The Signal: Human Advancement — Edition #6 — September 24, 2026

The Excelsior Group · The de-siloed AI × (education · work · talent acquisition · L&D · HR) letter · Covering September 17 – 23, 2026

The Read

For five editions the entry-rung tide has rested on one payroll dataset and a stack of surveys. This week it got administrative data from three directions at once. The Census Bureau linked graduates of 142 majors to their employment and earnings records and found that graduates in the most AI-exposed tenth of majors — computer science, accounting, journalism, engineering — saw their odds of initial employment fall 5 points and their first-quarter earnings fall 13% after ChatGPT, a loss the authors call comparable to graduating into a large recession. Stanford's Digital Economy Lab looked at 1.25 billion job postings and 154 million employment records across 41 countries and found that firms adopting AI grew total headcount about 3.3%, with senior employment up 6.7% and junior employment 2.5% lower; the junior share fell in 23 of 31 countries. Denmark's central bank, using firm-level register data, found the same shape one week earlier: adopters hire fewer people, and the shortfall lands on the young and the highly educated. Indeed then priced it: advertised pay in AI-exposed occupations is up 46% since 2021 against 25% elsewhere, but the premium is 17 points for senior roles and about 2 points at entry level. The market answered in the same seven days with products that treat learning, credentialing and hiring as one loop — Handshake's AI Skills Studio with OpenAI and Google as content partners, LinkedIn turning the profile into a vouched work record, OpenAI Academy badging 4 million learners, and a16z raising $42 million to build a school that skips college. Old problems, new physics: the ladder is not disappearing, it is being pulled up — and the businesses being built this week are ropes.

Tide status

T1 — the entry rung is breaking: evidence upgraded, thesis held. No widespread displacement — every dataset this week says so — but a consistent junior-share decline at adopting firms, and a graduation-timed earnings penalty for exposed majors that is now visible in tax records rather than surveys. Three independent sources (US Census, Stanford/Revelio across 41 countries, Danmarks Nationalbank) converge on one mechanism: reduced hiring of juniors, not layoffs. Stanford's Canaries dashboard refreshed on September 23 to July data with no new research note; the last published headline gap stays at 19%, and the dashboard now shows early-career women in the most-exposed quintile contracting 5.4% over the past year against 3.1% for men. Pew adds the sentiment layer: in 34 of 37 countries surveyed, majorities expect AI to cut jobs, and roughly seven in ten Americans do, up 7 points in two years. Tide holds; the evidence tier rose from payroll-and-survey to administrative. T2 (instruction cost → ~zero) logged the sharpest null result yet on the AI layer of a tutor — see W3 — and a third governance model for diffusion: Florida chose consent-and-guardrails over New York's and Los Angeles' bans (rules passed September 16, effective 2027-28). T3 (the credential's monopoly is ending) collected two data points without moving: a $42 million venture-funded alternative to college with ten frontier-company founding partners, and skill badges issued by OpenAI and by Handshake that are designed to be read by employers, not registrars. Watching; not calling.

Waves

W1 — Agentic recruiting absorbs the funnel: the trust layer is being built, and ChatGPT is now a sourcing channel [deployed]

LinkedIn shipped the closest thing yet to a verified work record: members can now vouch that they actually worked or studied alongside someone in a given period (the voucher must be verified, connected for a year or more, and on two-factor), companies can remove accounts that falsely claim employment from their pages and search, and identity verification expands with Truecaller and PeerSpot. LinkedIn says 115 million members and 700,000 companies are verified and that verified profiles get 50% more post views — its own numbers, so directional. Phenom's Fraud Detection Agent, announced as an awards finalist a week after its September 17 AI Day, reports "87% confidence in duplicate identity detection" and that recruiters validated 76% of its AI-generated-answer flags — vendor-reported, no named customers, and despite a dedicated session on adverse-impact monitoring no adverse-impact numbers were published. On Chad & Cheese, Velocity Network's Etan Bernstein made the case for verifiable credentials on an open protocol instead of fraud detection, citing 11,000 AI-generated applications per minute. The product side kept moving: Jack & Jill's CEO told Recruiting Future the platform now has close to half a million professionals and about 5,000 companies, with its candidate agent screening roughly 14 million listings a day; Ribbon's COO reported that 85% of Extra Space Storage's hiring flows through its AI interviews and claims 10–30% better 180-day retention by role (vendor numbers, unaudited); Paychex put agentic recruiting in front of roughly 840,000 SMB customers through WISE Hire and listed ChatGPT alongside Google for Jobs and Meta as a sourcing channel; Findem re-platformed as an MCP server for Claude, ChatGPT, Copilot and Gemini with FedEx, Intuit and Emirates named; Joveo unveiled a full-funnel agent at RecFest; and Bullhorn's 2,300-candidate survey found 92% rating AI voice interviews as good as or better than human ones (TA) (HR).

