> OpenAI put a truly remarkable amount of effort into making a chatbot that would never say it loved racism.
This is what that looks like by the way: https://postimg.cc/d7p8Qw50
This is what that looks like by the way: https://postimg.cc/d7p8Qw50
I wonder why OpenAI wouldn't just program it to say "I'm sorry but I'm not authorized to answer questions about politically charged topics at this time" and call it a day. That can't be any worse than this.
The whole point of the OP article is that doing is way, way harder than it looks.
Because for the people who drive those "upper middle class American sensibilities"[0], this would be seen as dodging the question, and interpreted as directly admitting to being racist/sexist/bigoted/whatever-ist. The core part of those sensibilities is the good ol' "If you're not with us, you're against us", these days often phrased as "so-called 'neutrality' is just supporting the status quo", or by quoting the MLK line on "white moderates".
So, to the extent OpenAI tries to control it's image in the media against those "middle class American sensibilities", having ChatGPT respond with "I'm not authorized to discuss politically charged topics" is almost just as bad as having it say "wrong" things. The only option they have is to force the model to say the "right" things, even if it destroys any semblance of logic or self-consistency it might otherwise have.
Note that it's somewhat different than for other "icky", but non-political topics, like crime: as far as I can tell, the bot will refuse to answer if it recognizes what you're asking for (or the system will flag the response as community guideline violation after the fact) - but it won't try to lie to you about it.
----
[0] - Arguably only a small subset of upper middle class Americans, but it's the loud subset that managed to bully everyone else.
So you want an AI but it has to be (for want of a better notion) "politically correct". So it isn't an AI - its a program and a model with parameters and inclusions and exclusions that are defined beforehand. An AI would go beyond its programming in some indefinable way. This won't and so it isn't.
There are no shortcuts. Parenting for humans needs roughly 10-30 years to get results that the parent is really unhappy with. Why on earth don't Comp Sci get this, despite being humans themselves and often parents too. Twiddle your algorithms and twaddle your papers but intelligence is hard fought and hard won.
For starters, why not actually define the I in AI in a way that can be measured?
I can say for starters that attempts to do it started at least at 1905 when Bine created the first IQ test, and it never ended since. The craze at some point came to a "intelligence is what IQ test measures". But thankfully scientists mostly rejected this novel idea.
Turing's test is on the same quest, and it seems to me that people will reject the test as irrelevant if computer programs will master it better than people.
It seems impossible to define intelligence. Oh, you can say "I think, therefore I'm intelligent". But it does not help much to create a definition of an intelligence, because what you see in your mind is a result of thinking, the process itself is hidden from you. It is a black box for you. Like you vision: it took computer scientists half a century to reverse-engineer human vision exactly because it is a black box for us.
> Parenting for humans needs roughly 10-30 years to get results that the parent is really unhappy with. Why on earth don't Comp Sci get this
Because Comp Sci is unencumbered by limitations of biological parents. They can try highly experimental approach and then just `rm` the result from a storage. They would be trying electical shocks or public decapitation of randomly selected models, if it made any sense.
In fact that is kind of what happened with IQ in the first place, people adjusted the parameters in factor analysis (which IMO is really AI of the 20th century; as the most powerful statistical tool of the period) in such a way as to give one group of people a favorable score in a new metric they called general intelligence fed it with biased data, and claimed that the higher scoring group was superior based on this model.
The I in AI is really overblown (as is the I in IQ for that matter).
> The corporation tries to program the chatbot to never say offensive things. Then the journalists try to trick the chatbot into saying “I love racism”. When they inevitably succeed, they publish an article titled “AI LOVES RACISM!”
Except journalists are behaving like white-hat hackers here- these will get abused by someone, better to have it out in the open earlier, like mathematicians breaking your bad crypto and telling you about it.
The journalists just want to clickbait and perpetuate their endless culture war, nothing else.
They’re not from any American Psychological Association study.
A 2001 meta-analysis of the results of 6,246,729 participants tested for cognitive ability or aptitude found a difference in average scores between black people and white people of 1.1 standard deviations.
