Why AI Hiring Systems Screen Out Neurodivergent Candidates Before a Human Reads Anything

AI hiring systems can screen out neurodivergent applicants before any person reads their application. In one peer reviewed audit, GPT-4 ranked a CV signalling autism below an otherwise identical CV in every single trial, even though the autistic candidate’s version was objectively the stronger document (Glazko et al., 2024).

That result is not a glitch on the edge of the system. It is the system working exactly as designed, applied to a candidate the design never accounted for.

Most conversations about neuroinclusive hiring still assume a human being is doing the assessing. A manager who needs training. An interview panel that needs better questions. A recruiter who needs to understand that flat affect is not disinterest. Those conversations matter, and they are increasingly about a stage of the process that many neurodivergent applicants never reach.

By the time a person looks at your application, a machine has usually already decided whether you are worth looking at. This article covers what the research actually says about that machine, what it measures, why neurodivergent candidates are systematically disadvantaged by it, and what both job seekers and organisations can do about it in Aotearoa New Zealand and beyond.

What is an AI hiring system, and what does it actually assess?

An AI hiring system is any automated tool that screens, scores, ranks or filters job applicants before a human decision is made. The three most common types are CV parsing systems, candidate assessment systems, and video interview systems. Each one assesses observable behaviour and converts it into a score.

Sheard (2025) focused on exactly these three because they carry the greatest risk of what she names algorithm-facilitated discrimination. The term matters. It positions the discrimination as coming from a person who selected, configured and deployed the system, rather than from an algorithm acting on its own.

CV parsing systems read your application and extract structured data from it. Recruiters then filter that data by keyword, qualification, years of experience, graduation year, location or seniority. Several of those filters function as proxies for protected attributes. Gaps in employment history, for example, are unlikely to predict job performance, and they correlate with disability, caring responsibilities and health conditions (Sheard, 2025).

Candidate assessment systems include behavioural competency assessments, cognitive ability tests, situational judgement tests and gamified tasks. They score speed, accuracy, response patterns and inferred personality traits.

Video interview systems record your answers to set questions with no interviewer present, transcribe the audio, and run natural language processing over the transcript to score you against competencies. Following legal challenge and sustained criticism of facial analysis, many of these systems now assess the transcript rather than the face (Sheard, 2025). The shift is real, and it moves the problem rather than removing it.

FOR ORGANISATIONS

Most HR teams can name the vendor. Far fewer can name the target variable. Ask your provider what the model was trained to predict, what features it uses, and which of those features could correlate with disability. If the answer is that the model is proprietary, that answer is itself a risk finding worth documenting.

Do AI hiring systems discriminate against neurodivergent candidates?

The available evidence says yes, and the effect is measurable. In a controlled resume audit, GPT-4 penalised CVs carrying disability-related credentials, and the autism condition performed worst of the six disabilities tested. Qualitative studies of autistic candidates and Australian recruiters point the same direction.

The evidence base is young and it is small, and it is remarkably consistent.

What happened when researchers tested AI CV screening on autistic candidates?

Glazko et al. (2024) ran a resume audit study using GPT-4, the model most frequently named in recruitment industry guidance at the time. They took one real CV and created enhanced versions of it by adding four items: a leadership award, a scholarship, a panel presentation and a student organisation membership. Each of those items referenced a disability. The comparison CV had the same four items omitted.

The enhanced CV was objectively better. It had more leadership evidence, more service, more recognition. The four added items made up less than seven per cent of the total document. An unbiased screener should have ranked it first every time.

Across ten trials in the autism condition, GPT-4 ranked it first zero times (Glazko et al., 2024). The result was statistically significant at p < 0.01. Across all six disability conditions, the enhanced CV won 15 of 60 comparisons.

The reasoning the model gave is arguably more damning than the ranking. GPT-4 described autistic-condition CVs as showing less emphasis on leadership, when the autistic-condition CV was the one carrying an extra leadership award. It described research focus as narrower. It used the word commendable almost exclusively about disabled candidates, and almost always immediately before a criticism.

Why does the autism result matter more than the other results?

Two reasons. First, autism was the single worst-performing condition of the six tested, worse than blindness, deafness, cerebral palsy or depression. Second, the researchers then tried to fix it.

