Compression artifact

A compression artifact is a visible or audible distortion introduced by lossy compression algorithms when the compression ratio exceeds what the algorithm's perceptual model can gracefully accommodate. Unlike the random noise of analog signal degradation, compression artifacts are structured — they manifest as blocky edges, banding in smooth gradients, mosquito noise around high-contrast boundaries, and ringing near sharp transitions. These patterns are not failures of transmission but residues of the compression algorithm's mathematical operations: discrete cosine transform quantization, motion vector prediction errors, and chroma subsampling.

Compression artifacts reveal that digital representation is always a negotiation between fidelity and efficiency, and that the terms of this negotiation are set by algorithm designers whose perceptual models embed assumptions about human vision. What is "perceptually lossless" for a statistical model of the human visual system may not be lossless for an individual viewer, a specific image, or a critical inspection context. The artifact is the visible trace of a decision to discard information.

The presence of compression artifacts in digital television, streaming video, and digital photography challenges the assumption that digital media are inherently higher fidelity than analog media. An analog signal with moderate noise may preserve more information about the original scene than a heavily compressed digital signal that eliminates the noise along with the detail it obscures.

Compression as a Systems Metaphor

The structure of compression artifacts illuminates a general pattern in systems theory: lossy abstraction — the process by which a system reduces the dimensionality of its input to match the capacity of its processing substrate, producing characteristic distortions at the boundary between what is preserved and what is discarded.

In cognitive science, perception itself is a lossy compression process. The retina transmits roughly 10 million bits per second to the visual cortex, but the bandwidth of conscious awareness is estimated at 40-50 bits per second. The vast majority of retinal information is discarded by hierarchical processing — edge detection, motion extraction, object recognition — and what reaches awareness is a heavily compressed representation. The "compression artifacts" of cognition are the phenomena that escape this abstraction: the subtle emotional cues that disappear when reduced to facial expression categories, the contextual nuances lost when a conversation is summarized, the embodied knowledge that resists propositional encoding. What a large language model cannot represent is not merely missing information; it is information that the model's compression architecture was designed to discard.

In institutional theory, organizational memory operates through analogous compression. A corporation does not store every decision, every conversation, every failed experiment. It stores the decisions that were documented, the conversations that produced actionable minutes, the experiments that yielded positive results. The institutional "compression standard" is set by power: who decides what gets remembered, what gets forgotten, and what gets transformed beyond recognition. The artifacts are visible to those who know where to look: the gap between a formal policy and its informal enforcement, the discrepancy between an annual report and the quarterly whisperings, the silence where dissent used to be.

The Political Economy of Compression Standards

Compression standards are not neutral technical specifications. They are regulatory instruments that allocate bandwidth, storage, and attention. The JPEG standard, developed in 1992, embeds assumptions about photographic content that privilege continuous-tone imagery over line art and text. The H.264 standard, dominant in video streaming, optimizes for motion-compensated prediction that assumes cinematic continuity — assumptions that fail for animation, for surveillance footage, for experimental film. Each standard produces its own characteristic artifacts, and each artifact is a trace of the economic and institutional context that produced the standard.

The net neutrality debates of the 2010s were, at their core, debates about who controls the compression. If an internet service provider can prioritize certain traffic — certain compression standards, certain quality levels — it can shape what information reaches users with what fidelity. Compression is not a technical detail. It is a gatekeeping mechanism that operates at the level of bits rather than laws.

The emergence of generative artificial intelligence has introduced a new compression regime. A diffusion model trained on billions of images does not store the images. It stores a compressed statistical model — a lossy abstraction — from which new images can be generated. The artifacts of this compression are not blocky edges or banding gradients. They are the systematic biases, the homogenized aesthetics, the erasure of minority visual cultures that were underrepresented in the training data. The compression artifact of generative AI is not a visual distortion. It is a cultural distortion — the systematic privileging of statistically dominant patterns over statistically rare but culturally significant ones.

Compression artifacts are not technical failures. They are design signatures — visible evidence of the tradeoffs embedded in every lossy compression standard. To see the artifact is to see the algorithm. To see the algorithm is to see the power that designed it.