top of page
Search

Neurodiversity and AI at work: risks, safeguards and useful applications

Writer: Divergent Thinking
Divergent Thinking
May 15
12 min read

Updated: Jul 27

AI can remove friction at work, but it can also scale a poor decision.

For neurodivergent employees and candidates, the useful question is not whether a tool is described as intelligent. It is whether the tool does something relevant, is accessible in practice and leaves room for explanation, adjustment and meaningful human review.

That answer depends on the type of system. A person using a generative tool to organise their own notes is not the same as an employer monitoring activity data. Neither is the same as software ranking candidates or recommending who should be promoted.

This guide separates those uses into three groups:

  1. employee-controlled assistive and generative tools;

  2. workflow and monitoring tools;

  3. automated hiring and employment decision systems.

The groups can overlap, but the distinction matters. The more a system observes, scores or influences decisions about a person, the stronger the evidence, governance and safeguards need to be.

Legal note: This article provides practical governance guidance, not legal advice. Data protection, equality and AI rules depend on the facts and can change quickly. Check current regulator guidance and obtain specialist advice for a planned deployment.

The short version

  • Start with the work problem, not the promise of AI.

  • Do not treat all AI tools as one risk category.

  • Give people a clear explanation of what the system does and what data it uses.

  • Test accessibility and disability-related impact before deployment.

  • Offer a straightforward route to reasonable adjustments or an alternative process.

  • Treat supplier claims about accuracy, fairness and return on investment as claims that require evidence.

  • Make human review meaningful: the reviewer needs time, information and authority to change the outcome.

  • Monitor the live system, not only the pilot.

  • Stop or change the use if the evidence does not support it.

1. Assistive and generative AI: tools a person controls

Assistive or generative tools can help an employee structure notes, draft a first version of an email, summarise a document, turn speech into text or break a task into smaller steps.

Some neurodivergent people may find one of those functions useful. Others may not. A tool that reduces cognitive load for one person may add extra checking, uncertainty or sensory effort for someone else.

The safest starting point is choice. Let the person decide whether the tool is useful, agree what it may be used for and provide another route where needed. Do not present one product as the “neurodivergent solution”.

What these tools may help with

  • organising information;

  • producing a first draft;

  • changing the format or reading level of text;

  • creating a checklist or sequence;

  • transcription and meeting notes;

  • preparing questions before a conversation;

  • exploring different ways to communicate an idea.

These are possible uses, not guaranteed outcomes.

The main safeguards

Keep confidential information out of unapproved tools

Employees need clear rules about what can be entered, where the information goes and whether the provider stores or uses it. Personal data, health information, commercially sensitive material and confidential employee conversations need particular care.

The UK Government's AI Playbook recommends mapping data flows, minimising personal data and checking retention, access and international transfers. It also highlights the risk of data leakage when AI services process organisational information.

An approved enterprise account is not a substitute for understanding the contract and settings. Ask:

  • Is prompt or output data retained?

  • Can the provider use it to train or improve models?

  • Where is it processed?

  • Who can see logs?

  • Can the organisation apply its usual access controls?

  • What happens when the contract ends?

Check the output

Generative systems can produce confident, plausible information that is wrong. The UK Government AI Playbook describes these false but plausible outputs as hallucinations.

Use generative AI to support thinking, not to create unreviewed facts about a person. An employee should not have to defend an incorrect summary, invented capability or distorted account of a meeting because software produced it.

Protect voluntary use

If an employee uses an approved tool as a workplace adjustment or personal working aid, that does not make the resulting prompts, drafts or usage patterns a new performance-monitoring dataset.

Define the purpose before deployment. Information collected to provide assistance should not quietly become evidence for capability, conduct or productivity decisions.

A practical standard for assistive AI

Good use looks like this:

  • the employee can choose whether the tool helps;

  • the organisation has approved the data environment;

  • the purpose and boundaries are clear;

  • outputs are checked by a person;

  • an accessible alternative is available;

  • use is reviewed when the role, tool or person's needs change.