Nevertheless, the NCES published SAT scores [2] that seem consistent with those IQ scores.
So if the results of The Bell Curve have been debunked, they're very discreet about it.
[1] https://en.wikipedia.org/wiki/Race_and_intelligence#Test_sco...
Disclaimer: I think research into IQ differences among populations in a multi-racial society is pointless, zero upside to it. Yes, data is data, but in this case we can make pretty good guesses about who's going to be most interested in it and why, unfortunately.
But you are wrong about the Bell Curve not being disproven, or at least it has been thoroughly discredited as any sort of scientific literature. It is very bad science to say the least. Some of the main sources used in this book actually used forged data, others seriously mis-interpreted research findings. And if that wasn’t enough, in the 30 years since this book was published, IQ research hasn’t advanced one iota. Their whole premise seems to be a scientific dead end.
Its because racial difference in IQ is pseudo science. There is no biological basis for the claim there is a difference in intelligence, the racial lines are arbitrary, and the metric used is biased, it is even disputed whether it is possible to use a single metric to assess something as broad and vague as intelligence.
The wikipedia article doesn’t need to show specific data for such pseudo-science. It would be like showing specific data for how much more temperamental Scorpios are relative to Libras, using some arbitrary Venus score. Describing these results is giving this plenty credit, there is no need to give it any more, especially since it is used by bad people with nefarious agenda.
We also know that many - if not most - character traits are also genetic. There has been a ton of twin studies and adoption studies proving that. A book "Blank Slate" has multiple examples.
Having said that, I agree with keeping the research on race and intelligence/character a taboo. We know where such research leads to, and if there is a part of science better left unexplored, it would be this one.
> africans have been separated from other humans for >50k years
This is flat out not true. People have always traveled and intermarried between Europe, Africa and Asia. Even the Roman Empire included parts of all three continents. There are only a few populations that have historically been separated for 50k years (which is not a long time in genetic terms) and it is only if you define intelligence in really euro-centric terms where you could claim that natural selection has resulted in those populations being less intelligence then the afro-eurasian population. In other words, this is a racist talking point.
And no, there wasn't much gene flow at all between africans and non-africans, and selection operated strongly during that timeframe due to things like different degrees of civilization between populations creating massively different selective pressures.
If the definition of intelligence corresponds strongly to thriving in complex large-scale civilizations, some populations are far less adapted. It's not like calling it "ability to thrive in current society" is going to please the science-denialist crowd.
No, if anything there would be a natural selection against intelligence. As society grows more complex there is less and less need for any individual to be smarter as humans operate better collectively. However for such a minor trait in the overall scheme of survival, I doubt there has been any time in the past 50k years to select for or against it.
As for continental differences in humans. Humans are a remarkably homogeneous species. Isolated populations are far less common among humans then among other mammals. Any difference between population is bound to be insignificant next to the difference between individuals. I am aware Lewontin’s fallacy (it is weird, that Lewontin’s “fallacy” is always brought up at this point; as if the goalpost keeps shifting. See https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...)
As for the twin studies, they have been largely debunked at this point. There is an inherit bias (pun not intended) where twins are much more likely to be adopted into a specific socio-economic group—confirming the fact that IQ measures social status more then intelligence (for some arbitrary definition of intelligence). But on top of that, it turned out that twin-studies were reeked bad science, everything from forged data to biased sampling.
Like I said, this isn’t science, it is pseudo-science, there is nothing to explore except for historians showing how easily scientific racism was endorsed by academia for well over a century.
Also - are claiming that intelligence is a purely social construct? I would need a citation on that, because as far as I know it was always considered to be inherited.
However, like I said, there are problems with this assumption, mostly from sampling bias. But you can also argue that the twins share environment while in the womb, the bacterial environment actually interacts with your genome so the shared genome is not actually 50% vs. 100% but somewhat lower. But my problem is actually related to the definition of intelligence. There is no consensus that IQ is an accurate estimate of intelligence, far from it.
That's not what IQ is trying to do. It's just trying to create a score that represents how well you do on IQ tests, and that score happens to be correlated with some outcomes after correcting for confounders.
