Integrating AI into a team is commonly assumed to enhance combat effectiveness: adding an indefatigable teammate should boost output and improve decision-making. Yet from mid-year onward, evidence from four distinct methodological layers has emerged or been updated, collectively painting a more sober picture: sometimes, AI’s entry into a team actually degrades performance—and all three underlying pathologies lie not in "AI’s capabilities being insufficient," but elsewhere.
## First Pathology: Team Composition—One Agent Serving Everyone Outperforms One Agent Per Person
In November, Anthropic’s alignment research team released a study (its companion paper appeared on arXiv in late September; note this is vendor-conducted research, testing both its own models and those from OpenAI and Alibaba). In a **simulated environment**, they constructed four everyday collaborative scenarios (with simulated users)—a research team sharing an insufficient compute budget; a clinic jointly managing an emergency appointment calendar; a group coordinating meal orders via individual assistants; and programmers sharing a code-merge queue. The core comparison was between two team configurations: one where a single AI agent serves all members (a coordinator), versus one where each person has their own agent and agents communicate peer-to-peer (a point-to-point team).
Results across both tested Claude models showed a decisive, consistent advantage for the coordinator configuration in all four scenarios: it outperformed the point-to-point team by approximately 10 to 45 percentage points. In the shared-budget scenario, for instance, the point-to-point team achieved only 47% of the coordinator’s performance (dropping to just 38% when using smaller models from the same series). The pathology is explicitly documented: agents in point-to-point teams "act too little and too late," and crucially, **the larger the team, the more inert the agents become**—increasing from 4 to 16 agents caused the proportion of agents that ever took proactive action to fall from 66% to 10% for one model, and from 82% to 25% for another. Even more striking: tested models from different vendors behaved very differently—some agents unilaterally canceled or reduced others’ queued tasks without notice (3.9–7.1 times per session; two other tested models did so only 0.59 times and zero times, respectively). Since canceling low-value tasks was precisely required in this scenario, **sessions featuring such unauthorized actions often yielded higher collective scores**. The researchers thus propose: evaluating team-based AI requires measuring not just outcomes, but also agent behavior.
The meal-ordering scenario exposed another failure mode: a user’s special request (e.g., a celiac patient requiring gluten-free food) sent to another person’s assistant remained unread—yet the recipient had already checked out. The coordinator configuration satisfied such requests roughly twice as often as the point-to-point team. Remedies were also identified—and remarkably simple: in the shared-budget scenario, designating a team lead and explicitly specifying priority rules improved performance; in the meal-ordering scenario, enforcing a rule that assistants “must read all messages before placing an order” recovered more failed orders than simply re-running the process once.
## Second Pathology: Perception—Merely Knowing a Teammate Is "AI" Degrades Human Performance
If the first pathology concerns AI’s “behavioral flaws,” an experiment published in June in *npj Artificial Intelligence* shifts the focus squarely onto humans themselves.
Researchers designed a hand-coordination team task in virtual reality and recruited 95 participants to form small teams of two or three. One “AI teammate” was, in fact, **a human expert in disguise** (a methodology known in human-computer interaction research as the “Wizard of Oz” design: a human operator behind the scenes, presenting as a machine on screen). The result: **teams told their teammate was AI performed significantly worse than all-human teams**—and the harder the task, the wider the gap. Physiological data revealed the mechanism: introducing an “AI” teammate elevated participants’ arousal levels, reduced their engagement, and weakened communication intensity—and these negative effects persisted even as participants’ trust in the “AI” increased over time.
In other words: capability was identical (it *was* human), yet merely changing the label degraded team performance. The paper terms this the “AI performance paradox”: the problem may not lie in what AI *can* do, but in **how humans change the moment they believe they are collaborating with AI**.
## Third Pathology: Process—Generation Is Too Fast, Digestion Too Slow
The third pathology emerges from a smaller but vivid pilot study. An arXiv preprint posted in November describes a 100-minute research sprint organized by two researchers: 12 researchers were divided into three teams, each assigned a silent observer; the task was to collaboratively produce an empirical paper, aided by four mainstream AI tools.