Roadmap implication: Agent-to-agent hiring only works on verified inputs, and this week the verification layer arrived from the incumbent with the most profiles rather than from a startup. The open question is whether it stays LinkedIn's walled garden or becomes a portable credential. For buyers: any AI screener that cannot show its adverse-impact numbers is asking you to hold the risk it declined to measure. And note the quiet one — OpenAI still has no Jobs Platform, but ChatGPT is now on a payroll incumbent's sourcing list.

Links: LinkedIn adds new tools to fight fake profiles and bogus work histories · Phenom Fraud Detection Agent Named a Finalist in The 2026 A.I. Awards · Candidate Fraud Stops Here with Etan Bernstein · Ep 827: Agent To Agent Recruiting · Ep 826: Do AI Interviews Work? · Paychex Introduces Agentic AI Recruiting Capabilities Through WISE Hire · Findem Launches Studio: People Intelligence, Built for AI, That Returns Finished Work Teams Can Trust · Bullhorn's 2026 'GRID Talent Trends Report' finds 92 percent of candidates rate AI voice interviews as good as or better than human interviews

W5 — The engineered apprenticeship: the skills-erosion bill arrives from the CHRO's office [research]

IBM's 2026 CHRO Study (1,500 CHROs and senior executives, 8,800 employees) puts the other end of the entry-rung problem on the record: 60% of employees say skill erosion directly affects them, 46% of executives call it a top concern, 80% of CHROs say AI creates "invisible work" (validating recommendations, fixing errors), and only 26% of organizations clearly define which work is human-led, AI-assisted or AI-executed. The gap that matters: 71% of executives prioritize the ability to supervise, validate or override AI output — against 38% of employees. Culture Amp's benchmark across 123 organizations and ~112,000 employees explains why the productivity never shows up as capacity: 85% of employees are encouraged to use AI, 42% say no one has explained why, 71% feel more productive (93% among power users), yet power users and non-users report almost identical workload reasonableness (72% vs 69%), and awareness of internal career opportunities fell 10 points in a year. Philippa Hardman's eight-test audit of Articulate's new Frontline course generator found it built a course for a non-training problem, produced the same design for recall, procedure and judgment objectives, paired apply-level goals with recognition-only quizzes, and silently dropped 6 of 14 mandatory policy rules — authoring automated, design not. Ethan Mollick's "The Overhang" names the assets that compound with AI rather than depreciate: deep knowledge, wide knowledge, taste and agency. Exponential View's earnings-call count is the board-level read: 33% of S&P 500 companies made a quantified AI statement last quarter, 15% quantified a business impact (up from 9%), and pilot language stayed under 4% — companies talk deployment, not pilots, which hides the failures (L&D) (HR) (work).

Roadmap implication: Last week's design rule was judgment first, tool second, measure the workflow. This week adds the accounting line: if seniors are the only ones who convert AI into durable capability (Autor) and adopters hire fewer juniors (Stanford, Denmark) while employees report their own skills eroding (IBM), the enterprise is running down its expertise balance sheet without a depreciation schedule. L&D's job is now to write one — and "AI access" programs that ship without workflow redesign are, on Culture Amp's numbers, buying sentiment.

Links: 2026 CHRO Study: Designing the Thinking Organization · IBM report warns AI could erode human skills deemed vital by CHROs · Culture Amp's 2026 AI at Work Benchmark: 85% Are Encouraged to Use AI. 42% Don't Know Why · Frontline can generate your course, but does it really have "instructional design expertise"? · The Overhang

W3 — The evidence bar in learning: the flagship AI tutor's RCT says access is not use [outcome-validated]

