For the Ethicists — From Imperative to Indicative

Framing

A Small Shift With a Potentially Massive Impact

Technology has reshaped the economy, the political order, and the private formation of individual people, and the curve is steepening. The role of the technology ethicist has never been more important — or more poorly equipped for the moment.

The last decade tells a consistent story. Market incentives crowded out ethical efficacy on a curve far gentler than the one we are now on. If well-argued principles could not constrain the platforms of 2015, they will not constrain the systems of 2030.

Section 1

A Short History of Technology Ethics

Technology ethics is not new. The first serious treatments appeared in the early twentieth century, alongside electrification and mass industrialization: labor displacement, the concentration of productive power, the cultural effects of mass media. Norbert Wiener's The Human Use of Human Beings (1950) named the problem of systems that amplify human will without amplifying human wisdom — a generation before the systems he described existed at scale.

The modern field took shape in the late 1980s and 1990s with the personal computer and then the internet. Computer ethics established foundational concepts around privacy, access, and intellectual property. Bioethics offered procedural models — institutional review, informed consent, the distinction between research and clinical practice — that technology ethics tried to import. The IEEE published codes for engineers. Universities built centers and curricula. The field was well-organized and intellectually serious.

Then the platforms arrived. Between roughly 2005 and 2015, a small set of social media companies became the dominant channels of public discourse for first hundreds of millions and then billions of people. The ethical literature kept pace in volume: AI ethics, algorithmic accountability, data rights, the ethics of persuasion, of recommendation, of surveillance capitalism. Think tanks and digital-rights nonprofits published continuously. In 2016 Google, Microsoft, Amazon, Apple, and IBM formed the Partnership on AI. By 2020 more than eighty AI ethics principles documents had been published worldwide.

Platform behavior moved in the opposite direction over the same window. Time on app rose. Engagement-optimized recommendation deepened. The most extreme content received the most amplification. Adolescent mental health diverged sharply from the pre-smartphone baseline. The ethical output was voluminous, legitimate, and substantially ineffective by its own measures.

Section 2

Imperative vs. Indicative — The Difference

The distinction is grammatical, but it points at something real. An imperative claim takes the form: platforms should respect human dignity; AI systems ought to be transparent; companies must not exploit children's attention. These claims are made in the mood of moral demand. They address the system as if it were a person who might choose to comply. They are easy to write, legitimate to hold, and — in the main — impossible to operationalize.

An indicative claim takes the form: a default privacy setting determines roughly ninety percent of subsequent user behavior; recommendation systems trained on engagement signals amplify emotionally charged content measurably more than neutral content; entry-level hiring in knowledge industries contracted by X percent in the twelve months following the release of general-purpose language models. These claims are made in the mood of description. They address systems as they actually function. They are harder to write, demand evidence, and — when correct — are possible to act on.

The argument here is not to eliminate imperatives. It is to reposition them. The imperative is the destination: we want systems that do not exploit children's attention. The indicative is the diagnostic: here is the specific mechanism by which that exploitation operates, here is the incentive that sustains it, and here is the point at which a different input would produce a different output. Skip the indicative step and the imperative floats free — a legitimate aspiration with no mechanical connection to the system it is trying to change. The cost of doing the indicative work is real: it requires sitting with market structure, reading product roadmaps, learning what a recommendation system actually does. That cost is what the field has been unwilling to pay.

The practical consequence of skipping the indicative is that ethical energy accumulates in the wrong place. Statements are issued when an interface change is what the situation requires. Reports are published when a regulation is what the situation requires. The work is real; the leverage is absent.

Section 3

How Technology Grew Unethically — and What Made It Possible

The ethical deterioration of the last fifteen years did not happen because the people building these systems were unusually bad. It happened because the incentive structures they worked inside were misaligned with the interests of the users they served, and no mechanism existed to impose the cost of that misalignment on the people producing it. This is not a moral failure of ethicists. It is a methodological gap: the field arrived with imperatives without first paying the indicative price of understanding the market in which its imperatives had to live.

The Engagement Trap

The dominant business model of the consumer internet — attention sold to advertisers — creates a direct financial incentive to maximize time on platform. Time on platform tracks emotional activation, and emotional activation tracks content that produces fear, outrage, and social comparison. A system optimized for time on platform migrates, without malice, toward content that produces fear, outrage, and social comparison. The engineers were not trying to harm users. They were optimizing a metric their employment depended on, inside an architecture that made harm the byproduct of success.