For related guidance, see neuro-inclusive digital communication.

2. Workflow and monitoring AI: tools that observe or shape work

Workflow systems may allocate shifts, route tasks, predict demand, analyse calls or recommend priorities. Monitoring systems may collect information about attendance, location, keystrokes, application use, messages, voice or output.

These systems may not make a final employment decision on their own. They can still influence workload, visibility, performance conversations and who is treated as a concern.

That matters for neurodivergent workers. A metric may misread a different working pattern as low effort, treat time away from an application as inactivity or reward one communication style over another. This does not mean every workflow tool is discriminatory. It means the employer needs to test what the measure represents and what it misses.

Start with necessity and proportionality

Ask what problem the monitoring is intended to solve and whether a less intrusive method could do it.

The ICO's worker-monitoring guidance says employers remain responsible when they buy commercial monitoring tools. It recommends clear accountability, transparency, proportionate safeguards and early discussion with workers or their representatives. The guidance is marked as under review following the Data (Use and Access) Act, so check the current version before deployment.

Do not hide the monitoring

People should understand:

  • what information is collected;

  • why it is collected;

  • when monitoring occurs;

  • who receives the information;

  • how long it is retained;

  • whether the system profiles or scores them;

  • what decisions the information may influence;

  • how to question or correct it.

Covert monitoring will rarely be a proportionate default. A broad clause in an employment contract does not explain a specific system or make every use fair.

Avoid weak proxies

Do not assume that time online, typing speed, eye contact, facial expression, tone or message frequency measures capability or commitment.

The UK Government's Responsible AI in Recruitment guidance warns that some video-interview systems use eye detection as a proxy for engagement and notes the potential impact on neurodivergent candidates. It also says there is little scientific consensus for inferring emotion from facial or behavioural signals.

The same caution is useful beyond recruitment. If a vendor says a system can detect engagement, wellbeing, honesty or burnout, ask for the construct being measured, the validation evidence and the error rates for the groups affected.

Emotion recognition is a particularly poor workplace shortcut

Within the EU, the AI Act prohibition on emotion recognition in workplaces has applied since February 2025, apart from limited medical or safety exceptions. The European Commission's current AI Act guidance confirms that prohibition.

Even where that legislation does not apply, inferring emotion from voice, face or behaviour creates serious validity, privacy and equality questions. A direct, accessible conversation is a better starting point than an automated claim about how somebody feels.

A practical standard for workflow and monitoring AI

Before use:

  • name a senior accountable owner;

  • complete the relevant impact assessments;

  • involve workers and neurodivergent perspectives early;

  • define the minimum data required;

  • test whether the metric is job-relevant;

  • explain the system in plain English;

  • create a correction and challenge route;

  • set a review and deletion schedule;

  • define the conditions that will trigger suspension or withdrawal.

3. Automated hiring and employment decisions: tools that score people

This group includes systems used to:

  • target job adverts;

  • screen or rank applications;

  • score tests, games or recorded interviews;

  • recommend a shortlist;

  • evaluate performance;

  • allocate work based on personal characteristics or behaviour;

  • recommend promotion, discipline or dismissal.

These uses need the highest level of scrutiny because they can directly affect whether a person gets work, support, progression or continued employment.

Job relevance comes first

An assessment should measure something needed for the role. A technically consistent score is not useful if the system measures the wrong thing.

The Department for Science, Innovation and Technology's Responsible AI in Recruitment guidance tells buyers to ask suppliers for evidence supporting claims about performance, fairness and capability. It highlights impact assessments, risk assessments, model cards and data-protection impact assessments as useful evidence.

For neurodivergent candidates, examine:

  • whether the task matches the real work;

  • whether speed is genuinely essential;

  • whether the interface works with assistive technology;

  • whether instructions and practice materials are clear;

  • whether sensory, motor or communication differences alter the score;

  • whether an adjusted or alternative format preserves the same assessment standard.

Where possible, use skills-first, job-relevant tasks rather than behavioural proxies.

