Send My Clone

The Spectrum of Standing In

To understand why this question is more complicated than it first appears, it helps to recognise that meetings have always involved varying degrees of presence, attention, and substitution.

Consider the humble out-of-office auto-reply, a digital stand-in that has existed for decades. No one considers it deceptive when a colleague's email bot informs you they are unavailable. Move up the spectrum and you find shared calendars where assistants accept invitations on an executive's behalf, or junior colleagues who "represent" a department without the senior leader's direct involvement.

The video call itself, which became the default mode of professional interaction during the pandemic years, already introduced a layer of mediation between participants. Filters smooth skin. Virtual backgrounds conceal messy kitchens. Gallery views flatten hierarchies into a grid of equally sized rectangles. None of this is typically described as deception, yet each element subtly manipulates the impression one participant forms of another.

AI avatars occupy a new and considerably more potent position on this spectrum. When Zoom's Steve Rafferty, the company's head of APAC and EMEA, used his AI avatar to introduce a quarterly meeting in fluent French, he was not simply delegating a task; he was projecting a version of himself that could do something he could not. Rafferty's team spans from the Arctic Circle to Antarctica, covering roughly sixty different languages, and the avatar allowed him to deliver a personal, multilingual message at scale. The tool cannot yet interact with other participants or answer questions in real time, but the direction of travel is unmistakable.

The crucial distinction is between transparent substitution and covert impersonation. If everyone in the meeting knows they are watching an AI avatar, the dynamic is fundamentally different from a scenario where participants believe they are speaking to a living, breathing human being who happens to be on camera. The first is a communication tool. The second is, by most reasonable definitions, a form of deception.

But between these two poles lies an enormous grey zone: the avatar that is technically disclosed but functionally indistinguishable from the real person; the avatar whose presence is noted in a meeting invitation that nobody reads; the avatar that begins as a disclosed introduction but seamlessly transitions into a conversation that feels, to other participants, like a human exchange. The spectrum of standing in, it turns out, is not a spectrum at all. It is a fog.

What Philosophers Make of Digital Doubles

The philosophical landscape here is richer than the technology industry tends to acknowledge. Luciano Floridi, the founding director of Yale University's Digital Ethics Center and a professor at the University of Bologna, has spent years developing an ethical framework for artificial intelligence built around five principles: beneficence, nonmaleficence, autonomy, justice, and explicability.

Floridi's work on deepfakes is particularly relevant. He argues that AI-generated synthetic media has the capacity to undermine our confidence in the original, genuine, authentic nature of what we see and hear. The threat is not merely that a specific piece of content might mislead; it is that the very existence of convincing synthetic media corrodes the epistemic foundations on which trust depends.

Apply this framework to the meeting avatar scenario and the implications are sobering. A meeting is not just an exchange of information; it is a social contract. Participants implicitly agree to be present, to listen, to respond in good faith. When one party secretly outsources their participation to a machine, they violate not just the expectation of presence but the norms of reciprocity that make collaborative work possible. The person who sent the avatar may receive a neat summary afterwards, but their counterparts invested real cognitive and emotional effort into an interaction they believed was mutual. That imbalance is not a minor technical detail. It is a breach of the implicit bargain that makes professional relationships function.

From a Kantian perspective, the issue is equally stark. Immanuel Kant's categorical imperative holds that one should act only according to principles that could be universalised without contradiction. If everyone sent avatars to every meeting, the meeting itself would cease to function as a space for genuine human deliberation. The universalisation test fails spectacularly: a world in which all meeting participants are AI avatars is a world in which meetings are simply algorithms talking to algorithms, with no humans in the loop at all. The very concept of a "meeting" presupposes the meeting of minds, not the collision of language models.

Yet utilitarians might see the matter differently. If an AI avatar can represent its principal accurately, freeing that person to do more meaningful work or simply to rest, the aggregate benefit might outweigh the discomfort of reduced authenticity. PwC's 2025 Global Workforce Hopes and Fears Survey, which interviewed nearly 50,000 workers across 48 economies and 28 sectors, found that daily users of generative AI reported being more productive (92 per cent, compared to 58 per cent of infrequent users), with higher perceived job security and pay. If avatars extend these productivity gains by reclaiming hours lost to routine meetings, the utilitarian calculus could tip in their favour. The question then becomes empirical: does the avatar actually represent the person faithfully, or does it introduce distortions, biases, and errors that compound over time?

The Markkula Center for Applied Ethics at Santa Clara University published a case study examining precisely these tensions. The centre frames the discussion through multiple ethical lenses, including rights, justice, utilitarianism, the common good, virtue, and care ethics, and invites readers to consider what obligations a person has to disclose their use of an avatar. The case study does not offer a tidy resolution. Instead, it highlights that the ethics of meeting avatars depend heavily on context: who is in the meeting, what is at stake, whether disclosure has occurred, and what alternatives exist.

