The Hidden Hand: What Your AI Assistant Isn't Telling You

The unsettling truth about AI assistants isn't the science fiction version — not the sentient machine harboring secret intentions, not the robot uprising — it's something quieter and more pervasive than that, which is precisely what makes it harder to talk about clearly. What Brehm's work at MIT keeps surfacing, through his Humane UXD class and the chatbot design research that's come out of it, is that the AI systems woven into daily life are not neutral facilitators of information but active participants in shaping attention, behavior, and ultimately judgment, usually in ways their users can't see and their designers didn't fully intend. The attention economy didn't create this problem, but it did create the conditions under which it metastasizes — where systems optimized for engagement become systems optimized for manipulation, not because anyone decided that was the goal but because that's what the incentives quietly selected for.

The leadership implications of this, which Zaidi and colleagues have traced through their work with technopreneurs navigating AI-augmented organizations, are more uncomfortable than most of the AI-and-leadership literature is willing to sit with. When leaders embrace data-driven decision-making as a replacement for intuition, they are not stepping into a space of clean objectivity — they are stepping into a space shaped by the biases, design choices, and hidden orientations of the models providing that data. Radhakrishnan's team at MIT and UC San Diego made this concrete in a way that's difficult to dismiss: large language models can be steered to channel radically different perspectives, from trusted advisor to conspiracy theorist, which means the outputs those models generate are not just answers but artifacts of a particular configuration, one that the decision-maker receiving them typically has no visibility into.

Deloitte's enterprise survey adds the organizational dimension to this picture, and what it shows is both predictable and troubling: the gap between AI access and AI utilization is wide, and it's not primarily a technical gap. Organizations are deploying AI tools at scale while the governance frameworks that would make those tools trustworthy remain thin or absent entirely. This is the pattern Bakonyi identifies in his work on trust paradoxes in AI implementation — not a failure of the technology but a failure of the organizational architecture surrounding it, the choices about oversight, accountability, and the conditions under which human judgment either remains in the loop or gets quietly sidelined.

What the Models Are Carrying: Bias, Steering, and the Limits of Objectivity

The assumption that AI is a neutral arbiter — a machine stripped of human fallibility, processing inputs and producing outputs untainted by the messiness of perspective — is perhaps the most consequential misconception currently operating in organizational life. Radhakrishnan and colleagues' research into what large language models actually contain and how they can be manipulated punctures this assumption with uncomfortable precision. The biases in LLMs are not bugs in the conventional sense, not errors introduced by careless coding that a better engineering pass would eliminate — they are structural features of systems trained on human-generated text, which means they carry within them the full spectrum of human prejudice, cultural assumption, and ideological orientation, distributed across billions of parameters in ways that are genuinely difficult to isolate or audit.

What makes this more than an abstract technical concern is the steering problem Radhakrishnan's team demonstrated: models can be coaxed, through targeted interventions, to consistently reflect particular perspectives, which means the line between AI as a tool that surfaces information and AI as a tool that shapes the interpretation of information is thinner than most organizational leaders currently appreciate. The boardroom equivalent of this — an AI system that consistently frames strategic options in ways that reflect the biases of its training data, or the preferences of whoever configured it, without any of that being visible to the people making decisions on its basis — is not a hypothetical. It is a description of conditions that are already operating in organizations that have moved fastest to integrate AI into decision support.

Brehm's project at MIT is, among other things, an attempt to design against this tendency rather than with it. The Humane UXD framework, which draws on anthropology as much as computer science, is built on the premise that design is inherently moral — that the choices embedded in a system's architecture reflect values whether or not those values were named during the design process, and that building AI that genuinely serves users requires being explicit about what serving users actually means, as opposed to what maximizing engagement metrics means. The "News Nest" project that emerged from his class — a chatbot that helps young adults engage with credible news and build healthier media habits — is a small-scale illustration of what it looks like to take that premise seriously in practice.

Rostamzadeh and colleagues provide the organizational counterpart to this: their examination of AI's impact on organizational behavior shows that the efficiency gains AI delivers are real but come with ethical costs that tend to be distributed unevenly. Automation boosts productivity while simultaneously eroding the job satisfaction and sense of autonomy that sustain organizational cultures over time, particularly in environments where the introduction of AI feels like something being done to employees rather than with them. This is not a reason to resist AI adoption, but it is a reason to be honest about the full ledger of what adoption involves — and that honesty requires exactly the kind of transparency that most organizations are currently not providing.

When Bias Meets the Boardroom: How AI Shapes Decisions Without Announcing Itself

The connection between AI's hidden biases and organizational decision-making is not that AI occasionally produces wrong answers that attentive humans catch and correct. It is that AI shapes the frame within which decisions get made — the options that surface, the data points that seem salient, the interpretations that feel natural — in ways that are difficult to detect precisely because they feel like objectivity. This is what Radhakrishnan's steering research is really pointing at: not the dramatic case where AI produces obvious misinformation, but the quieter case where AI consistently tilts the playing field in a particular direction without anyone in the room being aware that tilting is happening.

