Many of the statistics you're repeating here could be a case of correlation rather than causation, so by themselves they don't mean much. But that's not even the point. Using sexual or racial stereotypes (even when they are realistic) to make important decisions denies due process to the individuals involved. And I don't want to live in a world where AIs are denying people due process.
On Tue, Jun 27, 2023 at 8:26 AM Matt Mahoney <mattmahone...@gmail.com> wrote: > The relevant topic here is not who on this list is racist. It is how > racism should be handled by AGI. A language model learns to predict or > compress text by constructing a model of the world as it is known to > humans. Humans are universally racist. People prefer friends, neighbors, > and usually mates who share their race, culture, and language. People learn > to predict behavior using easily observable traits associated with age, > sex, and ethnicity. For example, people never confuse my preferred pronouns > (he/him). When I travel, people usually recognize me as American and speak > to me in English. > > Human interaction would me much more difficult if we were all blind to > race, age, and sex in an attempt to be fair. In the US, police kill about > 1000 people per year. The victims are 98% male, 50% Black, and 5% unarmed. > Police must use Bayesian reasoning to make a split second about whether a > person is a threat, which is not always clear. They implicitly know, like > everyone else, that 95% of homicides are committed by men and 50% by Blacks > (who are 15% of the population), and mostly by young adults around age > 15-30. In an attempt to reduce police bias, some departments train using > interactive shooting ranges with pop up targets of little old ladies > pointing guns at them. So far this has not reduced the number of unarmed > Black men killed by police, and probably increases the risk to little old > ladies. They are already subject to the same invasive airport security > screenings as people who are actually a threat in an attempt to be fair. > > ChatGPT was trained to be non racist this way, by curating the training > data to make it race and gender blind. This is stupid. Television has been > doing this since the 1970s, portraying an imaginary world where whites and > blacks are friends and equals, which would be nice if it actually worked. > One look at a census map shows that it hasn't. We all still live in > segregated neighborhoods. > > The issue for any AI used in business or government is obeying civil > rights laws. Some discrimination is legal. For example, children do not > have the same rights as adults in any country. In many countries, including > the USA before 1973 and Switzerland and Israel today, there is compulsory > military service for males but not females. Discrimination by country of > birth is legal and widely practiced in every country. > > It is not the job of AGI to eliminate bias. If there is a genetic basis > for the stereotypes that everyone knows, then it is impossible. Male > aggression is not limited to humans. Chimpanzee killings are also 95% male. > Race is more controversial. Asians consistently score higher than whites on > academic tests, and whites higher than blacks. Same with life expectancy. > Asians in the US live 5 years longer than whites, and whites live 5 years > longer than blacks. The lowest crime rates are in Asia. All of the > countries with majority Black populations in Africa and the Caribbean have > poor economies. The controversy over genetics is that if we accept it, then > all of our policies to eliminate racism are doomed to fail and we should > abandon them and return to legalized segregation. Therefore, any such claim > is racist and must be punished. > > An LLM trained on unfiltered data should learn to be biased, learn how to > recognize bias, and learn the laws relevant to discrimination, just like a > human would. This was the point I made at the start of this thread about > constitutional learning. When you tell it not to say anything racist (which > would be bad for your business), it understands what that means. > *Artificial General Intelligence List <https://agi.topicbox.com/latest>* > / AGI / see discussions <https://agi.topicbox.com/groups/agi> + > participants <https://agi.topicbox.com/groups/agi/members> + > delivery options <https://agi.topicbox.com/groups/agi/subscription> > Permalink > <https://agi.topicbox.com/groups/agi/Tcf822b60238b0592-M926d02f8edf791e00787a8bb> > ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Tcf822b60238b0592-M1195a553bc3db7fd5d67353b Delivery options: https://agi.topicbox.com/groups/agi/subscription
