In the last post I told you about the candidate I loved who didn’t exist — the FAANG-pace answers, the “internet issues,” the moment I finally asked the recruiter whether we’d verified he’d worked anywhere he claimed.

What I didn’t tell you is what I went looking for next. I wanted a way to take everything a candidate hands you — résumé, LinkedIn, portfolio, and the interview itself — and ask one simple question: do these four things tell the same story?

Not “is this person AI?” Not “score this candidate.” Just: does the résumé match the LinkedIn, does the LinkedIn match the portfolio, and does the interview match all three? When a real person tells you about the launch they led, it shows up in their portfolio, their title lines up, the dates reconcile, and the details get more specific when you dig. When a story is borrowed, it’s the opposite — it gets vaguer under pressure and the sources quietly disagree.

There’s a name for this in research: triangulation — using independent sources to converge on the truth. I started calling it Signal Triangulation, and the thing it produces is a grounding effect: every claim in the interview either gets grounded in evidence the candidate already put on the table, or it doesn’t.

Two surprises followed. First: the tools sold to do this are terrible at it. Second: the moment you reach for AI to help, you walk straight into the single most litigated question in hiring right now. Both are worth your time.

Why HR tools are horrible at grounding

You’d think this is a solved problem. It isn’t, for a few structural reasons.

Your ATS matches keywords, not stories. Applicant tracking systems were built to rank résumés against a job description and move volume. They have no concept of “this portfolio corroborates this résumé claim.” They can tell you the word “Kubernetes” appears; they cannot tell you whether the project the candidate described in the interview is the same one on their GitHub, or whether the dates make sense.

Background checks verify the wrong layer. A standard background check confirms that an employment record exists — dates, title, maybe a degree. It does not check whether the narrative is real. The KnowBe4 hire — the North Korean operative who cleared four video interviews on a stolen identity (KnowBe4) — passed the background check. The identity was real. The person wasn’t.

Nothing cross-references across sources. Résumé, LinkedIn, portfolio, and interview live in four different tools (or four different tabs), and the comparison happens, if at all, in a tired recruiter’s head at 5pm. There is no system that reads all four and flags “the interview claims they led this, but the portfolio shows them as a contributor, and the LinkedIn dates don’t overlap.”

And the AI that HR does buy is the dangerous kind. The AI tooling that’s actually deployed in hiring is overwhelmingly the selection kind — software that scores, ranks, and rejects candidates. Which brings us to the legal problem, because that is precisely the category now getting sued into the ground.

What the law actually says (and what it doesn’t)

Let me be clear up front: I’m a hiring manager, not your lawyer, and this is not legal advice. But you cannot reason about AI in hiring without knowing where the lines are, and the lines are clearer than the noise suggests.

The federal “rollback” changed the guidance, not the law. In January 2025 the EEOC quietly removed its AI hiring guidance, and an April 2025 executive order told agencies to deprioritize disparate impact enforcement (K&L Gates). It’s tempting to read that as “the coast is clear.” It is not. As multiple firms have stressed, Title VII, the ADA, and the ADEA didn’t change — the agency just stopped explaining how they apply, and applicants can still sue (Cooley). Removing the road signs didn’t remove the cliff.

The private suits are where it’s happening. In Mobley v. Workday, a federal judge in May 2025 let a nationwide collective action proceed alleging Workday’s screening algorithms discriminated by age, race, and disability — and, critically, found the AI vendor could be liable as an “agent” of the employer (Holland & Knight). Every selection-AI vendor and every employer using one is now on notice. The thing being sued is AI making or shaping the reject decision.

The state laws all hinge on the same word: “decision.” This is the part hiring managers miss. The regulations don’t ban AI in hiring — they regulate AI that decides:

• New York City (Local Law 144) regulates an “automated employment decision tool” — defined as AI that issues a score, classification, or recommendation used to “substantially assist or replace discretionary decision-making.” A tool counts if it’s the sole determinant, the primary factor, or can override human judgment (NYC DCWP). Use one, and you owe a published bias audit.
• Illinois (HB 3773, effective January 1, 2026) requires notice when AI is used “for employment decisions,” bars using zip codes as proxies for protected classes, and flatly prohibits AI that discriminates (Seyfarth).
• California’s automated-decision-system regulations took effect October 1, 2025, with disclosure and opt-out duties when automated tools replace human decision-making (Seyfarth).
• Colorado hit reset — its broad AI Act was repealed and replaced, now effective January 2027 (Troutman) — but employer risk under existing discrimination law didn’t pause.

