Source smarter, screen honestly, and know exactly where the AI stops and your judgment has to start.
LinkedIn’s AI-Assisted Search and Hiring Assistant are real productivity tools: LinkedIn’s own data shows an 18% InMail acceptance lift from AI-Assisted Search and a 69% lift for Hiring Assistant’s charter customers.[1][2] ChatGPT and Claude are genuinely useful for drafting outreach and screening questions in minutes instead of hours. But AI resume screening carries real bias risk: a 2024 University of Washington study found AI models favored white-associated names 85% of the time versus 9% for Black-associated names.[3] If you’re screening for a New York City-based role, you likely need a bias audit under [Local Law 144](https://www.nyc.gov/assets/dca/downloads/pdf/about/DCWP-AEDT-FAQ.pdf). Use AI to save time on sourcing and drafting. Keep a human making every screening decision.
Post a mid-level product manager role on LinkedIn on a Monday, and by Wednesday you’re staring at 380 applicants, a Recruiter inbox full of half-warm InMail replies, and a hiring manager asking when the shortlist is coming. That’s the actual problem AI is solving in recruiting right now: not some abstract future of robot interviewers, but the very real math of too many resumes and too few hours. Most HR teams have already made the call. A 2024 ResumeBuilder.com survey of 948 business leaders found 82% of companies already use AI to review resumes, and 51% use AI somewhere in their hiring process today.[6]
LinkedIn knows this too, which is why it’s built AI directly into Recruiter instead of leaving it to third-party plug-ins. In the HR workshops I run, it’s common to hear someone in Talent Acquisition quietly admit they’ve been using two or three AI tools for months without anyone in Legal knowing, which tells you how far ahead of policy this has moved. There are two genuinely different jobs AI gets asked to do here, and mixing them up is where most of the confusion, and most of the risk, comes from. Sourcing is finding people: searching LinkedIn’s network for candidates who fit a role, including ones who never applied and aren’t actively looking. Screening is evaluating people: ranking, scoring, or filtering candidates already sitting in your pipeline. LinkedIn’s AI tools are genuinely strong at the first job. The second one is where you need to slow down.
Quick framing before we go further: sourcing AI helps you find more of the right people, faster. Screening AI decides who gets a callback. Treat them with very different levels of trust.
Open Recruiter, click the search bar, and instead of building a Boolean string by hand you can paste your actual intake notes, a job description, or just type what you’re looking for in plain English, something like “senior backend engineer with distributed systems experience who’s led a team of 5+.” AI-Assisted Search reads that and translates it into a structured search, matching not just on keywords but on qualifications that show up in context. LinkedIn describes this as matching skills that aren’t explicitly listed on a resume but are implied by someone’s actual experience.
LinkedIn’s own product data shows AI-Assisted Search sessions get an 18% higher InMail acceptance rate than searches built with manual filters.[2] That’s a meaningful lift for a change that mostly just saves you the time of building the search itself. One recruiter using it, Victoria Östryd Söderlind at Toyota Material Handling Europe, told LinkedIn that a search that used to take 15 minutes now takes about 30 seconds.[2]
Here’s the part worth being straight about: AI-Assisted Search is still searching LinkedIn’s existing member data. It doesn’t find people who aren’t on the platform, and it doesn’t fix a badly written prompt. If your intake notes describe “culture fit” in vague, coded language, the AI will happily search on that vague language too. Garbage prompt in, biased-adjacent shortlist out. Write your search prompts the way you’d want a sharp researcher to interpret them: specific, skills-based, and free of proxy language for age, gender, or background.
Hiring Assistant is LinkedIn’s first AI agent built specifically for recruiters. It launched to a charter group of customers in October 2024 and went globally available in English by the end of September 2025.[1] It’s a paid add-on to Recruiter, not a free upgrade, and it’s built to take over more of the repetitive middle of the sourcing process: running searches based on an intake conversation and surfacing candidates for you to review.
According to LinkedIn, early adopters including Siemens, Expedia Group, and Aurecon are saving 4+ hours per role, reviewing 62% fewer candidate profiles, and seeing a 69% improvement in InMail acceptance rates.[1] A Senior Talent Acquisition Partner at Siemens put it this way in LinkedIn’s own case study: instead of spending an hour sourcing for one project, she can now source for five or more projects in 10 to 15 minutes.[1]
Worth saying plainly: those numbers come from LinkedIn’s own case studies of its own charter customers, the kind of testimonial that shows up in a product launch press release, not an independent audit. That doesn’t make them false. It does mean you should treat a “69% improvement” as directional, not as a guarantee you’ll see on your own req. Hiring Assistant makes the most sense if you’re running high volume: multiple open roles at once, constant sourcing pressure. If you’re a hiring manager filling one role a quarter, the AI-Assisted Search already built into Recruiter gets you most of the benefit without the add-on cost.
