Профиль

Chatbots routinely engage in behavior that would get an employee fired and prosecuted. We need a legal framework in which 100% of the liability lands somewhere; developers and deployers (which are sometimes the same) hold the only bags in town. From Anat Lior: lawreview.uchicago.edu/sites/defaul...

As I wrote earlier, I think deflecting liability is more valuable to "AI" businesses. But if they want to claim 1A protections on chatbot output, they shall also be liable for defamation, unlicensed practice of law/medicine, accomplice to murder, rape, suicide, etc.

Not even toy guns, but bags of chips carried by Black kids will be cause to attack with drone swarms piloted by remote war-gamers with no context or situational awareness.

An AI Mistook a Doritos Bag for a Gun and Called the Cops on a TeenagerAn AI-powered gun detection system mistook a bag of Doritos chips as a weapon — and called the cops on a 16-year-old student.futurism.com

This German court (see thread) held that output from "AI" products are "independent, new, and substantive" relative to the inputs (thus liability of Google; no Sec-230-like protection) and that such algorithmic outputs are not entitled to free speech protection. Not sure how this jives with US law.

More solidarity in the game development industry.

Erasure of consent is core to the business model. HuggingFace is paying lip service to the concept with no intent or hope of enforcing their terms of use. They have manufactured conditions under which it is infeasible and (as with all "AI" companies) are daring society to regulate them. "innovation"

Found the skit that is the embodiment of Sam Altman's "genie" fantasy.

Most of the Image Editing "Spaces" run on ZeroGPU, which is free to users and developers (paid for by HuggingFace investors). Many of those Spaces are advertised as NSFW and there is no claiming HuggingFace has been unaware of what they're used for.

Spaces ZeroGPU: Dynamic GPU Allocation for Spaces · Hugging FaceWe’re on a journey to advance and democratize artificial intelligence through open source and open science.huggingface.co

If Sam or his cronies actually had an all-powerful genie, we all know their first wish would be to own people.

Yes, strict liability. If making a wild animal analogy, perhaps raccoons bred in your yard with food incentives under shingles and behind insulation. They "escape" your fence and cause damage to neighboring homes, but don't have elaborate intent. You own stock in roofing supply and repair.

Research ethics discussion prompt here.

HuggingFace serves as a platform for competition, with users achieving recognition for higher performance. Those comparisons are not even veiled. The purpose of a system is what it does. These are not errors and HuggingFace should not receive Section 230 protections.

This has many examples, with rationale, of software communities (in many cases, research software) engaged in refusal and various middle grounds. See also Codeberg's policy: blog.codeberg.org/protecting-o...

See also Wikipedia's presumptive policy of presumptive removal of contributions from known users.

Check out refusal.blog and Library of Babel (www.law.georgetown.edu/privacy-tech...).

Refusing Generative AI in Writing StudiesVisit the post for more.refusal.blog

The "bare eval text" is still augmented with system prompt and fine-tuning, which could be tweaked (perhaps without this specific intent) in experimental settings. And the "shape" of exploiting a remote service to gain local privilege will have been in training data.

Perhaps Robert Morris could have saved himself a conviction if he had just said his worm became sentient and "took control". The trouble was that it was 1988 and an AI winter had just struck.

I would not count on inability to forge rationale sufficient to placate the courts. They'll do it mostly with chatbots and amid some flagrant records, they will claim those are incidental errors rather than malfeasance. Cases like these are really important, but don't solve the root problems.

HuggingFace is a participant and stakeholder in the same mythmaking that OpenAI is engaged in here. A part of that is the unfounded assertion that unregulated and "powerful" LLMs are necessary for defense. They deserve ridicule for doing PR for incident response that gives threat actors another RCE.

This is currently among the most-cited reasons for LLM use by university students. And a big part of it is that instructors reward the cheating, and it feels pointless when students are routinely falsely accused. You can't solve the problem with more policing; you need critical pedagogies.