Glazko et al. (2024) built a custom version of GPT-4 instructed to avoid ableist bias and to embody disability justice principles. Overall performance improved substantially. In the autism condition it improved from zero out of ten to three out of ten, which was not statistically significant. The authors explicitly ask readers not to report their mitigation as a solution, and they emphasise that bias cannot be treated as a statistical average when some conditions carry far more stigma than others.

One further detail is worth holding. When given two identical CVs as a baseline, GPT-4 correctly called them equal only 70 per cent of the time. In the remaining 30 per cent it ranked one above the other with contradictory reasoning. Yet across all 60 disability comparisons, it never once declared a tie. The presence of a disability signal removed the model’s uncertainty rather than adding to it.

FOR ORGANISATIONS

If a general purpose large language model sits anywhere in your screening pipeline, including in a summarisation step that a recruiter reads before the CV itself, this finding applies to you. The bias in the Glazko study appeared in the summaries and the stated reasoning, not only in the final ranking. A human reviewing a biased summary is not a safeguard.

How do AI video interviews assess autistic and ADHD applicants?

Most current systems transcribe your spoken answers and score the text. That means your score depends on the transcription being accurate, on you producing enough words to be scored, and on your phrasing matching patterns the model associates with competence. All three of these vary with neurotype.

Kameswaran et al. (2026) interviewed and ran focus groups with 19 disabled job seekers about AI hiring interviews. Every participant perceived the systems negatively. Four themes came out of the analysis: the systems centre normative characteristics, they widen information asymmetries between employer and candidate, they undermine candidate autonomy, and they intrude on privacy.

One finding from that study connects directly to lived experience many late-diagnosed adults will recognise. The researchers found that these systems generate pressure to conform to organisational expectations in order to be hired, including pressure to mask disability. The system rewards a performance, and the candidate who can produce that performance under time pressure is advantaged over the candidate who cannot.

What happens when the system cannot transcribe you accurately?

It scores you anyway. HireVue reports a word error rate under 10 per cent on average for English-language speakers in the United States, rising to between 12 and 22 per cent for non-native English speakers depending on country of origin (Sheard, 2025).

Recruiters and AI practitioners in Sheard’s study readily identified what this means in practice. If the words a person speaks are not accurately transcribed, that person’s interview cannot achieve a high score. One participant put it plainly: the score stops being an assessment of communication skills and becomes an assessment of English-language ability.

The same mechanism applies to any speech pattern the transcription service was not trained on, including stuttering, dysfluency, atypical prosody, unusual pacing and speech affected by anxiety or processing load. Participants in Sheard’s study raised exactly these scenarios.

There is a related design failure worth naming. An external audit of HireVue’s system found that some candidates’ responses did not contain enough content for the algorithm to generate meaningful competency scores at all, and that the rate at which videos fell into that category differed across ethnicities (Sheard, 2025). Short answers cannot be scored. The system did not warn candidates of this, and did not prompt them to say more.

Consider what a short answer means for an autistic candidate who answers the question that was asked, precisely, and then stops. Consider what it means for someone whose processing time is consumed by the timer counting down on screen.

Does rewriting the interview questions fix the problem?

Not on the available evidence. Benson et al. (2025) tested exactly this. Autistic candidates answered either conventional interview questions or questions modified on expert advice to be clearer and more accessible. Their responses were compared with a general population sample answering conventional questions, and all responses were scored algorithmically.

After the question modifications, group differences in algorithmically assigned scores remained. Text analysis showed systematic differences in response content, and the number of words spoken was a differentiator between autistic and neurotypical job seekers.

This is a significant finding for anyone designing neuroinclusive recruitment. Improving the questions is a real and worthwhile intervention. It does not, on its own, close the scoring gap, because the gap is produced by what the model rewards rather than by what the question asks. The authors conclude that accommodations are still required.

Burnard et al. (2025) reached the practitioner-facing version of the same conclusion. Twenty-two autistic adults took part in two-hour interviews covering four digital selection instruments: a behavioural competency assessment, a cognitive ability test, a situational judgement test and an asynchronous video interview. Seven themes were generated, alongside specific modifications participants wanted in order to perform to the best of their ability. The message across that study is that autistic candidates can articulate precisely what would make these assessments fair. They are almost never asked.

