Two Boston University law professors have proposed a three-part legal definition for "AI slop" that could reshape how courts, regulators, and law firms draw the line between acceptable automation and output that shifts unacceptable costs onto its recipients. The draft paper, posted to SSRN on Sept. 6, 2026, arrives as the European Union's AI Act transparency obligations and California's AI Transparency Act both became operative on Aug. 2, 2026, making the definitional question immediately relevant for compliance.
Jessica Silbey and Woodrow Hartzog define AI slop as "any output of a generative probabilistic automated system produced with little exertion that asymmetrically burdens recipients and tends to degrade cultural domains." The definition is deliberately narrower than anything a machine touched, and that narrowing is the point. It creates room for useful automation while giving policymakers a framework to target output that externalizes costs onto people who did not ask for it.
Three dials, not three switches
The test has three components, and the authors treat them as dials rather than switches. Negligible exertion asks how much thought, revision and judgment went into the output. Asymmetrical imposition asks who absorbs the shortfall - the recipient, who reads twice, fills gaps and corrects errors the sender declined to catch. Domain degradation asks whether the output, at scale and over time, corrodes the practices of the field it lands in.
Output is "more or less 'sloppy,'" the authors write, depending on how much of each component is present. That spectrum is the part practitioners should sit with, because compliance frameworks want a binary and this one refuses to supply one. The framing draws on developer Simon Willison's May 2024 observation that "if it's mindlessly generated and thrust upon someone who didn't ask for it, slop is the perfect term for it." Silbey and Hartzog keep the emphasis on imposition and drop Merriam-Webster's focus on quality - a deliberate choice that lets clumsy first drafts pass while catching the polished report nobody stands behind.
What is already on the books
Governments have mostly reached for disclosure. Article 50 of the EU AI Act, which began applying Aug. 2, 2026, requires providers of systems that interact directly with people to inform them they are interacting with an AI system. Providers of systems generating synthetic audio, image, video or text must mark outputs in machine-readable format and make them detectable as artificially generated. Deployers publishing AI-generated text to inform the public on matters of public interest must disclose that fact, unless a human has reviewed the text and holds editorial responsibility. Penalties reach 15 million euros or 3 percent of total worldwide annual turnover, whichever is higher - though for SMEs and start-ups, the comparison reverses to whichever is lower.
California's AI Transparency Act, SB 942, became operative the same day. It covers generative systems with over 1 million monthly visitors or users that are publicly accessible in California. For image, video and audio content, it requires a free detection tool, optional visible disclosure, and latent disclosure conveying the provider's name, system name and version, timestamp, and a unique identifier. Text sits outside that scope - a direct difference from Article 50, where synthetic text is expressly covered. Civil penalties run $5,000 per violation, with each day counting as a discrete violation.
For legal professionals navigating these requirements, understanding the operational distinctions between jurisdictions has become a compliance necessity. The AI for Legal landscape now includes overlapping transparency duties that differ on whether text output is covered, what counts as human review, and which carve-outs apply to editing workflows.
The enforcement problem nobody is solving
Detection tools carry error rates that make them hazardous at scale. MIT Sloan Teaching & Learning Technologies publishes guidance flatly titled "AI Detectors Don't Work. Here's What to Do Instead." OpenAI withdrew its own detector for poor accuracy. Even a 98 percent accurate tool misclassifies 20,000 documents across a million - and the people behind those documents are the ones who must answer for it.
Silbey and Hartzog warn that bolting a policing layer onto institutions that ran on presumed sincerity compounds the damage rather than repairing it. That warning should land with anyone weighing detection tooling for internal investigations or review quality control. The paper also traces what happened when one security team refused to carry the cost any longer: the curl project ended its bug bounty on Jan. 31, 2026, after AI-generated reports pushed the share of confirmed vulnerabilities below 5 percent. Reporting moved to a private channel with no reward attached.
The paper's gap is telling. Courts get almost no treatment across 70 pages, despite the fact that Damien Charlotin's AI Hallucination Cases database listed 2,022 decisions as of Sept. 7, 2026, with close to 68 percent sitting in U.S. courts. The Administrative Office of the U.S. Courts broadcast interim AI guidance on July 31, 2025, recommending - not requiring - that users review and independently verify AI-generated output. For what arrives from counsel, the judiciary points to Rule 11 of the Federal Rules of Civil Procedure and Canon 3B(6) of the Code of Conduct for United States Judges. Rule 11 does not appear anywhere in the paper.
Why this matters for legal professionals
Organizations treating AI acceptable-use as an HR memo rather than a records policy are storing the problem instead of solving it. The plausible-looking report nobody will vouch for still enters the record, still gets retained, still gets collected, and still must be defended by someone who did not write it. For paralegals and legal operations teams managing document review workflows, the question is not whether a tool was involved but whether the sender kept the judgment that makes output worth another person's time - or pushed that work downstream. The AI Learning Path for Paralegals addresses these distinctions directly, focusing on when AI-assisted work product requires human verification and when automation can appropriately stand on its own.
Silbey and Hartzog close where policy arguments usually open, arguing the window for treating AI slop as nascent is closing. Their bet is that naming the thing precisely is what makes it governable. For any legal team, the operational question follows directly: if someone asked tomorrow how much of your AI-assisted work product a named human is prepared to stand behind, could anyone answer?
Your membership also unlocks:
