Does AI make experience less valuable in 2026?

New Instacart study: AI rewards experienced workers

No — the opposite. A randomized experiment with nearly 6,000 Instacart shoppers, published in Information Systems Research in 2026, found AI guidance raised productivity 3.16% on average, but experienced workers captured the most value, especially on complex tasks. That is Naval Ravikant's 'specific knowledge' in action: AI multiplies judgment, and judgment is earned, not downloaded.

Every few months a client asks me some version of the same nervous question: if AI keeps getting better, does my fifteen years of experience still count for anything? This week an unusually clean piece of evidence landed, and the answer it gives is more interesting than the hype in either direction.

The study — "Human-Algorithm Collaboration in Gig Work: The Role of Experience, Skill Level, and Task Complexity," published in Information Systems Research and covered by Phys.org this week — ran a randomized controlled field experiment with nearly 6,000 Instacart shoppers across roughly 160 US grocery stores. Shoppers who received AI-powered navigation and product-location guidance showed a 3.16% increase in productivity, a 3.29% improvement in picking speed, and a 3.83% reduction in refund rates compared with the control group. The number that surprised me most was a 32.5% increase in cross-store shopping: the AI didn't just make workers faster, it made them more flexible about where they could take work at all.

Notice the sizes. Three percent, not ten times. That matches every credible field measurement I've seen this year, and it's the honest baseline you should carry into any AI purchase decision. But the finding that actually matters is who captured the gains. Less experienced shoppers leaned on the AI most heavily and did not consistently get better outcomes. Experienced workers pulled ahead — because they could tell when to trust a recommendation and when to override it — and their advantage widened precisely where stakes rise: complex trips and heavy workload. Study author Benjamin Knight put it bluntly: there's a growing assumption that AI levels the playing field by making expertise less important, and the findings suggest something different.

The timing matters because using AI is no longer a differentiator. The European Central Bank's blog post of August 26, 2026 reports that 52% of workers now use AI on the job, up from 26% in 2024, and the median user saves about three hours a week — 7.7% of working time. When half the workforce has the same tool, the tool is table stakes. The edge has moved entirely to what you bring to the tool.

This is the cleanest field validation I've seen of an idea Naval Ravikant has been repeating for years in The Almanack of Naval Ravikant: specific knowledge — "knowledge you can't be trained for" — is what leverage multiplies. The Instacart AI held a perfect map of the store; it could not hold the judgment about when the map was wrong, which aisle reorganization it hadn't seen, which substitution a customer would actually accept. Naval's formula says wealth comes from judgment amplified by leverage. AI is the cheapest leverage in history, and cheap leverage mathematically raises the price of the one input it can't supply. Experience isn't being devalued. It's being repriced upward.

But only a specific kind of experience — the kind that produces calibrated overrides. Years of doing a task the same way don't count; knowing when the algorithm is wrong does. Here is the practice I now run myself and assign to clients, which I call the Override Log:

Default to the AI, but log every override. Each time you reject an AI recommendation — a draft, an analysis, a plan — write one line: what it said, what you did instead.
Record the outcome, not the feeling. A week later, mark whether your override actually beat the machine's suggestion. Most people never close this loop and remember themselves as better than they were.
Name the pattern monthly. The overrides that consistently win are your specific knowledge, stated in words. Mine cluster around client emotional context and pricing; yours will differ.
Steer toward complexity. The study found experience paid most on complex, high-load work. Volunteer for exactly the work where your override rate is high — that's where AI makes you most valuable, not least.

The honest caveats: this is one study, in gig work, on structured physical tasks — grocery picking is not strategy, and effects may not transfer cleanly to knowledge work. The lead author works at Maplebear, Instacart's parent company, which doesn't invalidate a randomized design but deserves naming. And small average gains mean AI still isn't a substitute for hiring well. What the study kills is the comfortable idea that you can skip the years of judgment-building because the machine will cover for you. On the evidence so far, it covers best for the people who need it least.

Sources: Phys.org, "AI helps workers most when paired with experience, new study finds" (August 2026); Knight et al., "Human-Algorithm Collaboration in Gig Work," Information Systems Research, DOI 10.1287/isre.2024.1664; European Central Bank blog, "AI adoption and the productivity promise: what workers report" (August 26, 2026); Eric Jorgenson, The Almanack of Naval Ravikant.