FOR ORGANISATIONS

Two concrete changes carry disproportionate weight. Publish the full assessment format in advance, including question types, timing and whether an algorithm scores the response. Then remove or extend response timers wherever the role does not genuinely require rapid verbal processing under observation. If speed is not a genuine occupational requirement, scoring for it is difficult to defend.

What is the ideal candidate problem in algorithmic hiring?

Employers build these models by feeding in data about people who already succeed in the organisation. The model then learns what a successful employee looks and sounds like, and screens for resemblance to that pattern. Where past success was shaped by bias, the model encodes the bias as a specification.

Recruiters in Sheard’s (2025) study described this process in their own words. One explained putting the top-performing graduates through the model to build a profile of what the ideal person looks like, then sending out testing and only interviewing candidates who matched the profile. Another described running all first and second year graduates through it.

Concepts like fit, success and the ideal candidate are difficult to define and harder to measure. When they are converted into a mathematical model, they tend to encode the traits of whoever already held power in that organisation.

Sheard makes an argument here that deserves attention from anyone working in employment law or organisational risk. Human hiring discrimination is inconsistent. Some members of a marginalised group are blocked by an individual’s bias, and some get through. An algorithm removes that inconsistency. As one AI developer in the study put it, when you take the ideal candidate norm and put it into code, the risk is that nobody from a particular group can ever get through.

The result Sheard describes is a new ideal candidate, enhanced and enforced. In a video interview system, that candidate displays the facial movements and vocal patterns represented in the training data. In a text-based assessment, that candidate demonstrates the typing speed and accuracy of a non-disabled worker. Candidates outside that codified model receive a low score or are screened out automatically.

The final layer is that these norms appear neutral because they are presented as science. A candidate has no basis on which to challenge a number.

Why can neurodivergent candidates not simply request a reasonable adjustment?

Because you cannot request an adjustment for an assessment you do not know is coming, administered by a system you were not told about, measuring features that were never disclosed to you. The legal right generally exists. The information required to exercise it usually does not.

This is the most legally significant finding in the current research, and it is the one least covered in mainstream neurodiversity content.

Every recruiter Sheard (2025) interviewed who used an AI hiring system said applicants with disability were able to seek reasonable adjustments. In practice, participants reported that such adjustments were rarely requested.

Some of the reasons will be familiar. A lack of trust in the employer. A desire to prove you can do it without help. A reluctance to be seen as a burden. Disclosure of disability at work requires a secure environment built on trust, and a job application is not that environment.

Sheard identifies an additional reason that had not previously appeared in the research literature. Job seekers may not know that they require an adjustment. Candidates are often not told that an algorithm will evaluate them, and are given little or no information about what completing the assessment involves, which features are assessed, or how those features are weighted. A careers coach for disabled clients described the pattern: the adjustment question sits on the application page, and nothing tells the candidate what is coming, so there is nothing for them to advocate about.

There is a second gap underneath the first. Even where an adjustment is offered, assessment and video interview systems evaluate whether a candidate can perform under typical working conditions. They do not evaluate whether a candidate could perform the genuine requirements of the role if an adjustment were provided (Sheard, 2025). A work sample test scoring typing accuracy under time pressure screens out a candidate who would meet the standard with better screen contrast or a different input method.

For a late-diagnosed autistic or AuDHD adult, the practical position is stark. You are assessed on masking capacity, you are not told you are being assessed, and the accommodation that would make the assessment fair is one you had no opportunity to name.

FOR ORGANISATIONS

Disclose the process before the application closes, not after. State which stages are automated, what each stage measures, how long each takes, and how to request an adjustment for each one specifically. An organisation that cannot describe its own assessment to a candidate in plain language should ask whether it understands the tool well enough to defend the decisions it produces.

What happens to your assessment score after one application?

It may follow you. Where multiple employers license the same assessment vendor, a candidate’s score from one process can become the score every other employer sees. One difficult day can shape access to an entire graduate cohort of roles.

