Hiring Assistant 2 Learns From Your Past Hires. Colorado May Call It ADMT

LinkedIn Hiring Assistant 2 starts reaching existing English-language Hiring Assistant customers in November, and it brings a memory that learns how you have hired before. On October 26, Colorado's Attorney General holds the rulemaking hearing for SB26-189, a law that names "rankings or scores" as covered output. If you source Colorado candidates, those two dates land on your desk together, and what you log between now and January 1 matters more than which tool you use.

What Hiring Assistant 2 actually changes in November

Hiring Assistant 2 adds two things that matter for compliance. One is a memory that reorders candidates based on your past hiring patterns. The other is a "candidate interest" score built on how people behave on LinkedIn. Both shape the shortlist before you look at it.

LinkedIn announced the update at Talent Connect in New York on September 29. Dan Reid (VP Product) and Prashanthi Padmanabhan (VP Engineering) wrote the announcement. Existing English-language Hiring Assistant customers get it as a free, automatic update starting in November. The agent also moves past sourcing into screening and scheduling.

The parts worth understanding:

• Memory at two levels. It learns at the recruiter level and the company level, and it reaches into connected ATSs: Greenhouse, Lever, SmartRecruiters, Zoho Recruit and Tracker today, with iCIMS and Workday coming.
• Inferred preferences. LinkedIn's own example is a product-manager search that draws on the recruiter's hiring history, "favoring candidates close to the office, from the retail industry and at mid-level experience."
• Candidate interest score. This estimates how receptive someone is likely to be, using platform engagement, Open to Work details and how often the person answers InMail.
• Scale. More than 20,000 companies use LinkedIn's agentic hiring tools, and recruiters are 27% more likely to contact a Hiring Assistant candidate than they were six months ago.

That last figure is the important one. A 27% rise in contact rate suggests the ranking is steering who you reach out to. Steering is exactly what the Colorado statute is written to catch.

Does a memory-driven shortlist count as ADMT under SB26-189?

Probably yes. A Hiring Assistant 2 shortlist produces rankings and scores, which the statute names directly. The only real question is whether it qualifies for the "present information for human review" carve-out, and the word "solely" makes that hard to claim.

Automated decision-making technology (ADMT), under SB26-189, is technology that produces "predictions, recommendations, classifications, rankings or scores that are used to make, guide or assist a decision concerning an individual." The law was signed on May 14, 2026, with obligations starting January 1.

The carve-out covers tools "used solely to summarize, organize, translate, draft, route or present information for human review." A results page sorted by date fits that description. A list reordered by what you hired last quarter does not, because the system has already applied a judgment before any human sees the output.

The real fight is the exemption, not the definition

Nobody serious will argue that a ranked shortlist isn't a ranking. The argument will be about whether it "materially influences" the decision or just informs a human who decides. The AG has to define "materially influence" in these rules, alongside the post-adverse-outcome disclosure requirements. An interim draft came out on October 6. The comment period closes at 11:59 p.m. Mountain Time on October 26 unless it is extended.

That definition settles the question for every memory-driven sourcing tool, not only LinkedIn's. If it reads broadly, a shortlist you acted on is covered. If it reads narrowly, you might argue the human review was independent. Plan for the broad reading, because the 27% contact lift is the kind of number a regulator would use to show influence.

A ranker that learned from your last ten hires is not presenting information. It is applying your past judgment to new people.

Memory turns your past hires into criteria you never typed

The biggest compliance risk in Hiring Assistant 2 is not the ranking. It is that the ranking runs on criteria you never wrote down, and the law expects you to describe them to a rejected candidate.

SB26-189 requires a 30-day adverse-outcome notice, and that notice has to say what the tool actually did. Ogletree frames the notices as a two-stage framework, and this post-adverse-outcome stage is the one that bites. Michael Hoog's practical guidance on SB 26-189 notices is blunt: a vague line saying "automated tools were used" does not qualify, and "you cannot describe what the system did if the vendor will not tell you."

Look at LinkedIn's own example again. "Near the office, retail, mid-level" are three filters the recruiter never entered. Proximity can stand in for neighbourhood demographics. Industry history can replicate whoever happened to apply years ago. Mid-level can work as an age proxy. If a candidate who lived farther out and came from healthcare got dropped, and all you have on file is the job title you searched, you cannot write an honest notice.

The fix is boring and works: log which inferred criteria you saw and which you accepted. That log is not overhead sitting next to the compliance work. It is the compliance work.

Two more details make this harder:

• Memory can't be exported. HeroHunt notes that teams who want a different talent profile have to tell the agent explicitly to ignore its learned patterns. You can't hand an auditor a file of what the memory believed.
• It changes as you work. Each hire teaches it something. The shortlist you got in March is not reproducible in June.

This is where explicit criteria pay for themselves. With Refolk, you describe the person you want in plain English and get a ranked list back. The sentence you typed is the criteria, so there's no hidden layer of learned preferences to reconstruct later. It doesn't remove your obligations. It does mean that what you searched for and what the tool ranked on are the same text.

