When AI Is Better Than You — and When It Isn't

For the past two years, the standard advice has been "use AI more." More tasks, more tools, more workflows. The problem with that advice is it skips the question that determines whether AI actually helps you: is this specific task one where AI will outperform you, match you, or fall short? Ethan Mollick's new framing — "Co-Existence" — is built around that question. It changes what you do with every task, not just your AI strategy at large.

What changed

On June 4, Mollick announced his new book "Co-Existence" (out October 2026) with a post arguing the era of human-directed AI is over. The earlier framing — "you're the manager, AI is the tool" — made sense when AI was reliably worse than you at domain work. It's less accurate now. AI exceeds human performance at specific knowledge work tasks in specific domains. The mental model that says "I direct AI" doesn't hold when AI direction could reasonably go the other way.

The frame Mollick proposes: Co-Existence means working alongside something that is sometimes better than you, sometimes worse, and developing the judgment to know which is which. His earlier research ("Choosing to Stay Human," May 26) is what makes this concrete. He found that students who used AI as an answer machine scored better on AI-assisted work and worse on independent tests taken immediately after. The capability gap opened fast and quietly.

What to do differently Monday morning

Three things change if you take this frame seriously:

Map your tasks by AI performance, not by habit. For each task you regularly use AI for, ask whether the AI output is better than what you would have produced yourself. Not faster — better. If you haven't compared them recently, you might be over-supervising work where AI already outperforms you, or under-supervising work where it reliably misses.

Set your supervision level deliberately. AI outperforming you doesn't mean zero oversight — it means lighter oversight calibrated to the actual risk. AI falling short doesn't mean never use it — it means you stay in the loop at the right point.

Test without AI periodically. If you've been delegating a task to AI for three months, do it yourself once. The gap between what you produce and what AI produces tells you something about whether your own skills are still intact.

Nadia: two tasks, two supervision levels

Nadia is a content strategist at a 35-person B2B software company. She writes positioning copy — brand voice, landing pages, campaign briefs. After six months of working alongside AI, she discovered something specific: AI generates more headline variants in ten minutes than she would write in an hour, and roughly one in six of those variants is better than anything she would have written herself. On first-draft headlines, AI is better than she is.

On deciding which variant ships, she's better. She knows what the sales team has already heard from customers, which framings have landed in previous campaigns, and which words make their specific audience stop reading. The AI doesn't. She now writes the brief, takes the AI's headline variants, and makes the final selection without consulting AI again. Different task, different supervision level — and the split is based on actual evidence, not instinct.

Tom: when knowing AI is better isn't enough

Tom runs operations at a 120-person logistics firm. He started using AI to review vendor contracts — flagging unusual terms, missing clauses, liability exposure. The AI was better at this than his initial read. He reviewed its output, made edits where he disagreed, and sent contracts with fewer problems than before. That part worked.

Eight months in, he ran a test. He read a vendor contract without AI first, noted what he'd flagged, then checked what AI flagged. AI caught four things he'd missed. Two of those things would have mattered. He realised he'd stopped reading contracts carefully because AI was catching the issues — and he hadn't noticed the degradation while it was happening.

Knowing AI is better than you at a task is useful information. Not knowing when your own capability has eroded as a result is a separate problem, and it's harder to catch because there's no obvious failure moment until there is.

The question isn't "how much AI should I use?" — it's "on this specific task, right now, is AI better than me?" Answer that honestly and your supervision level is no longer guesswork.