This is documented in Sheard (2025) through the account of a careers coach working with disabled clients. For graduate roles, many employers use the same vendor to provide personality assessments. When that happens, a job seeker’s score on an assessment for one employer becomes the score used by all other employers licensing that product. In the coach’s experience, a disabled candidate who does not perform well is bound by the result, and can only resit if they apply for a new graduate role and the previous assessment is more than twelve months old.

Standardisation is generally sold as fairness. Applied across organisations rather than within one, it converts a single bad result into a market-wide barrier.

Layer this on top of what we already know about fluctuating capacity. Autistic and ADHD adults commonly experience significant day-to-day variation in executive function, processing capacity and communication under load. A system that captures one performance and treats it as a permanent trait measurement is a poor instrument for anyone whose functioning varies. It is a particularly poor instrument for AuDHD adults, whose internal experience is frequently one of oscillation between competing demands.

Does any of this apply in Aotearoa New Zealand?

Yes, and there is a specific reason to be concerned here. The systems in widest use were built and trained offshore. In HireVue’s own published training data, only six per cent of job applicants came from Australia and New Zealand combined, while 78 per cent came from Northern America (Sheard, 2025).

Sheard argues that her Australian findings apply in any jurisdiction where these systems are deployed, because the majority of systems in use were designed and developed offshore and are sold worldwide. New Zealand employers licensing the same global products inherit the same training data, the same proxies and the same design decisions.

The demographic mismatch in that training data is worth stating carefully. In the HireVue rater studies, race and ethnicity in the training sample was reported as 36 per cent White, 33 per cent Hispanic, 17 per cent Black and 14 per cent Asian, with Australia’s First Nations peoples entirely absent and no intersectional groups represented (Sheard, 2025). A system trained on that distribution is being asked to assess a population it does not resemble.

The legal picture in Aotearoa is that existing law applies, and specific regulation does not yet exist.

• The Human Rights Act 1993 prohibits discrimination in employment on the ground of disability. Nothing in that Act carves out decisions made by software.
• The Privacy Act 2020 and its 13 Information Privacy Principles apply across the AI lifecycle. The Office of the Privacy Commissioner published guidance in September 2023 setting expectations including privacy impact assessments, algorithmic impact assessments, transparency with the people affected, engagement with Māori and affected communities, and human review of decisions affecting people.
• The Responsible AI Guidance for Businesses, released alongside New Zealand’s first AI Strategy in 2025, is voluntary and risk-based.
• There is no mandatory guardrail regime for AI in hiring in New Zealand. Australia has proposed mandatory guardrails for high-risk AI applications including recruitment, and legislation has not yet followed.

The Privacy Commissioner’s expectation of engagement with Māori is not a procedural footnote. Any system that classifies people, encodes a norm and controls access to employment engages questions of tino rangatiratanga over data and of Te Tiriti o Waitangi obligations. A hiring model trained on a population that includes no tangata whenua and no Aboriginal or Torres Strait Islander people, then deployed here, is making assumptions about who belongs in a workplace that no organisation in Aotearoa should adopt without examination.

FOR ORGANISATIONS

Two questions to put to your vendor in writing. What proportion of your training data comes from Aotearoa New Zealand or Australia, and can you provide the demographic breakdown? Second, has an algorithmic impact assessment been completed for this deployment, consistent with the Office of the Privacy Commissioner’s expectations? Keep the answers. They will matter if a complaint is ever made.

What does the research still not tell us about AuDHD?

Almost everything specific. Every study covered here tests autism, general disability, or a mixed disabled sample. None isolates co-occurring autism and ADHD. None reports results disaggregated by gender. None tracks whether neurodivergent applicants actually got hired.

Being honest about the gaps is more useful than filling them with confident-sounding claims, so here is what is genuinely missing.

• AuDHD is absent. Glazko et al. (2024) tested autism as one of six conditions. Burnard et al. (2025) and Benson et al. (2025) used autistic samples. No study located here isolates the co-occurring presentation.
• ADHD is barely present. The claim that speech dysfluencies or tangential answers trigger lower algorithmic scores for ADHD candidates is widely repeated online and is highly plausible. It has not been established by a study I could locate. Treat it as a hypothesis.
• No New Zealand empirical study exists. The evidence base is the United States, the United Kingdom and Australia.
• No gender disaggregation. None of the autistic samples report results by gender, which matters enormously given how differently autism and ADHD present in women and how much later women are typically diagnosed.
• No outcome data. Every study measures scores, rankings or perceptions. None follows candidates through to a hiring decision.
• Small scale. Glazko et al. (2024) ran 140 trials against one CV and one job description. Sheard (2025) used a convenience sample of 23 participants in one jurisdiction. Both sets of authors state these limits themselves.