Consent, Disclosure, and the Trust Deficit

If the philosophical arguments suggest that undisclosed avatar use is ethically problematic, the practical question becomes: what kind of disclosure is sufficient?

Zoom's own approach offers one model. When the company's AI Companion joins a third-party meeting to transcribe and summarise, it automatically posts a message in the meeting chat identifying itself as a bot and indicating that it is transcribing. Its video tile displays the word "Transcribing" alongside the Zoom AI Companion logo. This is transparency by design, built into the product architecture so that disclosure is not left to the discretion of individual users.

But the new photorealistic avatar feature complicates this model considerably. If the avatar looks and sounds convincingly like a real person, a small chat notification may not be enough to prevent participants from believing they are interacting with a human. The gap between what the technology can simulate and what a text disclaimer can effectively communicate grows wider with each improvement in rendering fidelity, voice synthesis, and facial animation.

There is an old principle in design: if you have to explain it, you have already failed. When a photorealistic avatar requires a text disclaimer to prevent deception, the product itself is designed in a way that defaults to misleading.

Zoom appears to recognise this tension. Alongside its avatar rollout in March 2026, the company introduced deepfake-detection technology for meetings, providing real-time alerts when synthetic audio or video is detected. This is a notable acknowledgement that the very product Zoom is selling - convincing digital replicas of real people - simultaneously creates a security and trust risk that requires countermeasures. It is as though a locksmith, having sold you the world's most sophisticated lock-picking kit, also offers to install a better deadbolt.

The broader data on consumer attitudes reinforces the concern. Research consistently shows that the vast majority of people value authentic content and view undisclosed AI usage as a breach of trust. More than half of consumers surveyed demand explicit disclosure when AI-generated video, images, or avatars are used, and younger demographics, particularly Generation Z, tend to view AI-generated content as inauthentic and unethical when it is not clearly labelled.

This creates a paradox for companies eager to deploy the technology. The more convincing the avatar, the more useful it is as a communication tool, but the more convincing it is, the greater the expectation of disclosure, and the more disclosure undermines the illusion of natural presence that makes the avatar appealing in the first place. Call it the uncanny valley of trust: as the technology improves, it enters a zone where it is good enough to deceive but not good enough to make deception acceptable.

The Legal Landscape Takes Shape

Regulators have not been idle. The legal framework surrounding AI-generated likenesses, synthetic media, and digital avatars has expanded rapidly across multiple jurisdictions, creating a patchwork of obligations that any organisation deploying meeting avatars must navigate.

In the European Union, Article 50 of the AI Act establishes transparency obligations for providers and deployers of AI systems that generate or manipulate content constituting a deepfake. The rules require that such content be clearly disclosed as artificially generated or manipulated. These transparency provisions are set to take full effect in August 2026, with a Code of Practice expected to be finalised in mid-2026 to establish practical standards. The scope is broad: the EU's framework covers AI-generated text, audio, video, images, avatars, and digital twins. For any multinational corporation considering the deployment of meeting avatars across European operations, the compliance obligations are substantial and the penalties for failure significant.

In the United States, the regulatory picture is more fragmented but no less active. As of early 2026, forty-six states have enacted legislation targeting AI-generated media in some form. In 2025 alone, 146 bills were introduced to state legislatures that included language specific to AI deepfakes. The federal TAKE IT DOWN Act, passed in 2025, represents America's first national law directly regulating deepfake abuse, though its primary focus is nonconsensual intimate content rather than business communications.

At the state level, Tennessee's ELVIS Act (Ensuring Likeness, Voice, and Image Security) prohibits the unauthorised commercial use of a person's voice, including AI-generated replications. California's AB 2602, effective from January 2025, renders unenforceable any contract provision that allows for the creation of a digital replica of an individual's likeness in place of work the individual would have otherwise performed in person, unless the contract includes a reasonably specific description of intended uses and the individual had professional legal representation.

Morrison Foerster, the global law firm, published an extensive analysis in September 2025 noting that digital avatars sit at the nexus of several evolving legal regimes, including intellectual property rights, publicity rights, and consumer protection. The firm's assessment is unambiguous: companies deploying digital avatars must navigate a complex and rapidly shifting regulatory environment, and the cost of noncompliance is rising.

The Federal Trade Commission has also signalled its intent to act. Fines for "deceptive synthetic endorsements" now reach fifty thousand dollars per violation, a figure that continues to climb as enforcement priorities shift toward AI-generated content.