Is this why we are solving motorcycle, stairs, fire hydrant, etc. captchas?
To help blind drivers, no. To help AI, yes.
I am missing a small amount of context - is reddit randomly prompting users to describe images in posts? Or is it prompting you to describe your own image at upload time?
Context aside, I definitely think that providing image descriptions is something we should do in spite of the fact that its definitely going to be used to train AI. Choosing to not do so is throwing our blind peers under the bus to reduce the amount of training data for ai fractionally.
I haven’t been there in a while but I remember there was a sub of volunteers that were around for years that went around just describing images, way before AI LLM were really a thing.
I’m assuming this is something new being pushed by reddit itself, but as you said, it’s a good thing regardless.
As long as, even if reddit is using it to train LLM, they are actually still using the descriptions to add accessibility to those images, which I don’t take for granted
At this point, any request for information could potentially be used as training data. That includes things like captchas.
I recommend everyone have an extremely literal interpretation of “labor”. Unless you have tremendous insight into where your data is going and how it is being used (and perhaps even then), then assume any ask is ultimately an ask for unpaid labor.
Obviously you can’t avoid things like captchas, but you can avoid things like this.
Edit: and it should go without saying, but anything you upload to socials is probably automatic training data at this point. The best approach is simply not to engage with corporate social networks.
Though Lemmy is not corporately controlled, the information is publically accessible, so even this post is potential training data to be scraped. That is harder to avoid, lest we stop using the internet altogether, but at least avoiding the corpo routes is a good start.
Captchas have been for training ai for years that’s nothing new. Iirc the reason you do two is one to confirm you’re human, one for training data
I use an app called Be My Eyes to help the visually impaired.
You’ll get a random notification that a person needs your help. If you’re the first to respond, you’ll be paired up. Their phone camera is displayed on your screen, and you can talk to each other.
I always have a great experience when I use it.
Just fyi, the vast majority of people don’t use Apple products
https://play.google.com/store/apps/details?id=com.bemyeyes.bemyeyes
iOS is the majority in the US, sadly.
Globally it’s about 12% though
I dont think the majority of lemmy users are from the US though
They’re not winning over the blind again after they limited access to screen-reader-accessible apps.
On the flip side, training ai for image recognition has the potential for auto labelling images for the blind
Could be either the website owners themselves generate them if a human written one isn’t provided, or a browser extension that auto labels any unlabelled images on the screen
They’re probably going to make a deal with Google to improve Google Lens. Yes, it will eventually help the blind but Reddit’s shareholders will be getting even richer from people’s donated time.
I promise that I’m leaving this here in good humor, although it is also true.
OP did actually use it correctly tbh, a rare case
kinda. Text on the right is supposed to be the same both times.
It’s supposed to be similar, not necessarily the same. The idea is something like this
- Anakin: something
- Padme: but [normal thing] right
- Anakin: …
- Padme: but [more basic thing] right??
- (implied that not even that)
The implication here is that not only is the primary purpose not for the blind, but it won’t help them at all
(am I overthinking this?)
Context, please?
AI steali…I mean harvesting, pardon learning
I’m sure blind people are happy to have the models that are built with this data, and since both the image and the description are public facing, anyone can use them including open source.
Except you need to pay Reddit to use their data for training
Helping self-driving cars and drones.
