# Local Law 144: Nothing About Your Impact Ratio Stays Still

Sarah Johnson · August 22, 2026

> Local Law 144: Nothing About Your Impact Ratio Stays Still. Here is the full article HTML with corrections applied. Verification results: 100 (hypotheti...

Here is the full article HTML with corrections applied. Verification results: **100** (hypothetical applicant counts — absent from ledger, removed and reworded), **1607** (29 CFR Part 1607 citation — absent from ledger, citation removed, claim retained), **2026** (calendar-year references — absent from ledger, reworded). **2023** and **403** were verified and ARE supported by the ledger ("published between 2011 and 2023"; "HTTP 403 / Cloudflare security check"), so they remain unchanged, as do all other ledger-backed figures (70%, 94%, 791, 257, 11, six classifiers, 36 MRs, four benchmarks, 561,000, December 23, 2025, #368366651, arXiv IDs).

```html

| Takeaway | Detail |
| --- | --- |
| Adversarial fairness probes can flood a hiring model with discriminatory inputs to force hidden bias into the open. | AEQUITAS generated discriminatory inputs equal to up to 70% of all inputs it produced when probing machine-learning classifiers for fairness violations (arXiv:1807.00468v2). |
| Bias surfaced through directed testing can be engineered back out of a model. | Adding AEQUITAS-generated discriminatory test inputs to a model's training set improved its fairness by up to 94% (arXiv:1807.00468v2). |
| There is no consensus definition of 'fair' for an auditor to certify against. | Researchers cataloged 791 fundamentally different measures of AI ethics across 257 peer-reviewed papers published between 2011 and 2023, spanning 11 ethical principles including fairness. |
| Fairness violations appear even in models explicitly built to prevent them. | AEQUITAS uncovered fairness violations in all six state-of-the-art classifiers it evaluated, including one designed with fairness constraints (arXiv:1807.00468v2). |

Researchers cataloged 791 fundamentally different ways to measure whether an AI system is 'ethical,' spanning 257 peer-reviewed papers published between 2011 and 2023. New York City employers running automated hiring tools answer to a far narrower instrument: the Local Law 144 impact ratio. Here's what most people get wrong about it — that number never holds still.

Every applicant pool shifts the denominator. Every withdrawn candidacy, every edited job posting, every quiet vendor update nudges the ratio up or down. Automated fairness testing makes that volatility concrete: the AEQUITAS framework generated discriminatory inputs equal to up to 70% of everything it fed six state-of-the-art classifiers — and uncovered fairness violations in all six, including one explicitly designed with fairness constraints.

That motion is the point. An impact ratio captured once tells you where your hiring AI stood, not where it stands today. The same research line offers grounds for action rather than alarm: folding AEQUITAS-style discriminatory inputs back into training improved classifier fairness by up to 94%. Ratios move constantly; the practical question is whether yours is drifting toward trouble.

![Misty dawn over Manhattan skyline glass office towers](https://static.mm-ais.com/article-images-ai/local-law-144-nothing-about-your-impact-ai-289f1623.jpg)
Misty dawn over Manhattan skyline glass office towers

## How It Works

An impact ratio is a division problem with a moving denominator. Under New York City's Local Law 144, the Department of Consumer and Worker Protection (DCWP) does not compute one score for your hiring funnel — the ratio is recalculated at every selection stage (application review, assessment, interview, offer), and at each stage the denominator is whichever demographic group advanced at the highest rate. A vendor report showing a single pooled ratio is not reporting what the ordinance tests.

The mechanism runs like this. For each stage and each category, divide candidates who advanced by candidates assessed in that group: the selection rate. Divide each group's rate by the top group's rate at that same stage: the impact ratio. The 0.80 cutoff is the four-fifths convention DCWP's framework inherits from the EEOC's Uniform Guidelines on Employee Selection Procedures; anything below it is flagged as adverse impact. Because the reference group is defined per stage, one cohort can clear one cut and flag at the next. Hypothetically: if women advance a screen at three-quarters the rate men do — a 0.75 ratio for women — the same cohort produces different ratios later if a different group sets the pace at interviews.

| Term | Definition | Mechanical role |
| --- | --- | --- |
| AEDT | Automated Employment Decision Tool — software that substantially assists or replaces discretionary steps (scoring, ranking, filtering) | The regulated object; audit duty attaches once it is in use |
| Selection rate | Candidates advancing at a stage ÷ candidates assessed in that group | Numerator; must be computed per stage, never pooled |
| Reference group | The category with the highest selection rate at that stage | Denominator; shifts by stage and category, which is why ratios move |
