Google Scrambles After Software IDs Photo of Two Black People as “Gorillas” | Hacker News Reader

Google Scrambles After Software IDs Photo of Two Black People as “Gorillas”

slate.com

21 pointslingben15 comments

> Ultimately, Google applied a secondary solution, reworking the system so that it wouldn’t tie photos to the “Gorilla” tag at all. Zunger writes that it is also working to develop a number of “longer-term fixes,” including identifying “words to be careful about in photos of people” and “better recognition of dark skinned faces.”

I hope photos of gorillas are not now marked as "black people" - since they've blocked the gorilla tag being used.

> Ultimately, Google applied a secondary solution, reworking the system so that it wouldn’t tie photos to the “Gorilla” tag at all

Looks like dirty fix.

"Like I understand HOW this happens; the problem is moreso on the WHY"

Eh, what?

They apparently are clever enough to understand the identification algorithm, but not the social implications.

This is the problem with automatic image processing and categorization. It's good, but mistakes like this can happen and embarrass an entire company.

Yeah, shit happens sometimes.

Especially for AI algos.

So assuming this was a mistaken classification that happened naturally and not the result of someone purposefully training the system to do this... does this mean we need to teach these systems to be politically correct? Confusing a dog and a goat doesn't carry the same potential to cause extreme negative publicity that an error like this does, especially when the system works well enough that most other classification attempts are correct.

We are affected by this mistake because we are socialised to be offended and outraged by labeling we perceive as offensive. We do this to substitute an impossible external enforcer (of common norms) with an internal, distributed one.

It is tough, emotional work to understand that there is no need to punish an algorithm. I guess it takes time and additional norms (e.g. devs saying they are sorry, etc..) to adapt to this new existence of semiautomatic affronts.

I think a possible root to the problem is the general lack of diversity in tech. A bug that confuses white people for goats will be caught sooner than a bug that mistakes black people for gorillas because there are more white people around to notice that problem. It wouldn't surprise me if the initial test data included few to no non-white people.

I'm curious what white people are mistaken for when you have close ups of their faces. Potatoes?

Great algorithm accuracy: 98,25% of DNA shared between gorillas and humans. I could forgiven the error easily is someone tagged me as gorilla.

Modify tags with positive adjectives that work both ways like "Strong as a gorilla" or "Happy as a chimp" and the 90% of the current problems will dissapear.

If I called somebody a gorilla out of anger, I might end up apologizing to the gorilla.

Gorillas are much better behaved than primates in general so I wouldn't be insulted if somebody called me a gorilla.

> Gorillas are much better behaved than primates in general so I wouldn't be insulted if somebody called me a gorilla.

Then again, you likely don't belong to a race that was subjugated for centuries by people comparing them to monkeys or gorillas to intentionally dehumanize them.. Black athletes in foreign countries still get bananas thrown at them on the field. I can assure you, this isn't meant as a compliment.

    Eberhardt described an experiment based on the
    Rodney King case. Participants viewed a video of
    police officers beating a Black suspect. Participants
    who saw words associated with apes before
    watching the video believed the officers’ actions 
    were more justified, compared to participants who 
    did not see the animal imagery. In a separate study, 
    Eberhardt found that people are more likely to 
    consider animal language (such as “barbaric,” 
    “animalistic,” etc.) appropriate and necessary in
    court cases of Black defendants compared to White 
    defendants. News articles describing Black 
    defendants are more likely to use animal language 
    than articles about White defendants.

    In addition, news articles of Blacks who received a 
    death sentence contain more animal language than 
    the articles of those with life sentences. “So not only 
    are Blacks associated with apes, but this 
    association is linked to justifications of violence and 
    death,” Eberhardt concluded. “It’s almost as though 
    the rules for what moral treatment is get shifted for 
    Black suspects and defendants.”

http://gender.stanford.edu/news/2011/continued-dehumanizatio...

I'd say say that homo sap is on the list of worse behaved apes than Gorillas. Chimpanzees and Gibbons will tear off your face, but it is a big mistake to suppose animals are in a lower ethical state than humans, as animals don't

* have 20 cops beat up one suspect * deal crack * attack people with assault rifles out of hate * invent poison gas * find demeaning labels for one another, etc.

Look at animal behavior and it seems that animals are basically rational in that they do things in the social space that raise their utility function. This is not so clear in the case of humans (an herbivore will always express satisfaction if you feed it, a human might not.)

I think you may be missing the point here.

Reply on news.ycombinator.com