Privacy by Editorial Team

EFF: AI Moderation Is Now Default, but Appeal Has Not Caught Up

EFF's two-part series argues automated content moderation is now permanent at platform scale, while transparency, human review, and appeal have not kept up.

Two people in conversation at a dimly lit table, a visual metaphor for the asymmetry EFF describes between platforms that automate speech decisions at scale and individual users left without a meaningful human appeal.

On July 10, 2026, the Electronic Frontier Foundation closed a two-part series arguing that automated content moderation is no longer a pandemic-era emergency measure that platforms keep promising to roll back. It is, in their words, the permanent default. That shift has not been paired with the transparency, human review, or appeal machinery that would make it tolerable, and the people paying the cost are the ones whose languages and identities the systems were never built for. For intelligibberish readers who track privacy and AI governance, the EFF’s framing is the cleanest public articulation so far of a problem that has been building for five years: speech decisions that used to require a human now happen in software, and the recourse paths have not caught up.

What the EFF actually argues

The series is co-authored by Jillian C. York and Corynne McSherry at EFF. Part 1, published on July 7, opens with a single observation: what was meant to be temporary has become the default mode of speech governance. The post quotes a 2025 joint statement from the United Nations, OSCE, OAS, and African Commission on Human and Peoples’ Rights that “the use of AI content moderation can lead to over-removal, discrimination and censorship.” That sentence is not from EFF - it is the position of four international human rights bodies, and EFF treats it as the settled baseline.

The historical thread runs from Meta’s 2018 push to automate violent-extremism review, through the Frances Haugen disclosures in 2021, to today’s default-on classification layers. Per the Part 1 post, Meta disclosed in 2020 that its automated systems incorrectly deleted nonviolent Arabic-language content 77 percent of the time. EFF picks this thread up in Part 2, arguing that “paths to remedy have all but collapsed.” GLAAD, quoted in Part 1, warns that “when moderation systems lack nuance, transparency, and human oversight, they can fail to curb harassment and wrongly suppress legitimate LGBTQ content.” The pattern is consistent: the systems scale, the failure modes scale with them, and the appeal paths do not.

What Part 2 adds

Part 2, published July 10, turns from diagnosis to prescription. EFF restates the Santa Clara Principles 2.0 - a set of due-process norms for content moderation that EFF, the Center for Democracy and Technology, and other civil-society groups drafted and have updated since 2018 - and reads them as a checklist for AI systems rather than a wish list. Part 2 lays out eight recommendations:

  1. Automate to assist humans, not replace them.
  2. Publish transparency reports that disclose how much moderation is automated.
  3. Run regular bias audits, especially for low-resource languages, vulnerable communities, and conflict zones.
  4. Route user appeals through human moderators.
  5. Publish human-rights impact assessments.
  6. Audit third-party vendors the platform pays to do moderation work.
  7. Lawmakers should not mandate automated moderation.
  8. Policymakers should not dictate platforms’ technical or design choices about expression.

The most consequential items for intelligibberish readers are numbers 2, 4, and 5. Transparency reports, in their current form, do not distinguish between automated and human takedowns at a level useful for audit. Human appeal of an automated decision is, in practice, a black-box model overriding a second black-box model. And human-rights impact assessments are not a requirement at any major US platform.

The Palestinian content case is the load-bearing example. EFF documents “systemic suppression” of Palestinian content across Meta properties in Part 2, citing Human Rights Watch’s “Meta’s Broken Promises” report (December 2023) and linking to a 7amleh (the Arab Center for Social Media Innovation) post on overzealous moderation in the region. EFF’s Part 1 separately cites the Center for Democracy and Technology’s recent series on content moderation in the Global South, which documents persistent inequities across four “low-resource” languages. EFF’s argument is that when the training data, the annotator pool, and the appeal reviewers are all concentrated in a handful of languages and jurisdictions, “AI moderation” is a euphemism for an Anglo-American classifier making speech decisions in places it has never been tested.

Why this matters for intelligibberish readers

Three threads connect this argument to the rest of intelligibberish’s coverage. First, the same platforms running these moderation stacks are the ones shipping AI assistants, image generators, and code agents into the rest of the stack. The pattern of “ship first, build the appeal path later” is the same pattern that drove the openclaw supply-chain failures in agent tooling and the audio-scam detection misses covered in our Savi write-up. The corporate incentive is the same in both places: get to default-on, get to scale, deal with the recourse gap later.

Second, the low-resource language problem EFF describes is the same problem in disguise for AI assistants. A moderation system that misclassifies Maghrebi Arabic at 77 percent error rates is built on the same labeled-data pipeline that produces a ChatGPT answer in Kiswahili. The bias shows up in the moderation log and in the assistant output, and the fix has the same shape in both places: native-speaker annotators, public audits, and a human-review path that the platform actually staffs.

Third, the EFF’s framing lines up with what users already experience. The Meta takedown numbers, the Discord image-classifier false positives reported by TechCrunch on July 7, and the GitLost disclosures from the same week are not separate incidents. They are the same incident, told in three different product surfaces. The EFF series is the attempt to give that pattern a name and a policy framework before the platforms ship the next layer of defaults.

The Bottom Line

EFF’s argument, in two posts, is that automated moderation has moved from emergency measure to permanent default without the transparency, bias audit, human review, or meaningful appeal that would make it defensible. The fix is not a single rule; it is a checklist of eight norms drawn from the Santa Clara Principles 2.0, and the platforms shipping these systems have not signed up to any of them. For readers who watch AI governance, the two-part series is the clearest public case yet that the appeal gap - not the moderation error rate - is the structural problem.