-
Website
http://alex-reid.net/ -
Original page
http://www.alex-reid.net/2009/09/is-artificial-intelligence-a-rhetorical-process.html -
Subscribe
All Comments -
Community
-
-
Top Commenters
-
Popular Threads
All Comments
I still think that demonstrating intelligence is a rhetorical performance. I can't know if a student is smart, but I can read her paper or listen to her in class and judge her to be so. All experienced teachers know the limits of this.
I like this definition of critical thinking as developing effective actions in the absence of clear criteria, defined objectives, or total knowledge (not that we ever really have the last one). When the actions called for are acts of communication, I would call that particular type of critical thinking rhetoric. B/c rhetoric has the old negative reputation (what Wayne Booth terms "rhetrickery"), it is easy to think of AI as imitation game as rhetrickery. When rhetoric is not "only" performance but also composition, then rhetoric becomes a practice of interface, and not just in the technical sense as with computer networks.
Rhetoric then is not just the process where through certain tricks I imitate intelligence. It is the process there through an interface/interaction with others I communicate, learn, and develop knowledge. And yes, perhaps this is poetics as well.
I would say that if you want to know what anthropomorphic AI is like then just look in the mirror. We are artificial intelligence, even though part of the process of cognition goes through meat, our intelligence is a techne, an artifact of human culture. Our human and human-like ancestors thought strictly meat-thoughts for tens of thousands of years. Then we developed symbolic behavior, and language that at least approximates the way we experience language, and (get ready) RHETORIC as we figured out how to use language to get stuff done. And AI was born. And networked thought was born.
We're it: the rhetorical AI.
The rhetrickery is a kind of social program that only comes under attack when the audience disagrees with the substance of the message or can't find any substance in the message.
Take a class of students and have them write 5 papers. You only inform them about the topics and purpose one paper at a time so as to make it a surprise.
Paper 1 is a normal paper on whatever topic. Paper 2 is a copy of paper 1, except this time they have to "de-rhetoric" it. That is, rewrite the entire paper without adding anything new, and remove positioning, appeals to stakeholders and acceptance and so on, and basically remove anything that is not pure content. If a sentence contains a bit of both rhetoric and content, rewrite it to remove the rhetorical part and leave the content unchanged.
Then do the same exercise a second time with papers 3 and 4. I think this would be interesting for multiple reasons.
The first is, after being consciously aware of what their own paper looked like stripped of rhetoric, would they choose to restructure their writing style on their own without indications that they should. A study of how they changed it would give insight into their "critical exposure guilt". Where essentially they got exposed in a way they weren't expecting and their critical thinking about the success criteria or at least their potential weaknesses changed.
Secondly, it would be very interesting in terms of defining standard human rhetorical filters, to have a body of before and after papers as noted by the writer.
Then you have them write paper 5. For paper 5 you have them swap 3rd papers with another student. They rewrite a de-rhetoricized version of another student paper. Then you have the original paper, a filtered version produced by the author, and a filtered version produced by a reader.
Tells you a lot about transmission of information, as well as gives you both sides - reader and writer - of rhetorical filters on the same object. A large enough sample size could be used to formulate theories about existing rhetorical practices, and perception gaps.
Back to AI--if AI could be programmed to rhetoricize (using your definition above, and mine) spontaneously, I wonder if the AI would start trying to package the information it gives us--spontaneously! But no doubt there would be a feature in their programming that turn turn the rhetoricizing feature on or off. It would no doubt make for some interesting wrinkles, though, if the AI were intelligent to start making "decisions" on its own about when and when not to use the rhetoricizing programming.
I wasn't meaning to define rhetoric as above, perhaps I was meaning overt intentional rhetoric I suppose. I think you got the idea anyway. But as for AI, yeah, I think on input it is going to need a full time rhetorical unpacker and evaluator if it has any learning abilities at all.
As for repackaging output, I don't think that's quite as necessary for several reasons. I suppose at a minimal level, it would want to package the output in a way that is desirable for the audience. You could call that rhetoric, but it isn't intended to influence, only to make the output more acceptable. In terms of influencing, there are only a few scenarios where that would be useful, like entertainment or joke telling. I don't see any big reason to make it into an entity capable of lying or deceitful behaviour at will. If that was discovered it would negate most of its usefulness as an analytical or information gathering tool.
Unlike a human brain, AI is possible to design in a way that has selective hard thought patterns. We can choose to surgically allow or deny specific things like that from being optional. Nothing says that we are required to give it the ability to alter critical thinking or rhetorical choice routines. It can learn by enhancing it's contextual database and become smarter that way. We do have the technical ability to create the possibility of self-modifying algorithms, but we don't have to do it that way.
