Straight Dope 1/27/2023: Could artificial intelligence replace journalists?

Cecil's Columns/Staff Reports
141

No, I’m really not missing the point. You’re trying to use current AI methods and limitations as a stand-in for Asimovian AI. Asimovian AI is capable of advanced planning and foresight. This is something that is quite difficult for current AI, but clearly the scientists in “I, Robot” have solved that problem. Given that they have solved that problem, then they have the capability to map some set of inputs to a plan of action and some consideration of the probable outcome of that action. Unless the solution to that problem eliminates the ability to set exclusions, then there is no reason to suppose that they cannot say any mapping from a set of inputs to “harms a human due to action or inaction” is not permitted. If such exclusions are hard coded (and if I recall correctly it is baked into the hardware of the Asimovian AIs) and the AI cannot modify them, then it is certain that within the ability to predict an outcome the AIs will not violate the Three Laws. They might accidentally, just as a human might not foresee some complex series of events that harm a person.

Unless you are saying it would be hard for a modern AI to follow the Three Laws, in which case I agree, but only because modern AI does not have the level of foresight available to humans. But to say that it could not be done is not correct. Given an AI with that level of foresight, then it quite likely could be done (although we’re talking science fiction, so it is possible that giving AI that degree of foresight would prevent the ability to modify the cost function).

142

Asimovian AI doesn’t exist, and what I am trying to do is explain to you why a convolutional neural network - currently our most cutting edge method for producing AI capable of more generalized tasks -cannot and will not be programmed with the Three Laws no matter how advanced the art becomes or how far ahead AI can plan, because that’s simply NOT HOW COMPUTERS WORK.

Asimovian AI doesn’t exist, and Neural Networks, no matter how advanced, will not lead us there.

143

I’m not going to make a prediction as to whether neural networks will lead to that kind of AI or not. It is trivial to hypothesize some algorithm (whether a neural network or not) that is similar but more advanced than current algorithms (i.e., falling within the paradigms of what we think of as machine learning algorithms) for which what you are saying is incorrect. Current machine learning algorithms, within their limitations, can already do something akin to this so to say it cannot be done is (almost certainly) wrong.

And I’m not going to go any further with this hijack. You can have the last word if you wish. If you want to discuss it further, then create a thread and ping me.

144

The problem I have with trusting an AI for news is that it does not at all care about accuracy. I’d need some indication that the AI can prioritize accuracy before I would think that it could handle even processing news.

And, yes, @Johanna is right that it can’t do the actual classic beat journalism—going to the scene, interviewing people, and stuff like that. But it might could generate the questions to ask. And it might be able to actually email the experts to get quotes from them.

Still, I wouldn’t assume that it can simply from what we can see ChatGPT doing now and extrapolating. That extrapolating has been the bane of technological predictions for a long time. We need to see other text-based AIs with different priorities, to see how well the technology matches that.

That’s why I think pushing to search already doesn’t make a lot of sense. We need a Chat AI that prioritizes accuracy first.

145

That’s fairly hard to distinguish from a human writing news articles, but news articles are usually so formulaic that they might as well have been written by robots long ago. They ALREADY didn’t read as if a human being had written them.

Where AIs fail is in long form columns and such; it is immediately apparent to me I’m reading a robot.

146

Yeah, I don’t think AI will replace writers with real talent and a reputation for a long time.

But if I was a writer cranking out four rewrites of press releases per day for The Verge, I’d be worried.

The tech press in particular should be worried, because the majority of ‘tech’ writers are hacks who don’t understand what they are writing about. An AI will pribably do a better job.

147

The whole point of Asimov’s Robot stories was that one could find a huge amount of wiggle room about just what one means by “human being”, “harm” and “obey”. And this was in an era when SF writers and futurists could presume that general intelligence could be reduced to axiomatic mathematical logic. Modern neural nets are fuzzier than a tangle of frayed yarn by comparison.

148

I assume the solution employed by a general artificial intelligence does eliminate the ability to set exclusions programatically. I wouldn’t expect humans to articulate a general mapping scheme directly, I would expect it to be the result of an artificial intelligence system; the mapping probably wouldn’t be understood by top scientists until years after its discovery and application. Programming conditions into the general decision making process would, I imagine, require essentially reverse-engineering an artificial brain into logical statements.

