Avatar
An avatar is the representation of a person in a digital space: the character a player controls in a game, the figure that stands for a user in a virtual world or video call, or the profile picture that stands for an account. The word comes from the Sanskrit avatara, the descent of a deity into earthly form, and it was adopted for computing in the 1980s because the idea fits: a user's presence takes on a body inside a system. Avatars range from a 2D icon to a stylized cartoon to a photoreal 3D head driven by cameras, and what defines one is that it stands for a particular human who operates it, rather than for a character or system that acts on its own.
The computing history of the word is short and traceable. Richard Garriott's Ultima IV in 1985 cast the player as the Avatar, a character on a quest for virtue; Lucasfilm Games' Habitat, built by Chip Morningstar and Randall Farmer in 1986 for an online service, used the word for the graphical figure that represented each user in a shared world; and Neal Stephenson's novel Snow Crash in 1992 popularized it for a virtual-reality embodiment. Online worlds then made avatars an economic and social space: Second Life, launched in 2003, let residents build and trade items for their avatars, and multiplayer games turned avatar customization into a major business through cosmetic items. In these settings the avatar is identity as much as interface, and choices about appearance, name and behavior are part of how people present themselves.
Photoreal codec avatars are the research frontier of the avatar as telepresence. Meta's Reality Labs researchers, including Shih-En Wei, Jason Saragih and Yaser Sheikh, developed a line of work from about 2018 in which a person is captured in a multi-camera dome to train an encoder-decoder, and at runtime headset cameras that see the eyes and mouth supply the codes that drive a realistic face. The aim is to make a video-call-like presence in virtual reality feel like being in the same room, and Mark Zuckerberg showed a demonstration in a 2023 podcast interview. Apple introduced a lower-fidelity version called Personas in the Vision Pro in 2024, built from an on-device scan. Newer representations such as neural radiance fields (Mildenhall and colleagues, 2020) and 3D Gaussian splatting (Kerbl and colleagues, 2023) are now widely used for building head and body models from far less capture than a studio dome.
Talking-head synthesis is a related but different problem, in which a video of a face is generated from audio, text or a single photo rather than tracked from a live person. Early methods such as Wav2Lip (Prajwal and colleagues, 2020) retimed lip movements to match speech, and the First Order Motion Model (Siarohin and colleagues, 2019) animated a still portrait with the motion of a driving video. Later diffusion and transformer systems, including Microsoft's VASA-1 and Alibaba's EMO shown in 2024, generate head movement, expression and blinks from audio and a portrait, and companies such as Synthesia and HeyGen sell presenters for training and marketing videos. These are used for localization and corporate communication, and they also underlie scams and impersonation: in 2024, Arup confirmed that an employee in Hong Kong was deceived by a deepfake video conference into authorizing transfers, an example of an avatar standing in for a person who never agreed to it.
The uncanny valley, proposed by the roboticist Masahiro Mori in 1970, remains the main design constraint at the realistic end. Near-human faces that are slightly off, in eye contact, micro-expressions or lip timing, can provoke discomfort more than clearly stylized ones, which is one reason many mainstream platforms use cartoon avatars on purpose. A misconception is that more realism is always better; stylization can convey emotion reliably, runs on cheap hardware and avoids sharing a person's literal face. Another is that an avatar is the same as a digital human: an avatar is operated by a user as their representative, whereas a digital human is a standalone synthetic character, often driven by an AI system as a presenter or agent, and the two blur when a realistic avatar is puppeted by software.
Identity and consent issues grow with fidelity. A scanned face and voice are biometric data that can be misused if stolen, and platforms must decide who may use a person's likeness, whether avatars must be labeled as synthetic, and how users can prove they are who their avatar claims. Harassment in virtual spaces, bodily personal-space norms, and accessibility, where avatars allow people to express themselves in ways their physical situation does not, are all part of the design problem.
Read avatar as a person's stand-in inside a system: from a game sprite to a codec-driven face, what changes with realism is not the idea but the stakes, because the closer the stand-in looks and sounds like its owner, the more it matters who controls it and whether others can tell.