Google offers to help others with the tricky ethics of AI
arstechnica.com
arstechnica.com
Isn’t google a big part of why there is so much distrust to begin with? Seems like asking your oxy dealer for detox advice.
The actual problems with AI ethics are down to "it isn't ready for that morons" and "It would be obviously abusable even if it was jackass."
That's literally the "killer" application that 100% has obsessed many military planners ever since they saw the Atari playing demos.
And whether people recognize it or not, those are the killer robots. They are built into fighters, defense and planning systems.
Regardless of any objections or how many people sign letters against killer robots, the military strategists have no choice but to include the most sophisticated software available.
Now I assume that there is some "failsafe" in most of not all operational systems that prevent actions that are not authorized. But it is also necessary to make that authorization fairly vague in some circumstances. And there is quite possibly a way to turn that failsafe off in the case of a war or scenarios where they feel there is no time to wait for the human (which seems guaranteed assuming the other side has AI also).
Same reason someone might render medical aid to a slave. Doesn't make it a charitable act.
Note that this is a different issue than data privacy, which Google rightly takes flak on. AI bias is referring to questions like “if I train a network to (sentence criminals, price insurance, ...), how do I detect if the resulting predictions are racially biased?”
The conversation on this subject by lay folks is rife with statistical ignorance, and Google has done good work communicating and clarifying the conversational starting points. This is a hard issue because it takes ethical trade-offs and forces you to specify mathematically exactly how you want to handle inequality, which is a subject that most people haven’t thought through rigorously, and would rather hand-wave away with virtuous sound bites.
For example, see http://research.google.com/bigpicture/attacking-discriminati...
As their example shows, there're two traditional approaches: maximize profit and group-unaware.
* Maximizing profits gets the most profit, but treats people differently based on their group.
* Group-unaware treats everyone the same regardless of their group, but can generate far less profit.
The example presents two alternatives: "demographic parity" and "equal opportunity". Presumably the authors would argue that these may be superior choices because they generate nearly as much profit as profit-maximization while having a plausible argument for being socially responsible.
This seems a bit off.
Fundamentally, we might say that there're 2 kinds of discrimination: fair and unfair. For example, it might be fair to discriminate for objective reasons, but it'd be unfair to discriminate for non-objective reasons (e.g., bigotry).
The advantage of profit-maximization is that it takes full advantage of fair-discrimination while fully avoiding unfair-discrimination; the drawback is that it does discriminate.
The advantage of group-unaware is that it fully avoids all discrimination; the drawback is that it sacrifices fair-discrimination, causing it to yield the lowest profits.
The two alternatives proposed in that example seem to get the best of both worlds because they're basically just cloning profit-maximization, but with slight concessions to plausible-sounding criteria for equality to dodge perceptions of unfair-discrimination.
Here're the tricks:
* In "demographic parity", everyone has the same odds regardless of group. This would appear to be the same thing as profit-maximization if the risk/rewards were the same, but since it ignores them, it ends up being basically "profit maximization, but ignoring different risk/rewards".
* In "equal opportunity", both groups get the same true-positive rate. This again seems to sacrifice some of the risk-vs.-reward information, but with a slightly different skew.
So for the privileged group (Orange) vs. the disadvantaged group (Blue):
* Profit maximization: $32,400 from 50 vs. 61
* Group-unaware: $25,600 from 55 vs. 55
* Demographic parity: $30,800 from 52 vs. 60
* Equal opportunity: $30,400 from 53 vs. 59
To me, that looks like 3 ways to discriminate, all yielding roughly the same profit and thresholds -- using either of the 2 proposed alternatives gives up a little bit of the profit in exchange for a pleasant-sounding rationale.
What I dislike about this is that it seems entirely superficial. The proposed alternatives engage in roughly the same level of fair-discrimination (and none of them engage in unfair-discrimination, which wasn't given in the example at all) to generate roughly the same level of profit.
---
> The conversation on this subject by lay folks is rife with statistical ignorance, and Google has done good work communicating and clarifying the conversational starting points. This is a hard issue because it takes ethical trade-offs and forces you to specify mathematically exactly how you want to handle inequality, which is a subject that most people haven’t thought through rigorously, and would rather hand-wave away with virtuous sound bites.
That's exactly what this looks like!
In this case, the virtuous sound-bites are "Demographic Parity" and "Equal Opportunity". They both worked out to be mostly the same as simple profit-maximization, but if someone in a disadvantaged group protests that they're being discriminated against, they'd probably find it difficult to follow the math far enough to sustain their complaint.
At the object level, I think you perhaps oversimplify with “ Fundamentally, we might say that there're 2 kinds of discrimination: fair and unfair.” What we think is fair is the crux of the whole issue, and there is not agreement on that distinction. For example, if we price risk at the “true” risk of default, and this means that fewer black people get loans, is that fair? Some think it is, and some think it is not, even though I hope most would agree there is no racist intent in the decision itself. The economic context in which these decisions are made already contains the impact of past racism, which can be perpetuated or or restituted by decisions made now.
