[Foundation-l] Image filtering without undermining the category system

Andrew Crawford acrawford at laetabilis.com
Fri Oct 14 04:24:29 UTC 2011

> It need be no more complex than
> http://en.wikipedia.org/wiki/Special:Preferences#mw-prefsection-watchlist

In my opinion, it will need to be far simpler than that if it is to address
the needs of the casual reader. We simply can't expect them to make any
significant investment in understanding the process, unlike a casual editor
who can be expected to make that investment. It needs to "just
work". Primarily this is "just" an issue of interface design but that has
not been a historical strong point for us ;)

> Subject to gaming, well it's bound to be. But vulnerable to gaming,
> hopefully not.  Fans of penises are welcome to add their preferences.
> That's
> why I didn't include the option "Hide all images except those that a fellow
> filterer has whitelisted".

Well, to be fair, you did initially include that option - but even without
it, the system will be gameable as long as "Hide all new images unless they
have been OK'd by a fellow filterer with similar preferences to me" exists
as an option. Yet, that is the most powerful and potentially most popular

I had typed out a long description of precisely how this might be gamed
here, but I started boring even myself, so let me just state the principle:
if you create enough accounts that have preferences matching a particular
statistical user cluster (easy to do by blocking all images of Mohammed),
you can exercise disproportionate statistical control over which new images
those real users see by having your army of ringers deliberately green-light
new images.

 Well of course any of the editors could themselves have the filter set on
> and would know what the score was relative to their preferences. But
> otherwise the information would be secret. I don't see how we could give
> editors access to the reputation information without it leaking to censors,
> or indeed divulging it generally.

Although the per-user score might be the important one for the operation of
the algorithm, it's obvious that per-image data will exist, or at least
could be calculated. We *could* calculate statistics for each image like
"likelihood of actually being seen by a viewer" without revealing any
personal information. It would also be possible to show "likelihood of being
seen by readers who have blocked image X" for image Y.

This information *could* be useful to censors, but if it was available only
to logged-in users it would be practically difficult for them to access in
an automated manner.

> As for "legitimate editorial reasons", I think it would be quite
> contentious if anyone started making editorial decisions based on the
> filter
> results, so best not to enable that - but I'll clarify that in the proposal

I certainly don't think it would be more contentious than having the filter
in the first place! There are certainly legitimate editorial reasons for
wanting to know some of the information I mentioned. I personally believe
that images are content, that they hold information, and that they are not
mere decoration in an article. There are plenty of situations where the
question "what percentage of readers of this page will not be shown this
image" would inform the decision about whether to use the image at all, and
inform decisions about the wording of the article copy.

Anyway, as I have said, I think your basic idea here is the only practical
option if a filter is to be implemented. I am personally not in favour of
doing so, for both idealistic and practical reasons (specifically, the
amount of effort it will involve vs. the amount of benefit it will deliver,
with special consideration for the fact that almost no one seems to have
asked for an opt-in filter, but rather for Wikipedia just not to have those
images in the first place).

Be all that as it may, if we are to have a filter, we need to have a working
filter. We need to have the best filter we possibly can, with the least
increase in workload for editors and the most functionality for readers. In
my opinion the system you have outlined is not only the best option so far,
but also the best option possible. I do encourage others to read your design
(http://meta.wikimedia.org/wiki/User:WereSpielChequers/filter) and to get
involved in figuring out how it might work.


Andrew (Thparkth)

More information about the foundation-l mailing list