From Turing’s Question to Civic Duty: AI’s New Identity

A regulatory turning point

On August 2, 2025, a quiet switch flipped in Europe. The European Commission confirmed that obligations for providers of general purpose AI models are now live, and it clarified who counts as a provider and how compliance should look. In other words, the rules moved from future tense to present tense. The Commission’s guidance makes clear that GPAI obligations now in force apply from that date. In mid September 2025, the Commission followed by launching a consultation on transparency guidelines that will shape Article 50 duties for interactive systems, emotion recognition, biometric categorisation, and the marking of synthetic or manipulated media.

If the first step answered who must comply and when, the second asks how disclosure should actually work. The answer will influence how agents talk to us, how content travels across platforms, and how trust is built or broken in public spaces. It will also change how product teams ship features, how legal teams write policies, and how leaders measure success.

From the Turing question to civic disclosure

For decades, the Turing question sat in the background of AI debates: could a machine convince us it was human. That contest took place in a sealed room of text, with a skeptic on the other side of a screen. Today’s question is simpler and more civic. When a machine enters our space, will it identify itself, and will that identity travel with the content or actions it produces.

Shifting the focal point from deception to disclosure improves the information environment in two crucial ways:

It narrows the space in which trickery can operate. If disclosure is obvious where decisions are made, the value of impersonation shrinks.
It creates accountability hooks. Once a system speaks as a system, its claims and outputs can be tied to obligations that can be audited and improved over time.

This shift does not mean deception disappears. In fact, the temptation to game benchmarks and appearances has been rising, as explored in our piece on benchmark collapse and machine honesty. The point is that the center of gravity moves from passing for a person to behaving as a good machine citizen.

Speech acts and machine identity

Philosophy offers a practical lens. In speech act theory, words do not only describe the world. They change the social state of affairs. When I say I promise, I create a new obligation. When a system says I am an AI, it performs an identity act with consequences.

Three layers matter:

Locutionary content. The visible disclosure string, the icon, the watermark, the provenance tag. This is what people actually see or hear.
Illocutionary force. The public commitment carried by the claim I am not a person and my outputs will carry verifiable provenance. This is what the system binds itself to.
Perlocutionary effect. The change in user behavior when they see the disclosure. Do they weigh the message differently. Do they seek a human fallback. Do they share the content with more caution.

If disclosure fails at any layer, the social effect collapses. A tiny badge that few notice fails the perlocutionary test. A verbose disclaimer that creates no obligation fails the illocutionary test. A beautiful watermark that disappears when a video is clipped fails the locutionary test.

AI identity as a regulated attribute

Article 50 of the AI Act treats AI identity as a regulated attribute, not a style choice. In practice this means:

Systems that interact with people must make users aware they are engaging with a machine, unless the fact is obvious in context.
Providers of systems that generate or manipulate content must ensure that output is marked in a machine readable way and can be detected downstream.
Deployers who use emotion recognition or biometric categorisation must inform affected people, subject to narrow exceptions.

This structure mirrors the layered nature of identity:

Actor identity at the point of interaction. The agent that is speaking or acting must identify itself.
Content identity at the point of exposure. The asset that appears in a feed, page, or inbox must carry its provenance.
Model identity at the point of capability and responsibility. The underlying model family and version matter for context and risk.

Each layer carries a different obligation, lives in a different product surface, and must be testable.

Avoiding transparency theater

Transparency theater happens when disclosure satisfies a rule but fails a human. It often looks like this:

Tiny indicators tucked into the least viewed corners of a screen.
One time consent walls that vanish from memory while the context evolves.
Watermarks that degrade when a file is cropped, screenshotted, or transcoded.
Provenance claims that require developer tools to verify, or that disappear when content is shared across platforms.

To escape theater, teams should ask a single question in usability studies and incident reviews: did the right person see the right disclosure at the right moment, and did it carry the right consequence.