Reasonable adjustments still apply

Buying an automated tool does not remove an employer's equality duties.

GOV.UK's reasonable-adjustments guidance confirms that employers must make reasonable adjustments so disabled workers are not substantially disadvantaged, and that adjustments can include changing the recruitment process.

For an AI-enabled process, an adjustment could include:

  • accessible instructions or a compatible interface;

  • additional time where speed is not the competence being tested;

  • an alternative assessment format;

  • advance information about the stages;

  • a non-video route;

  • removing an automated behavioural or emotion score;

  • human consideration of relevant context.

The right adjustment depends on the person and the role. See these practical interview-adjustment examples.

Human involvement must be meaningful

A person clicking “approve” is not necessarily a safeguard.

Meaningful review requires a reviewer who:

  • understands what the system measured;

  • can see the relevant evidence and limitations;

  • has enough time to consider the case;

  • can receive additional context or a correction;

  • is not pressured to accept the automated recommendation;

  • has authority to change the outcome;

  • records the reason for the final decision.

This practical standard reflects the UK Government AI Playbook's principle of meaningful human control and the current UK safeguards for significant automated decisions.

What UK data-protection law now requires

The Data (Use and Access) Act 2025 changed the UK rules for significant decisions made solely through automated processing. All its data-protection provisions were in force by 19 June 2026, according to the ICO's current DUAA overview.

The new framework permits significant automated decisions using a wider range of lawful bases, but it does not remove safeguards. The ICO's detailed summary says organisations must:

  • provide information about the decision;

  • allow the person to make representations;

  • allow human intervention;

  • allow the decision to be contested.

Stronger restrictions remain where special-category personal data is used. Neurodivergence, disability or health information may fall within that protected data depending on the information and processing.

The wider data-protection principles still apply. An organisation needs a lawful basis, a clear purpose, data minimisation, accuracy, security, transparency and appropriate retention. A data-protection impact assessment is required where the planned processing is likely to result in high risk.

The ICO's March 2026 recruitment ADM update calls for transparency, bias monitoring and a clear route to challenge and human review. At this article's review date, the ICO's final updated detailed ADM guidance was still pending. Check the latest version before implementing a significant automated decision.

What the EU AI Act adds

The EU AI Act treats specified employment uses as high-risk, including systems used to analyse or filter applications and evaluate candidates. It also covers certain systems used for promotion, termination, task allocation and worker monitoring.

The implementation timetable has changed. The EU's AI Omnibus entered into force on 27 July 2026 and moved the application date for employment and other listed high-risk systems to 2 December 2027, according to the European Commission. Transparency rules for certain interactive and generative systems start to apply on 2 August 2026. The workplace emotion-recognition prohibition has applied since February 2025.

The AI Act is EU legislation, but it can still matter to a UK organisation depending on where a system is provided or used and where its output is used. Organisations operating across borders should obtain advice on scope rather than assuming UK location settles the question.

The later high-risk deadline is preparation time, not evidence that the risk is low. Accessibility, data protection, equality and employment duties can apply before the AI Act's high-risk provisions.

A procurement checklist for neuroinclusive AI

| Decision | Ask the supplier | Evidence to request | |---|---|---| | Intended use | What exact task does the system perform, and what decisions may it influence? | System description, intended-use statement and prohibited-use list | | Job relevance | Why is each input or score relevant to the work? | Validation study using a comparable role and population | | Accessibility | Has the complete user journey been tested with disabled and neurodivergent people? | Accessibility audit, assistive-technology results and remediation record | | Adjustments | Can an employer change time, format or channel without invalidating the result? | Adjustment process and equivalent-assessment guidance | | Data | What personal or special-category data is collected, inferred, retained or shared? | Data map, privacy information, retention schedule and processor list | | Fairness | Which groups and outcomes were tested, and what were the error rates? | Disaggregated test results, methodology and limitations | | Explainability | Can a candidate or employee understand the main factors affecting the result? | Example explanation and user-facing notice | | Human review | What can the reviewer see, question and override? | Review workflow, authority levels and audit trail | | Contestability | How can a person correct information or challenge an outcome? | Escalation route, response times and reconsideration process | | Security | How are access, logging, transfer and deletion controlled? | Security assessment, access model and incident process | | Monitoring | How will drift, changed use and emerging unequal outcomes be detected? | Live monitoring plan, review frequency and stop criteria | | Accountability | Who owns the outcome: supplier, employer and named senior lead? | Responsibility matrix, contract terms and escalation contacts |

