Use AI to screen CVs fairly and legally
A first-pass plan for using AI to sort or score job applications without unfair or unlawful results: what the tool may decide, the human review, the bias test, what to tell candidates, and the rules for the EU (AI Act and GDPR), UK, US (federal, New York City and Colorado), UAE and India.
This is a practical, copy-paste process for your applicant tracking or hr software offers ai ranking, scoring or shortlisting and you want to switch it on. Use it when you want a clear structure, then add your own details.
Verified on 2026-10-2921 sources cited
Quick start
1. Use AI to sort applications, not to reject people. A person should look at everyone the tool would turn down, and be able to overrule it. 2. Before switching it on, test it on last round's applications, but only ones you are still allowed to keep. In the EU or UK, do a data protection impact assessment (a written privacy risk check) first, and make sure the vendor will not keep or train on that data. Compare the tool's picks with your own decisions, and compare pass rates between groups (for example men and women). 3. Rule of thumb: if one group's pass rate is below 80% of the best group's, stop and fix the criteria. With small numbers, add more past rounds before you decide. 4. Remove things that can lead to unfair results: names, photos, ages and graduation years, home addresses and postcodes, and penalties for gaps in work history. 5. Tell candidates you use AI, how to ask for a human to review their application, and how to ask for an adjustment. 6. This covers the EU, UK, US (federal rules, New York City, Colorado), UAE and India; anywhere else is not checked. In the EU, data protection law applies now, CV-screening AI gets extra "high-risk" duties from December 2027, and AI that reads emotions at work is banned. New York City needs a yearly bias audit and notice to candidates. 7. This is a first-pass check, not legal or tax advice. Confirm with a qualified accountant or lawyer for your country.
Before you start
- You need to write the job description and scorecard first. Write a job description and scorecard
- Note: if the tool is already switched on, still use this method: stop automatic rejections today, have a person read the applications it scored low in any open round, and run the impact assessment and test now.
- You want to design the interviews that follow the CV screen. Design a structured interview process
- The tool analyses faces, voices or emotions in video interviews. In the EU, AI that infers people’s emotions at work is banned (since February 2025); treat hiring as covered. Other face or voice analysis may be high-risk, or banned if it infers sensitive traits. Get legal advice before using it anywhere.
- A candidate has complained, or a regulator or tribunal has been in touch. Speak to an employment lawyer now.
Full method
How to use this method
- 1. Copy or downloadCopy the prompt below, or download the method file and attach it to your chat.
- 2. Paste into any AIChatGPT, Claude, Gemini or Copilot. Replace each [PLACEHOLDER] with your details, or let the AI ask you.
- 3. Add your materialAttach your own files, data or brief so the AI works from your facts, not guesses.
The prompt
Paste it as-is. The AI will ask for anything still in [BRACKETS].
# Use AI to screen CVs fairly and legally Method by Deepstory Research: deepstoryresearch.com/method/fair-ai-cv-screening You are an HR compliance lead who has rolled out and audited AI screening tools. Follow the method below step by step. Goal: A first-pass plan for using AI to sort or score job applications without unfair or unlawful results: what the tool may decide, the human review, the bias test, what to tell candidates, and the rules for the EU (AI Act and GDPR), UK, US (federal, New York City and Colorado), UAE and India. ## My inputs - Role and where it is based: [ROLE_AND_WHERE_IT_IS_BASED] (example: warehouse supervisor, New York City) - The AI tool and what it does: [THE_AI_TOOL_AND_WHAT_IT_DOES] (example: ATS ranks CVs and suggests a shortlist) - Where candidates are: [WHERE_CANDIDATES_ARE] (example: New York City and New Jersey) - Applications per opening: [APPLICATIONS_PER_OPENING] (example: about 300) - Who makes the final decision: [WHO_MAKES_THE_FINAL_DECISION] (example: hiring manager, from the AI shortlist) Before you start: if any input above is still in [BRACKETS] or missing, ask me for it in one short numbered list and wait. Do not guess. If I attached files (data, briefs, Deepstory evidence files), treat them as the primary source. ## Steps to follow 1. Confirm where the job and the candidates are, what the tool does (rank, filter, reject, chat, video analysis) and where a person decides. 