The DMCA takedown has become a censorship tool, and the burden of proof sits on you

The Digital Millennium Copyright Act was written in 1998 to do two things at once: give creators a fast route against piracy, and give platforms a safe harbour from being sued over what appears on them. Section 512 delivers the second by conditioning it on the first. A search engine or a host keeps its protection so long as it removes flagged material quickly.

Quickly is the word doing the damage. Removal requests now arrive in the millions every month, which means software processes them rather than people, which means a page can fall out of an index without anyone having read the claim against it. A mechanism built to defend intellectual property has become a serviceable instrument for burying inconvenient reporting.

The clearest recent demonstration involved the search industry's own press. Press Gazette published an investigation into Clickout Media, a digital media firm it reported had been buying up established news domains and filling them with AI-written gambling content — the practice known in search circles as site reputation abuse. Search Engine Land followed with its own report on the findings.

Inside two days, both articles were gone from Google worldwide.

The complaint came from something calling itself US Webspam, and alleged that Search Engine Land had reproduced its material word for word, images included. Three things were wrong with that. The article carried no images at all, so there was nothing to have taken. The filing asserted that repeated good-faith attempts had been made to settle the matter directly, and no such message ever reached the editorial team. And the domains it named as the original source of the reporting had nothing to do with it.

None of which stopped the request going through. Google's automated pipeline removed the article from the index. Getting it back took a formal counter-notice, sustained pushback from analysts in the industry, and enough press attention to make the case impossible to ignore — several days, for a publication with the reach to mount that kind of response. A niche blog or a small business has none of those levers.

The reason bad-faith notices work is not a bug in anyone's software. It is the incentive the statute creates. A platform's safe harbour survives only if it acts, in the statute's phrase, "expeditiously to remove, or disable access to" the material once a claim lands. Taking a page down carries essentially no legal exposure; leaving it up carries some. Verification is slow, and sits on the wrong side of that trade.

The deeper problem is where the burden falls. Section 512 asks the accused to establish innocence rather than the claimant to establish ownership. The page comes down first and the argument happens afterwards.

Three patterns of abuse follow. In the first, an attacker copies a genuine article onto a throwaway site, backdates the timestamp so their version appears to have come first, and files a claim against the original. In the second, the notice goes in under an invented organisation or in the name of an unrelated blog, leaving nobody to pursue. In the third, a business facing bad press aims claims at investigative work, unflattering reviews or customer complaints, timed to the window when that coverage would hurt most.

If it happens to you, the first move is to read the actual notice rather than guess at it. Google reports the removal in Search Console under Security and Manual Actions and sends an email alongside it, and the filing itself is usually archived in the Lumen Database, a public repository of legal requests to remove web content. What matters there is who is named as the owner, which URL they claim is the original, and precisely what text or media they say was copied.

The second is evidence that your version came first, assembled before anything is filed. Publish logs exported from your CMS carry creation dates and edit history. The Internet Archive can show your page live on a date preceding whatever the claimant asserts. RSS records timestamp distribution to the second. If the site is built from version control, commit history dates the file itself. This is also the argument for keeping raw footage, unedited photographs, audio and drafts somewhere off the production server: they document a process a scraper cannot manufacture.

Third comes the counter-notice, which is a formal document with required contents. It must identify what was removed and where it lived, carry a statement made under penalty of perjury that you believe in good faith the removal was a mistake or a misidentification, and give your name, physical address, telephone number, email and signature. It must also consent to the jurisdiction of the relevant federal district court — the one covering you, or the one covering your host if you are outside the United States.

Then the clock starts. The platform is required to pass your counter-notice to the claimant, who then has ten to fourteen business days to show they have actually gone to court for an order restraining you from publishing. If nothing arrives inside that window, the platform is obliged to put the content back.

Two weeks is survivable for an evergreen page and ruinous for time-sensitive reporting or a launch campaign. Where a claim is transparently fraudulent — image theft alleged against a page containing no images — publicising the details and tagging the search platforms' own representatives has been known to move a manual review up the queue. And Section 512(f) does provide a remedy: anyone who knowingly misrepresents that material is infringing is liable for damages, including the costs and legal fees of the person they falsely accused. Litigation is slow, but a formal litigation hold letter sent to a claimant who used a real legal identity has a way of producing a quiet withdrawal.

Most of the useful defence happens before any of this. Archive each article to the Internet Archive automatically at publication, through a webhook or the API. Carry proper structured data on every page — Article or NewsArticle markup with datePublished, dateModified and an author — so the record of who published what, and when, is machine-readable. Route Search Console alerts to an inbox somebody opens daily, because a missed notification eats directly into the window for responding.

The uncomfortable conclusion is structural rather than technical. Automated copyright enforcement was built to protect very large platforms from liability, and at that job it works. It was never built to determine who wrote a piece of text, and it does not. Until filing a notice requires verified identity, or a false one carries a penalty worth fearing, the asymmetry holds: making a page disappear is cheap and instant, and getting it back is neither.