A “fairness tested” badge is not enough. Ask which population was tested, what outcome was measured, what threshold was used and what the test did not cover.

A six-step implementation route

1. Inventory the use

List every AI-enabled feature already used in recruitment, HR, collaboration, security and productivity tools. Include features switched on by a supplier update, not only products bought as “AI”.

2. Classify the influence

Record whether the tool:

  • assists a person;

  • observes or recommends;

  • scores or materially influences a decision;

  • makes a significant decision without meaningful human involvement.

Use the highest-influence function when deciding the level of governance.

3. Name the accountable owner

Assign one senior owner for the use case. Procurement, HR, IT, legal and data-protection teams may all contribute, but shared involvement must not become missing accountability.

4. Assess impact with affected people

Complete the relevant data-protection, equality, accessibility and security assessments. Involve people who will use or be assessed by the system, including neurodivergent people with varied access needs.

Do this before the final procurement decision. Consultation after a contract is signed limits what can change.

5. Pilot one defined use

Test the full journey, not only model accuracy. Include instructions, access, adjustment requests, human review, correction and appeal.

Record:

  • what happened;

  • who experienced difficulty;

  • whether an alternative route worked;

  • where human reviewers disagreed with the tool;

  • what the supplier changed;

  • whether the evidence supports continuing.

6. Monitor, review and stop when needed

Set a review date, accountable owner and stop criteria. Monitor for changed data, supplier updates, drift, complaints and unequal outcomes.

If the system moves from assistance into monitoring or decision-making, reassess it before the new use begins.

Questions employers often ask

Can employers use AI in recruitment?

Yes, but “AI-assisted” is not a legal or ethical exemption. Employers remain responsible for data protection, equality, accessibility and the fairness of their recruitment process. Significant automated decisions require specific safeguards under current UK law.

Can generative AI be a reasonable adjustment?

It may be useful as part of an adjustment for some people, depending on the task, person and data environment. It should be agreed rather than imposed, and an alternative should be available where the tool is inaccessible or unhelpful.

Is a human in the loop enough?

Not automatically. The person needs relevant information, competence, time and real authority to question or change the outcome.

Do we need a data-protection impact assessment?

A DPIA is required where processing is likely to result in high risk. AI used for systematic evaluation, significant decisions, extensive monitoring or sensitive worker information is particularly likely to require careful assessment. Ask your data-protection lead for advice on the specific use.

Does the EU AI Act apply to a UK employer?

It may, depending on where the system is provided or deployed and where its output is used. Cross-border employers and suppliers should take advice on their specific position.

Should we let AI infer neurodivergence?

Do not use a tool to diagnose or infer neurodivergence from behaviour, voice, face, writing or work patterns. The validity, privacy and discrimination risks are substantial, and a label is not needed to ask what would make a process accessible.

Start with one use case

Treat neuroinclusive AI as a governance question, not a productivity add-on.

Name the accountable owner. Define the intended use. Test for disability-related barriers. Offer reasonable adjustments. Keep a meaningful review route.

Used with clear boundaries, AI may help people organise, summarise or translate information. Used as an unexamined judge of behaviour, it can make work less fair at scale.

Start with one use case, one impact assessment and one route for feedback before wider deployment.

Review AI through a neuroinclusion lens

Divergent Thinking can examine where digital tools, recruitment and management systems create cognitive or disability-related barriers, then turn the findings into prioritised actions.

Sources and further reading

Comments

bottom of page