The Cost of Asking the Question
On hesitation, responsibility, and decision-making in the age of AI.
No one wants to be the one who asks the question.
Not because the question lacks legitimacy - most people recognise that it should be asked - but because once it’s voiced, ignorance is no longer available as a defence. To ask is to acknowledge that an answer might exist, and that the answer might be prohibitive, or at least meaningfully complicate progress. It introduces friction into environments optimised for momentum.
In contemporary technology development, and particularly in AI, this creates a subtle psychological and organisational tension. It is not that engineers or data scientists are indifferent to potential harm. Rather, the relationship between action and consequence is often indirect, delayed, and distributed across many actors. What emerges downstream rarely feels traceable to any single upstream decision, even when, in aggregate, those decisions matter.
Under these conditions, responsibility becomes difficult to locate. Harm, when discussed, is framed not as an event but as a risk: probabilistic, modelled, abstract. This framing is not dishonest, but it does make moral weight easier to defer. When combined with real constraints - delivery timelines, competitive pressure, job security - the path of least resistance is often to proceed while assuming that any truly serious concerns will be identified and addressed elsewhere in the system.
Responsibility for harm is so diffuse, so downstream, that it becomes cognitively survivable to defer it.
In theory, this is where escalation mechanisms and whistleblowing protections are meant to operate. In practice, speaking up is rarely experienced as neutral. Even when formal protections exist, the informal costs can be significant: reputational risk, career stagnation, social friction, or the quiet sense of becoming “difficult”. These costs are borne by individuals, while the benefits of preventing diffuse or speculative future harms are collective and uncertain. It is therefore unsurprising that many concerns are expressed cautiously, indirectly, or not at all.
What results is not malice, nor negligence, but a form of structural quietude. Each actor assumes that governance, leadership, or regulation will intervene if necessary; governance assumes industry has some capacity for self-correction; regulators operate with partial visibility into systems already in motion. Responsibility is not rejected, but continually displaced.
Is this the bystander effect industrialised?
This is not a story of villains or heroes. It is a story of systems that distribute decision-making in ways that make hesitation expensive and action easy. Many of the most consequential choices about AI systems are made incrementally, embedded in technical design decisions that appear locally reasonable but whose global effects are difficult to anticipate or contest.
In this context, the cultural framing of AI matters. When technologies are presented as inevitable, transformative, or intrinsically progressive, caution risks being interpreted as obstruction. Hesitation can be misread as a lack of imagination, rather than as a legitimate response to uncertainty.
Perhaps what is needed, then, is not alarmism, nor moral grandstanding, but a recalibration of how these technologies are discussed, developed, and implemented. Less emphasis on spectacle and inevitability; more space for provisionality, limits, and genuine uncertainty.
Perhaps AI needs to be made less sexy.