The Speed Asymmetry

Technology compounds faster than institutions. A feature that reaches a billion users can ship in weeks. A regulation that addresses it takes years. Academic literature that frames the regulation takes longer still. The gap between what technology can do and what ethical or legal frameworks can say about it has widened monotonically for thirty years, and the curve of current AI capability suggests it will widen substantially in the next decade.

The Authority Migration

Decisions about what an AI system refuses to do, what content reaches a minor's feed, and what speech a platform will host have migrated — largely without public debate — from public institutions to private firms. The state has been outpaced. Academic ethics now functions downstream of decisions already shipped. The authority that ethics depends on to be actionable has moved to the precise location where the financial incentive to ignore it is strongest.

These three dynamics are the structural context in which technology ethics operates. An ethics that does not take them as settled premises — that does not begin by mapping them carefully — will be well-argued and inert. An ethics that starts from them, and only then asks what should be done, has at least a chance of moving something.

Section 4

What Has Actually Worked

The counter-examples are already in the record. Where ethical argument has been converted into mechanism, behavior has changed.

The UK Age Appropriate Design Code

In force since 2021, the UK's Age Appropriate Design Code produced an audited ninety-one changes to platform design — covering default privacy settings, geolocation, profiling restrictions, and the elimination of dark-pattern nudges that steered children toward weaker protections. YouTube, TikTok, Snapchat, Instagram, and Google Search changed behavior in measurable, audited ways. The Code succeeded where decades of exhortation had not, precisely because it operated in the indicative. It did not tell platforms to respect children. It told them which switches had to be in which position, with a financial penalty attached to non-compliance.

GDPR and Consent Architecture

The EU's General Data Protection Regulation changed the architecture of consent across the European internet within a single compliance cycle. Companies that had ignored ethical arguments about data collection for years redesigned consent flows in months — not from moral persuasion, but because a cost was attached to the prior architecture. The argument had been made for a decade. What changed behavior was the mechanism, not the argument.

Default-Off as a Design Principle

Sunstein and Thaler's work on defaults established empirically what platform engineers already knew operationally: default settings determine roughly ninety percent of subsequent user behavior. An intervention that changes a default is more powerful than any amount of user education aimed at the same behavior. The indicative framing — the default is the behavior — points directly at the intervention.

Each example shares a structure. A specific mechanism was identified. A specific point of intervention was located. A specific change was made to the incentive or architecture at that point. The ethical value was the destination. The indicative work — here is the mechanism, here is the lever — was what made the journey possible.

Section 5

The Tool Belt and the Role

There are three categories of intervention, and they are not interchangeable. The dominant failure mode in this field is to reach for the third when the situation requires the first or second. Knowing which tool fits which problem is the first operational skill of indicative ethics, and it depends on first understanding the market structure the tool will operate in.

The role of the ethicist has been poorly defined relative to the people who actually ship decisions. Ethicists are often hired as a signal — a visible commitment to responsible AI — and then deployed at lever iii: reviewing outputs, writing blog posts, advising regulators on principles, managing reputational risk. These are legitimate functions, but they are downstream of the moments where systems are actually shaped. The indicative method requires upstream placement. If the binding constraint is a missing technical capability, the ethicist belongs in the room where the roadmap is written. If the binding constraint is a policy gap, the ethicist belongs with the regulatory affairs team, not communications. Role alignment to lever is the operational precondition for effectiveness.

The cross-functional partnership this requires is not natural and not yet common. Engineers are trained to ask what is technically possible. Ethicists are trained to ask what is morally permissible. The indicative method requires a third question: what is the mechanism, and where is the lever? Neither profession is currently trained to ask it in a form the other finds useful. Building that capacity is the long institutional work. It starts inside companies, with ethicists embedded in product teams rather than attached to communications functions. It continues in universities, where ethics and engineering remain almost entirely separated. It matures when the two communities share enough vocabulary to diagnose a problem jointly and choose a lever together.

Section 6

What We Should Do

The case against imperative-mode ethics is not a case against ethics. It is a case against a particular deployment — the declarative, aspirational, audience-facing mode that has dominated the field for a decade — and an argument for an alternative grounded in mechanism, diagnosis, and the disciplined matching of problems to levers.

The stakes are not abstract. Pure economic optimization in the age of AI does not stabilize. It compounds. The technological moment now underway is the largest test of human disposition since the industrial revolution, and by most measures considerably larger. A handful of operators, by accidents of timing and capability, will accumulate leverage over the future that historically required armies. The quality of the ethical thinking that can actually reach those operators is a first-order question about the next century.

