The promise everyone repeats is that humans and AI together will outperform either alone. It gets said so often that it has stopped sounding like a claim.
Patrick Hemmer, Max Schemmer, Niklas Kühl, Michael Vössing and Gerhard Satzger treated it as one. Their paper in the European Journal of Information Systems in 2025, volume 34, is peer reviewed and unusually direct about the state of the evidence.
Their term for the promise is complementary team performance: a level of performance that neither the human nor the system can reach individually. Their observation about it is that it has rarely been observed empirically.
Not never. Rarely. Which means that the arrangement almost every organisation is now paying for is one that does not reliably produce the thing it is bought for, and that the reasons are worth understanding properly.
Potential and effect are different quantities
The paper's first useful move is to separate two things that get run together.
Complementarity potential is the theoretical ceiling. If you could always tell, case by case, whether the human or the system would be more accurate, and always went with the better one, how much better than either alone would you be?
Complementarity effect is what you actually get. It is the improvement realised by the arrangement you have built.
The gap between those two is where most of the disappointment lives. An organisation can have genuine potential sitting in its team and capture almost none of it, because capturing it requires knowing which one to trust on this particular case, and nothing in the workflow establishes that.
Notice that potential has to exist before effect is possible. If the human and the system would give the same answer with the same confidence in every case, there is no ceiling to reach. So the prior question is where potential comes from at all.
The two sources are not equivalent
Hemmer and colleagues identify two.
Information asymmetry. The human and the system hold different information. The system has the structured record. The person has the conversation last Tuesday, the thing the customer said that was never logged, the reason the previous attempt failed.
Capability asymmetry. Both hold the same information and process it differently, arriving at different answers with different error profiles.
Their studies test each separately. In one, on real-estate price prediction where the model had tabular data and photographs were available to people, information asymmetry raised both the ceiling and what was actually realised. In another, on image classification with the model's capability varied deliberately, capability asymmetry did the same.
Both sources are real. But they are not equally available to you.
Capability asymmetry is largely bought. It is a property of which model you use, how it reasons, what it was trained on. Vendors compete on it, it improves without you doing anything, and it is available to your competitors on the same terms. As models converge, it also gets smaller.
Information asymmetry is yours. It is a property of what your people know that is not in any system, and it is the one thing in the whole arrangement that nobody else can acquire.
The standard workflow throws the useful asymmetry away
Picture how AI actually enters a decision in most organisations.
The system produces a recommendation, a summary, a draft, a score. A person receives it. The person is asked to review it, approve it, or flag concerns.
Look at what that arrangement can and cannot use.
It can use capability asymmetry, weakly. The human might process the same material differently and catch something.
It cannot use information asymmetry at all. Nothing in the interaction asks the person what they know that the system does not. The question posed is: is this output acceptable? The question that would unlock the value is: what do you hold about this situation that is not in here?
Those produce very different responses. The first invites a judgement on someone else's work, which people mostly answer by looking for errors in what is in front of them. The second invites contribution, which requires the person to go and get something from their own experience.
And the review framing has a second cost. The output arrives first, complete and fluent. Whatever the person might have thought unprompted is now anchored to it. The asymmetry has not just gone unused. It has been partly destroyed by the order of operations.
If you want to look at where the unique knowledge in your organisation is currently being asked for, and where it is not, talk to us.
Unique information does not surface by itself
Suppose you accept the argument and start asking people what they know. There is a further problem, and it is older than any of this.
Groups systematically fail to pool information that only one member holds. Garold Stasser and William Titus demonstrated the shared information bias in 1985. The meta-analysis by Lu, Yuan and McLeod in 2012 covered 65 studies and 3,189 groups. It found that groups discuss commonly held information at roughly twice the rate of unique information.
While Hemmer et al. (2025) call for exploiting information asymmetry (that humans should bring to the table unique knowledge that AI does not have), Stasser and Titus' research shows why this does not happen in practice. When humans and AI form a "team", three critical problems arise: 1) Consensus-seeking: Humans blindly trust the general knowledge provided by AI (because it feels the most valid and "shared") and fail to bring their own tacit knowledge or local context to the process. 2) Rejection of unique information: The blind trust of employees in algorithms (27% trust AI more than their colleagues, as reported by UpGuard in November 2025), exacerbates the phenomenon. If a person notices something that conflicts with the AI's analysis, they remain silent because, like peer pressure, the mass of AI data devalues the individual's own observation. 3) Agent blind spots: When companies build what Gartner calls autonomous agents, they create systems that operate only on the shared data fed to them. Because they can't ask a human for the "knowledge that only one member has", they make decisions with incomplete information.
So information asymmetry between a person and a system has an exact analogue between people, and it fails in exactly the same way. The unique thing does not get said, or gets said once and not picked up.
That means asking is necessary and nowhere near sufficient. The structure has to make the unique contribution survive after it arrives.
What makes an asymmetry usable
Three requirements follow directly.
The ground has to travel with the contribution. Information asymmetry only helps if you can tell what kind of information you have received. Someone repeating a figure from the same report the model read is not asymmetric information. Someone reporting what they observed in a plant last month is. Stored as plain text those look similar, which is why so much of the value evaporates in aggregation.
This is why every Spark on Hunome carries its characterisation, for example knowtype, the contributor's own account of how they know what they are contributing: research, expert fact, lived experience, values or gut feel. It is not a label for tidiness. It is what makes the asymmetry legible, and therefore usable.
Contribution has to come before synthesis. The refinement position, not the ideation position. If people see the machine's framing first, what they contribute afterwards is a reaction to it, and reactions do not contain the information the framing missed.
The relationships have to be kept. One person's unique contribution changing another person's position is the actual mechanism of value here. A system that stores contributions in parallel and summarises them has recorded the inputs and lost the event.
A SparkMap holds the perspectives, their grounds and the connections between them. The Lens is how each SparkMap’s insights page reads what is there: where meaning clusters, which chains of building produced something new, where different ways of knowing show up across the map. What comes out of the process is deliberative intelligence, which is the full picture rather than an average of it.
The asymmetry is the whole asset
Strip the argument back and it is quite simple.
There are two reasons a person and a system might beat either alone. One of them improves on its own, is sold to everybody, and shrinks as models converge. The other is specific to your organisation, cannot be bought, and is currently being discarded by the way most workflows are arranged.
Hemmer and colleagues found that the arrangement rarely produces the promised result. That is not an argument against putting humans and AI together. It is an argument that the arrangement matters more than the components, and that the arrangement most organisations chose is the one that uses the source they do not own and wastes the one they do.
The information your people hold and the system does not is the entire basis of any advantage available here. Building the process that gets it into the decision is the work.