AI hiring recordkeeping requirements (2026): which laws make you keep records of automated hiring decisions, and for how long
On this page
- Which laws require you to keep records of AI hiring decisions?
- What is the status of the Uniform Guidelines' recordkeeping?
- What counts as an AI hiring record?
- Which period wins when several laws apply?
- Do AI interview notes and transcripts count?
- What about deletion rights and data minimization?
- What should staffing agencies keep?
- A copy-ready AI hiring record retention schedule
- Checklist
- Questions people ask
No single law sets one retention period for AI hiring records, as of November 2026. The specific rules are: California requires employers to keep automated-decision system data for four years (2 CCR § 11013(c), since October 1, 2025); Colorado will require deployers of covered automated decision-making technology to keep compliance records for three years after each decision (C.R.S. § 6-1-1703, from January 1, 2027); New York City requires the bias audit summary to stay posted for six months after the tool's last use (6 RCNY § 5-303); California's privacy rules require risk assessments to be kept for five years (11 CCR § 7155(c)); and the EU AI Act requires deployers of high-risk recruitment AI to keep system logs for at least six months (Article 26(6), from 2 December 2027). Under all of them sits the federal floor of one year for hiring records (29 CFR 1602.14). Keep each record for the longest period that applies to any place you hire, and stop deletion once a claim is filed.
Not legal advice. Periods and sources below were checked against the linked statutes, regulations and agency pages as of November 2026. Several of these rules are new or not yet in force, and some federal rules are under challenge. Have employment counsel set your final schedule.
Which laws require you to keep records of AI hiring decisions?
The general retention rules for applications and interview notes (the federal one-year floor, federal contractors' two years, California's four and Colorado's five-year personnel-records rule) are set out in how long to keep interview notes. The table below covers only the rules that reach automated tools, their outputs and the evidence around them.
| Rule | Who it covers | What it requires | In force from | Source |
|---|---|---|---|---|
| EEOC recordkeeping (Title VII, ADA, GINA) | Employers with 15 or more employees | Any personnel or employment record made or kept, including applications and records about hiring, for one year from the record or the personnel action, whichever is later; all relevant records until final disposition once a charge is filed | In force | 29 CFR 1602.14 |
| ADEA recordkeeping | Employers (20 or more employees) and employment agencies | One year for applications, test papers from employer- or agency-administered tests, job orders, referrals and job ads | In force | 29 CFR 1627.3; 1627.4 |
| Uniform Guidelines on Employee Selection Procedures | "Users" of tests and other selection procedures | Records showing the impact of each selection procedure by sex, race and ethnic group, with adverse impact checked at least annually; simplified records for employers with 100 or fewer employees | Still printed; standing challenged in 2026 (see below) | 29 CFR 1607.4; 1607.15 |
| Federal contractors (VEVRAA, Section 503) | Covered federal contractors | Two years (one year below 150 employees or the contract threshold), and the listed records include "tests and test results" and "interview notes" | In force; the EO 11246 rule was removed October 26, 2026 | 41 CFR 60-300.80; 60-741.80 |
| California FEHA regulations | Employers with 5 or more employees and other covered entities, including agents and employment agencies | Four years for personnel and employment records, expressly including selection criteria and "automated-decision system data"; after a complaint, until the later of the civil-action deadline or final resolution | October 1, 2025 | 2 CCR 11013(c) |
| California CCPA regulations (risk assessments) | Businesses covered by the CCPA whose processing triggers a risk assessment, including ADMT for significant decisions | Keep risk assessments, original and updated, for as long as the processing continues or five years after completion, whichever is later | January 1, 2026 (ADMT duties January 1, 2027) | 11 CCR 7155(c) |
| New York City Local Law 144 | Employers and employment agencies using an AEDT in the city | Bias audit summary and distribution date kept posted at least six months after the last use; data types, sources and your retention policy available on request within 30 days | Enforced since July 5, 2023 | Admin. Code 20-871; 6 RCNY 5-303 |
| Illinois Artificial Intelligence Video Interview Act | Employers using AI to analyze applicant-recorded video interviews for Illinois-based positions | Delete the videos within 30 days of the applicant's request, including copies others hold; employers relying solely on AI to pick who gets an in-person interview report race and ethnicity data to the state every December 31 | 2020 (reporting 2022) | 820 ILCS 42/15, 42/20 |
| Colorado SB 26-189 | Deployers and developers of covered ADMT doing business in Colorado | Deployers: records reasonably necessary to show compliance, at least three years after the decision. Developers: three years after each record is created | January 1, 2027 | C.R.S. 6-1-1702(4), 6-1-1703 |
| Texas Responsible AI Governance Act | Developers and deployers of AI systems in Texas | No retention period, but the attorney general's civil investigative demand can ask for the system's purpose, training and input data, outputs, performance metrics, known limitations and post-deployment monitoring | January 1, 2026 | Bus. & Com. Code 552.103 |
| EU AI Act | Deployers (employers, agencies) and providers (vendors) of high-risk recruitment AI | Deployers keep automatically generated logs under their control at least six months; providers keep logs six months and technical documentation ten years | 2 December 2027 for Annex III systems | Art. 26(6); Art. 18; Art. 19 |
What is the status of the Uniform Guidelines' recordkeeping?