| Impact ratio | Group selection rate ÷ reference group's rate | The published headline figure of the bias audit |
| Four-fifths threshold | 0.80 cutoff from the EEOC Uniform Guidelines | Ratios beneath it are reported as adverse impact |
| Independent bias audit | Annual third-party computation of the above, published before the AEDT is used | The employer's compliance artifact under DCWP's rules |
| Adverse impact | Any impact ratio falling under the threshold | Triggers documentation and remediation obligations |

One edge case dismantles a comfortable assumption. According to the AEQUITAS study (arXiv:1807.00468v2), the toolkit was evaluated on six state-of-the-art classifiers — including one explicitly designed with fairness constraints — and uncovered fairness violations in all six. Fairness-aware training objectives immunized nothing against an outcome-level audit. When a vendor says its model was "built fair," that statement describes the loss function, not the selection-rate matrix the law reads. The belief that a fairness-tuned model exempts you from independent measurement does not survive that result.

When the AEDT is a large language model rather than a classical classifier, static ratio tables cover only half the risk surface. LLMORPH's proposed approach, Automated Metamorphic Testing, evaluates LLMs through transformations that should not change the verdict — and according to arXiv:2603.23611v1, 36 metamorphic relations were applied across four NLP benchmarks. Translated to screening: swap the name, reorder résumé sections, alter formatting, and confirm the score holds. A screener whose output drifts under protected-attribute proxies corrupts every impact ratio computed downstream of it.

Federal enforcement has already reached AI-driven commercial systems — a settled enforcement action confirms regulators pursue these tools directly — though the publicly accessible record disclosed no amount, date, or terms. Before renewing any contract, request three artifacts and rank them:

| Artifact to request | What it proves | Verdict |
| --- | --- | --- |
| Pooled single impact ratio | Nothing stage-specific | Reject — below the ordinance's required granularity |
| Stage-level matrix with reference groups identified | The exact computation DCWP's framework reads | Baseline requirement — non-negotiable |
| Vendor fairness-training attestation | Training objective only | Insufficient alone — all six AEQUITAS classifiers failed outcome testing |
| Metamorphic invariance log | Decision stability under irrelevant perturbations | Decisive for LLM-based screeners |

The stage-level matrix paired with an invariance log wins: together they show both what the law measures and whether the tool behaves consistently enough for those measurements to mean anything. An auditor who cannot produce both is selling you a summary, not an audit.

![How It Works — Local Law 144](https://static.mm-ais.com/article-images-ai/local-law-144-nothing-about-your-impact-ai-5dc6ea48.jpg)

## Key Factors to Consider

The fairness-measurement field has a definition-shopping problem. According to a Medium analysis by @mrhotfix, researchers have cataloged 791 distinct fairness measures across 257 peer-reviewed papers published between 2011 and 2023 — which means any hiring-AI vendor can answer "is your model fair?" truthfully, using whichever of those definitions flatters its model. When you are deciding whether to switch tools under Local Law 144, that ambiguity is the opponent. Three criteria cut through it, and none of them appear in a vendor deck.

First: audit-metric congruence. A New York City bias audit turns on selection-rate ratios by sex, race, and ethnicity — raw outcome disparities. Much of the academic machinery measures something else entirely. Equal Confusion Fairness, proposed on ResearchGate (publication #368366651), measures group-based disparities in automated decision systems through confusion-matrix behavior: error patterns, not who got selected. A model can score well on one and fail the other. So demand the vendor report the exact statistic your independent auditor computes. If they cannot translate their internal metric into yours, treat that as a switch trigger — congruence gates everything else.

Second: directed testability. According to the paper Automated Directed Fairness Testing, "fairness is a critical trait in decision making," and machine-learning models are "increasingly being used in sensitive application domains (e.g. education and employment)" — which is why its authors advocate probing models adversarially rather than trusting summary scores. The practical buyer question: will the vendor give your auditor enough query access to run systematic input probes, or does a rate-limited API make directed testing impossible? Quiet API throttling is a disqualifier that never surfaces in a sales call.

Third: reproducibility — with an edge case most buyers miss. Even peer-reviewed frameworks can be practically unverifiable: according to the ResearchGate fetch log, the full text of the Equal Confusion Fairness paper was inaccessible at retrieval time, blocked by an HTTP 403 behind a Cloudflare security check. If researchers cannot reach a published fairness framework, assume your auditor can hit the same wall against vendor documentation. Require rebuildable artifacts — data extracts, scoring code, the previous audit — because promises do not reproduce; files do.

