Stand up for your work. Cloak it from the bots.

Web scraping is almost as old as the web itself. In the mid-1990s, the first search engine crawlers politely walked from link to link, indexing pages so humans could find them. That original bargain was simple and mostly fair: a crawler read your page, sent visitors your way, and — if you asked it to stop through a robots.txt file — it usually listened. For nearly three decades that gentleman's agreement held. Bloggers, photographers, illustrators, and journalists published freely, understanding that their work would be read by both people and machines, but that the machines were largely there to help their audience discover them.

The rise of large language models and diffusion-based image generators broke that bargain. Beginning around 2020, industrial-scale scrapers began sweeping the open web not to index it, but to ingest it — pulling billions of images, essays, poems, forum posts, artist portfolios, and photo sets into training datasets used to build commercial AI products. Unlike a search crawler, a training crawler does not send visitors back to the source. It absorbs the work, learns from its style and substance, and enables downstream tools that can reproduce that style on demand, often without credit, license, or compensation to the original creator.

For most independent creators, the first sign that something had changed was personal. A novelist discovered their prose inside a chatbot's output. An illustrator watched a generator produce images that mimicked their signature palette. A photographer found their compositions echoed in AI-generated stock. The legal system is still catching up — class actions are pending in multiple jurisdictions, and licensing standards are being debated at every level from individual platforms to national copyright offices. In the meantime, a growing ecosystem of client-side, creator-first defenses has emerged. CloakMyWork belongs to that ecosystem: a small, honest, browser-only tool that lets any creator add friction to the ingestion pipeline in under a minute, without waiting for the courts or the platforms to solve the problem for them.

Adversarial noise is the technical name for a family of tiny, deliberate modifications that are invisible to a human observer but disruptive to a machine learning model. The core idea comes from a decade of academic research showing that image classifiers, text encoders, and generative models are surprisingly brittle: small, carefully chosen perturbations at the pixel or character level can meaningfully change how a model represents and learns from an input, without changing how a human perceives it.

For images, the technique CloakMyWork applies is a lightweight pixel-jitter pass performed entirely inside your browser's canvas. Every pixel receives a small, bounded random shift in its red, green, and blue channels — well below the threshold your eye can detect on a screen, but enough to nudge the numerical fingerprint that a vision model would extract during training. A scraper's automated pipeline that expects clean, statistically well-behaved pixels sees inputs that are subtly off-distribution. Over millions of training samples, this kind of noise raises the cost and lowers the quality of unauthorized ingestion. It is not a cryptographic lock, and no one should pretend otherwise — but it is a real, measurable speed bump.

For text, the equivalent trick is even simpler and, in some ways, more elegant. Modern language models tokenize text into small subword units before learning from it. By weaving zero-width Unicode characters between visible glyphs, we shatter the tokenizer's expected token boundaries: what looked like the word 'protect' to a human becomes a long, unfamiliar sequence to the model. Because zero-width characters are ignored by browsers, email clients, screen readers, and search engines, your human readers and your SEO ranking are unaffected. Only the pipelines that expect clean, well-tokenized prose have to work harder — which, at scale, is the entire point.

Neither transformation is a silver bullet. Determined, well-funded adversaries can preprocess your work to strip noise or normalize characters. But the vast majority of scraping today is opportunistic, off-the-shelf, and volume-driven. Making your work even modestly harder to ingest cleanly is often enough to route the scraper toward easier prey. That is the honest, practical value of adversarial noise, and it is exactly what CloakMyWork delivers.

If you write for a living or for love, the tension is familiar. You want your work to be read, shared, quoted by real people, and found through search — but you did not sign up to donate your voice to a training corpus that will, in turn, be sold back to the market as a synthetic replacement for writers like you. Text cloaking exists to preserve the first set of goals while quietly resisting the second.

Cloaking works best on the surfaces you fully control. Your personal blog, your Substack or Beehiiv newsletter, your portfolio site, the long-form 'About' pages on your author site, and the article body of guest posts you can edit — all of these are ideal candidates. Paste the finished draft into CloakMyWork, copy the cloaked output, and publish that version instead of the original. Readers will not notice a difference. Search engines will index your prose normally. Screen readers will read it aloud correctly. Only automated ingestion pipelines will feel the friction.

There are surfaces where cloaking is less useful. Anything you publish inside a platform that re-encodes or sanitizes text — some social networks, some CMS editors, some comment systems — may strip zero-width characters on the way in. In those cases the tool still causes no harm; the cloaking simply doesn't survive. A pragmatic workflow for a working writer is to keep two versions of any important piece: an un-cloaked master in your own drafts folder for editing and archiving, and a cloaked publication copy for the public web. Poems and short prose benefit most, because every character carries weight. Long-form journalism benefits too, especially for reported pieces that took real time and expense to produce.

One thoughtful practice: pair cloaking with a plainly worded license or usage note at the foot of your posts. A single sentence — 'This work is published for human readers; automated ingestion for model training is not permitted' — is not itself legally binding everywhere, but combined with a technical deterrent it strengthens both your ethical and, in some jurisdictions, your legal position. Cloaking is one layer. Clear terms are another. Robots.txt is a third. Together they add up to a real posture, not a performative one.

Photographers and illustrators face a slightly different threat model than writers. Your work is inherently visual, so it can be lifted with a right-click and dropped straight into an image-to-image pipeline that reproduces your composition, palette, or subject matter with no attribution. Watermarks help with attribution but rarely stop ingestion — a diffusion model happily learns from watermarked training data and simply generates outputs without the mark. Cloaking targets the layer beneath the watermark.

The recommended workflow for a visual artist is to run every export through CloakMyWork before it hits your portfolio, print shop preview, or social feed. Save your original master file locally, upload it to the image station, download the cloaked PNG, and use that as the public-facing version. The visible quality is preserved for clients and collectors reviewing your work; only the automated ingestion pipelines encounter the noise. For photographers publishing large galleries, this can be batched into your standard export routine — cloak once at the point of publication, then move on.

Two supporting habits multiply the effect. First, publish at reasonable but not maximal resolution: a 2048-pixel-wide portfolio image is more than enough for a human reviewer on a retina display, but denies scrapers the ultra-high-fidelity captures they prefer for training. Second, take advantage of the Portfolio Protection Badge available in the image station: a small, transparent PNG you can place at the foot of a portfolio page or gallery landing to signal, publicly and to any human editor reviewing your work, that you have taken deliberate steps to opt out of AI training. It is a small gesture, but a growing number of platforms and clients are starting to notice.

As with writing, cloaking is one layer of a healthy posture, not a replacement for the others. Combine it with a properly configured robots.txt on your portfolio domain, clear licensing language on your commissions page, and — where the medium allows — larger-scale defenses such as Glaze or Nightshade for illustrators specifically concerned about style mimicry. No single tool solves the problem. Used together, they turn your work from easy pickings into a genuinely inconvenient target, and that is very often enough.