The Uniform Guidelines are the only federal text that tells employers to keep adverse-impact records for each selection procedure, which is what most AI bias testing is built on. They are still printed at 29 CFR part 1607 as of November 2026. But the Office of Personnel Management's interim final rule of July 31, 2026 says it removed references to them from federal civil service regulations to conform to a June 9, 2026 Justice Department Office of Legal Counsel opinion "finding the UGESP unlawful" (91 FR 48234), and the Labor Department's copy at 41 CFR part 60-3 was rescinded with its other Executive Order 11246 rules, effective October 26, 2026 (91 FR 54444). Title VII's disparate impact provision is unchanged, and state rules (California's anti-bias testing evidence, New York City's bias audits) still need the same data. Keeping impact records remains the only way to answer a claim. The federal background is in EEOC guidance on AI in hiring and Title VII disparate impact and AI tools; the calculation is in the four-fifths rule.
What counts as an AI hiring record?
California's definition is the broadest written one. Section 11008.1(d) defines automated-decision system data as "any data used in or resulting from the application of an automated-decision system, such as data provided by or about individual applicants or employees, or data reflecting employment decisions or outcomes," and "any data used to develop or customize an automated-decision system for use by a particular employer." Colorado's list for deployers is shorter: version identifiers, changelogs and documentation of material mitigation changes. In practice, plan to keep six kinds of record for every tool that scores, ranks, filters or recommends:
- Inputs: the application, resume, answers, assessment responses or recording the tool processed.
- Outputs: each score, rank, tag, recommendation or pass/fail result, with the date and the tool version.
- Configuration: selection criteria, cutoffs, weights, knockout rules, prompts and any customization for your roles.
- The human decision: who reviewed the output, what else they considered, and what they decided.
- Notices and requests: the notices you gave, opt-outs, accommodation requests, alternative-process requests, data and deletion requests and your responses.
- Testing and vendor evidence: bias audits or adverse impact analyses, the data behind them, and the vendor's documentation of the tool's purpose and limitations.
New York City adds a practical reason to keep the data the tool generates: DCWP's FAQ says bias audits must use "historical data" from your own use of the tool, that you can rely on a multi-employer audit only if you supplied your historical data (or it is your first use), and that imputed or inferred demographic data cannot be used. If you delete outputs quickly, next year's audit has nothing to work with. Details are in the Local Law 144 bias audit guide.
Which period wins when several laws apply?
The longest one that applies to any job or candidate in the process, and watch where each clock starts: the federal and California rules run from the later of the record or the personnel action, Colorado's from the decision, New York City's from the tool's last use, and California's risk assessment rule from the end of the processing. A pending charge or complaint suspends deletion for the relevant records under 29 CFR 1602.14 and 2 CCR 11013(c).
Worked example (illustrative only)
A staffing firm uses a resume-ranking tool for a role open in both its Los
Angeles and New York City offices. A candidate is rejected on March 2, 2027.
Federal floor (1602.14): keep to at least March 2, 2028
California FEHA (11013(c)): keep ADS data to at least March 2, 2031
NYC LL144 (6 RCNY 5-303): audit summary stays posted until at least
6 months after the firm last uses the tool
California CCPA risk assessment: keep while the tool is in use, plus 5 years
after the assessment is completed
Result: one schedule, set to California's four years for candidate-level
records, with the risk assessment and audit summaries kept longer.
Do AI interview notes and transcripts count?
Usually yes. The EEOC rule covers "any personnel or employment record made or kept," and the federal contractor rules name "interview notes" and "tests and test results" expressly. Where the note taker also scores or recommends, its output is likely automated-decision system data in California. Before you switch on automatic deletion in a recording or note-taking tool, check the setting against the longest minimum period; see AI note taker auto-delete settings and Zoom recording retention settings. What goes into the notes matters as much as how long they last: what not to write in interview notes and protected information in AI interview transcripts.
What about deletion rights and data minimization?