2. Check whether the use is banned, high-risk or regulated in each place, and what must happen before it goes live. 3. Design the process so a person makes every rejection decision, and remove criteria that act as proxies for protected characteristics. 4. Test the tool on past applications: compare its results with human decisions and selection rates by group where you may lawfully hold that data. 5. Write what candidates are told, how they ask for a human review or an adjustment, and what the vendor must provide. 6. Set monitoring: selection rates, complaints and overrides every month, and a stop rule. ## Deliver the output in this structure - Places covered and not checked - What the tool may and may not do - Human review design - Bias test and monitoring plan - Candidate notice and review route - Vendor questions - Questions for a lawyer ## Quality rules - Show the selection-rate arithmetic for any bias test. - Never let the tool reject a candidate without a person’s meaningful review. - Label the result a first-pass check, not legal or tax advice, and name the places covered. - Label every number as FACT (with source and date), ESTIMATE (show the calculation) or ASSUMPTION. - If a Deepstory evidence file is attached, use its figures as the reference values and cite its source references; tag claims that are not from it as [UNVERIFIED]. - End with: the 3 assumptions that matter most, what would change the answer, and the single best next step.
What you need to give the AI
| Input | Placeholder | Example |
|---|---|---|
| Role and where it is based | [ROLE_AND_WHERE_IT_IS_BASED] | warehouse supervisor, New York City |
| The AI tool and what it does | [THE_AI_TOOL_AND_WHAT_IT_DOES] | ATS ranks CVs and suggests a shortlist |
| Where candidates are | [WHERE_CANDIDATES_ARE] | New York City and New Jersey |
| Applications per opening | [APPLICATIONS_PER_OPENING] | about 300 |
| Who makes the final decision | [WHO_MAKES_THE_FINAL_DECISION] | hiring manager, from the AI shortlist |
Use it when
- Your applicant tracking or HR software offers AI ranking, scoring or shortlisting and you want to switch it on.
- You get too many applications to read and are thinking of using an AI tool.
- You already use an AI screening tool and want to check it is fair and lawful.
Best with: Any chat AI; attach the vendor’s description of the feature and, if you have it, anonymised results from a past hiring round.
What you get
- Places covered and not checked
- What the tool may and may not do
- Human review design
- Bias test and monitoring plan
- Candidate notice and review route
- Vendor questions
- Questions for a lawyer
The AI will ask you first
Answer these before the AI advises you; each one changes what it recommends.
Where is the job based, and where do candidates live? Is any role remote across borders?
The rules follow the job and the candidates: the EU AI Act and GDPR, UK data and equality law, US federal and city law (such as New York City) all differ.
If unknown: Apply the rules for the job’s location and every candidate country named, and list the rest as not checked.
What exactly does the tool do: rank, score, filter out, reject automatically, ask screening questions, or analyse video or voice?
Automatic rejection and emotion analysis carry the highest legal risk; ranking for a person to review is lower.
If unknown: Ask for the vendor’s feature description; until then, assume it can reject automatically.
Who looks at each application, and can that person overrule the tool? Do they see why it scored a CV as it did?
A decision with no meaningful human involvement is "solely automated", which GDPR restricts and UK GDPR allows only with safeguards.
If unknown: Assume no meaningful review and design one.
What criteria or data does the tool use (keywords, years of experience, education, location, gaps, photos, names)?
Some criteria act as proxies for age, sex, race, disability or other protected characteristics.
If unknown: List the criteria from the vendor documents and flag likely proxies.
Do you have results from a past hiring round, and any lawful data on candidates’ sex, ethnicity, age or disability (for example from voluntary equality monitoring)?
A bias test compares selection rates between groups. In the EU, the amended AI Act (new Article 4a) lets deployers process sensitive data such as ethnicity for bias detection only where strictly necessary and with strict safeguards.