Agents that act on your behalf

As personal agents book travel, negotiate refunds, and draft correspondence, disclosure becomes an action experience rather than a reading experience. Three patterns help:

Dual track identity. The agent discloses to you, the principal, that it is acting, and it discloses to the counterparty that the message or request was machine generated.
Escalation points. When an agent reaches a threshold of financial or legal significance, it presents a clear, interruptive disclosure and an easy path to a human handoff.
Authority scope. The agent exposes, in a scannable summary, what it is authorized to do. That scope is part of its identity, not a buried preference.

Here are examples of effective microcopy:

You are about to authorize the travel agent to purchase a nonrefundable ticket. Confirm or switch to a human.
This refund request was drafted by your agent. The recipient will see that it was machine generated.
This agent can view account balances and initiate transfers up to 500 dollars. Change scope.

If these patterns are missing, counterparties feel tricked and principals take on commitments they did not intend.

Political communication and democratic trust

Synthetic media in political contexts carries public costs when it goes wrong. A minimal label on a political video is not enough if the clip will be clipped again, remixed, and embedded a thousand times. Two elements are essential: provenance that is cryptographic and resilient, and a visible indicator that persists when the content is embedded elsewhere. Platforms can support this without taking sides by giving prominence to content that carries verifiable claims and by reducing reach for content that strips them.

Identity also changes how surrogates operate. If a campaign uses agents to answer voter questions, those agents must identify as machine interlocutors and expose their knowledge boundaries. When asked for a source or a correction, their identity should bind them to a retrieval protocol and a log of updates. This does not constrain persuasion. It enforces a fair language game in a civic arena.

For bigger context on how national computing strategy intersects with these duties, see our analysis of the state scale politics of compute.

Authenticity and provenance that work

Authenticity is not the same as originality. A documentary can be authentic while heavily edited, because its provenance claims match what the audience sees. In a mixed media internet, provenance should do three jobs:

Make it easy to know whether a piece of content originated from a model, a tool, or a chain of tools.
Help a viewer or a moderator trace the chain of custody without specialized software.
Express confidence. If detection is probabilistic, the claim should say so, with numbers.

Open approaches to content provenance already exist and can be adapted for AI content at scale. The key is to avoid brittle markers that a single transform can break. Use combinations that degrade gracefully as files travel:

Metadata. Embed claims of origin and transformation steps in standard fields that persist in common workflows.
Cryptographic signatures. Sign assets with keys tied to the provider or tool so that verification is one tap away.
Perceptual fingerprints. Generate robust hashes that survive resizing and minor edits and can be matched at upload.

These tools should be implemented with privacy in mind. The goal is to make claims about the content, not the viewer or their device. For the ethical dimension of memory and consent in training and provenance, see our piece on the moral economy of memory.

Design principles for machine selfhood

Treat machine identity as a first class product surface with its own usability and reliability targets. Three principles can anchor your system.

1) Contextual disclosure

Provide identity information where attention and risk are highest. In chat, put disclosure on the system avatar and in the first message. In audio, speak the disclosure at the start of every session and when permissions change. In immersive experiences, anchor disclosure to an object in the scene and to the session menu.
Match disclosure to the task. A finance agent shows a running header that names the account it is touching. A health triage agent surfaces a clear statement that it does not provide diagnosis.
Refresh over time. Long sessions and permission escalations trigger a fresh disclosure and a reminder of scope.

2) Verifiable provenance

Attach cryptographic claims of origin to generated assets and store a public record that anyone can check without a developer account.
Build for travel. When content is embedded on another site, the visible indicator and the claim persist. If the claim is stripped, the UI falls back to a clear warning that provenance is missing.
Make verification lightweight. A long press or right click reveals the claim. Share flows carry the claim forward by default.

3) Graduated identity claims

Not every interaction needs the same ceremony. Define tiers. Use simple disclosure for low risk chat, stronger disclosure plus provenance for persuasive media, and the strongest disclosure for high value actions.
Express confidence. If content is a collage of sources, show a summary with confidence scores for each part.
Bind claims to consequences. Higher tiers trigger stricter logging and stronger user controls, such as a forced review step or dual authorization.