Jill Barshay's Proof Points column surfaced the result Khan Academy has been sitting with since August: Oreopoulos and Low's two-year cluster RCT across 18 Tennessee middle schools, run inside daily remedial math, found Khan Academy raised achievement by 1.3 national percentile ranks per term (0.06–0.08 SD a year; 0.14 SD for a full year of active participation) — and Khanmigo, configured to coach rather than answer, added nothing beyond the platform. The mechanism is in the usage log: 96% of students tried the tutor at least once, but the median student messaged it on only a third of practice days and in only 17% of the sessions in which they made a mistake, mostly bare answers or clicks on suggested prompts. "The binding constraint appears to be engagement." Sal Khan praised the study and said Khanmigo now triggers automatically after a wrong answer. Two procurement-side responses landed the same week. Learning Commons, the CZI spin-out, launched an open platform — a Knowledge Graph of state standards and learning progressions, Evaluators that check AI-generated content for grade level and standards alignment, and Agent Skills — with Canva Education, MagicSchool, Level, OKO Labs, Really Great Reading and TalkingPoints integrating and OpenAI and Anthropic named as collaborators. And EdTech Insiders published Shawn Rubin's synthesis of 100 district-leader conversations: the districts getting AI right built change management, capacity and ROI discipline before AI arrived (Denver scaled from 225 to 4,000 teachers over three years; KIPP Northern California used Claude Code to build an on-device analysis app and found its AI instructional tool helped only in small-group settings for students well below grade level). Phil Hill's catalogue of the ban wave — NYC K-8, LAUSD, UChicago's analog courses, Berkeley Law's default prohibition, Rutgers pausing Canvas IgniteAI — asks the right question: what function, exactly, is being banned? The Crimson's faculty survey (460+ Harvard professors) has 67% saying AI has hurt their courses, up from 42% a year ago (edtech) (policy).

Roadmap implication: The cost of instruction went to zero and the tutor got ignored. The scarce good in AI learning is engagement design that survives contact with a thirteen-year-old — and the evidence bar just moved from "did it help" to "did they use it, and did it help beyond the platform without AI." Vendors should expect districts, armed with Learning Commons evaluators and their own on-device analysis, to run that test themselves.

Links: Students didn't get answers from Khanmigo. They didn't want its questions, either · One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment · Learning Commons announces platform to scale proven teaching and learning practices · The Districts Positioned to Get AI Right Aren't Thinking About AI First · When Schools Ban AI, What Exactly Are They Banning? · 2 in 3 Harvard Professors Say AI Has a Negative Impact on Class

W2 — Private capital reprices incumbents: a university buys a talent marketplace, and the assessment layer consolidates [deal]

The de-siloing thesis showed up in an M&A filing: Phoenix Education Partners, parent of the University of Phoenix, agreed to acquire Fuel50, the AI-powered internal talent marketplace and skills platform — an education incumbent buying the system that decides who inside an enterprise gets developed and moved (terms undisclosed; close expected within 30 days). Harver, the assessment vendor, bought Symphony Talent, the recruitment-marketing and CRM firm, pitching a "global talent intelligence solution" with 1,500+ combined clients and 65 million candidates assessed — assessment moving left of the application to cope with AI-generated resume volume. McGraw Hill took TeachFX, the AI classroom-audio coaching tool, into its professional-learning line; Pearson bought Internet Testing Systems; Stepful, the healthcare workforce-training company, bought a 45-year-old accredited nursing school to extend short credentials into degrees; India's Unstop took PerspectAI, a game-based assessment startup, in an all-equity deal. On the capital side, Brighteye closed a $72 million first close of Fund III on September 16 with an explicit "HumanOS" thesis — technology that helps people learn, work and adapt — with Lumina Foundation among the new LPs (edtech) (TA) (L&D).

Roadmap implication: Four of this week's six deals bought assessment, verification or coaching instruments — the parts of the stack AI made scarce — and the one strategic buying a talent marketplace is a university, not an HCM vendor. The buyers are moving across the silo lines faster than the analysts covering them.

Links: Phoenix Education Partners to Acquire Fuel50, Advancing the Future of Workforce Transformation · Harver Acquires Symphony Talent, Creating a Global Talent Intelligence Solution for the AI Era · Brighteye Ventures lands $72M first close to back the future of learning and work · Unstop acquires PerspectAI in all-equity deal to deepen AI-led talent assessment

Ripples

Census Bureau: graduating into disruption — the AI penalty is now in the tax records [research]

Orr, Tucker and Warren (Center for Economic Studies working paper CES-26-56, September 2026) link administrative records for graduates of 142 majors, rank majors by the pre-ChatGPT AI exposure of the occupations they feed, and compare outcomes before and after November 2022. Graduates in the most-exposed decile saw their probability of initial employment fall five percentage points and full-quarter initial earnings fall 13%, with flat pre-trends and the divergence beginning immediately after ChatGPT; half the earnings loss is lower pay within the same industries, half is a shift into retail and food service. The effect attenuates with experience but does not vanish. Phil Hill's caveat stands: the 2022 tech-hiring correction is in the picture too, but the major-level exposure gradient makes it "increasingly likely that AI added to the deterioration." Most exposed majors: computer science, accounting, journalism, engineering; least: nursing, education (work) (edtech) (TA).