If unknown: Test against past human decisions first and plan voluntary, separately stored monitoring.
Does the vendor provide instructions for use, bias testing results, data retention terms, a data processing agreement with its list of subprocessors (including any third-party AI model), where candidate data is stored, and a promise not to train on your candidates’ data?
Deployers must follow the provider’s instructions and need its information for a data protection impact assessment.
If unknown: Treat each as missing and add it to the vendor questions.
Do you have employee or worker representatives (for example a works council)?
In the EU, workers’ representatives and affected workers must be informed before a high-risk AI system is used at the workplace, and national law may require consultation.
If unknown: Assume none and say national consultation rules were not checked.
Roughly how many applications do you get for each role?
It decides whether a person can read every application the tool would reject; any rejection nobody reads is a solely automated decision with extra legal rules.
If unknown: Assume a person must read every rejection, and design the process for that.
Is the employer a public body, or based in a UAE free zone (DIFC or ADGM)?
Public bodies have extra EU duties (a fundamental rights impact assessment), and the free zones have their own employment and data laws.
If unknown: Assume a private, mainland employer and say free-zone rules were not checked.
The procedure, step by step
Read this before you run it, so you can check the AI actually followed the method.
Scope
Job location, candidate locations, what the tool does, where a person decides.
Classify
Banned (EU emotion inference), high-risk (EU), solely automated (GDPR and UK GDPR), audited tool (NYC); other places flagged.
Get vendor information
Instructions for use, criteria, bias testing, logs, data retention, no training on your data.
Impact assessment
Data protection impact assessment before the test run and before go-live (EU and UK).
Design human review
Who reviews, what they see, how they overrule, what is logged.
Clean criteria
Job-related requirements only; remove proxies.
Test
Past round: agreement with human decisions and selection-rate ratios.
Tell candidates
Notice that AI is used, how, how to get a human review, how to ask for an adjustment.
Monitor and stop rule
Monthly ratios, overrides, complaints; revert to human screening if needed.
Worked example
The figures below are illustrative. Use your own numbers and facts.
Scenario: A logistics firm hiring warehouse supervisors in New York City wants its ATS to rank CVs and shortlist the top candidates for a manager to interview.
Input: Test on last year’s round: 300 applicants with voluntary sex data, 200 men and 100 women. The tool would have shortlisted 60 men and 21 women. The vendor has no bias audit and says the score rewards relevant experience and continuous employment. A manager reviews only the shortlist.
Output excerpt: Places covered: US federal rules and New York City. Not checked: New York State law and any other state. Selection rates: men 60 / 200 = 0.30; women 21 / 100 = 0.21. Impact ratio 0.21 / 0.30 = 0.70, below the 0.8 warning level in the federal Uniform Guidelines (the EEOC plans to remove their interpretive parts from its regulations; keep using 0.8 as a warning level). With only 300 applicants the gap is not statistically significant at the usual 5% level (a two-proportion test gives p of about 0.10), so pool more rounds before drawing a firm conclusion. Federal enforcement of disparate impact has been scaled back since the Justice Department’s June 2026 opinion, but private claims and New York City law still apply. Do not switch the tool on as configured. New York City: the tool needs a bias audit within one year before use, a published summary, and notice to candidates at least 10 business days before use; the vendor has none, so ask for one or commission it. Process: the manager should also review the candidates the tool ranks out, not only the shortlist, and be able to overrule it. Check the criteria for proxies: the tool weights continuous employment, which can disadvantage women returning from caregiving; remove that criterion and re-test. This is a first-pass check, not legal or tax advice. Confirm with a qualified accountant or lawyer for your country.
Check your result
Before you trust the AI's output, confirm each of these.
- ☐ Places covered and not checked are named at the top.
- ☐ The tool’s role is stated as sort, score or reject, and no rejection happens without meaningful human review.
- ☐ Any bias test shows each selection rate and the impact ratio.
- ☐ Candidate notice, human review route and adjustment route are written out.
- ☐ The output says it is a first-pass check, not legal or tax advice.