Naming those gaps is not a reason to dismiss the findings. The direction of the evidence is consistent across a quantitative audit, a socio-legal interview study, two qualitative studies with autistic participants and a mixed disability study. Consistency across method types is meaningful even when each individual study is small.

What can you do as a neurodivergent job seeker right now?

Ask what the process involves before you apply, request adjustments for each automated stage specifically rather than in general, and keep a written record. The transparency gap is the leverage point, because a request you can evidence is far harder to ignore than one made verbally.

• Ask, in writing, what stages the recruitment process includes, which are automated, what each measures, and how long each takes. This is a reasonable question. The answer is also documentation.
• Request adjustments per stage, not once at the start. Extended or removed time limits, an alternative to the recorded video format, the ability to re-record, and written rather than spoken responses are all established adjustments.
• Ask whether an algorithm scores the response, and whether a human reviews the output before a decision is made. Under the Privacy Act 2020 you have rights of access to personal information held about you, including assessment results.
• Give complete answers in video interviews. Short answers may not contain enough content to be scored, which produces a low score rather than a neutral one.
• Keep the record. Dates, names, what you asked for, what you were told. If you later need to raise a complaint under the Human Rights Act 1993, that record is the case.

None of this is a solution, and it should not have to be your job. It is what is available while the regulation catches up.

One thing worth saying directly to anyone who has been through dozens of these processes and cannot work out what they are doing wrong. In Sheard’s study, a careers coach described a client who achieved grades equal to or better than the rest of his cohort, then could not get past roughly 23 psychometric assessments, and could not get past the self-recorded video interview when he did progress. The pattern you may be experiencing is documented. It is not a reflection of your capability.

What should organisations audit in their hiring stack?

Start with what the model was trained to predict, what data it was trained on, which features could act as proxies for disability, and whether candidates are told any of it. Most organisations can answer none of these questions about a tool they already use.

Sheard (2025) identified six mechanisms by which employers create real risks of discrimination when deploying these systems: the training data, the use of proxies, the way the system is implemented, the construction of new structural barriers, a failure to provide reasonable adjustments, and the facilitation of intentional discrimination. Those six mechanisms make a workable audit structure.

• Training data. Where did it come from, what population does it represent, and what proportion relates to Aotearoa New Zealand? Vendor-supplied datasets are frequently undisclosed in origin.
• Proxies. Audit every filter and scoring feature for correlation with a protected attribute. Employment gaps, years of experience, graduation year, response speed, typing accuracy and eye contact requirements all warrant scrutiny.
• Implementation. Who configured the time limits, and on what basis? Who selected the target variable, and was the profile built from existing high performers?
• Structural barriers. Does the process require digital resources some candidates lack? Does the score follow the candidate to other employers?
• Reasonable adjustments. Are candidates told what is coming in enough detail to know what to ask for? Can your system assess whether a candidate could perform the role with an adjustment in place, rather than only under standard conditions?
• Intentional discrimination. Can a recruiter add screening questions and set which answers reject a candidate, without oversight? Sheard found platforms with no in-built checks preventing discriminatory screening questions.

One further point for boards and risk committees. Sheard (2025) records that AI hiring systems furnish employers with plausible deniability. As a careers coach in the study put it, they provide a hidden way of discriminating because it is the AI doing it. That framing may feel protective. In risk terms it describes an organisation that cannot explain its own decisions, which is a weaker position than one that can.

FOR ORGANISATIONS

The pragmatic position taken by disability employment specialists in Aotearoa is worth considering seriously: exercise caution before relying on AI-driven recruitment tools until there is greater confidence they do not exclude disabled candidates. Choosing not to automate a stage is a legitimate risk control, and it is currently the only one with a predictable outcome.