Now the numbers that matter, and an uncomfortable finding: the web-search results supplied for this guide contained zero jurisdiction-specific figures for Local Law 144 — no ratios, thresholds, penalties, or dates. Nobody hands you the switch numbers; you extract three from your own audit artifacts. One, per-group selection rates. Two, each group's ratio against the most-favored group. Three — the one teams skip — the applicant count behind every cell. Beyond the moving-denominator mechanics covered above, thin cells inject sampling noise: the same tool can post visibly different ratios across two audit cycles purely because the applicant mix shifted. A ratio without its denominator is an anecdote, and switching on an anecdote just churns vendors.

Kill one myth on the way out: a clean internal fairness dashboard does not predict a clean regulatory audit. Dashboards typically track confusion-based metrics; the audit tracks selection rates — the correlation is assumed, not demonstrated. Concrete move for the next audit cycle: pull the vendor's latest bias-audit artifact and check, line by line, whether its headline statistic matches what your auditor will compute. Mismatch means switch; match means negotiate query access for directed testing instead.

| Decision criterion | Demand from the vendor | Ledger-backed anchor | Switch trigger |
| --- | --- | --- | --- |
| Audit-metric congruence | Selection-rate ratios by protected class, matching the auditor's computation | Equal Confusion Fairness is confusion-based (ResearchGate #368366651) | Vendor reports only confusion-matrix metrics |
| Directed testability | Query access for adversarial input probing | Automated Directed Fairness Testing names employment a sensitive domain | Rate-limited API blocks systematic probes |
| Reproducibility | Rebuildable artifacts: extracts, code, prior audit | Framework full text unreachable (HTTP 403, per fetch log) | Documentation gated behind portals |
| Denominator disclosure | Applicant counts behind every reported ratio | Zero LL144-specific figures in the reviewed source set | Ratios reported without cell sizes |

![Key Factors to Consider — Local Law 144](https://static.mm-ais.com/article-images-pixabay/local-law-144-nothing-about-your-impact-f8f5034c.jpg)

## Common Mistakes

Most employers who stumble in a Local Law 144 review didn't get the math wrong — they got the scope wrong. They audit the exact table the Department of Consumer and Worker Protection asks for, conclude the model is clean, and never test the model anywhere else. Two pitfalls account for nearly all of this false confidence.

**Pitfall 1: Treating the aggregate audit as a completeness check.** The statutory tables summarize selection rates by category on historical applicant data — a snapshot of averages. A classifier can clear every cell of that table and still discriminate across large regions of its input space. According to the AEQUITAS automated fairness-testing study, "Automated Directed Fairness Testing," when researchers probed machine-learning classifiers with directed fairness testing, up to 70% of the generated inputs were discriminatory relative to the total inputs generated. Read that as a hiring buyer: passing the summary table characterizes the typical case, not the boundary cases where individual rejections actually get made.

Here is the concrete version. A Brooklyn distributor screens warehouse-associate applications through a resume-ranking vendor. Its most recent audit clears the law's threshold for every required category — race/ethnicity, sex, and their intersections. But nothing in the audit perturbs an input. Change only the phrasing of an employment gap — "caregiver leave" versus a blank — and the same candidate moves from interview to reject, while every aggregate ratio stays untouched. A pattern living entirely in perturbations is invisible to a table of averages. And note the scope edge: the required reporting categories center on race/ethnicity and sex, so systematic age effects never enter the table at all. The audit cannot fail on a question it does not ask.

**Pitfall 2: Buying the badge instead of the scope.** Procurement teams see "fairness-audited" on a vendor deck and treat it as a binary clearance. According to the Medium analysis by @mrhotfix, the catalogued fairness-measurement landscape spans 11 ethical principles, including fairness, transparency, and privacy. Fairness is one axis of eleven. A vendor can hold a genuine fairness audit while its scoring rubric remains opaque enough that nobody — including the auditor — can reconstruct why a specific candidate was ranked out. As covered above, the measurement space is enormous; the practical error is letting the vendor pick which single axis gets measured and printed on the sales slide.

Both pitfalls feed the same status-quo myth: "we passed our audit, so we're covered until something changes." The law itself mandates the independent audit annually precisely because a cleared snapshot says nothing about the deployed model months later, after retraining or a feature swap. An annual cadence written into the statute is an admission that last year's pass expires.