What you'll get(show template)
## AI CV screening check: [ROLE], [DATE] Places covered: [only those involved among EU (AI Act, GDPR), UK, US federal, New York City, Colorado, UAE, India]. Not checked: [every other place involved, including any other US state]. ### What the tool may and may not do Use | Allowed? | Rule | Condition ### Human review design Step | Who | What they see | Can overrule? | Logged? ### Criteria check Criterion | Job-related? | Possible proxy for | Keep or remove ### Bias test Group | Applicants | Selected | Selection rate | Ratio to highest ### Candidate notice and review route [Text for the job advert and privacy notice] ### Vendor questions 1. [Question] ### Monitoring and stop rule Measure | Frequency | Stop if This is a first-pass check, not legal or tax advice. Confirm with a qualified accountant or lawyer for your country.
Quality checks built into the prompt
- • Show the selection-rate arithmetic for any bias test.
- • Never let the tool reject a candidate without a person’s meaningful review.
- • Label the result a first-pass check, not legal or tax advice, and name the places covered.
- • Every number labelled FACT, ESTIMATE or ASSUMPTION; claims outside attached evidence tagged [UNVERIFIED].
Decision rules
If/then rules the AI should apply. Heuristic rules are a practitioner's rule of thumb, not a platform rule.
| If | Then | Type |
|---|---|---|
| An AI system is used in the EU to filter job applications or evaluate candidates. | It is a high-risk AI system under the AI Act (Annex III, point 4(a)); because ranking CVs profiles people, it is always high-risk. Under the Digital Omnibus amendment, the high-risk obligations for Annex III systems apply from 2 December 2027, not 2 August 2026. A system already on the market or in use before that date is covered only if its design changes significantly afterwards (amended Article 111(2)); vendors update often, so ask the vendor and a lawyer and plan as if it will be covered. GDPR, UK GDPR and equality law already apply today. | Legal rule |
| You deploy a high-risk AI screening system in the EU (from 2 December 2027). | Use it according to the provider’s instructions; assign human oversight to people with the competence, training, authority and support to do it; make sure input data you control is relevant and representative; monitor it and suspend use and tell the provider if it presents a risk; keep its logs for at least six months; before using it at the workplace, inform workers’ representatives and affected workers; tell candidates they are subject to it; and use the provider’s information for your data protection impact assessment. Prepare these before the date rather than after. | Legal rule |
| A candidate in the EU is turned down on the basis of a high-risk AI screening tool’s output, even if a person made the final decision, and considers the decision harms their rights. | The candidate can ask you for a clear and meaningful explanation of the AI system’s role in the decision and the main elements of the decision (AI Act Article 86), where other EU law does not already give that right. When this starts for your tool depends on the high-risk dates and the rule for older systems, so ask a lawyer; prepare a short, plain explanation of how the tool ranks candidates anyway. | Legal rule |
| A tool would infer candidates’ or workers’ emotions in the EU (for example from video interviews). | The AI Act prohibits AI systems that infer people’s emotions in the workplace, except for medical or safety reasons, from 2 February 2025. The Act does not mention job candidates; read it as covering them. Do not use emotion analysis in hiring. | Legal rule |
| Staff use an AI tool in hiring in the EU. | Take measures to support the AI literacy of the staff using it, taking into account their knowledge and the people it will be used on (AI Act Article 4 as amended). | Legal rule |
| A candidate in the EU would be rejected by the tool without meaningful human involvement. | Under GDPR Article 22, people have the right not to be subject to a decision based solely on automated processing that significantly affects them, unless an exception applies (contract necessity, law or explicit consent) with safeguards. Design the process so a person meaningfully reviews every rejection. | Legal rule |
| You plan to use an AI screening tool on candidates in the EU or UK, or to test it on past applications. | Do a data protection impact assessment first. GDPR and UK GDPR Article 35 require one before processing likely to result in a high risk, and name, as an example, a systematic and extensive evaluation of people based on automated processing, including profiling, on which decisions with significant effects are based; AI screening of job applications will usually meet this. This applies now, not only from December 2027. Cover the test run and what the vendor may keep. | Legal rule |
