notes on Humans Commons
Update: Update
on CC Signals: What Changed and Why - Creative Commons But as we
began consulting with our community, it became clear that this approach
was not enough. The feedback was direct and consistent in stating that
preference signals without enforcement do not meaningfully shift power.
Signals alone cannot create agency in a system that many people did not
choose to participate in. That feedback forced us to confront some of
our own assumptions.
Previously: fair use alignment chart
It’s pretty clear that we need some kind of open content license that
allows for Creative Commons-style re-use by other writers or artists,
but also includes some kind of protection from machine learning
training. We’re not getting a noai
license from the original
Creative Commons organization, though. Instead, they’re offering a
separate signals
system with at least
one major loophole that would make it ineffective in practice. There
is no requirement in the license to preserve the
signal, so an AI firm could encourage someone else to mirror
your site without the signal, then crawl the mirror. Not a big obstacle
for an industry that’s already running at supervillain scale.
Humans Commons is
offering a set of licenses that help define how IP can be interacted
with by both Humans and non-Human entities, such as AI systems and
automated tools.
Their approach addresses some of the concerns that
are keeping me from using a Creative Commons license on this site. Right
now, what’s missing from Human Commons is a short list of practical web
nerd features, and an update to help protect authors from having their
content used for surveillance advertising. These are my notes on AI0-BY-SA,
which is the license that’s closest to what I would want for this
blog.
4.1.1 Can the attribution requirement be extended to
include a requirement to preserve rel=author
and rel=canonical
links if a full page of the work is used in a medium where it is
possible to preserve them? (SEO is a practical issue with CC-licensing
web content today)
4.5.2 Include a requirement to preserve any machine-readable AI opt-out signals present on the original site? (this would fix the CC license/CC signal loophole linked to above.)
9 Termination if the licensee sues the author or host for patent infringement? Borrow from open-source software licenses? (Not sure which license currently has the best patent clause, some research is needed.)
Human Commons is also an opportunity to get ahead of AI privacy threats. Right now there is a widely recognized client-to-server opt out preference signal, but we’re missing an opt-out preference that goes the other way, from server to client (see SPC: an X-Robots-Tag for privacy control).
While you’re taking steps to deny the use of your content for AI slop, it would also make sense to deny permission to use it for surveillance. A privacy section in the Human Commons license could include
Author withdraws consent to processing of any personal information contained in or inferred from this work
Author objects to processing of any personal information contained in or inferred from this work
Author opts out of sale and/or sharing of any personal information contained in or inferred from this work
Author opts out of the use of any personal information contained in or inferred from this work for behavioral and/or personalized advertising.
And yes, none of the current state privacy laws explicitly cover a
server to client OOPS but it should be possible to include you must
respect my OOPS
as a condition of the license.
Bonus links
Microsoft
tracks your Copilot AI usage in its workplace spyware by David
Gerard. The numbers track Copilot in Teams, Outlook, Word, Excel,
PowerPoint, and Copilot Chat. Microsoft will be adding numbers on total
meetings summarised, total hours summarised, and various classes of
prompts.
Three apps that
will not change your life by Justin Duke. [Helium is] a faster,
more performant version of Chrome, minus all the awful and annoying
stuff. (A fun fact: I discovered Helium by spelunking around for
alternatives after discovering that Chrome now ships a 10gb LLM model to
your hard drive and immediately crashes all of your local data if you
try to delete it.)
A
new breed of analyzers by Daniel Stenberg. You know we have seen
more than our fair share of slop reports sent to the curl project so it
seems only fair that I also write something about the state of AI when
we get to enjoy some positive aspects of this technology.