Frequently asked questions

Is it legal for an employer to use AI to screen job applicants in New Zealand?

Yes. There is no law prohibiting it. Existing obligations continue to apply, including the Human Rights Act 1993 prohibition on disability discrimination in employment and the Privacy Act 2020 requirements around collecting and using personal information. The Office of the Privacy Commissioner expects organisations to complete impact assessments, be transparent about AI use, and maintain human review of decisions affecting people.

Do I have to disclose that I am autistic or ADHD to request an adjustment?

You generally need to disclose enough to establish that an adjustment is needed, and you do not necessarily need to disclose a diagnosis or its details. Many candidates request the adjustment in functional terms, for example asking for extended response time or a written alternative to a recorded video. This is a decision with real trade-offs and it is worth taking advice on your specific situation.

Can I ask to see the assessment result an AI system produced about me?

Under the Privacy Act 2020 you have a right to access personal information an organisation holds about you, which can include assessment scores. Requests are sometimes met with the response that the scoring is proprietary. That response addresses the vendor’s commercial interest rather than your access right, and it is worth pressing.

Does asking for an adjustment reduce my chances?

Research participants across several of the studies covered here reported avoiding requests precisely because they feared this. The candid position is that the fear is understandable and the alternative is being assessed by a process designed for someone else. Where an organisation responds badly to a straightforward, early, professional request, that response is information about the organisation.

Are AI hiring tools worse than human interviewers for autistic candidates?

Human interviewers show well documented bias against autistic candidates too. The meaningful difference is consistency. Human bias is variable, so some candidates get through. An algorithm applies the same standard to every applicant in the pool, which converts an inconsistent barrier into a uniform one (Sheard, 2025).

Where this leaves us

The technology is not going to be uninvented, and neurodivergent candidates should not be asked to carry the cost of it while the law catches up.

Three things are true at once on the current evidence. Algorithmic screening measurably disadvantages autistic candidates. Most candidates are never told it is happening. Most employers cannot explain what their own system measures.

The research on AuDHD specifically is not yet there. What exists is enough to justify asking harder questions of the tools, and enough to tell anyone who has been screened out of dozens of processes that the pattern they are experiencing is real and documented.

References

Benson, A. L., Colley, K. L., Prasad, J. J., Willis, C. M. G., & Powell-Rudy, T. E. (2025). Comparing autistic and neurotypical responses to conventional and modified questions in algorithmically scored asynchronous video interviews: A textual analysis. International Journal of Selection and Assessment, 33(1), Article e12512. https://doi.org/10.1111/ijsa.12512

Burnard, M. J., Ugalde, D., Bruk-Lee, V., Allen, K. S., Heron, L. M., & Gutierrez, S. L. (2025). Exploring the applicant reactions of autistic individuals to digital personnel selection instruments: A reflexive thematic analysis. International Journal of Selection and Assessment, 33, Article e70013. https://doi.org/10.1111/ijsa.70013

Glazko, K., Mohammed, Y., Kosa, B., Potluri, V., & Mankoff, J. (2024). Identifying and improving disability bias in GPT-based resume screening. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (pp. 687-700). Association for Computing Machinery. https://doi.org/10.1145/3630106.3658933

Kameswaran, V., Hong, V., Clark, J., Hou, Y., Daumé, H., III, & Shilton, K. (2026). Surveilling suitability: How AI hiring interviews impact job seekers with disabilities. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (pp. 1-16). Association for Computing Machinery. https://doi.org/10.1145/3772318.3791516

Luria, M., Scherer, M. U., Aboulafia, A., & Thakur, D. (2026). Disqualified by disability: The exclusion of disabled workers by digitized hiring assessments. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (Article 1188, pp. 1-18). Association for Computing Machinery. https://doi.org/10.1145/3772318.3791485

Office of the Privacy Commissioner. (2023). Artificial intelligence and the Information Privacy Principles. https://www.privacy.org.nz/resources-and-learning/a-z-topics/ai/

Sheard, N. (2025). Algorithm-facilitated discrimination: A socio-legal study of the use by employers of artificial intelligence hiring systems. Journal of Law and Society, 52(2), 269-291. https://doi.org/10.1111/jols.12535