Quadrant and National Artificial Intelligence Centre (2025). “Australian Responsible AI Index 2025.” https://www.fifthquadrant.com.au/content/uploads/Australian-Responsible-AI-Index-2025_Full-report.pdf
- Salazar de Leon, L. and Rafi, M. (2026). “What everyone gets wrong about the modern job search, and what actually works,” The Conversation, 5 July. https://theconversation.com/what-everyone-gets-wrong-about-the-modern-job-search-and-what-actually-works-285582
- Civil Rights Litigation Clearinghouse (2026). “Mobley v. Workday, Inc., 3:23-cv-00770 (N.D. Cal.).” https://clearinghouse.net/case/44074/
- Callaham, S. (2026). “A Federal Judge, A 1967 Law And A Billion Rejected Job Applications,” Forbes, 29 May. https://www.forbes.com/sites/sheilacallaham/2026/05/29/a-federal-judge-a-1967-law-and-a-billion-rejected-job-applications/
- Office of the New York State Comptroller (2025). “Enforcement of Local Law 144: Automated Employment Decision Tools,” 2 December. https://www.osc.ny.gov/state-agencies/audits/2025/12/02/enforcement-local-law-144-automated-employment-decision-tools
- Ogletree Deakins (2026). “Illinois Steps Up AI Regulation in Employment: Key Takeaways for Employers.” https://ogletree.com/insights-resources/blog-posts/illinois-steps-up-ai-regulation-in-employment-key-takeaways-for-employers/
- European Union (2026). “Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated text as amended by Regulation (EU) 2026/1744 (Digital Omnibus on AI), in force 27 July 2026.” EUR-Lex. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:02024R1689-20260727
- Court of Justice of the European Union (2023). “Case C-634/21, SCHUFA Holding (Scoring), OQ v Land Hessen, judgment of 7 December 2023.” EUR-Lex. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62021CJ0634
- Office of the Australian Information Commissioner (2026). “Consultation on Guidance for Transparency in Automated Decision Making.” https://www.oaic.gov.au/engage-with-us/consultations/consultation-on-guidance-for-transparency-in-automated-decision-making
- Montreal AI Ethics Institute (2026). “AI Policy Corner: From proposed mandatory guardrails to the National AI Plan, AI governance in Australia.” https://montrealethics.ai/ai-policy-corner-from-proposed-mandatory-guardrails-to-the-national-ai-plan-ai-governance-in-australia/
- Attorney-General's Department, Australian Government (2025). “Robodebt Class Action Appeal Settlement,” 4 September. https://ministers.ag.gov.au/media-centre/robodebt-class-action-appeal-settlement-04-09-2025
- Alex (2025). “We raised $20M to help AI hire more humans.” https://www.alex.com/blog/we-raised-20m-to-help-ai-hire-more-humans
- Sydney Morning Herald (2026). Report on the use of artificial intelligence in Australian recruitment, including the account of jobseeker Amanda Bowler, the remarks of a senior human resources leader on automated screening, and the predictions of Abhijay Arora Vuyyuru, 29 August. https://www.smh.com.au/
Tim Green UK-based Systems Theorist & Independent Technology Writer
Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.
His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.
ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk
Listen to the free weekly SmarterArticles Podcast