The market will not self-regulate the moral dimensions of this transition. The test of that proposition was the last ten years, and the test was failed — on a curve gentler than the one ahead. The work falls to those willing to take the indicative seriously: to study systems as they actually operate, to diagnose at the level of mechanism, and to match each problem to the lever that can move it. That sequence — indicative first, imperative second — is what the practice now requires.

Concretely, this looks like four moves.

For ethicists inside technology companies: Push for upstream placement. Ask to sit in the product review where defaults are chosen. Write the spec for what auditable compliance would look like before the feature ships, not after the criticism arrives.

For ethicists in academia and policy: Write for the lever, not the principle. A regulation proposal that specifies which switches must be in which position is more valuable than a framework document that names the values the switches should reflect.

For ethicists in civil society: Identify the one policy change that would move the incentive, and pursue it single-mindedly. The Age Appropriate Design Code was not a comprehensive statement about children and technology. It was one very specific set of default requirements. That specificity is what made it work.

For all: Measure outcomes, not outputs. The question is not whether a principles document was well-written. The question is whether a metric moved.

Picture the next decade two ways. In one, technology ethics continues as it has — eighty more principles documents, another generation of conferences, the same aspirational vocabulary applied to systems whose incentive structures it has not bothered to learn. The metrics that matter — adolescent well-being, the share of public discourse mediated by recommendation systems, the concentration of decision-making in a handful of firms — continue their current trajectories, and the ethical literature continues to describe them. In the other, a cohort of ethicists sits in product reviews and regulatory drafting rooms, writes specifications instead of statements, and measures success by whether a default moved. The first decade is the one we are on. The second is still available, but only to those willing to pay the indicative price first.

Read Next — The Companion Pieces

The three problem spaces below apply the indicative method to specific harms. Each starts from market structure and works toward the lever.

Appendix A

References and Source Data

On the History of Technology Ethics

• Wiener, N. (1950). The Human Use of Human Beings: Cybernetics and Society. Houghton Mifflin.
• Moor, J. H. (1985). What is Computer Ethics? Metaphilosophy, 16(4), 266–275. The foundational framing of computer ethics as a field.
• Floridi, L., et al. (2018). AI4People — An Ethical Framework for a Good AI Society. Minds and Machines, 28, 689–707. A representative example of the principles-document era.
• Jobin, A., Ienca, M., & Vayena, E. (2019). The global landscape of AI ethics guidelines. Nature Machine Intelligence, 1, 389–399. Documents the proliferation of AI ethics guidelines and the convergence of their content.

On the Failure of Imperative-Mode Ethics

• Mittelstadt, B. (2019). Principles alone cannot guarantee ethical AI. Nature Machine Intelligence, 1, 501–507.
• Metcalf, J., Moss, E., & boyd, d. (2019). Owning Ethics: Corporate Logics, Silicon Valley, and the Institutionalization of Ethics. Social Research, 86(2).
• Hagendorff, T. (2020). The Ethics of AI Ethics: An Evaluation of Guidelines. Minds and Machines, 30, 99–120.

On Regulation as Mechanism

• Information Commissioner's Office (UK). (2021–2025). Age Appropriate Design: A code of practice for online services. Implementation and audit reports.
• Lessig, L. (2006). Code: Version 2.0. Basic Books. The canonical statement that architecture is regulation.
• Sunstein, C. R., & Thaler, R. H. (2008). Nudge: Improving Decisions About Health, Wealth, and Happiness. Yale University Press. On the determinative power of defaults.

On the Structural Dynamics of Platform Ethics

• Zuboff, S. (2019). The Age of Surveillance Capitalism. PublicAffairs. The authority migration and its structural logic.
• Haidt, J., & Rausch, Z. (2023). The Anxious Generation. Penguin Press. Empirical grounding for the harm-at-population-scale claim.
• ROOST (Robust Open Online Safety Tools). (2025). Founding announcement. Discord, OpenAI, Roblox, Bluesky founding partners. An example of lever i work — building the infrastructure that policy can then mandate.

Companion Pieces

• EconFaithAI: Social Media and Children — Problem space i.
• EconFaithAI: The Morality of Language Models — Problem space ii.
• EconFaithAI: Workforce Displacement — Problem space iii.
• EconFaithAI: For the Innovators — Moral Restraint Moves Us From Greed to Generosity