The viral AI avatar app Lensa undressed me – without my consent | Hacker News Reader

The viral AI avatar app Lensa undressed me – without my consent

technologyreview.com

19 pointsabsqueued13 comments

A big part of AI is weaponising our biases, it's not the first time, won't be the last.

Feed it shit and it'll give you shit, too bad the training set (internet) is already and open air dump.

We'll be in a very bad downward spiral really soon when AI trains on AI generated content

AI of our era so far is working solely based on a trained model and haven't evolved beyond the "data" behind training. All the material fed to train the AI machine leads to this mess.

I heard Sam (OpenAI founder) talk about current state of AI, we are far from AI making its own concious decisions.

I don't really know how AI works. Is it possible for AI to train itself on its own generated content? Wouldn't this be like some kind of lossy recursive reproduction, like an audio tape loop continually degrading itself?

Yes, it can, and it could be degrading on itself. Let's use an example of AI art. A lot of AI art is pretty terrible. But if it was trained on data containing AI art that was 1. the description of an image 2. and the image itself, and 3. the image turned out how exactly how the user wanted, then this could, in some cases, actually produce decent training data, particularly with weird, super specific data. But again, it would have to be exactly what was wanted from the description (which could be the original prompt), otherwise it would be bad training data. Really it's just a matter of how well the photo matches its description.

Disclaimer: I don't know that much about AI, take this with a big grain of salt. Maybe some other people can chime in here.

If the data is sparse, some researches used AI generated content to train another network. E.G. https://arxiv.org/pdf/1802.06739.pdf

In some cases like here it also has privacy issues.

3 different banners/popups showed up at the same time and blocked me from reading the content:

* Cookie policy consent banner

* "Subscribe now" banner (before I even got to read any of the content)

* "This is your first complimentary story" banner

I never heard of "the viral AI avatar app Lensa" until I saw this article being promoted. I honestly wonder if this isn't a sick marketing attempt.

And I don't mean one the author is necessarily in on - manufacturing threats or abusive material of someone anonymously, then contacting the victim pretending to be concerned about it, is a tactic I've seen trolls employ.

It was a genuine viral moment. I moderate on an AI related Facebook group and the release of Lensa created a spike in our amount of users of activity that lasted several weeks and was bigger than almost all of the activity for the year. A lot of people now have those avatars instead of those regular profile pictures even outside AI related spaces.

It's because it doesn't have a free tier. So barely anyone cares it exists.

nah. it blew up last year when people were using it for ai profile pics

Yup, I actually got it on a newsletter, intesting take though about the marketing stunt. I trust the MIT Tech review and the HN.

There is likely a lot of anime stuff in LAION-5B, as well as photos of Asian women who consented to sexualized photoshoots.

I think this tells us something about the current aesthetics of the internet, Asian women are idealized to some degree, just look at a random cyberpunk style picture.

But it's also really easy to ask Stable Diffusion to generate non-sexualized images by adding what you don't want to see to the "negative prompt".

So it's on the app makers to make their app less NSFW, I don't buy the company representative responses

> I think this tells us something about the current aesthetics of the internet, Asian women are idealized to some degree.

Fetishising Asian women predates the internet[1]. Seeking Asian Female[2] is a documentary on the matter, by an Asian woman.

[1]: https://en.wikipedia.org/wiki/Asian_fetish

[2]: https://www.imdb.com/title/tt2320029/

Reply on news.ycombinator.com