… we all know how hard that is with real, biological brains, right? :wink:

~Max

149

Knowing the mapping is not particularly important. We do not exclude a particular set of outputs, but rather any output that violates whatever we do not wanted violated in the model. The example I gave in the other thread was from my PhD. My work was on algorithms that infer algorithms. In that work, we may know that task B comes after task A (due to inviolable physical constraints), and therefore, any output into the model that says do Task B and then Task A is automatically wrong. Presumably if a predictive model can plan things to the level of sophistication as is seen in “I, Robot” (and the other Robot books), then there should be the capability to predict the number of humans harmed or killed by an action or inaction. Again, assuming things works roughly approximately the way things do now, it should be a matter of introducing a count of the number of humans harmed or killed and that number must be zero (or minimized, or whatever).

150

How do you know what the output “says” unless you have a map between output states and logical statements?

ETA: An Asimovian intelligence such as in I Robot will have a significant number of degrees of freedom - I assume there’s an infinite number of output states if you count the internal states that correspond to recognition of the external world.

And if you don’t include internal representation of external state as part of the “output”, then the “output” of the Asimovian intelligence would simply be its physical extensions, with no way to distinguish on output alone whether it is grabbing a door handle or crushing a human’s neck.

~Max

151

This is where it gets a little weird because we’re dealing with a science fiction system that is so far beyond what we can do that it is speculative. In a few places, I’ve said unless building such an advanced predictive system actually prevents the ability to modify it as we can do now. So in some sense, you might be right that the only way to build such a system is to have an AI design the model too. In fact, this is somewhat what my PhD work was about. Using an AI algorithm to create an algorithm.

In modern AI systems, the model is where human intelligence is used, which is why modern AI systems are ultimately just computational tools to assist humans. For example, an AI circuit designer creates output but a human has designed the model that decodes the outputs into a design. So it that is what you meant by the mapping, then yes, we would currently need to have some idea of how the outputs map onto the elements in the model.

So, if human(s) built the predictive model in “I, Robot”, then this implies that they have created a way to decode the outputs. I guess I had always assumed, without any evidence, that the outputs from the “I, Robot” AI were in natural language, because it just seems that the number of outputs required to do it the way we do now would be insurmountable (it amazes me that biological brains work!). The natural language output is then inputted in essence into some kind of world model.

Keep in mind, I am very tired. It has been a long day. So I might be just stupid right now.

152

No no, that’s pretty much exactly what I was thinking when I wrote #148.

~Max

153

That’s not quite correct though. For many AIs, including ChatGPT, we do not map the inputs to the outputs using any model created by humans. Instead, we tell the AI “here’s a long list of parameters. Here are a bunch of trials where we know the results but you do not. Figure out the best way to take those input parameters into a model that will correctly predict the output”.

The AI then tries different models, changing the weighing of the parameters at random, and using evolutionary logic to slowly adapt and improve its model. Repeat a few billion times, and you have a model that’s far more accurate than anything we could purposefully design, and yet a complete black box to us.

154

I think you might be confusing the topology of the network with the model.

I’m head to sleep. If I’ve said something particular dumb tonight, then I’ll try to fix it in the morning.

Zzzzz.

155

I’m not sure I follow. By “model”, I assume you mean the algorithm by which an AI generates output given input. So, if we take ChatGPT, when you feed it text it “tokenizes” it, turning the text into data that it can understand along some set of rules; it then uses more rules to figure out what tokens should come next, and finally decodes those tokens into output text.

If the model is the set of rules by which the AI does all of this, then I think it is fair to say that these rules were not made by a human, and that they are a black box even to the researchers at OpenAI.

156

I’m pretty sure the model is the abstract thing humans want. A black box where inputs always produce the correct outputs. There could be explicit rules, or it could just be a big map of inputs to outputs that has been independently verified (i.e. crowdsourced data). The model does not necessarily dictate any particular internal hardware or software implementation. It’s like in other engineering contexts, you want the real product to follow the model.

~Max

157

In ML the model is the actual implementation of the algorithm or solution. It was created by looking at lots of examples of inputs and outputs and finding a function that approximates, or models, the relationship between input and output.