I also disagree with your characterization of these types of parity as sound-bytes. My definition of a sound-byte is a conceptually thin phrase that sounds good, but does not contain much content. These are the opposite; they are specific jargon with a precise meaning that is explained in detail in the article. You could call them “Foo” and “Bar” and their usefulness in the conversation would be the same. (Sure, “equal opportunity” has been deployed as a sound-byte elsewhere without any rigorous definition, but I hope you’d agree that this article at least fleshes out one possible concrete definition, even if you don’t think it is the right one. I’m sure they are others - and now we can further the field by writing them down, citing this piece, and making an objective critique!)
Yeah, same. And to my eye, this whole thing's paper-thin; it tested my suspension of disbelief to not dismiss it as an obvious joke or bad attempt at a PR stunt, and I'm honestly still undecided.
Part of the issue is that its conclusions are absurd. We can invent ethical constraints to arrive at max-profit by pulling things out of thin air; the procedure's analogous to p-hacking. For funsies:
1. Come up with a huge portfolio of ethical-sounding constraints. Include basically any ethical-sounding argument imaginable, even if it seems stupid or bad for business (the numbers can be hacked; we just need the pretty words).
2. Come up with model-transform strategies. For example, instead of calculating the threshold, calculate the log-threshold or the threshold-entropy. (Again, don't think logically -- we're hacking the math, so just come up with random things.)
3. Come up with model-hybridization strategies. For example, construct a meta-model that uses 50% of the threshold + 25% of the inverse-log threshold + 20% of the inverse-entropy threshold + a flat 5% (because, hey, why not?).
4. Whenever we want to maximize profit, just do so with simple profit-maximization. No "ethical" constraints.
5. Repeat Step (4) for millions (or more) different hybridized models, combining all sorts of different model-transforms and ethical-constraints at random. Optimize each for approach to simple-maximization.
6. Stop upon finding an "ethical" solution that effectively equals simple profit-maximization.
7. Generate a description of the "ethical" solution, and write a report on how your company's using that particular ethical strategy as part of its latest, on-going campaign to promote social equity.
8. Just do simple profit-maximization. Or the "ethical" solution; whatever, they're literally the same thing. But if anyone asks, be sure to tell them it was for ethics.
While we're at being ethical, we could even shift some paradigms to envision new modes of engagement with socioeconomically disadvantaged stakeholders to promote a synergistic realignment between our commitment to ethical outreach and disrupting the zeitgeist with our social media blitz to rapidly accumulate surprisingly liquid social capital, indelibly scrawled across the human blockchain of love and mutual respect, validated in the proof-of-work of billions of sapient hearts yearning to emerge from the trials-and-tribulations of our modern era into the welcoming bosom of a new tomorrow.
But who cares about pointlessly defining jargon for logical inconsistencies, though? It's hallow and meaningless; at most, it might fool people before they realize that it's logically inconsistent, but that'd just be lies. It doesn't seem to have an actual use.
One might suggest that Google could create partnerships with medical facilities in order to do the development. Of course, as soon as the news of any partnership with Google arrives, there's an immediate outcry over the data privacy issues. It seems impossible for them to act in this space without receiving flak.
I believe there's also the risk that, since they're going to be blamed for all possible evils whether they do them or not, they may as well go ahead and reap the rewards of such bad behavior.
This doesn't seem to imply that Google is serious about security which was the original point.
Being able to count on probably 5 years of OS and security updates instead of the mishmash of abandoned handsets or potential security nightmares with hobbyist firmwares is another nice feature.
We need the ethics of AI to be an offshoot of open source. It's tough to suggest that in an environment where enough people can pressure a group to adopt objectively unethical policies, but at the very least open source communities do not have the trigger of financial risk.
Google is going to need to make money from this "Ethics as a Service" eventually. And that is an obvious, all too obvious conflict of interest for what Google envisions.
Maybe they can do this. I'm not 100% objecting to the idea. But I'd like to suggest they start with YouTube. Ads on the platform have quadrupled since November 2019. Creators are playing a lottery with monetization every other upload, and they're making as little as 1-5% compared to YouTube's ad-share program 10 years ago. Start there.
I'm sure people may interpret things differently but doing the right thing would mean Google making AI that has a positive net impact on society, while "don't be evil" could simply mean good for Google but neutral to everyone else.
Some fun history: I thought the phrase originated with the Christian bible:
> "when the blind lead the blind, both shall fall into the ditch"
but Wikipedia https://en.wikipedia.org/wiki/The_blind_leading_the_blind traces it to the Upanishads:
> "Abiding in the midst of ignorance, thinking themselves wise and learned, fools go aimlessly hither and thither, like blind led by the blind."
If Google'll act as any other company in a free market, then this'd seem like a neat offering. Maybe it'll be a helpful service, or maybe it'll be a flop; but, worst-case scenario, just don't use their service if it's no good.
But the language of "auditing" and "charg[ing] for ethics advice" sounds almost like someone's laying the groundwork for regulatory authority over AI.
An ethical AI needs to be able to function without gorging itself on users' private data
Step 2. JK