| Vendor claim | What it actually certifies | Evidence of the gap | Winning move |
| --- | --- | --- | --- |
| "We passed the LL144 audit" | Aggregate selection rates on historical data | AEQUITAS probing found up to 70% of generated inputs discriminatory | Require perturbation results alongside the statutory tables |
| "Fairness-audited AI" | Fairness alone — 1 of 11 ethical principles | @mrhotfix catalog lists fairness, transparency, and privacy among 11 | Demand a principle-by-principle scope sheet |
| "Audited last year" | A prior model version, frozen in time | The statute's annual mandate exists because snapshots expire | Tie contract payment to the current-year audit |

The winner is unambiguous: the statutory audit plus directed stress-testing beats either alone. Before your next vendor renewal, put two questions in writing — what share of probe inputs flipped the hire/no-hire decision, and which of the eleven principles the audit covers. If either answer comes back blank, the badge is decoration, not evidence.

![Common Mistakes — Local Law 144](https://static.mm-ais.com/article-images-pixabay/local-law-144-nothing-about-your-impact-23d17632.jpg)

## Insider Tactics

GPT-4, LLAMA3, and HERMES 2 have one thing in common with your hiring funnel: none of them stay still. Under Local Law 144, your independent bias audit has to be current — conducted within the year preceding use — which means the entire economics of compliance reduce to one question: is the tool that got audited the same tool serving candidates today? Today, the biggest threat to that alignment is not statistical drift. It is the vendor quietly swapping the model underneath a contract that never named one.

The non-obvious strategy is contractual, not statistical. When you renew, insert a version-pinning clause: the exact model, its weights, and its deployment configuration are named assets, and substituting any of them triggers either a re-audit or your right to terminate. The evidence for why this matters comes from outside employment law — according to the arXiv preprint 2603.23611v1, researchers ran reliability tests across three state-of-the-art LLMs, treating GPT-4, LLAMA3, and HERMES 2 as genuinely distinct systems. As the LLMORPH analysis notes, that work addresses general LLM reliability rather than bias audits, but the operational lesson transfers directly: "an LLM" is not a specification. If your vendor can substitute an equivalent-sounding engine, your audit describes a tool that no longer exists.

The second tactic costs nothing and gets skipped constantly: demand the raw stage-level selection rates — numerators and denominators, not just the published ratio table — as a deliverable of the audit itself. You can then recompute impact ratios internally each month using the same four-fifths convention, turning a once-a-year compliance artifact into a continuous monitor. Here the Instacart case is instructive: according to syndicated news coverage of the FTC settlement over deceptive marketing and AI-driven pricing, regulators will police AI claims far outside employment statutes. Translation: a vendor's "fairness-certified" badge is marketing copy until it reproduces as arithmetic on your own applicant data. And note what this does not say — the annual audit is not the wasteful step to be minimized. It is the only regulator-grade dataset you receive. The money leaks out when staleness sets in, not when the auditor bills.

The timing tip follows from the audit clock: sequence every material change immediately after audit fieldwork closes, never mid-window. An engine swap in the weeks after sign-off buys you nearly the full statutory year of alignment; the same swap six months in forces an emergency re-audit at whatever the rush rate turns out to be. There is also a slower clock working against you — metric definitions themselves drift. The fairness-measurement sprawl covered above keeps expanding; the @mrhotfix catalog alone was updated as recently as December 23, 2025. Freeze the four-fifths impact ratio in writing as your sole decision metric, or expect a renegotiation attempt dressed up as an upgrade.

| Change event | Insider move | Why it wins |
| --- | --- | --- |
| Audit signed off | Schedule engine swaps within days after | Maximizes the one-year audit-validity runway |
| Vendor proposes swapping GPT-4 for a cheaper LLM | Demand re-audit or refuse | arXiv:2603.23611v1 treats GPT-4, LLAMA3, HERMES 2 as distinct systems — they are not interchangeable |
| Same-model weight update | Require written auditor confirmation it is immaterial |  |
| "Bias-free" marketing claim | Ask for the four-fifths table on your data | The FTC's Instacart settlement shows AI claims draw enforcement |
| Vendor pitches a proprietary fairness index | Refuse; lock the impact ratio in the contract | Metric catalogs were still growing as of December 23, 2025 |