LinkedIn’s own AI can draft InMail messages for you inside Recruiter, but plenty of recruiters, and just about every hiring manager without a Recruiter seat, reach for ChatGPT or Claude instead, either to draft the first message or to punch up whatever LinkedIn generated. Used well, this is one of the highest-leverage things AI does in recruiting: it turns a twenty-minute stare-at-a-blank-message-box problem into a ninety-second edit.
The trick is feeding it specifics, not a vibe. Don’t ask AI to “write a friendly outreach message.” Paste the candidate’s actual headline, their most recent role, and one specific detail from their profile worth referencing: a project, a publication, a skill combination that’s genuinely rare. Then ask for a short, direct message that references that detail in the first line, states the role and one concrete reason it might interest them, and ends with a low-pressure ask. Tell it explicitly: no flattery filler, no “I came across your impressive profile,” no exclamation points.
I’ll be blunt about something that bugs me: a huge amount of LinkedIn InMail right now reads like it was generated from the same three prompts, because it was. “I hope this message finds you well” followed by a paragraph of generic praise is instant delete territory for anyone who gets recruiter messages regularly. AI didn’t invent InMail spam, but it made low-effort spam a lot cheaper to produce at volume. The fix isn’t avoiding AI, it’s using it to go narrower and more specific per message instead of blasting the same draft to two hundred people. If you wouldn’t send the message to a friend without changing a word, don’t send it to a candidate either.
Once candidates are in your pipeline, AI is genuinely useful for building the structured part of screening: turning a job description into a consistent set of questions every candidate gets asked, rather than whatever the interviewer thinks of on the spot. Paste your job description into ChatGPT or Claude and ask for screening questions split into two groups: must-have qualifications that disqualify a candidate if missing, and preferred qualifications that differentiate between candidates who clear the bar. Ask it to phrase each question behaviorally (“tell me about a time you…”) rather than yes or no, since yes-or-no answers are easy to fake and hard to score consistently.
Also ask it, directly, to flag any question that touches age, disability, family or marital status, national origin, or anything else that could raise a Title VII or ADA issue before an offer is made. AI is decent at catching the obvious version of these (“how old are your kids?”) but not the sneaky version (“what year did you graduate?” as a proxy for age), so a human still needs to read the final list. The EEOC’s specific AI guidance was pulled from its website in January 2025, but the underlying law didn’t go anywhere: Title VII and the Uniform Guidelines on Employee Selection Procedures still apply to AI-assisted hiring the same way they apply to a human interviewer.[5]
Yes, and the data on this is more specific than most people expect. Researchers at the University of Washington tested three widely used large language models, from Mistral AI, Salesforce, and Contextual AI, by running more than 550 real resumes against over 500 real job listings across nine occupations, producing more than 3 million resume-to-job comparisons. They varied only the names on the resumes, swapping in names associated with white and Black men and women, and kept everything else identical.[3]
The models preferred white-associated names 85% of the time versus 9% for Black-associated names. They preferred male-associated names 52% of the time versus 11% for female-associated names. And in a detail the lead researcher called a unique harm, the models never once preferred a Black male-associated name over a white male-associated name, across the entire study.[3] This isn’t a hypothetical, one-off finding. It’s three different commercial-grade models, tested at scale, all showing the same pattern.
Here’s what genuinely frustrates me about this: companies aren’t blindsided by the risk, they’ve told researchers about it themselves. In that same ResumeBuilder.com survey, 9% of business leaders said their AI hiring tools always produce biased recommendations and another 24% said they often do, meaning close to a third of companies using AI to screen resumes are knowingly running a tool they believe is biased, often. And 21% let AI reject candidates automatically at every stage of hiring, with no human reviewing the rejection.[6] Nobody stumbled into that setup by accident. Reviewing every AI rejection takes time, and skipping it is faster, which is exactly why teams keep doing it even when they suspect the tool is biased. If you’re going to use AI to screen resumes, the minimum bar is a human looking at every rejection before it goes out, not just every hire.
If you’re hiring for a role based in a New York City office, including a remote role tied to an NYC office, and you use software with machine learning or AI to substantially assist candidate screening, you’re likely covered by NYC’s Local Law 144. The law requires an independent bias audit within the past year, calculating selection rates and impact ratios by sex, race and ethnicity, and intersectional categories, a public summary of the results posted on your website, and at least 10 business days’ notice to NYC-resident candidates before the tool is used. Penalties start at $500 for a first violation and climb to $1,500 per day after that.[4]
Outside New York, the federal picture got murkier, not because the law changed but because the guidance explaining it disappeared. The EEOC published detailed technical guidance on AI and Title VII adverse impact in May 2023, then removed it from eeoc.gov in January 2025 as part of a broader rollback of prior policy pages. Title VII itself didn’t go anywhere: it still prohibits AI hiring tools from producing discriminatory outcomes, guidance page or not.[5] The clearest sign this is a live legal issue, not a theoretical one, is Mobley v. Workday, a federal case in California where a rejected applicant argues that Workday’s AI screening tool functions as an “agent” of the employers who use it and can be held liable for discriminatory outcomes. The court refused to dismiss the case in 2024 and certified a nationwide age-discrimination collective in May 2025. Workday disclosed in court filings that its software has rejected 1.1 billion applications, which gives you a sense of the scale a single vendor’s tool operates at.[5]
Since the federal guidance came down, California, Illinois, Texas, and Colorado have all passed their own AI employment laws, and they don’t agree with each other: California and Illinois use a disparate-impact standard, Texas requires proof of intent to discriminate (a much higher bar for plaintiffs to clear), and Colorado uses a “reasonable care” standard with a safe harbor for companies following NIST’s AI risk framework.[5] And enforcement of the law that’s been on the books longest isn’t exactly aggressive: a December 2025 New York State Comptroller audit found the city’s own enforcement agency received just two complaints in two years under Local Law 144, and when auditors checked 32 employer websites themselves, they found at least 17 with likely violations the city had missed.[5] The honest takeaway: the legal risk is real, but nobody is going to catch you automatically. Compliance here is mostly still on the honor system, which is exactly the kind of setup that ages badly.
