On May 19, 2026, Google refreshed its Workspace suite icons, moving away from uniform four-color geometric shapes to product-specific gradient palettes and distinct silhouettes. Mainstream design commentary notes two core reasons for the change: it fixes the human usability issue of identical-looking icons blending together in browser tabs, and the softer, fluid gradients serve as a branding signal for Google's shift into the Gemini AI era.This raised a practical question: beyond the visual branding, did Google's redesign also make the icons more distinct for machine vision?Measuring object separability in embedding spaces is standard practice in robotics and physical AI. To see how it applies to these new desktop UI assets, I ran a quick script to measure pairwise cosine distances across the old and new icon sets using three open-source vision encoders.The short answer: Yes, solving visual ambiguity for users fixes the problem for machine vision, too.The SetupI gathered the pre-redesign and post-redesign icons for 8 core Workspace apps (Gmail, Drive, Docs, Sheets, Slides, Calendar, Forms, Tasks) and evaluated them across three different architectures:OpenCLIP ViT-L/14 (The baseline text-aligned encoder)SigLIP-SO400M (Google’s strongest public text-aligned encoder)DINOv2 ViT-L/14 (Self-supervised, looking purely at structure and texture without language grounding)To reduce background noise, all icons were composited onto a neutral gray background before encoding. This avoids cases where transparency defaults (white/black) influence embeddings more than the icon itself.The evaluation is straightforward: compute L2-normalized embeddings, derive cosine distances, and analyze pairwise separation across all icon pairs.import numpy as npimport torchimport torch.nn.functional as Fdef evaluate_encoder_separation(model, processor, images, device="cuda"): inputs = processor(images=images, return_tensors="pt").to(device) with torch.no_grad(): embeddings = model.get_image_features(**inputs) embeddings = F.normalize(embeddings, p=2, dim=-1) sim_matrix = embeddings @ embeddings.T sim_matrix = torch.clamp(sim_matrix, -1.0, 1.0) dist_matrix = 1.0 - sim_matrix idx = torch.triu_indices(dist_matrix.size(0), dist_matrix.size(1), offset=1) return dist_matrix[idx[0], idx[1]].cpu().numpy()Separation ResultsAcross the board, the new icons pulled further apart in embedding space.| Encoder | Old Mean Distance | New Mean Distance | Distance Shift (Δ) | Cliff's delta (Effect Size) ||----|----|----|----|----|| OpenCLIP | 0.341 | 0.453 | +0.112 | +0.59 || SigLIP | 0.240 | 0.309 | +0.069 | +0.52 || DINOv2 | 0.394 | 0.481 | +0.087 | +0.30 |The most notable difference appears within the document editors (Docs, Sheets, Slides, Forms). Under the old design, these were constrained to identical four-color palettes and highly similar geometric boxes, causing them to sit close together in the text-aligned models' spaces.Look at the jump in DINOv2 pairwise distance after the redesign:Docs ↔ Sheets: 0.128 → 0.287 (+123%)Docs ↔ Forms: 0.098 → 0.515 (+425%)Docs ↔ Slides: 0.188 → 0.326 (+73%)\The DINOv2 heatmaps show this shift on a purely structural level. In the old matrix, tight, cooler zones reflect how closely the model grouped the icons when they shared identical geometric containers. In the new matrix, the plot warms up across the board, indicating better pairwise separation. Because DINOv2 is entirely self-supervised and has no language grounding, this visual shift confirms that the redesign didn't just change semantic context—it altered fundamental textures and silhouettes, giving the model cleaner geometric boundaries to work with.TakeawayA sample size of 8 icons is obviously tiny, and production systems powering tools like Claude's computer use or GPT-4V use proprietary vision heads. However, open-source models are usually highly directional for these types of checks.While embedding separability is a foundational concept in physical AI, it is highly applicable to desktop software as agents become more visual and DOM-free. Checking the embedding separation of your own core UI assets is a cheap, quantitative way to ensure an autonomous agent can cleanly distinguish your product components. In this case, optimizing for human readability naturally optimized for the model.\n \