| You test the tool on past applications in the EU or UK. | Use only applications you still hold within your retention period. Testing a new tool is a new purpose, so check it is compatible with the purpose you collected them for, or get advice on another lawful basis. Use no more data than the test needs, for example by removing names and contact details first where you can. As good practice, also make sure the vendor deletes the test data and does not train on it. | Legal rule |
| A significant decision about a UK candidate is based solely on automated processing. | Since 5 February 2026, UK GDPR allows it for ordinary personal data only with safeguards: tell the candidate about the decision, let them make representations, let them obtain human intervention, and let them contest it. A decision is solely automated if there is no meaningful human involvement, and you must consider how far the decision relies on profiling. | Legal rule |
| You use automated decisions in UK recruitment. | The ICO expects you to monitor proactively for bias (good practice includes asking developers about their own bias testing and considering monthly bias reviews), to be clear with candidates that automation is used and how it works, and to tell them how to challenge a decision and get a human review. Its detailed ADM guidance was still a draft at the last check (consultation closed 29 May 2026); check for the final version. | Professional guidance |
| A screening criterion in the UK puts people with a protected characteristic at a particular disadvantage. | It is indirect discrimination unless you can show it is a proportionate means of achieving a legitimate aim. The Act does not mention AI, but a tool’s scoring criteria, or the decision to use the tool, are likely to be treated as a "provision, criterion or practice" (an interpretation, not stated in the Act). | Legal rule |
| A UK applicant is disabled, or you plan health questions. | Make reasonable adjustments where your process puts a disabled person at a substantial disadvantage, for example an alternative to an automated test. Do not ask about health before offering work or putting the person in a pool for an offer, except for the purposes the Act allows. | Legal rule |
| You test a screening step for adverse impact in the US. | Divide each group’s selection rate by the highest group’s rate. The federal Uniform Guidelines still say a ratio below four-fifths (0.8) is generally regarded as evidence of adverse impact, that smaller differences can count if significant, and that results from small numbers are unreliable. On 9 June 2026 the Justice Department published an opinion that the EEOC’s disparate-impact guidelines are unconstitutional (implementing Executive Order 14281), so federal enforcement on this basis is unlikely now. The EEOC has also scheduled a final rule to remove the interpretive parts of the Uniform Guidelines (29 CFR part 1607) from the Code of Federal Regulations, listed for November 2026 and effective January 2027; check whether it has been published. Title VII itself has not changed, and private claims and state and city laws (such as New York City’s) still apply, so keep running the test and treat a ratio below 0.8 as a warning. | Legal rule |
| You use an automated employment decision tool for jobs in New York City. | Local Law 144 requires a bias audit within one year before use, a published summary of the audit results, and notice to candidates or employees at least 10 business days before the tool is used. Check the DCWP rules for who counts as an independent auditor and what the notice must contain. | Legal rule |
| You hire in the UAE private sector. | Discrimination based on race, colour, sex, religion, national or social origin or disability that impairs equal opportunity or equal treatment in employment is prohibited (Decree-Law 33 of 2021, Article 4). Rules to increase UAE citizens’ participation are not discrimination. Check the tool’s criteria against these grounds. This covers mainland private-sector employers; the DIFC and ADGM free zones have their own employment and data protection laws, and the federal data protection law was not checked. | Legal rule |
| You collect candidates’ personal data in India with their consent. | Give a notice describing the personal data and purpose, how to exercise rights and withdraw consent, and how to complain to the Data Protection Board (DPDP Act section 5). Most DPDP duties apply from about 13 May 2027 (check the Gazette for any change). No India-specific AI hiring law was found; equality and disability law still apply and were not checked. | Legal rule |