Put it all together and a workflow that actually holds up looks less like “let AI handle recruiting” and more like AI handling the repetitive first pass, with a human making every judgment call that affects a real person’s shot at a job.
| Tool | Best For | What the Data Shows | Watch For |
|---|---|---|---|
| LinkedIn AI-Assisted Search | Sourcing passive candidates | +18% InMail acceptance vs. manual filters[2] | Only searches LinkedIn’s existing member data; reflects your prompt’s bias |
| LinkedIn Hiring Assistant | High-volume sourcing and first-touch outreach | 62% fewer profiles reviewed, 69% higher InMail acceptance, 4+ hours saved per role (charter customers)[1] | Paid add-on; results are from LinkedIn’s own case studies |
| ChatGPT / Claude | Drafting outreach and screening questions | General-purpose LLMs; no standardized recruiting benchmark | Never let it auto-score or rank real candidates against protected traits |
| AI resume screening tools | High-volume resume triage | Preferred white-associated names 85% of the time vs. 9% for Black-associated names in independent testing[3] | May trigger bias-audit requirements under NYC Local Law 144[4] |
Sources: LinkedIn Newsroom and LinkedIn Talent Solutions product data; University of Washington (Wilson & Caliskan, 2024); NYC Department of Consumer and Worker Protection.
Notice the pattern: every tool on that list is more trustworthy the further left it sits (sourcing and drafting) and needs more human oversight the further right it sits (evaluating and rejecting). Finding people and describing a role well is a search problem, and AI is genuinely good at search problems. Deciding who’s good enough is a judgment call, and right now the data says AI’s judgment call carries real bias risk baked in from its training data. Use it to widen the top of your funnel and speed up the parts that are pure busywork. Keep your own eyes on every single rejection.
If you only change one thing after reading this: stop letting any AI tool auto-reject a candidate with zero human review. Everything else here is optimization. That one is risk management.
AI-Assisted Search is a feature built into Recruiter that turns a plain-language description or job posting into a structured candidate search. Hiring Assistant is a separate, paid AI agent add-on that goes further: it runs the intake conversation, executes searches, and surfaces candidates with less manual input from you. Think of AI-Assisted Search as a smarter search bar, and Hiring Assistant as a smarter search bar plus an assistant running the rest of the sourcing loop.[1][2]
It’s proven, not overblown. A 2024 University of Washington study tested three commercial-grade AI models across more than 3 million resume-to-job comparisons and found they preferred white-associated names 85% of the time versus 9% for Black-associated names, and never once preferred a Black male-associated name over a white male-associated name.[3] That’s peer-reviewed, published research, not a hot take.
If the role is based in a New York City office, including remote roles tied to an NYC office, and you’re using AI or algorithmic tools to substantially assist candidate screening, yes: NYC’s Local Law 144 requires an independent bias audit within the past year, a public summary posted on your site, and 10 business days’ notice to candidates.[4] Outside NYC, check your state: California, Illinois, Texas, and Colorado have each passed their own AI employment laws with different requirements.[5]
Let AI draft first-pass outreach messages, screening question lists, job description language, and interview scorecards. Write yourself, or at minimum heavily edit, anything that decides whether a real candidate advances or gets rejected. AI is a strong first-draft tool and a risky final-decision tool.
Ask the vendor three things directly: what bias testing they’ve run, whether they’ll share the results, and whether a human can override every rejection the tool makes. If they can’t answer clearly, or the answer is ‘trust us,’ that’s your answer. LinkedIn at least publishes its own performance data and compliance white papers for its AI features, which is more transparency than most standalone AI screening vendors offer.[1][2]
This article draws on official LinkedIn product pages and press releases, a peer-reviewed 2024 University of Washington bias study, NYC’s Department of Consumer and Worker Protection FAQ on Local Law 144, a 2026 legal analysis of EEOC guidance and state AI employment laws, and a 2024 ResumeBuilder.com survey of business leaders. Every statistic here was pulled directly from its original source and checked on the date of publication.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.