Retention minimums sit next to rules that push the other way. Illinois requires deletion of AI-analyzed video interviews within 30 days of a request. The UK and EU GDPR set no fixed periods but forbid keeping data longer than necessary, and the EU AI Act's six-month log minimum expressly yields to data protection law. California's privacy law requires a disclosed retention period per category. Write down which legal minimum justifies each period, and when a deletion request conflicts with a retention duty, record the reason you kept the data. The UK position is in ICO guidance on AI in recruitment.
What should staffing agencies keep?
Agencies have their own duties, not just their clients'. The ADEA rules list what an employment agency keeps for a year (placements, referrals, job orders, applications and agency-administered test papers), California's retention rule applies to "any employer or other covered entity," which includes employment agencies, and Local Law 144 applies to employment agencies by name. Agree in the client contract who keeps which outputs, who holds the bias audit data, and who answers a data request. More in AI hiring laws for staffing agencies.
A copy-ready AI hiring record retention schedule
AI HIRING RECORD RETENTION SCHEDULE (example: adapt with counsel)
Tool / feature: [vendor, product, feature, version]
What it outputs: [score / rank / tag / recommendation / pass-fail]
Jurisdictions in play: [job locations; candidate locations]
Record Keep for Legal reason
Inputs processed [4 years if any CA role, else 1] 1602.14; 2 CCR 11013(c)
Outputs, with version + date [same] 11013(c); C.R.S. 6-1-1703
Configuration and criteria [life of tool + 4 years] 11013(c) "selection criteria"
Human review record [same as outputs] 1602.14; 11013(c)
Notices and candidate requests [same as outputs] LL144; C.R.S. 6-1-1704
Bias audit summary (NYC) [posted until 6 months after 6 RCNY 5-303(c)
last use; archive with data]
Risk assessment (CCPA) [while in use + 5 years] 11 CCR 7155(c)
EU AI Act system logs [at least 6 months, if deployer Art. 26(6)
of a high-risk system]
Video interviews (Illinois) [delete within 30 days of a 820 ILCS 42/15
request]
Legal hold: on any charge, complaint or demand, suspend deletion for the
candidate and all candidates for the same role and location.
Owner: [name, role] Last reviewed: [date] Next review: [date]
Checklist
- List every tool that scores, ranks, filters or recommends candidates, and every place you hire.
- For each tool, record which rows of the table above apply.
- Set each record type to the longest applicable period and note the legal reason.
- Keep outputs with the tool version and date, so you can show what the tool did at the time.
- Keep your own historical data if you use a tool in New York City, so future bias audits can use it.
- Align vendor auto-delete settings with your schedule, and confirm in writing what the vendor keeps and for how long.
- Write a legal hold procedure that covers AI outputs and logs, not just emails and files.
- Allocate retention duties in agency and client contracts.
- Re-check the schedule when Colorado's rules start on January 1, 2027 and the EU AI Act's high-risk rules on 2 December 2027.
For the policy wrapper around this schedule, use the candidate data retention policy template, and for the laws themselves see AI hiring laws by state, California's AI hiring regulations, Colorado's AI law and the CCPA ADMT rules. Which tools these rules reach is explained in what an automated employment decision tool is.
Questions people ask
How long do I have to keep AI hiring records in California?
At least four years. Since October 1, 2025, the Civil Rights Council's regulation at 2 CCR section 11013(c) requires employers and other covered entities to keep personnel and employment records, expressly including automated-decision system data, for four years from the date the record was made or the personnel action, whichever is later, and longer once a complaint is filed.
Does New York City's Local Law 144 say how long to keep bias audit records?
It sets one period: the summary of the bias audit results and the tool's distribution date must stay posted for at least six months after the employer or employment agency last used the tool for an employment decision. The law also requires you to make your data retention policy for the tool available, on your website or within 30 days of a written request.
What records does Colorado's AI law require employers to keep?
From January 1, 2027, a deployer of a covered automated decision-making technology must keep records reasonably necessary to show compliance for at least three years after each consequential decision, such as version identifiers, changelogs and documentation of material mitigation changes, or longer if another law requires it.
Is there a federal rule on keeping AI hiring records?
No AI-specific one. The general EEOC rules apply: employers covered by Title VII, the ADA or GINA keep hiring records for one year (29 CFR 1602.14), the ADEA rules require the same of employers and employment agencies (29 CFR 1627.3 and 1627.4), and records relevant to a charge must be kept until it is finally resolved. The Uniform Guidelines' adverse-impact recordkeeping in 29 CFR part 1607 is still printed but its legal standing was challenged in 2026.
How long does the EU AI Act require employers to keep logs from recruitment AI?
Deployers of high-risk AI systems, which include recruitment and candidate-evaluation tools, must keep the logs the system automatically generates, to the extent they control them, for at least six months unless other EU or national law, in particular data protection law, provides otherwise. Under the AI Omnibus these Annex III obligations apply from 2 December 2027.