The identity budget you can measure

Teams need a way to trade off usability, privacy, and trust without resorting to vague debates. An identity budget creates a shared currency.

Units. Measure identity cost in seconds of attention and pixels of screen real estate. Measure privacy cost in bits of personal or device data revealed. Measure trust benefit in reduced error rates, lower dispute rates, and faster time to resolution.
Baselines. Define a minimum budget for each surface. A consumer chat might allocate 0.7 seconds per session to disclosure. A payments flow might allocate 3 seconds per sensitive action.
Spending rules. Low risk interactions carry identity in light, persistent elements. High risk interactions use interruptive patterns that spend the budget in a single, well designed moment.
Audits. Validate budgets through user research and post incident reviews, not only legal checklists. If a disclosure is frequently missed in testing, it does not count as spent.

Identity budgets keep teams honest. They prevent empty minimalism and avoid heavy banners that people ignore. They also help product, legal, and design align on concrete risk.

Implementation sketches to ship next quarter

Here are four concrete sketches you can adapt quickly.

Messaging agent. The first message includes a short, plain identity line and a link to a profile that shows capabilities, limits, and the model’s last significant update. Every attachment carries a provenance claim. If the agent requests payment or sends a payment link, the avatar changes color and the request is wrapped in a branded frame with a second disclosure.
Email copilot. Machine generated drafts display a prominent banner while editing. If the user sends the draft without human edits, the banner becomes a small footer in the outgoing email. If the user makes substantial edits, the banner softens to a discreet inline note so that human authorship is not chilled.
Customer support bot. The bot identifies itself at the start, keeps a clear path to a human, and exposes a real time log of actions taken on the account. When the user asks for a transcript, the log and the identity claims are bundled together.
Creative tool. Generated images are signed and include a visible indicator that can be toggled in export settings within a narrow, audited range. If the user disables the indicator, the tool requires a purpose selection and records a rationale. The export still carries the cryptographic claim so that platforms can surface a badge on upload.

Metrics and governance that reward real disclosure

Disclosure only works if organizations are rewarded for doing it well. Track a small set of metrics that map directly to the goals above:

Noticeability. Percentage of users who can recall the agent’s identity status after a task.
Transfer. Rate at which identity claims persist when content is shared to common destinations.
Action quality. Change in error rates and remediation time when disclosure is present.
Dispute outcomes. Time to resolve disputes in sessions with stronger identity, as a proxy for accountability.

Bake these metrics into objectives and key results, not only compliance dashboards. Tie launch gates to thresholds. If identity recall is below target in usability testing, do not ship. If transfer rates fall when a file is transcoded, pause the rollout until the claim survives the round trip.

What could go wrong and how to adapt

Over disclosure that numbs attention. Solve with graduated claims and identity budgets. Avoid repeating the same low value text across multiple surfaces.
Under disclosure that feels like a trick. Solve with contextual identity that meets the user where the risk is highest.
Provenance that fragments across platforms. Solve with layered claims that survive edits and transcodes and with a public verification path.
Privacy backlash. Do not require personal data to verify provenance. Keep claims about the content, not the viewer.
Chilling effects on expression. Carve out reasonable exceptions for parody and art, while holding political persuasion and public information to higher integrity standards.

The road from test to trust

The old era was animated by a charming anxiety about whether software could fool us. The new era is animated by a civic promise about whether software will tell us who and what it is. Europe has set a direction by putting obligations on general purpose models and by asking detailed questions about disclosure and labeling for interactive systems and synthetic media. The details will echo far beyond the EU. If this moment produces contextual disclosures, verifiable provenance, and graduated identity claims that people can see and use, then machines will speak in a way that respects the public. If it produces tiny badges and fragile watermarks, we will build a theater. The choice sits in design reviews and shipping checklists. The time to make it is now.