There's a little bit of praxis that I do for these (when I can't just bounce from the site or service requesting it): purposefully get about 25% of it wrong, misidentifying things that aren't traffic lights as traffic lights, or failing to identify all traffic lights, for example. I'll sometimes have to go through a few rounds of captcha but I don't mind that as much when every one is an opportunity to have a tiny corrupting effect on their dataset.
I was fine with Google's captchas when they were designed to help translate scanned printed material into searchable text (providing a social good). I'm not fine with providing free labor for their self driving car project.
I’m confused about your reasoning. Translating scanned printed material into a searchable text is a social good, but providing safer transportation…that additionally may one day free drivers of the menial work of driving…that is not a social good?
Which will be available any day now, probably as early as 2018 if experts on the field can be believed. Some would even go so far as state that by 2025 private car ownership will be a thing of the past.
> that is not a social good?
Hype, empty promises and constantly moving deadlines. If this goes on we will have commercial fusion reactors before self driving has any positive impact outside of heavily supervised trials.
The problem, of course, is that the dataset is exclusive to Google, so while it may benefit society it's not creating a public good. But that's also true for text OCR.
I agree with you on this, and I'm glad you mention it. I think it might be a tangential topic to the relative positive impact to text-recognition / road-condition-recognition, though.
I was fine with Google's captchas when they were designed to help translate scanned printed material into searchable text (providing a social good). I'm not fine with providing free labor for their self driving car project.