Read those together and a pattern jumps out. Every one of them triggers on the same thing: AI that makes, substantially assists, or replaces the employment decision. Score a candidate, rank them, auto-reject them — you’re squarely inside the regulated zone, you owe audits and notices, and you’re exposed to the Mobley theory of liability.

So: can AI be used for grounding instead of deciding?

This is the question my recruiter friends keep asking, usually in a frustrated tone: “We’re being told we can’t use AI to screen people out because of bias. So what can we use it for?”

Here’s the distinction I’ve landed on, and I think it’s the whole ballgame.

The laws regulate AI that produces a score, classification, or recommendation that drives the hiring decision. They do not purport to ban AI that surfaces facts for a human to evaluate. There is a real, defensible difference between:

Selection: “The model rates this candidate 62/100. Reject.” (regulated, audited, litigated)

and

Grounding: “Three claims in the interview are corroborated by the portfolio and LinkedIn. One claim — ‘led the migration’ — conflicts: the LinkedIn dates don’t overlap the employer, and the portfolio lists them as a contributor. Here are the documents. You decide.” (verification / decision-support)

The first is an automated employment decision tool. The second is closer to what a diligent recruiter already does by hand — it just does it consistently and surfaces the receipts. It doesn’t score the person, rank them, or recommend an outcome. A human reads the underlying evidence and makes the call.

But — and this matters — “substantially assist” is a broad phrase, and the design is what keeps you on the right side of it. A “grounding” tool that spits out a trust score recruiters rubber-stamp into rejections is just a selection tool wearing a disguise, and a court will treat it that way. To actually stay in the verification lane, the discipline is non-negotiable:

1. Surface evidence, never a score. Output corroborations and contradictions with the source documents attached — not a number, ranking, or “recommend/don’t.”
2. Keep a human as the decider, reviewing the facts — not approving the AI’s verdict.
3. Apply it uniformly to every candidate. Running verification only on certain people is textbook disparate treatment.
4. No protected-class proxies (the Illinois zip-code rule is the canary here) and an auditable trail.
5. Absence is not evidence. A thin LinkedIn or no public portfolio is missing data, not a red flag — common for privacy-conscious people, career-switchers, and folks from places where these platforms aren’t the norm. Contradictions are high-signal; gaps are not. Score corroboration only where a signal exists.

That last point is also the fairness point, and it’s not a coincidence. The same speech and assessment AI that gets employers sued discriminates against people with accents, deaf candidates, and non-native speakers (ADA.gov). Grounding sidesteps that entire failure mode, because it checks the substance of someone’s record, not the polish of their delivery. A brilliant engineer who interviews in their third language grounds perfectly — their portfolio is real, their dates reconcile, their details deepen under questioning. The faker is the one whose sources quietly disagree. The discipline that keeps grounding legal is the same discipline that keeps it fair.

The grounding effect

Here’s the part I didn’t expect. When grounding is in the room, the interview itself changes.

Candidates who know their story will be checked against their own paper trail tell truer stories — the same way people drive better past a marked patrol car. And interviewers, freed from playing amateur detective, can do the thing they’re actually good at: have a real conversation, and dig where the evidence says to dig. The contradiction the system surfaces (“led” vs. “contributed”) isn’t a verdict — it’s a better follow-up question. You hand it back to the candidate: “Tell me more about your role in that migration.” Real owners light up. Borrowed stories wobble.

That’s where I’ve come out. AI’s safe, useful, legal home in hiring isn’t picking your candidate — it’s grounding the truth so a human can pick well. Score the person and you’ve bought yourself an audit, a lawsuit, and a system that punishes the wrong people. Ground the story and you’ve just given a careful human better evidence.

My fake favorite would have failed grounding in about ninety seconds. Not because a model flagged him as artificial — but because not one of his claims would have had a document behind it. That’s the check I actually needed. It turns out it’s also the only one I’m comfortable letting a machine help with.

---

Not legal advice — talk to employment counsel before deploying anything that touches a hiring decision, especially across NYC, Illinois, California, and the EU (whose AI Act classifies employment AI as “high-risk”).

Sources: EEOC guidance rollback — K&L Gates · Federal law still applies — Cooley · Mobley v. Workday — Holland & Knight · NYC Local Law 144 (AEDT) — NYC DCWP · State law roundup (IL/CA/CO) — Seyfarth · Colorado repeal/replace — Troutman · AI & disability discrimination — ADA.gov · KnowBe4 fake IT worker