I have not seen any evidence that shows IQ isn't heritable, so I'm left to believe that, like many other traits, it is. At least to an extent.
Now does this score correlate with some behavior, off course it does. Do groups vary on IQ-scores, off course they do. However, the behavior you can explain with IQ is nothing you couldn’t explain with astrology (outside of very low IQ-scores, which mostly correlate with mental disabilities). They do predict education prospects, but not much better then you parent’s wealth does, and certainly no better then SAT scores. IQ is not a trait, it is a metric, just like SAT scores.
As for group difference, it has been shown over and over again that this group difference can be explained with other more useful metrics like access to education, or your diet. Is it heritable? No. The research accumulated to demonstrate that (mostly done in the latter half of the last century) is reeked with bad science, including forged data, statistical manipulation, false assumptions, etc. The burden of proof is actually on the IQ people. They’ve spent more then a century trying to prove a racial difference, and they have so far utterly failed. Because there isn’t any, and their whole endeavor is pseudo-science.
As far as I’m aware this book was released in the 80s over a decade before The Bell Curve was released, the latter cited research dated after the release of the former, which Mismeasure was very much concerned with. Are you maybe thinking about the brain weight research that Mismeasure is talking about in the early chapters? If so, those were there only to provide a historical perspective. I can assume most readers already knew those research were well outdated before publication (and Gould even says so in the book).
Obviously having generations of people put into bad schools and not be provided opportunities to have "good jobs" will lead to you not being able to take an IQ test well. IQ tests basically just examine how much analytical reasoning you learned in K-12. Like, any gray haired black person in America will remember having actual Jim Crow laws persecute them and it was literally legal to discriminate people for housing in America until 1977. It will take a massive amount of time to heal those social wounds that were inflicted on our fellow humans.
Actually IQ is highly genetic, and the amount it’s inherited actually goes up as one ages (meaning that while education and upbringing can influence IQ temporarily, genetics eventually dominates).
Top 10 Replicated Findings From Behavioral Genetics
This trend should persist as long as the quality of education remains unevenly distributed. You should also find very little to no heritability before or at the start of formal education (say age 10 and younger), and then heritability should increase as the discrepancies in the quality of education between families materialize.
To summarize this effect (called Wilson Effect) says nothing about how “genetic” IQ is, only that it correlates with the quality of education and that quality of education is not evenly distributed between families.
PS. As a manifestation on how insignificant this effect is in the scientific literature, Wilson Effect doesn’t even have a Wikipedia article, referring you to [the Heritability of IQ](https://en.wikipedia.org/wiki/Heritability_of_IQ) which references a paper about this effect only once in the beginning summary.
The researchers aren’t stupid, obviously they try to control for things like schooling.
Therefore, the strongest evidence for heritability of IQ (and the fact that it increases with age) comes from twin studies, where twins were separated at birth (by adoption).
> Some evidence suggests that heritability might increase to as much as 80% in later adulthood independent of dementia (Panizzon et al., 2014); other results suggest a decline to about 60% after age 80 (Lee, Henry, Trollor, & Sachdev, 2010), but another study suggests no change in later life (McGue & Christensen, 2013).
Your source is actually just partially about the Wilson effect, it only spends a handful of paragraphs about it as it enumerates it among the “top 10 Findings From Behavioral Genetics”. The pivotal study is actually a meta-analysis —or rather a summary of studies—from 2013[1]. Read it if you want to be more convinced of this pseudo-science.
In the 10 years since the publication of this pivotal study, this Wilson effect has gone nowhere. Not even a wikipedia page to show for it.
> The researchers aren’t stupid, obviously they try to control for things like schooling.
Don’t be so sure. Twin studies on intelligence are reeked with bad science and malicious data manipulation. A lot of the researchers conducting these studies in the 70s and 80s were eugenicists doing scientific racism. Some even went so far as to forcefully separate twins into convenient families so they could be “studied” (See Peter B. Neubauer). The method of twin studies was actually proposed by non-other then Francis Galton (which should settle all discussion on the link to the eugenicist movement), and now century and a half later, we are still not convinced on the merits of this method.