Candidate interest is a behaviour score, not a fit score

The candidate interest score ranks people by how active they are on LinkedIn, not by whether they can do the job. If it pushes someone off a slate, the honest reason to record is "low platform engagement," and that is an awkward thing to put in a notice.

The inputs are engagement, Open to Work status and InMail response rate. Every good sourcer knows the people who matter most often score lowest on all three: the staff engineer who hasn't posted in years, the operator who ignores InMail on principle, the person who would never flip Open to Work while still employed.

Why it matters legally: InMail reply rate and platform engagement aren't job-related qualifications. Under SB26-189 you have to explain what the tool did. Under existing discrimination law, you have to defend why it mattered. Buchalter points to the Workday applicant case as a reminder that those claims are live today, whatever happens with the new statute.

My rule: use candidate interest to order your outreach, never to decide who gets cut. If a low score is why someone didn't make the slate, write that down in so many words. If you would be embarrassed to write it, don't let it make the cut.

Who this hits, and why you can't scope it out by geography

1,587 US recruiters, technical recruiters, TA staff and sourcers in Refolk's index list "LinkedIn Recruiter" as a skill. That is the closest proxy for the people getting the automatic November update. You can't reliably tell which of them, or which of their candidates, are in Colorado from profile keywords.

1,587

The population data, from Refolk's index of professional profiles:

The 289 Denver-headline profiles look like a Colorado cohort until you check where they actually sit. The top regions in the sample were mostly outside the state: Oxnard, Bentonville, Dallas. "Denver" in a headline can mean a past role, a client or a territory as easily as a home address. Require all three terms (Colorado, Denver, Boulder) and you get zero. People don't write their headlines like a location form.

The same applies to candidates. If you're deciding which searches fall under SB26-189 by keyword-matching profiles to Colorado, you will miss people. Two practical consequences follow:

1. Log every search, not just the ones you think are Colorado. It costs less than proving after the fact which candidates lived where.
2. Scope on location data, not headline text. Ask who is actually based in the state, then verify.

What to keep on record, whatever tool you use

Keep enough to rebuild any shortlist and explain any cut: your stated criteria, the inferred criteria you confirmed or rejected, the tool version, and why each dropped candidate was dropped. SB26-189 requires at least three years of retention, with ADMT version identifiers, changelogs and documentation of material changes.

A minimum log per search:

• Search criteria as entered. Exact text or filters, timestamped.
• Inferred criteria shown. Anything the tool added from memory (location radius, industry, seniority), and whether you accepted, edited or overrode it.
• Tool and version. Product name, version or release date, and a note when the vendor pushes an update. The November rollout is automatic, so the version can change without you doing anything.
• Scores relied on. If candidate interest or a fit score affected ordering, say so.
• Slate boundary. How far down the list you reviewed, against the total returned. A regulator will ask.
• Drop reasons. One line per candidate cut after review, in job-related terms. If the true reason was a score, write the score.
• Override record. Times you told the agent to ignore learned patterns, which HeroHunt says is the only way to reset a talent profile.

Don't count on the vendor to carry this. Employers can't escape liability for off-label use of AI tools, and contract clauses that shield a deployer from liability for its own discrimination are void. "LinkedIn built it" won't work as a defence.

Vendors like Warden AI are positioning themselves as evidence infrastructure for SB26-189, and that category will grow. A spreadsheet works too, as long as it's filled in at the time and not reconstructed afterwards.

Jan 1 may slip. Don't wait for it.

Enforcement probably starts later than January 1, but records created from that date still fall inside the three-year retention window. Waiting saves you nothing and leaves gaps you can't fill later.

The timing is messy. In xAI LLC v. Weiser, the Department of Justice intervened, the first time it has joined a challenge to a state AI law. xAI must file its preliminary injunction motion within 28 days after Colorado finalizes its rules, and enforcement stays suspended until 14 days after the court rules on that motion. Add the time it takes AG Phil Weiser's office to finalize rules after the October 26 hearing, and real enforcement could slip well past January 1.

None of that changes three facts:

• The retention clock covers decisions made while enforcement is paused.
• Discrimination claims are actionable under existing law today.
• Hiring Assistant 2's memory starts working in November, so the patterns it applies in January will have been built in your current hiring cycle.

Start logging with the November rollout. That way your first auditable shortlist is one you can actually explain.

FAQ

Does Colorado SB26-189 apply to small recruiting firms or solo founders?

Yes. The law has no small-business carve-out and no employee threshold, so a solo founder using an AI ranker to source Colorado candidates has the same notice and record-keeping duties as an enterprise TA team. The 30-day adverse-outcome notice and the three-year retention requirement apply regardless of headcount.

What has to go in an AI recruiting tool adverse action notice?

It has to describe what the tool actually did in that candidate's case, not just state that automated tools were used. In practice that means the criteria applied (including any inferred from your hiring history), the scores relied on, and the reason for the outcome. If you didn't log the inferred criteria at the time, you won't be able to write an accurate notice later.

Should I stop using Hiring Assistant 2 for Colorado roles?

Not necessarily. Use it for retrieval and ordering, log the inferred criteria it applies, and don't let candidate interest or memory-based ranking make the cut without a human writing down a job-related reason. The tool is not the liability. An unexplained cut is.