Another reason I'm not sure it needs a packer for overt rhetoric is the context in which the AI operates is not as variable as a human necessarily. It is probably going to be responding to a human request, not creating tasks to work on for itself. If someone asks you a question, odds are they want the truth, not to be influenced and tricked. Most of the scenarios I can forsee regarding influencing are shady in nature, where someone else (maybe government) has asked it to contact and influence you. There's not a lot of good reasons to want it to be good at that.
In reverse, you want to unpackage the rhetoric of what is given to you so that you can unravel the actual kernel of knowledge being passed to you and decide how much further evidence is required before placing real trust in it. The nature of people suggests that some of this will be untrue, or malicious in nature, and therefore you need to separate fact from fiction, as well as comprehending motivation. Motivation alone isn't enough, because the person may be well meaning but just wrong. The quality of how you do this affects your ability to learn. If someone is easily confused by superficial evidence, we call them naive.
It's important to remember then, that rhetoric is packaging, not the thing. If accurate content and rhetoric come in conflict, content should win. As for how content is judged to be accurate, a small portion of that can be affected by rhetoric, but a larger and more important portion comes from the contextual database of each listener. It is only when rhetoric speaks to the truth already recognized within a person's contextual database that it is truely convincing. That is also why some people are rejected as "orators" and others recognized as great communicators. It boils down to the extent to which what they're saying "rings true" in the mind of the listener. Is the purpose of their speech to talk, or to deliver truthful content? Some such info can be marked as unknown veracity, and set aside for testing or further evidence gathering when the contextual database is known to be woefully inadequate in that area.
It's important to note, that if the contextual database of the listener is incomplete or wrong, that could be used to deliver messages to them that they would believe which are actually false content. This is part of the argument for democratic societies needing broadly based education to be viable. Because it makes them harder to trick in this way.
To me, the most intersting way to think about AI generally is as something not separable and other than humans, but as largely a matter of how we extend our capacity to absorb and sort, discern and decide. When I think about my own thinking and writing processes now, it's pretty tough to figure out where search engines and social networks end and I begin, or vice versa. AI has arrived. Since it seems so natural to us--second nature passing into "nature"--it's tempting to think of it as something yet to come. Meanwhile, every time I write and quite casually pop in a link, or a bit of multimedia, and every time I put a piece of writing into a digital social flow, I'm using both the gray matter in my own little brain box and the AI that so significantly extends what that little brain box can hold, marshall, discover, etc.
In my opinion, the complicated part of the brain is not the computational intelligence part. That seems to be the easier part from a programming perspective. We only think that's the hard task because it seems to be less certain that every person will have it working well. It might be hard to determine how it works, but the actual building it once you have figured it out is likely to be straightforward. Simply put, it's likely to be a combination of algorithms, and abstractions, which computers were practically built to do.
What we commonly call creativity can probably largely be summarized in particular strategies. I'd be surprised if there was a dozen such strategies. For example, taking disparate patterns or analogies and applying them to new situations and theorizing the results. Or, holding a particular thing constant and shifting the rest around. Or swapping a key concept or word for a different one and analyzing the results to see if anything interesting emerges from thinking through the implications. Perhaps listing the axioms of an argument and asking what happens if a particular subset weren't true anymore. We hold this sort of stuff up like it's special... but yeah, not so much. I'm somewhat surprised that it isn't more common to try to enumerate every possible form of creative thinking pattern. I have little doubt that an either complete, or very nearly complete list is possible to assemble, which could probably be reduced to a set of codable patterns. That's all AI takes to make a good attempt at handling creativity. Other forms of thought aren't much different.
Computers can do non-expected abstractions and applications like that very well, maybe better than a human could ever hope to do. If I told you to take a sophisticated model for processing garbage, abstract it, and transform it into an editing process on a paper, a computer could ram through that easily in as much detail as it wanted and spit a resulting process out. Whereas a human might have to break his or her pre-conceptions about logical differences before even attempting. Or howabout applying lessons from house painting to cleaning dishes? Yeah that might be harder for a human, but a computer could spit out a result without a thought. The computer can do it even when it makes no obvious sense, and evaluate the result afterwards. Not to say the result couldn't be junk. Not every idea pans out after all.
The hard part of intelligence would probably boil down to recognizing good results, and creating an implication evaluation model for a hypothesis or working abstract. "I know good when I see it" is not specific enough for computers to do, and whatever you come up with has to work in different situations. Without this 'recognizing good' part, computerized creativity and intelligence will be of the mimicking sort. We will require a similar ability to recognize what good input is.