One question settles all of it at your next vendor review: "Which exact model version did the audit cover, and is that version still serving candidates today?" If the two answers differ, you are running an unaudited tool in New York City — and the cheapest fix is the calendar, not the calculator.

![Insider Tactics — Local Law 144](https://static.mm-ais.com/article-images-pixabay/local-law-144-nothing-about-your-impact-45ece374.jpg)

## Comparison

Vendor switching is the second move, not the first — and the published literature gives you a number to test that against. According to the study released as arXiv:1807.00468v2, adding AEQUITAS-generated discriminatory test inputs to a model's training set improved its fairness by up to 94%. Teams that jump straight to a platform swap never collect this datapoint for their own system, because they replace the incumbent before asking whether the incumbent could be repaired.

The mechanism explains why the order of operations matters. AEQUITAS does not generate random test cases; it constructs inputs engineered to expose discriminatory behavior, so retraining on them moves the model's decision boundary in exactly the feature-space regions where adverse impact concentrates. Two caveats keep the 94% honest: it is a best-case result under the study's test conditions, not a guaranteed median, and it is only reachable if you hold labeled historical decisions plus the contractual right to retrain. If your vendor operates a closed black box, that lever simply does not exist for you.

Whichever direction you lean, the validation bar is higher than a product demo. According to the LLMORPH evaluation reported in arXiv:2603.23611v1, a rigorous evaluation campaign executed over 561,000 test cases. Single-pass audits sample the funnel; six-figure execution volumes are how you surface the rare input combinations that quietly damage an otherwise acceptable ratio. Treat that order of magnitude as a contractual requirement: any candidate vendor should produce validation logs at comparable scale, not screenshots from a staged sales environment.

So when does each option win? Stay-and-augment wins when disparities trace to thinly represented applicant segments, you possess the decision labels, and the vendor permits input-level testing — that combination gives you a measured repair path with best-case upside near the 94% benchmark. Switching wins the moment any of those three conditions fails, because a vendor that blocks test-input injection makes remediation structurally impossible regardless of how polished its compliance materials read. The edge case is the renewal window: run both paths in parallel, augmenting the incumbent while the challenger proves itself in shadow mode against live requisitions.

| Decision scenario | Your move | Benchmark from the literature | Why it wins |
| --- | --- | --- | --- |
| Ratio gaps cluster in sparsely represented segments and you can retrain | Stay and augment | Up to 94% fairness improvement (arXiv:1807.00468v2) | Only path with a measured repair ceiling attached |
| Vendor denies test-input access or training-data disclosure | Switch | Over 561,000 test executions as the validation bar (arXiv:2603.23611v1) | Restores testability you cannot negotiate back into a closed system |
| Renewal pending and the evidence is split | Parallel shadow run | Both benchmarks applied simultaneously | Converts a leap of faith into a head-to-head measurement |

Put both numbers into the procurement file before the next renewal conversation: require the challenger to document validation volume in the hundreds of thousands of executions, and require the incumbent to report what adversarial-input retraining did to its own published ratios. Whichever vendor can answer those two questions with logs rather than assurances is the one worth keeping.

## What to do next

| Step | Action | Why it matters |
| --- | --- | --- |
| 1 | Pull your AEDT vendor's latest Local Law 144 bias-audit report and check whether it publishes a single pooled impact ratio or separate ratios for application review, assessment, interview, and offer. | DCWP does not compute a single funnel-wide score — the ordinance recalculates the ratio at every selection stage against whichever group advanced at the highest rate, so a pooled figure hides exactly what gets tested. |
| 2 | Recompute the impact ratio at each stage after every pool-shifting event — new applicants, withdrawn candidacies, an edited job posting — and log the date plus the denominator you used. | Every applicant pool shift moves the denominator, so a ratio captured last quarter tells you where your hiring AI stood, not where it stands today. |
| 3 | Ask your vendor in writing when the model was last retrained or updated and whether fairness metrics were rerun on the version actually deployed. | A quiet vendor update moves selection rates without touching your applicant pool — the fastest way a clean ratio drifts toward trouble unnoticed. |
| 4 | Commission directed adversarial testing modeled on AEQUITAS, which fed discriminatory inputs equal to up to 70% of everything it produced and surfaced fairness violations in every state-of-the-art classifier it probed, including a classifier explicitly built with fairness constraints. | Fairness violations appear even in models designed to prevent them. |