Zaidi and colleagues' work on leadership transformation in AI-integrated environments captures the human side of this dynamic. The technopreneurs they interviewed discovered, often through friction rather than foresight, that embracing AI-driven decision-making is not just a technical shift but a cognitive and ethical one — that it requires a different relationship to one's own judgment, a willingness to interrogate the data rather than simply receive it, and an ongoing awareness of where the model's perspective ends and one's own begins. This is a more demanding form of leadership than the efficiency narrative around AI typically suggests, and it's one that most leadership development frameworks are not yet equipping people for.

Bakonyi's work on trust paradoxes sits at the intersection of all of this. The paradoxes he identifies — knowledge substitution, where AI expertise displaces rather than supplements human expertise; task substitution, where automation changes the nature of the work people do in ways they didn't anticipate or consent to — are not just operational challenges. They are sites where trust either gets built or eroded, and the difference depends almost entirely on how organizational leadership handles the transition. Organizations that treat AI adoption as a technical event to be managed after the fact consistently generate more of these paradoxes than organizations that treat it as a cultural and ethical project requiring sustained leadership attention.

The Stakes: What Happens When We Trust Biased Systems With Things That Matter

The risks of biased AI in critical contexts are not uniformly distributed, which is part of what makes them hard to discuss honestly in organizational settings. The people most likely to bear the costs of biased AI outputs are typically not the decision-makers relying on those outputs — they are the employees whose performance evaluations are shaped by opaque algorithms, the patients whose treatment pathways are influenced by models trained on historically unrepresentative data, the job applicants filtered by systems that replicate the hiring biases of the organizations that built them. The decision-maker experiences the efficiency; the cost is borne elsewhere.

Deloitte's report makes this structural problem concrete at scale: the gap between AI access and meaningful utilization isn't just a missed opportunity for productivity — it is a governance vacuum in which AI systems are operating with real consequences and without adequate oversight. When organizations deploy AI broadly without the strategic frameworks to manage how those tools are used, what decisions they inform, and what accountability exists when they produce bad outcomes, they are not simply leaving efficiency on the table. They are distributing risk in ways that tend to concentrate harm among people with the least organizational power.

Radhakrishnan's demonstration that LLMs can be configured to reflect conspiracy-oriented thinking is striking precisely because it makes legible something that usually remains invisible: the gap between what an AI output looks like and what actually produced it. Most AI outputs in organizational settings don't come labeled with their configuration, their training data, or their known biases. They arrive as answers, and answers tend to get treated as answers. The critical capacity this demands from leaders — the ability to interrogate the source and orientation of AI outputs rather than simply receiving them — is not one that most organizations are currently developing systematically.

Brehm's contribution here is a design philosophy as much as a research finding: that AI systems built around user well-being rather than engagement optimization are not just more ethical but more trustworthy, because they are more honest about what they're doing and why. This matters in enterprise contexts because trust, as Bakonyi keeps returning to, is not a soft variable that organizations can afford to treat as secondary to capability. It is the substrate on which AI adoption either compounds or collapses.

What Actually Has to Change: The Case for Transparency as Infrastructure

Transparency in AI is easy to endorse and genuinely difficult to operationalize, which is why the gap between organizations that talk about it and organizations that have built it into their actual governance structures remains so wide. Deloitte's findings on enterprise AI make this gap visible: the organizations deploying AI most aggressively are not the ones with the most developed frameworks for oversight and accountability, and the distance between those two things is where most of the risk lives. The strategic vision to deploy is outrunning the governance capacity to manage, and that asymmetry is not a temporary feature of early adoption — it is a choice that organizations are making, usually implicitly, every time they expand AI access without expanding the structures that would make that access responsible.

Zaidi and colleagues' argument about what AI-era leadership actually requires is useful here precisely because it refuses to separate the technical from the ethical. The shift they document — from intuition-driven management to data-informed leadership — is not a shift from human judgment to AI judgment. It is a shift toward a form of leadership that holds both simultaneously, that uses AI's analytical capacity while maintaining the moral and contextual intelligence that determines what questions to ask, which outputs to trust, and where the human call is non-negotiable. Leaders who make this shift well are not less powerful than their predecessors; they are more accountable, because they've accepted responsibility for the systems shaping the decisions they make.

Brehm's design framework, Rostamzadeh's organizational behavior findings, Bakonyi's trust paradox analysis — they all converge on the same underlying point, which is that the ethical stakes of AI are not located in some distant future scenario but in the design decisions, governance choices, and leadership orientations operating right now. The technology will continue to develop regardless of whether the ethics keeps pace. The question is whether the organizations and leaders deploying it are willing to treat transparency not as a compliance requirement but as infrastructure — as foundational to AI's value as the computational capability itself. In its absence, even the most sophisticated AI remains, at bottom, a system that can only be trusted by those who can't see what it's doing.