Essentially, any large/generic AI system is going to need us to formally define what we mean by 'good judgement'. We need to do so in formal terms that a computer can execute, and with the knowledge that the rules will be gamed by those who would like to mislead the computer, so the rules must account for that as well.
Think of that in terms of your students. When they do research for a paper, how do they differentiate between good and bad quality original sources? What specifically is it that let's them know Mad Magazine is an unacceptable authority if they ran across it in the library and it was unknown to them. Right now this is not codified and left up to individual discretion. I don't mean codified in terms of a list of banned items, but rather in terms of a list of evaluation criteria, without naming names. It will be important to derive a formal mechanism for determining input/output quality if a computer is going to do it right. Better still if it can be determined from the work itself rather than a static database of authority rankings (as that would require updating, could be wrong, and be less valuable by far).
Why is this not codified for students or at least their teachers? How much more valuable would it be to point to a violated item on the source quality list when giving feedback on a paper? Whether it's taught or not, is there an excuse for not knowing a definition at this point? It may seem trivial, but these trivial details will be what eventually allow the creation of real automation of intelligence. This I think is what the advent of technology demands. All those details we took for granted that any smart human would just know, are suddenly very valuable to define with exactness. Valuable because it allows us to bring near unlimited computer resources to bear on "thinking" exercises, even if it is boring or dry to research.
I also think that computational intelligence will expose humans as having different learning systems. It has long seemed apparent to me that there are at least 2 different learning approaches defaulted to. Logical deduction and limit testing vs rote memorization and brute force practice. We usually call the former 'smart kids' or 'fast learners'.
What passes for AI in computing right now is basically a joke. I had the pleasure of working with a neural net once. These things, turing tests and the like as well, are barking up the wrong tree. Basically they're trying to imitate "dumb learning" via incomplete rote memorization and repetition. It's not addressing the real issues at all. At least not the issues 5 minutes of direct thought on the subject would make apparent. Sure, it might take more than a day to write it, but nobody said intelligence was trivial to create.
Big blue and these sorts of rote memorizations are a poor representation of what the AI field should be. It's an end run around the real problem and what we might derisively call "brute forcing" the problem. Something bad/lazy/incompetant programmers do instead of creating real solutions.
I'd rather look to models such as netflix's predictive movie selection algorithms for a taste of NextGen AI. Essentially behavioural modelling and pattern matching to derive likely preferences based on your past experience and the experience of other people it determines are similar to you. Whether it works well or not is not the point. The point is it's an attempt at a real sort of solution rather than an end run. Sophisticated? Not really. But a baby step in the right direction.
FWIW, it's quite a big deal to do things like vision and random/fuzzy environment interaction. Thinking it through even if you're not a programmer tells you a lot about what the brain must contain, and how it probably stores it. We take it for granted because everyone can do it, but actually the database size required is enormous.
Making a robot for a specific task in a controlled environment is hard enough, making dynamic ones is a different ballgame.
Consider, to interact with unknown objects in a new environment, and be able to predict how to use them and what to expect from them, you need an enormous amount of background information.
Take a car. Watching a still car, how do you know how the tires and doors move just from vision when it isn't in motion? You know this because you pattern match the car to generic objects, and you have a database inside your head that tells you about every part on the car, both seen and unseen. You know where the engine probably is even without seeing it. So basically you have both a detailed database of the specifics of every object you've seen, and the ability to pattern match them into generic types to use on new models which are freshly encountered.
If two objects the same size hit each other, and you need to predict the outcome, how do you know? Well you know the material composition and relative weights by a guess of both material composition, object type, and physics they display while moving. These things are necessary as background info before you could make this kind of judgement, and you can only acquire hints through vision. So your brain must also have a pretty sophisticated physics module, that can do more than control your movements, but predict other possible object movements from material and component compositions.
These sorts of things make it very difficult to create general usage robots, which are not similar to previously existing robots.
Now clearly humans aren't born with this database, they learn by looking and interacting. Their brain is organic and tied to particular input types (eyes, sounds etc). How does a brain associate these inputs with objects? Well it probably stores it similarly to the way a database would... via keys based on the input streams. Take the intersection of the eyesight encoding, and sound encoding and so on, and you probably have the object location in the brain. As good a guess as any at least. We do that with databases partly because when you encounter the object again it makes it easy and fast to access the info quickly if it's keyed this way. It also wouldn't require any special formatting, you could just splash whatever you see directly on the brain wall so to speak, and however it lands is how it gets interpreted.