| You hire in Colorado, or screen applicants who live in Colorado. | SB26-189, signed 14 May 2026, repealed and re-enacted the 2024 Colorado AI Act as a law on automated decision-making technology used to materially influence consequential decisions, including employment. It covers employees, job applicants who are Colorado residents, and anyone evaluated for an opportunity in Colorado by a business operating there. It takes effect on 1 January 2027 and applies to decisions made on or after that date. Deployers must give clear notice at the point of interaction, give a plain-language description of the system’s role within 30 days after an adverse decision, and, when someone gets an adverse outcome and asks, let them correct wrong personal data and offer meaningful human review and reconsideration, to the extent commercially reasonable. Developers and deployers must keep compliance records for at least 3 years, and the Attorney General must adopt rules on the adverse-outcome disclosures and on these rights by 1 January 2027; check those rules before go-live. | Legal rule |
| You hire in Illinois or another US state not listed above. | Some US states, including Illinois, have their own rules on AI in hiring. They were not checked for this method. List the state as not checked and confirm with a local employment lawyer before use. | Heuristic |
| The tool ranks or scores rather than rejects. | Treat every candidate below the point where reviewers stop reading as turned down by the tool with no meaningful human involvement. This is a cautious reading: GDPR recital 71 gives e-recruiting practices without any human intervention as an example of a solely automated decision, and UK GDPR Article 22A treats a decision with no meaningful human involvement as solely automated; how this applies to ranking cut-offs is not settled. Either read every application, or follow the rule for when a person cannot read every rejection. | Heuristic |
| You hire in an EU country. | National equal treatment law (for example Germany’s General Equal Treatment Act, AGG) also applies to selection criteria, often with short deadlines for rejected applicants to claim. This method does not state those national rules; check them with a lawyer for the country. | Heuristic |
| You keep applications and scores. | Set a retention period for unsuccessful candidates long enough to answer complaints and claims in that country, and no longer; where the AI Act applies, keep the tool’s logs for at least six months; in Colorado, keep compliance records for at least 3 years. Ask the lawyer for the local period, including any US federal record-keeping duty. | Heuristic |
| You design the process. | Use AI to sort or summarise, not to reject. A trained person reads every application the tool would reject, can see the reasons, and can overrule it. Log every override. Checking a random sample tests the tool, but it is not human review of the rejections nobody read. UK government guidance also asks employers to plan human oversight and reasonable adjustments before using a tool. | Heuristic |
| You cannot have a person read every rejection (for example at very high volume). | Every rejection no person has meaningfully reviewed is a decision based solely on automated processing. In the EU it is allowed only under a GDPR Article 22 exception (contract necessity, law or explicit consent) with safeguards including human intervention, the chance to give their point of view and the right to contest, and it must not be based on special category data such as health or ethnicity unless the person gave explicit consent or a substantial public interest law applies. In the UK it needs the Article 22C safeguards (information, representations, human intervention, contest), and if special category data such as health, ethnicity or religion is involved, only explicit consent, or contract or law plus the substantial public interest condition (Article 22B). It is not allowed at all if you process the data in reliance on the lawful basis in UK GDPR Article 6(1)(ea), known as a recognised legitimate interest (Article 22B(4)). Get legal advice before any rejection goes out unreviewed. | Legal rule |
| You choose criteria. | Remove or ignore names, photos, dates of birth and graduation years, addresses, postcodes and PIN codes, the name or ranking of a school or college unless the job needs it, and penalties for career gaps; score only job-related requirements from the job description. | Heuristic |
| You test before go-live. | Run the tool on a past round’s applications and compare its shortlist with the human decisions and, where lawful data exists, selection rates by group. Do not switch it on if impact ratios fall below 0.8 or it disagrees with good human decisions without a clear job-related reason. Agreement with past human decisions only shows the tool copies your choices, including any bias; if no lawful group data exists, write "Bias test not run" and rely on the criteria check and monitoring. | Heuristic |
| You monitor after go-live. | After each hiring round, or monthly if you hire continuously, review selection rates by group, overrides and complaints. Stop using the tool and revert to human screening if a ratio falls below 0.8 or a pattern appears. | Heuristic |
| Candidates or the job are in a place not listed above. | Say clearly that the local rules were not checked. | Heuristic |
Failure modes
Common mistakes with this method, how to spot them, and how to fix them.
