AI will give you a confident answer based on bad data. It won’t tell you the data is bad because it has no way to know. At IIM Indore’s CERE 2026 conference, the school’s director Himanshu Rai put it plainly, “AI has everything to do with information, but nothing to do with truth.”

What Rai is describing is a structural feature of how AI operates. It lacks what philosophers call epistemic humility, the capacity to recognise the boundaries of its own knowledge. It produces outputs but doesn’t interrogate them. That changes what managers are actually for.

Now in its 16th edition, IIM Indore’s annual Conference on Excellence in Research and Education grapples with AI’s implications not just for business, but for how business is understood.

“Research must go beyond processing information and focus on deeper understanding in an AI-driven world,” Professor Rai argues. “While AI excels at processing information, it does not engage with truth. Scholars should reclaim research as a pursuit of deeper understanding and meaning. AI has everything to do with information, but nothing to do with truth.

Ben Waber, visiting scientist at MIT Media Lab and keynote speaker at the conference, has spent his career mapping exactly the terrain Rai describes. His research, recognised by MIT Technology Review as a top emerging technology and by Harvard Business Review as a breakthrough idea uses real-time data flows to understand how organisations actually function: how teams communicate, how physical space shapes decisions, how collaboration patterns determine outcomes.

The distinction between information and truth is a central operational risk of the AI era.

Why AI gets it wrong with complete confidence

AI systems process vast quantities of data, identify patterns, and produce outputs calibrated to maximise a target – accuracy, efficiency, the probability of a given outcome. They do this faster and at greater scale than any human analyst. On those terms, they are extraordinary.

What they cannot do is step outside the objective and ask whether it was the right one. They have no independent concept of truth. They cannot distinguish a statistically robust pattern from a contextually misleading one.

They cannot tell you when the question they were asked was the wrong question.

Zillow discovered this at significant cost. Its algorithmic home-buying programme performed impressively against its targets, until market dynamics shifted. The system kept optimising on outdated assumptions and produced losses at scale, because nothing in the model could register that the model itself had become wrong.

Meta’s engagement-driven algorithms present a related problem. They surface content that maximises the objective they were given, which is attention, not accuracy. The result is a system that is extraordinarily good at its job and structurally indifferent to whether its outputs are true.

The common thread is not bad data or flawed engineering. It is the gap between what an AI system is asked to optimise and what an organisation actually needs.

When managers stop deciding

This would matter less if the humans working alongside these systems remained reliably sceptical. The evidence suggests they often don’t.

The pattern has a name in organisational research: algorithmic deference. It describes the tendency to accept machine outputs because they carry the implicit authority of objectivity. The output arrives dressed in data. It has processed more information than any individual could, and questioning it feels like preferring intuition to evidence.

Over time, managers who routinely defer to algorithmic outputs gradually exercise less independent judgment. The skill atrophies, and though people are still signing off on things. They have simply stopped deciding.

Amazon’s dynamic pricing systems illustrate how far this logic can extend. Prices shift constantly in response to demand signals, competitor activity and inventory levels, with no direct human intervention between the system’s outputs and the market. JPMorgan Chase’s AI-driven legal tools have reshaped what compliance lawyers do, not by replacing them, but by changing what they actually review, and therefore what they actually think about. In both cases the humans are still present, but their role has changed in ways that are not always visible to the people performing it.

Slowing down AI with friction

One response to this, discussed at length at CERE 2026, is that well-designed human-AI collaboration may require building deliberate friction into the process.

The corporate instinct runs the other way. AI promises to streamline decisions, remove bottlenecks, compress the time between data and action. Most organisations treat that as straightforwardly good. But removing friction entirely also removes scrutiny.

Microsoft’s internal guidelines require human review of AI-generated outputs before they enter consequential workflows. Goldman Sachs requires explicit sign-off on high-stakes algorithmic decisions. These are acknowledgments that the technology, left unchecked, will optimise efficiently toward whatever it was asked, and someone needs to keep asking whether it was asked the right thing.

The organisational implication is significant. If the value of human judgment in an AI-augmented organisation lies partly in its willingness to slow things down, then management practices, incentive structures and performance metrics need to reflect that. Speed and efficiency are easy to measure, but the value of a manager who knew when to stop and question the model is harder to quantify. That often means it tends not to get rewarded.

What AI can’t do is a management skill

The task Himanshu Rai is describing, and that Waber’s research helps operationalise is interpretation. It is distinct from analysis, and it is what AI cannot do.

Interpretation requires knowing what question was worth asking in the first place. It requires recognising when an output, however technically accurate, does not fit the context. It requires the ability to hold a model’s findings at arm’s length and ask what they leave out. And it requires the willingness to make a call, take responsibility, and be wrong in ways that cannot be blamed on the system.

None of this is new to management. What is new is the pressure working against it. Waber’s work has shown that organisations generate enormous quantities of behavioural data about how their people work, communicate and decide,and that this data, carefully analysed, reveals things that conventional management intuition misses. The risk is not that the data is wrong. It is that the data becomes the answer rather than an input into judgment. Better information, consumed without interpretation, produces faster versions of the same mistakes.

Information versus truth

Rai’s statement, “AI has everything to do with information, but nothing to do with truth” is not a complaint about technology. It is a precise description of what it does and does not do, and therefore a precise description of what human judgment is now for.

Trust, fairness, long-term resilience, organisational purpose: none of these reduce neatly to an objective function. They require decisions that are sometimes not optimal in the short term, made by people who understand the context deeply enough to know when the model is wrong.

The risk for organisations that lean heavily on AI is not that they become less capable. It is that they become capable only of measuring what they already know how to measure and mistake that for understanding.

That is the difference between information and truth. Right now, it is also the most important thing a manager can demonstrate.

Interested in this topic? You might also like…