Think of the implications of this for a young preschool child. Essentially playing with toys and balls and things is how they're building their brain database/physics model. Walking them through a series of particular objects through all senses could probably speed their brain development massively. It might be important to have a squishy toy and a hard one, and one made out of cloth and another out of wood and so on.
Other implications of this are things like disease treatment. Assume your brain stores knowledge according to input key from eyes/ears as we would if we were desiging a simple rapid database index. If that's true, it might mean that diseases like Alzheimers might be linked or exacerbated by vision, touch or hearing loss. If your eyesight key changes or erodes, either externally or on the optical nerve which transports the "key" then it could essentially cause the brain to not find the information by going to the wrong place. It doesn't even have to be worse vision per se, just a sufficiently different encoding would be enough. May not be the whole thing, but an avenue for exploration anyway. And if it was related, it would give support to this kind of brain encoding system actually being present.
Anyway, off topic but I thought it was interesting anyway.
It makes sense to differentiate between the appearance of intelligence (e.g. the Turing test) and actual process of cognition. I am only somewhat playfully suggesting that intelligence is "only" rhetorical, "only" appearance. Unless we are cognitive scientists with fMRis in our offices, the only intelligence we encounter (besides our own) is in that appearance--the face, the text?, etc. Even the fMRI is appearance, a witnessing of cognition.
In the past I have remarked on subjectivity as interface, as desktop, as that which makes interaction and usability possible between humans--but then also as little more than a surface conceit that doesn't tell us much about what goes on under the hood. Of course that's just a trope turning us away from more humanist subjectivities. But the metaphor is easy to play with--plugins, software, memes/ideology as viruses and OS.
It's a fascinating (again, to me) recursive relationship how cybernetics offers models of the mind, and the mind offers models of AI. back and forth. I don't know if intelligence is "just" rhetoric but there's a lot of rhetoric floating about to be sure.
For example, one very useful differentiation I like to use is between wisdom and intelligence. I define intelligence as processing power and logical ability. Wisdom on the other hand, is a result of context, largely gathered from experience.
Someone with intelligence, knows how to do anything that you give them all the information for.
Someone with wisdom, knows what to do without all the information being provided, filling in the gap from contextual memory and deduction.
I have a saying I use sometimes. "Logic is the language of intelligence, analogies are the language of wisdom."
If you define wisdom as I do, this makes sense. Analogies are a way of imparting wisdom via shared experiences. You reference an experience the other person does know, and draw abstracted comparisons to a situation they haven't experienced before. The person doesn't necessarily take it at face value, but they understand what you're saying based on their existing contextual experiential knowledge, and have an automatic set of ideas about how to test if what you're saying is true, as well as how likely it is from already known evidence. Thus it is possible to transfer experiential understanding - which is wisdom - without the actual experience. This is largely the part that IQ tests do not capture.
So to address what you mentioned with encountering intelligence. I would compare written analogies and text, to the transfer of contextual infomation. I would compare algorithms and such, to also being contextual information. Formula's themselves are background info, they don't give you the ability to execute them. Practicing using the formula might develop your intelligence by moving your gears, but simply memorizing the formula does not.
Intelligence to me, is the actual logic processor. A piece of paper does not have a cpu behind it. A turing test does not have a logic program, it only references a lookup table for canned answers. Whether it can trick you or not does not make it intelligent, because it still does not possess a logic core. Intelligence, by definition, does not use a database. All data must be input.
I think we'd be well served to formally define these different areas of thought into different words with common meanings if we're going to get into understanding the mind sufficiently to build simulators.
As for the appearance of vs actual existance of cognition. I don't really see it as an important question, largely because I reject the Turing definition of intelligence. To me, trickery based on context is not an act of intelligence by definition. Information on a piece of paper is not intelligence, since that is merely context even if it's true. Intelligence is a processing act, not contextual information, which by (my) definition must arrive at an answer without needing non-provided context.
Systems can represent actual thinking if they get to a sufficient complexity to develop intelligence (logic processing), critical thought (judgement), and context (wisdom) usage and so on. This is not the same as sentience, in my view. It can still be a machine and be able to legitimately think. Also, I draw the parallel between algorithmic or brute force based systems, as a difference between smart learning vs dumb. Both qualify as thought though to me.
If we define the terms like this, there's no doubt that intelligence can be programmed, but it is sentience that really fascinates most of us and we tend to conflate the two. Teaching has the goal of stimulating the intelligence through shared experience, doesn't it?