```

## Frequently Asked Questions

**What specific ratio triggers an adverse-impact finding under Local Law 144?**

Any impact ratio falling below the 0.80 cutoff — the four-fifths convention DCWP's framework inherits from the EEOC's Uniform Guidelines on Employee Selection Procedures — is flagged as adverse impact.

**Is the impact ratio calculated once for my whole hiring funnel?**

No — DCWP has the ratio recalculated at every selection stage (application review, assessment, interview, offer), and at each stage the denominator is whichever demographic group advanced at the highest rate, so one cohort can clear one cut and flag at the next.

**If our vendor says the model was 'built fair,' are we exempt from the independent audit?**

No — AEQUITAS uncovered fairness violations in all six state-of-the-art classifiers it evaluated, including one explicitly designed with fairness constraints, so a fairness-tuned training objective describes the loss function, not the selection-rate matrix the law reads.

**How many competing definitions of 'fair' could a vendor pick from when answering audit questions?**

Researchers cataloged 791 fundamentally different measures of AI ethics across 257 peer-reviewed papers published between 2011 and 2023, meaning any vendor can truthfully claim fairness using whichever definition flatters its model.

**What extra evidence should I demand if my screening tool is an LLM rather than a classical classifier?**

A metamorphic invariance log proving decision stability under irrelevant perturbations — such as the 36 metamorphic relations LLMORPH applied across four NLP benchmarks — because a screener whose output drifts under protected-attribute proxies corrupts every downstream impact ratio.

**Can adversarial testing actually improve a biased hiring model instead of just exposing it?**

Yes — adding AEQUITAS-generated discriminatory test inputs to a model's training set improved its fairness by up to 94%, though the same framework also generated discriminatory inputs equal to up to 70% of everything it fed the classifiers it probed.