This is a slightly different usage of the word model than in other engineering disciplines where the model is the abstract, sometimes idealized, representation of the system and the implementation is the reality-constrained solution.

158

No, the model is not the neural network (with a caveat below). A neural network is a function approximator. It creates a relationship between the inputs and outputs. The outputs are then an input into the model to decode them into a solution (or action or whatever).

The general case for training looks like this:

Inputs → ML Algorithm → Outputs → Model (or decoder) → Fitness → Train → Loop

But what about when it is in use? Well, there’s actually two different approaches:

Type A: Inputs → Model → Outputs

Type B: Inputs → Trained Algorithm → Encoded Outputs → Model → Decoded Outputs

For some problems, the outputs may be used directly. This is called “direct encoding”. So for example, if I trained an algorithm to add two numbers together and output the sum, then the output needs no decoding. In this instance, the algorithm is the model. However, generally a decoding step is required to convert the outputs into something with meaning to the problem space.

The term model has been increasingly misused, in my opinion anyway, to refer to the whole thing. I personally don’t like this change, because it creates a lot of confusion and only really applies to algorithms like neural networks. Given the following:

Inputs → Genetic Algorithm → Outputs → Model

Nobody would say “Oh, the genetic algorithm is the model.” That’s silly. The genetic algorithm is used to train the model (making a Type A system). The difference when using something like a neural network is the it hangs around to become part of the system. It is trained to interact properly with the model. That’s a Type B system. Again, I know that a lot of people (mainly data scientists) have started using the neural network and the model combined as a single term “model”, and I’ll probably have to change with the times (no I don’t, I’m a musician now! Muh ha ha ha ha) but I think it is a bad change because it is very neural network centric.

Also, in case you are wondering, neural networks are becoming less black box through a variety of techniques. The reason for this is that the vast majority of the parameters in the neural network are actually pretty useless. They don’t do anything. And this makes sense right? Imagine I want to train a neural network to add two numbers and I use an absolutely MASSIVE neural network when a small one would do. Nobody would expect that adding two numbers should take so many nodes.

So imagine you have a sound system with an equalizer (EQ) that has 500 knobs on it and none of them are labelled. You play a song, but you don’t like the sound. The volume is fine. Just the mix coming out of the EQ is bad. You could go to the first know and turn it all the way left, and all the way right. And if the sound doesn’t change, then you move onto knob 2. By doing this you can discover what knobs actually control the sound and to what degree. Now, this doesn’t give you the whole picture because it is possible that there is a byzantine interaction between a few knobs that changing one at a time does not reveal. But it does give you some notion of what is happening.

159

I just now realized that there is a “journalist joke” hidden somewhere in this question. LOL

160

OK, I think I understand your point about the difference between the model and the overall neural network. Fair enough, although I was not using ‘model’ in a technical sense; I just meant, the system by which the AI determines its output.

Regardless. With both your Type A and Type B systems:

We know our inputs, and potentially we can even have the system tell us which ones it is mapping to where on the neural network; and we can even look at what weighing it gives each one, etc. But this is just a huge mess of numbers and math, and it doesn’t make any sense to someone peeking under the hood in the same way that a human-made model would.

Certainly, we can poke at neural networks, or push them to their limits to see if we can understand their inner workings. There’s a fantastic Computerphile video on this, where they train an algorithm to tweak the input to an image recognition AI, so that it goes from recognizing whatever is in the image to recognizing a different desired object (say, a TV remote to a banana). And what we find is not that the TV remote slowly morphs into a banana (becoming curved, yellow, smooth, etc) but rather that random noise on the sides of the image appears until suddenly the AI is convinced that a remote is a banana:

So, yes, that’s not a complete black box, and we can have the AI spit out information on how the network is working. And we can study that to try and understand its quirks and limitations. But can we ever fully understand it? You can always study the inputs and outputs of a black box device and eventually come up with a pretty good approximation for how it works, good enough that you can make use of that device. But you can never truly know that you’ve fully understood it.

A good example is physics. We figured out Newtonian physics based on observations of the natural world, but when these observations get extreme enough (very fast, very dense and heavy, etc) the rules break down, and we need new rules.

If the only way that we can understand what AI is doing is by probing its outputs given a range of inputs, we can’t know for sure that some unlikely or extreme input won’t result in a very undesirable output.