Human-AI Teams Perform Worse Than Either Alone — And Your Martech Stack Is Proof

You bought the tools. Your team has access. Usage dashboards look healthy. And yet the campaigns aren’t sharper, the content isn’t better, and nobody can point to a specific outcome that justifies the spend. If that description fits, you’re not alone. You’re the norm.

The largest meta-analysis of human-AI collaboration ever conducted found something that should restructure every conversation you’re having about AI ROI: combining humans and AI produces worse outcomes than either working alone.

Not sometimes. On average. Across 106 studies. 370 effect sizes analyzed in the meta-analysis.

Vaccaro, Almaatouq, and Malone’s 2024 meta-analysis covered experiments from January 2020 through June 2023, and the headline finding was unambiguous. Human-AI (HAI) combinations performed significantly worse than the best of humans or AI alone, with a 95% confidence interval that didn’t cross zero. This isn’t a cherry-picked study from a lab with twelve undergraduates. It’s the field’s most rigorous assessment of a question the entire technology industry assumed it had already answered.

The Reaction Is Always the Same

When I present this finding to marketing leaders, the first response is disbelief. Then a pause. Then something closer to recognition: Actually… that explains a lot.

That moment matters more than the number itself. Every CMO who’s watched a team produce blander work with AI than without it, every ops leader who’s seen throughput metrics go sideways despite tool adoption, every content strategist who quietly rewrites everything the AI touches. They already know the finding is true. They just didn’t have the data to name it.

Usage is up but impact is not. Organizations report widespread AI usage but disappointing returns.

BCG AI Adoption Puzzle, 2025

BCG’s 2025 research confirmed the pattern at enterprise scale. The problem isn’t refusal. Eighty-five percent of leaders and 78% of managers use generative AI (GenAI). The problem is that usage and impact have decoupled. Your teams are logging into the tools. They’re generating outputs. They’re not generating value.

The Data Gets Worse Before It Gets Better

If the Vaccaro finding were limited to lab experiments, you could dismiss it. It isn’t.

Faros AI analyzed telemetry from over 10,000 developers across 1,255 teams. AI adoption correlated with a 9% increase in bugs per developer and a 154% increase in pull request size. Developers completed more tasks individually, but at the company level? No significant improvement in throughput, quality, or delivery metrics. The bottleneck shifted from writing code to reviewing code. Individual speed went up. Organizational output didn’t move.

Then there’s the perception gap. The METR study found that experienced developers using AI tools worked 19% slower on real projects in their own codebases, while estimating they were 20% faster. A 39-percentage-point gap between perception and reality. These weren’t beginners fumbling with new software. They averaged five years of experience in the repositories they were working in.

Two Stories in the Same Data

Here’s where most analyses of the Vaccaro findings stop too early.

The meta-analysis found enormous heterogeneity. Some human-AI teams dramatically outperformed. The average was negative, but the variance was massive. Two boundary conditions mattered most: task type and relative expertise. Creative and content tasks showed gains. Decision-making tasks showed losses. When humans brought genuine expertise the AI lacked, the combination created value. When the AI already outperformed the human, adding the human made things worse.

That heterogeneity is the whole game. If the outcome were uniformly bad, there’d be nothing to work with. But some teams are getting extraordinary results from the same tools that produce mediocre outcomes everywhere else. Same AI. Same access. Different results. The question isn’t whether AI works. It’s what distinguishes the teams that succeed from the ones that don’t.

The Variable Nobody Is Measuring

The Vaccaro team identified a critical mediator: whether humans could accurately assess when AI was adding value and when it wasn’t. The teams that succeeded had humans who knew when to trust, when to override, and when to bring their own expertise. The teams that failed had humans who either deferred to everything the AI produced or fought it reflexively.

This points to something no amount of tool procurement or prompt engineering training addresses. The variable isn’t the AI’s capability. It isn’t access. It isn’t even skill. It’s the mental model the human brings to the collaboration: the internal framework that determines whether they show up as a genuine contributor or a passive bystander. The H(x)AI Framework

H(x)AI

Human capability multiplied by AI capability, where x is the mental model multiplier. The relationship is multiplicative, not additive. When x approaches zero, it doesn’t matter how capable either side is.

I call this H(x)AI: Human times AI, where x is the mental model that determines how much of the human’s actual capability enters the collaboration. When x is high, both sides amplify each other. When x is low, the human doesn’t just fail to add; they actively degrade the AI’s output through poor calibration, uncritical deference, or reflexive resistance.

This isn’t motivational framing. It’s structural. The Wharton study (Bastani & Chung, PNAS 2025) tested the mechanism directly with nearly 1,000 students. Unrestricted AI access produced a 48% performance gain during use but a 17% decline when AI was removed: clear deskilling. Scaffolded AI access, designed to maintain cognitive engagement, produced a 127% gain with zero deskilling. Same AI. Same students. The only variable was how the human engaged.

What This Means for Your Stack

If you’re a marketing leader, the implications are uncomfortable but actionable. The gap between what your team gets from AI and what’s possible isn’t a training problem. Training correlates with usage, not impact. It isn’t a tool problem, because the tools are more capable than your team’s current use of them. And it isn’t a willingness problem, because your people are already using AI every day.

It’s a relationship problem. The psychological space between the human and the machine, where judgment happens, where expertise meets capability, where the value actually gets created or destroyed. That’s the layer nobody is working on.

The research points to four conditions that move the multiplier toward 1.0. Humans see themselves as the scarce input, not the AI. They maintain genuine cognitive engagement rather than coasting. Their professional identity feels clarified rather than threatened by the collaboration. And they treat calibration (knowing when to trust, when to push back) as an ongoing practice, not a one-time training event.

You can’t install those conditions with a workshop. But you can build them into how your team collaborates with AI every day: the review processes, the quality standards, the expectation that AI output is a starting point for judgment rather than a replacement for it.

Your AI investment thesis assumed the relationship was additive: combine humans and AI, get more. The data says it’s multiplicative. And when the multiplier is low, you’re paying for tools that make your team worse.

The question isn’t whether to keep investing in AI. It’s whether you’re ready to invest in the variable that actually determines the return.