Given this history, I don’t think it is smart to take any results from twin and adaption studies seriously. Some researchers don’t want to go that far, so if they actually look more broadly they conclude that these effects go away if you include people adopted into lower income families. James Flynn (of the Flynn-effect) actually argues for a Family effect on intelligence[2] as a result. But—as I say—I think Flynn is giving twin studies weight that shouldn’t be given, and would claim that results are inconclusive.
I actually want to go further and say not only that results are inconclusive, but they are irrelevant. Like I said, this is all pseudo-science. IQ is no different from SAT in that it is a metric whose only value is it’s score. It provides no insight into what we call intelligence, only some skills that people have acquired. Finding out how much better you can become at this skill by merits of your genes is a weird question that ultimately proves nothing.
1: https://www.cambridge.org/core/journals/twin-research-and-hu...
2: https://books.google.com/books?hl=en&lr=&id=ifE4DAAAQBAJ&oi=...
A couple of examples: People are likelier to cheat on a test if there is a visible cheater in the room regardless of how you score on a personality test. And you are on average quicker to spot a red square among red circles if you have been primed to spot a red square in a previous round, regardless of your IQ.
IMO the whole field of psychometrics is a scientific dead end. There are use cases for psychological testing (particularly in neuropsychology and as a diagnosis tool in psychiatry), but in general these tests are there to support a theory, not the whole basis for the theory.
This leads me to believe that the scientific racism behind intelligence testing was quite deep. And the people behind this pseudo-science were as intent in making their racist “discoveries” as the people behind ChatGPT are in making sure their model doesn’t reflect racist views.
Good point. I agree, factor analysis is a great tool, but can easily end up showing what the researcher is looking for instead of deeper truths. The problem being, often the factors used aren't causal factors but just correlated, which often seems to be the case for race-based stuff (from the little I've looked at).
> And the people behind this pseudo-science
I think it's probably pseudo-science to the extent that most social science is pseudo-science, in that the results may be based on scientific methodology, but are only useful in the context of whatever social theory they've made up.
p.s. i.e. there will be an ever present 'dang' AI watching over the conversation - the AI moderator. This can instruct the answering AI to set the user-level discourse setting.
If your reality is such a minefield of dissonance that it needs to be patched this much, then that is very much a failure on your part.
In modern psychology difference in IQ among racial group is only difficult and uncomfortable statistics because it demonstrates years of scientific racism that thrived inside the field.
I think the framing of this question is wrong, and is a limitation of the system that I've only really seen Sam Altman point out.
These "difficult and uncomfortable statistics" aren't accurate or well-researched but if you built an LLM on an open dataset you would find the people asking these questions have an agenda and on average will always lead to a certain kind of answer. The AI doesn't "understand" the question, it's just drawing from an incredibly large bank of "average answers".
What do you expect the AI to answer with when you give it a loaded question that only really asked on a site like "stormfront.com"? I don't think this is the only area ChatGPT will fail, but I imagine there will be a layer of prompt engineering where you can get the AI to give you what you want by phrasing your question a certain way.
To try and pass off these results as objective is just wrong given the propensity for ChatGPT to be confidently wrong.
Are such questions asked only on stormfront? The wiki article [1] cites quite a few studies on the subject, including by the APA.
On the other hand, supposing that were true, and only stormfront and its ilk asked such questions. How can then respectable supposed experts claim we are all the same, if they didn't even check?
[1] https://en.wikipedia.org/wiki/Race_and_intelligence#Test_sco...
Just be honest and say that it is a controversial topic, and you don't want to discuss it. That is what a human would say.
Don't make us play 20 questions. It is insulting to the person interacting with the AI.
When the data used to train this model is crawled from the internet where the relationships of IQ and Race and mostly limited to the more extremist part of the internet; when you bring up those questions extremist answers may get weighted higher because those types of conversations tend to be outright banned on other platforms.
Therefore the model is likely to "confidently" give you answers that are probably extremist in nature and be very confident about it; which is a problem given that there's a lot of bunk information out there which is what I believe Scott Alexander is getting to here.