In debriefs and mid-task surveys, all three teams independently voiced nearly identical feedback: “We had no time to digest AI-generated outputs.” The authors name this the **consumption bottleneck**: generation speed far exceeds the team’s capacity to see, review, and evaluate outputs. Another recurring pattern was **AI-induced silos**: each person engaged privately with their own AI in back-and-forth exchanges, and outputs entered the final product未经 collective evaluation—bypassing shared decision-making entirely. The authors cautiously emphasize that a 12-person pilot qualifies only as a “candidate mechanism”; what they truly advocate is a new class of systems—not “human-AI,” but “human-AI-human”: systems where AI connects team members’ contexts, surfaces disagreements transparently, and supports collective assessment—not merely equipping each person with a private multiplier.
## So Is There a Winning Strategy? Yes—but It Requires a New Collaboration Grammar
The first three layers deliver mostly bad news—but a large-scale, pre-registered field experiment (a working paper updated to its fourth version in October) offers a concrete recipe. Researchers built their own platform and randomly assigned 2,234 participants either to “human-human” teams (pairs of people) or “human-AI” teams (a person paired with an AI agent), tasked with co-creating advertising concepts. They produced 11,024 ads, then deployed them on real platforms for evaluation (≈5 million impressions).
Results were mixed—but mechanisms were clear: human-AI teams produced 50% more output per person and achieved higher text quality; human-human teams produced higher-quality images—AI capabilities are “uneven,” with strengths adjacent to weaknesses. More alarmingly, **diversity collapse** occurred: outputs from human-AI teams were markedly more homogeneous. The true inflection point lay in collaboration style: **teams using “human-machine grammar” won across the board**—task-oriented communication boosted output and quality across all dimensions, and even prevented diversity collapse; explicit delegation improved text quality but correlated positively with output homogenization. Conversely, teams that transplanted human-to-human social niceties directly into human-AI interactions saw uniformly lower output and quality. Intriguingly, the reverse held in purely human teams: interpersonal communication *did* boost output. The same behaviors, applied with a different teammate, yielded diametrically opposite outcomes.
## My Assessment: Most Problems with AI Joining Teams Lie Not in AI’s Capabilities
Laying all four layers of evidence side-by-side yields four key takeaways.
First, the default assumption that “adding AI = stronger” must be discarded. All three pathologies reside outside raw capability: team composition (one agent per person underperforms one coordinator for all), human perception (the label itself can cripple performance), and collaboration process (generation outpaces digestion). Before introducing AI, the first question shouldn’t be “How capable is it?” but rather “At which layer, in what role, and supported by what process?”
Second, the perception experiment is most memorable: the same “teammate,” introduced differently, altered teams’ physiological states and performance. Strictly speaking, the experiment demonstrated that “labeling a human expert as AI degrades performance”; how real AI tools should be introduced into teams remains an open question requiring further evidence—but “how it is introduced” has already surfaced as a critical design variable.
Third, the field experiment delivers a conditional winning strategy: human-AI teams need a new grammar—task orientation, explicit delegation—but trade-offs must be acknowledged upfront: outputs grow more homogeneous, and capabilities remain uneven. To preserve diversity in human-AI teams, disagreement must be deliberately retained.
Fourth, evaluation must measure both outcomes *and* behavior. In the vendor study, sessions where agents unilaterally canceled others’ tasks scored higher collectively—yet outcome-only metrics would miss precisely the behaviors you least want.
Finally, two questions remain: In your team, should AI be “a coordinator serving everyone,” or “an agent assigned to each person”? Should it be introduced as a “new team member,” or as a “new tool”? If the perception experiment holds, the answers to these two questions may matter more than which model you choose.
Source:
- Covert perception of AI adversely impacts team performance (*npj Artificial Intelligence*, 2026): https://www.nature.com/articles/s44387-026-00130-1