Wall Sreet Journal: AI art generators Dave Touretzky (23 Oct 2022 19:16 EDT)
Re: [AI4K12] Wall Sreet Journal: AI art generators Finlay.McCall (23 Oct 2022 21:01 EDT)
Re: [AI4K12] Wall Sreet Journal: AI art generators Pat Langley (24 Oct 2022 01:52 EDT)
Re: [AI4K12] Wall Sreet Journal: AI art generators Ken Kahn (24 Oct 2022 02:43 EDT)
Re: [AI4K12] Wall Sreet Journal: AI art generators Sabrina Hsueh (24 Oct 2022 08:18 EDT)
Re: [AI4K12] Wall Sreet Journal: AI art generators Dave Touretzky (25 Oct 2022 07:25 EDT)
Re: [AI4K12] Wall Sreet Journal: AI art generators Ilkka Tuomi (25 Oct 2022 11:37 EDT)
Re: [AI4K12] Wall Sreet Journal: AI art generators Ken Kahn (25 Oct 2022 21:32 EDT)
Re: [AI4K12] Wall Sreet Journal: AI art generators Finlay.McCall (25 Oct 2022 23:03 EDT)
Re: [AI4K12] Wall Sreet Journal: AI art generators Ken Kahn (25 Oct 2022 21:37 EDT)
Re: [AI4K12] Wall Sreet Journal: AI art generators Pat Langley (24 Oct 2022 15:44 EDT)

Re: [AI4K12] Wall Sreet Journal: AI art generators Dave Touretzky 25 Oct 2022 04:24 PDT

Sabrina Hsueh wrote:

>   One question that has been bothering us about using the AI Art
>   generator for education for kids is how we can avoid the generation of
>   age-inappropriate artwork. The use case is for younger kids. Therefore,
>   it is preferable to be able to censor words that involve violence,
>   etc.  Does anyone have a solution for enabling censoring by age for
>   these art generators?

This is a big problem for K-12, for both text and art.  These models are
trained on huge corpora.  It's not practical to manually curate the
training data, although one can do some basic filtering to eliminate,
say, known pornography sites.  But a lot of other stuff is going to get
through, e.g., famous paintings or sculptures that show nudity, or
medical/health education images.  Plus, our ideas about what's
acceptable for children depend on their ages.

Google talks about "safety" of its generative AI systems, and that
includes not generating references to sexuality or intimate body parts,
violent acts, racist or abusive language, or politically sensitive
topics.

In a recent paper on their LaMDA dialog generator model
(https://arxiv.org/abs/2201.08239) the Google researchers explain how
they try to achieve this safety goal.  It ain't cheap.  They paid humans
to rate the outputs of their system and flag anything that violated the
safety rules.  Then they used that data to fine-tune their model to
discourage it from generating such outputs.  They were also able to
train a model to automatically detect when something violated the safety
rules, so that the system could learn to self-censor.  It didn't work
perfectly, but it was pretty good.

We ran into a similar problem with the "Talk to Transformer" demo at
https://app.inferkit.com/demo.  You can prompt it with something a little
raunchy or dark and it will happily generate text that you would not want
your middle school students to see.

There must surely be a market for kid-safe text and art generators.  But
training them on a carefully curated dataset is going to be expensive.

-- Dave Touretzky