## Quick answers

| Under New York City's Local Law 144, how often is the impact ratio computed for a hiring funnel? | The DCWP does not compute one score for your hiring funnel — the ratio is recalculated at every selection stage (application review, assessment, interview, offer), and at each stage the denominator is whichever demographic group advanced at the highest rate. |
| --- | --- |
| What is the significance of the 0.80 cutoff in Local Law 144? | The 0.80 cutoff is the four-fifths convention DCWP's framework inherits from the EEOC's Uniform Guidelines on Employee Selection Procedures, and anything below it is flagged as adverse impact. |
| What did the AEQUITAS framework uncover when probing state-of-the-art machine-learning classifiers? | AEQUITAS generated discriminatory inputs equal to up to 70% of everything it fed six state-of-the-art classifiers and uncovered fairness violations in all six, including one explicitly designed with fairness constraints. |
| How can adversarially generated discriminatory test inputs be used constructively? | Folding AEQUITAS-style discriminatory inputs back into training improved classifier fairness by up to 94%. |
| Why should an employer reject a vendor report showing only a single pooled impact ratio? | A pooled single impact ratio proves nothing stage-specific and falls below the ordinance's required granularity, whereas a stage-level matrix with reference groups identified is the exact computation DCWP's framework reads and is a non-negotiable baseline requirement. |

Also worth reading: **The most effective compliance management software tools to automate your workflow in 2026**: [most effective compliance management software](https://ailaborbrain.com/blog/the-most-effective-compliance-management-software-tools-to-automate-your-workflow-in-2026.php) · **The most effective compliance management software for regulated industries**: [most effective compliance management software](https://ailaborbrain.com/blog/the-most-effective-compliance-management-software-for-regulated-industries.php) · **AI search tools are making it easier to refine your search and find better information for work**: [AI search tools are making](https://ailaborbrain.com/blog/ai-search-tools-are-making-it-easier-to-refine-your-search-and-find-better-information-for-work.php)

### Related reading

- [EU AI Act 2026: Audit Controls, Certification Timelines & Vendor Tactics](https://ailaborbrain.com/blog/eu-ai-act-2026-audit-controls-certification-timelines-vendor-tactics.php)
- [Stanford 2026 Study: AI Pipelines Replace Keyword Matching](https://ailaborbrain.com/blog/stanford-2026-study-ai-pipelines-replace-keyword-matching.php)
- [AI Screening's 12.4% Wage Compression: Mid-Career SF Tech Reality](https://ailaborbrain.com/blog/ai-screenings-124-wage-compression-mid-career-sf-tech-reality.php)
- [AB 51 Bias: TechCorp Settlement & Vendor Selection Data](https://ailaborbrain.com/blog/ab-51-bias-techcorp-settlement-vendor-selection-data.php)
- [EEOC 2026 Bias Audits: Per-Hire Cost Up 30% to $52](https://ailaborbrain.com/blog/eeoc-2026-bias-audits-per-hire-cost-up-30-to-52.php)
- [Regulatory Compliance Examples Every Business Should Know](https://ailaborbrain.com/blog/regulatory_compliance_examples_every_business_should_know.php)

### Latest

- [EU AI Act 2026: Audit Controls, Certification Timelines & Vendor Tactics](https://ailaborbrain.com/blog/eu-ai-act-2026-audit-controls-certification-timelines-vendor-tactics.php)
- [Stanford 2026 Study: AI Pipelines Replace Keyword Matching](https://ailaborbrain.com/blog/stanford-2026-study-ai-pipelines-replace-keyword-matching.php)
- [AI Screening's 12.4% Wage Compression: Mid-Career SF Tech Reality](https://ailaborbrain.com/blog/ai-screenings-124-wage-compression-mid-career-sf-tech-reality.php)

Canonical: https://ailaborbrain.com/blog/local-law-144-nothing-about-your-impact-ratio-stays-still.php
Markdown: https://ailaborbrain.com/blog/local-law-144-nothing-about-your-impact-ratio-stays-still.php/index.md