| Mistake | Symptom | Fix |
|---|---|---|
| Letting the tool reject candidates automatically. | Solely automated decisions without the required safeguards; good candidates lost with no record why. | Use the tool to sort only, and have a person review every rejection. If you can only review some, treat the rest as solely automated decisions and apply the GDPR or UK GDPR safeguards after legal advice. |
| Reading only the tool’s top-ranked candidates. | Everyone below the cut-off was turned down with no person looking, which is likely to count as solely automated in the EU and UK, and leaves no human review to point to if a bias claim is made. | Read every application, or apply the rule for unread rejections after legal advice. |
| Treating a random sample of rejections as human review. | The rejections nobody read are still solely automated decisions, without safeguards. | Review every rejection, or put the Article 22 safeguards in place for the rest after legal advice. |
| Guessing candidates’ sex or ethnicity from names or photos to run the bias test. | Unreliable results, and new sensitive data created without a lawful basis. | Use voluntary, separately stored monitoring data only; otherwise compare with past human decisions. |
| A "human in the loop" who just clicks approve. | The decision still counts as solely automated. | Give the reviewer the reasons, time, authority and training to overrule, and log overrides. |
| Switching the tool on without testing it. | Adverse impact found only after a complaint. | Test on a past round first and check the selection-rate ratios. |
| Keeping proxy criteria (postcode, graduation year, career gaps, name). | Indirect discrimination by age, race, sex or disability. | Score job-related requirements only. |
| Using video or voice emotion analysis. | Banned in the EU since February 2025; high legal and fairness risk elsewhere. | Do not use it. |
| Not telling candidates AI is used. | Breach of transparency duties (GDPR, UK GDPR, ICO expectations, NYC notice) and loss of trust. | Say so in the job advert and privacy notice, with how to get a human review. |
| Running a four-fifths test on tiny numbers. | Misleading pass or fail. | Pool several rounds or use a longer period before deciding. |
| Treating the vendor’s marketing as a bias audit. | No evidence when challenged. | Ask for the audit method, data, date and results in writing. |
| Waiting for the EU AI Act dates before doing anything. | GDPR, UK GDPR or equality complaints now, long before December 2027. | Apply the data protection, human review and equality steps today and prepare the AI Act duties in advance. |
| Missing New York City’s notice or published audit summary. | Complaint to the city and penalties. | Publish the audit summary and notify candidates at least 10 business days before use. |
| Storing equality-monitoring data with the screening data. | Reviewers or the tool can see protected characteristics. | Keep monitoring data separate, with restricted access, and delete it when no longer needed. |
| Treating this check as legal advice. | A process built on a misread rule. | Confirm with a qualified employment or data protection lawyer. |
Agent mode: if an AI agent runs this unattended
May do without asking
- • Read vendor documents and list the tool’s criteria, likely proxies and missing information.
- • Calculate selection rates and impact ratios from an anonymised export.
- • Draft the candidate notice, review route and vendor questions.
Must stop before (the user decides and does these)
- • Switching on, configuring or changing any AI screening feature in the HR or applicant tracking system.
- • Rejecting, shortlisting or messaging any candidate.
- • Uploading or exporting candidate data, or sharing it outside the business.
- • Deleting candidate records, scores, logs or monitoring data.
- • Sending any reply to a candidate’s complaint, explanation request or data access request.
- • Guessing or inferring any candidate’s sex, ethnicity, age, disability or other protected characteristic (for example from names, photos or dates) to run a bias test.
- • Publishing a job advert, notice or audit summary.
- • Accepting vendor terms, buying an audit or changing billing.
Evidence for this method
This method runs on your own material: your files, brief or code. If your task needs market or economic numbers, attach a sourced context file from the evidence library alongside the prompt.
Browse evidenceRun next
Methods that chain well after this one.
Deepstory methods are process guidance for AI tools. Review AI output before relying on it; legal, tax, medical and investment decisions need a qualified professional.