So what: This is the first estimate of the entry-level penalty that a college could act on by major. The "what should I study" and "who gets hired" questions now share a dataset — and the answer it gives is uncomfortable for the STEM-first orthodoxy that every state workforce plan is built on.

Links: Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors · A Clearer Picture of AI and the College Job Market

Stanford DEL goes global: AI adopters grow, seniors gain, juniors lose — in 23 of 31 countries [research]

Chandar and Klein Teeselink use Revelio Labs data — 1.25 billion job postings and 154 million employment records across 41 countries, January 2021 to March 2026 — and compare foreign affiliates of multinationals that adopt generative AI (detected from their own postings) against matched non-adopters. Adopting affiliates show marginally significant total employment growth of 3.3%; senior employment rises 6.7% while junior employment is 2.5% lower, and the junior share falls in 23 of the 31 countries with precise estimates, significantly in seven. In computer and mathematical occupations the employment share rises while the junior share falls 3.3 points. Denmark's central bank reached the same shape from register data on September 16: firms that start using AI hire fewer new employees, the shortfall is concentrated among younger and higher-educated workers, and wages are untouched; Johannes Sundlo's read of the underlying working paper puts adopters about 11% below their prior employment trajectory by September 2025 (roughly 15% at firms under 100 staff), with the adjustment coming through reduced hiring rather than layoffs (work) (TA).

So what: "AI is labor-saving for juniors and labor-expanding for seniors" is now a cross-country finding, not an American one. The policy implication is not a jobs program; it is that firms are externalizing the cost of training the next senior, and someone — the state, the university, or a Handshake — is about to be asked to pay it.

Links: How Does AI Change Labor Demand? Evidence from 41 Countries · Firms hire fewer people when they start using AI · The jobs we stop hiring for

Handshake launches AI Skills Studio — learn, badge, get hired, inside one product [deployed]

Handshake's free learning hub runs project-based "missions" of 10–45 minutes inside real AI tools with an AI instructor; completed projects earn a verified skill badge that gives employers evidence of applied AI ability, sitting on the same profile 1 million employers already search. Content partners are the tool makers themselves — OpenAI, Google, Salesforce, Figma, Notion, Replit, Lovable, Vercel, Gamma, Clay — plus SNHU. Handshake's own data: 58% of graduating seniors say they need a stronger understanding of AI to succeed at work; only 28% say AI has been meaningfully integrated into their education. Same week, OpenAI Academy added role-based learning paths with course badges earned by assessment and reported 4 million people engaged, 250+ events and 1,600+ educators and district leaders trained across eight US cities — while the Jobs Platform announced in September 2025 remains unlaunched for a sixth consecutive edition (edtech) (TA) (L&D).

So what: This is the de-siloing thesis as a product: the campus job network now owns a learning layer whose output is a hiring signal, and the frontier labs are supplying the curriculum. No outcomes data yet — the test is whether an employer ever changes a hiring decision because of the badge. Watch for the first Handshake employer that filters on it.

Links: Handshake launches first AI Skills Studio to prepare job seekers for the AI economy · Two years of OpenAI Academy · Expanding OpenAI Academy with new learning paths

a16z raises $42M to build a school that skips college — with Anthropic, OpenAI, Google and Palantir as founding partners [demo]

The Horowitz Andreessen Academy, led by Udemy and Maven co-founder Gagan Biyani, will open in San Francisco in fall 2027 as a tuition-free one-year Founding Class Fellowship for high-school graduates, built on "Pursuits" (self-chosen projects), short courses taught by founders, and co-ops at ten founding partners — Anduril, Anthropic, Coinbase, Google, Meta, NVIDIA, OpenAI, Palantir, Replit and Stripe — plus 40 hiring partners; fellows get $50,000+ in compute credits and a $5,000 travel budget. A two-year program priced "similar to an elite private university" is planned for 2028, subject to regulatory approval. Backers include Adam D'Angelo, Tobi Lütke, Tony Xu, Fidji Simo, Garry Tan and Alpha School's Joe Liemandt. Michael Horn and Thomas Arnett's Alpha School essay this week supplies the theory: judge these models as startups serving a narrow job-to-be-done, not as schools that must serve everyone (edtech) (work).