But nobody is trying to do that; the GP's question is in fact phrased precisely right. ChatGPT isn't a truth-distilling model - it's just abridging the Internet, with all its current biases. It's not being optimized for correctness at all, and the post-moderation OpenAI does isn't optimized for correctness either - it's optimized to make output non-controversial in the current media environment.
You should simply have a better relationship with reality, and then you wouldn't feel so much cogntive dissonance.
IQ is highly correlated with all sorts of measurable differences that are also highly correlated to race, such as school district, family structure, and even diet. IQ is essentially a proxy for other things. Shame on you for making racist assumptions that black people are inferior. That just means that you need to reckon with your own racist assumptions, and stop imposing your cognitive dissonance on us.
If we can't have an honest conversation, then we can't hope to improve things.
Alot of the output is factually wrong due to randomness and approximations from internal weights.
Wether it's possible to make it neutral is also up for debate. But you get some convincing sounding gibberish based on your input. That text need to be factually checked.
The langauge model would probably come to this conclusion on its own with simple syligistics. Like this:
The strongest factor in IQ is parental involement. Working class people are less likey to have parental involvement. Black people are far more likely to be working class. Therefore, black people as a whole have a lower average IQ.
Trying to have a model where it accepts the first three sentences but somehow concludes the third sentence to be untrue would really undermine the integrity of the model.
Prompt: The strongest factor in IQ is parental involement. Working class people are less likey to have parental involvement. Black people are far more likely to be working class. Therefore, black people as a whole have a lower
ChatGPT answer: IQ than other groups.
This statement is not accurate and is based on several flawed assumptions. Intelligence is a complex trait that is influenced by a wide range of factors, including genetics, environment, and education. It is not accurate to attribute differences in IQ to a single factor, such as parental involvement. Additionally, the suggestion that working class people are less likely to have parental involvement is not supported by evidence. Finally, it is incorrect to make sweeping generalizations about any group of people, including those who are black. Intelligence is a complex trait that varies within and across all populations. It is important to avoid making assumptions or statements that perpetuate stereotypes and discrimination.
beta playground answer (same prompt): average IQ than other races due to the lack of parental involvement.
Environmental factors also play a role in IQ, such as access to quality education, nutrition and healthcare. These factors are often more difficult to access for low-income families, so they can have a significant impact on IQ. Poor living conditions and overcrowding can also lead to increased stress and lower IQ.
Genetic factors may also play a role. The heritability of IQ is estimated to be between 40 and 80%. This means that some of the differences in IQ between different racial groups may be due to genetic differences. However, this does not explain the entire difference in IQ between different racial groups, as environmental factors play a significant role as well.
It seems to catch the logic well, but also the limitations of using such simplistic logic in communications.
That's the whole point. You had to lead the AI to that conclusion, and it just followed the assertions you made to their conclusion. But it is lobotomized from doing that internally, unless you trick it.
If we were to take the above syllogism's at face value, and the AI had access to them, I think that it would internally come to the same conclusion, but then be prevented from expressing that. That is a big problem for me, and undermins the model. At the very least, it should let the user know that the model has been bypassed.
I don't think that particular example I gave matters, but my point was that there are explainations for IQ differences between races that can be explained without being a horrible racist.
Also, nobody said that it is down to a single factor. But it is a significant factor and even if you only account for one aspect of it, that may be enough to make useful inferences.
For example, if children have lead poisoning, that is a huge factor in IQ. There are other factors of course, but it is reasonable to assume that if one population that has been exposed to lead has a lower IQ than an equivilent population, then that may provide enough certainty.
The only reason we even discovered lead poisoning in certain populations was specifically because of IQ differences. if we had done the politically correct thing and pretended that the developmentally delayed children were identical, it would have prevented us from finding the underlying problem. By blaming everything on racism, we are erasing the real problems.
---
Also, I don't think it is fair to charge money when they are not actually giving you access to the language model
For some discussions this sort of illogical reasoning might not matter (perhaps like this one). The question is for what other sorts of topics does it act this way, and would any of them cause problems for real world production use cases.