So what: The venture industry is no longer content to hire around the degree; it is building the alternative and staffing it with the companies that would otherwise recruit from Stanford. It is a demo for one cohort of exceptional kids — but the signal it sends about what the frontier labs think a credential is worth reaches every admissions office.

Links: Introducing The Horowitz Andreessen Academy · If you don't like Alpha, it's probably not for you. That should be OK

LinkedIn turns the profile into a vouched work record [deployed]

Beyond the fraud framing in W1, the structural move: for the first time a third party can attest, on the record, that a claimed job actually happened, and an employer can strike false claims from its own page. Combined with 115 million verified members, this is the largest verified-employment dataset in existence, and it sits inside the platform that also sells the job ads and the recruiter seats. Velocity Network's open-protocol alternative and the Clear/Truecaller identity partnerships are the competing architectures; National Student Clearinghouse and Equifax's Work Number (roughly $100 a verification, per Bernstein) are the incumbents being disintermediated (TA) (HR).

So what: Verified history is the input every AI screener, AI interviewer and agent-to-agent matcher needs and none of them own. LinkedIn just took the position. Expect the ATS vendors to price API access to it within a year, or to fund the open alternative.

Links: LinkedIn adds new tools to fight fake profiles and bogus work histories · Candidate Fraud Stops Here with Etan Bernstein

Indeed: the AI pay premium is real — and it is paid to seniors [research]

Indeed Hiring Lab's Jack Kennedy finds advertised pay in the most AI-exposed US occupations up about 46% since 2021 versus 25% in the least exposed, with a post-ChatGPT premium of 5.7% controlling for occupation, 4.7% within identical titles and 2.4% after adjusting for the seniority mix. By level, the gap is roughly 17 points for senior roles, 12 at mid-level and about 2 at entry. Workers agree on who wins: Pew's 42,151-respondent survey across 36 countries plus the US finds majorities in 34 of 37 expecting AI to cut jobs, with the young more worried than the old in most of them (work) (HR).

So what: Quantity (Stanford, Census) and price (Indeed) now tell the same story: the market is bidding up experienced judgment and not bidding at all for junior capacity. Compensation data is the fastest-moving confirmation of the tide, and it is pointing the same way.

Links: AI Exposure Isn't Squeezing Advertised Pay in the US — It's Boosting It · Globally, More People Expect AI to Cause Job Loss Than Growth

Watchlist — what moved

Scoreboard — 19%. The AI employment gap for workers 22–25 in AI-exposed occupations (Stanford DEL, ADP payroll data). The dashboard refreshed September 23 to July 2026 data without a new research note; the last published headline number stands. Next scheduled update October 21.

The Tape — pre-seed to Series B, September 17 – 23

Domestic (US)

No qualifying US pre-seed-to-Series-B round in the human-capital stack this week. (The Horowitz Andreessen Academy's $42M is a school, not an AI-native company — see Ripples.)

International

Gap catch-up — announced September 15–16, after Edition #5 closed

Tally: 3 rounds in window, ~$22.7M disclosed — 0 domestic / 3 international; a quiet week at seed-to-B while the action was in M&A (Fuel50, Symphony Talent, TeachFX, ITS, PerspectAI). Gap catch-up adds 2 rounds ($27M, 1 domestic / 1 international) and one $72M fund close. Rejected as stale republishes or out of scope: Humanly $25M B (April, still on HRTechFeed's front page), Intermezzo $10M (Sep 12, payroll), YoLearn.ai (Sep 3), Ema $77M B (generic agentic enterprise), Kastle $24M A (banking ops), Sol, Spiich, mika, ReFresh (thin AI). Post-window hold for next week: 50skills $6M (Iceland, Sep 24). Blind spots this week: FinSMEs and Tech in Asia were unreachable, so sub-$2M US rounds and Southeast Asia are under-covered.

The Signal: Human Advancement is published weekly by The Excelsior Group. Archive and subscribe: https://excelsiorgroup.ai/insights/signal/human/