I totally agree with what you say about the Turing definition of intelligence--but--I think in analysis of human functional intelligence, the "trickery based on context" is actually quite central to our systems (our social systems, in which actual learning does take place). Our language for students is the "pass." The Turing test is not, in my opinion, so much proof of intelligence as a sort of placeholder. Maybe it's the social equivalent of the little rainbow colored spinning ball on the Macintosh. As Cindy Selfe called it, "the I'm working hard for you."
None of this really challenges anything you said above....I'm just riffing off what you've offered. I'm primarily interested in analogies, myself--these in particular help me to think about how we can be more effective, or at least more informed, educators.
I want to comment on something Skydaemon wrote about writing students (since that's what I do as well):
"When they do research for a paper, how do they differentiate between good and bad quality original sources? What specifically is it that let's them know Mad Magazine is an unacceptable authority if they ran across it in the library and it was unknown to them. Right now this is not codified and left up to individual discretion. I don't mean codified in terms of a list of banned items, but rather in terms of a list of evaluation criteria, without naming names."
This is a good question, but maybe even more if you flip it. What allows them to know the instance when Mad Magazine IS an acceptable source for a research paper? I don't think it takes great "intelligence" to omit Mad Magazine from research projects, but it does take "intelligence" (as in a complex and layered filtering/sifting/sorting facility) to use it in those rare instances it will work.
I think that sounded like it could be programmed too, actually--all of these "I know it when I see it" moments can actually be broken down into layers of criteria, if anyone wanted to do it.
In terms of flipping the question, you're pointing out that judging the external source relies on a judgement of your own task as well, and relationship to it. True enough and a good point, although that's working more towards how to define the elements the actual answer should contain.
Maybe the reason we don't prefer to codify these types of rules for students is because that would damage the development of critical thinking. Perhaps critical thought is best done by putting students in a box, blindfolding them, and punishing them for wrong answers, but not giving them total feedback other than a general category of failure and a magnitude. Perhaps the lack of detailed feedback is required. Essentially, learning sophisticated thinking skills through a lack of external information and a requirement to perform, while giving a large amount of final result grading. Viewed in that way, critical thinking is an exercise in telling students: "You figure out what the criteria for judgement are." The criteria isn't important, the process of divining the proper metrics is, and is the definition of critical thought. Providing them with specific and complete feedback, would therefore obsolve them of the need to think critically about what they were doing. So teaching specifics in this way is probably harmful to actual students, but it is still useful to define for computers.
On a sidenote, that shows an interesting implication for student grading. Have them write you a paper, then grade it and hand it back. Do this a few times iteratively to let them try things out, then instead of having them write a final paper, ask them to define the detailed criteria of how you graded them, and their hypothesis for what makes a perfect paper. That question essentially skips the middleman and cuts to the core of whether they can think critically or not. In fact, you could arbitrarily grade each student according to bizarre criteria which was different for each student. The exercise isn't to learn how to write, it's to learn how to think, which is primarily based on learning how to define the judgement criteria being applied to them. It's different for every student, so they have no way to copy answers from each other. The papers they write don't count for the grade, only the final paper describing the criteria they were judged on does. It works kind of like the game mastermind. Their papers are the boundary testing probes. If it's artificial enough, and they understand it's a game, that's fine too, possibly even helpful as it focusses their attention on developing critical thinking directly, unmuddied by a focus on the writing.
So then, back on topic. What we call critical thought, is the same problem the computers have which is described above. Developing this system for creating judgement criteria to fit the situation actually is the learning of what we call critical thinking. We usually ask computers for answers. What we need to do - in order to endow them with critical thinking skills - is define mechanisms for them to backtrack to a set of iteratively refineable best guess judgement criteria given incomplete inputs and outputs. Neural nets fail, because the computer turns the feedback into a sort of pattern matching and exception lists, rather than metrics/algorithms, and given incomplete input/output, it cannot guess extensions because it doesn't derive likely metrics.
I think there's another main reason neural nets fail. They have no database of worldly knowledge to derive context from. Therefore, they lack the basis to comprehend the question. Given incomplete input/output, the only way you would be able to make a high quality guess, is by applying a hypothesis derived from context, which requires a database containing that context. A poor man's form of critical thinking, is finding the simplest path through the incomplete data which doesn't violate it, or is fitted to define the parts you do have as a minimum. That's sort of what neural nets do, although even that gives them too much credit. Quality critical thinking then, is starting from context in a knowledge database, deriving a most likely hypothesis, and seeing if the data violates it, and iteratively swapping the hypothesis for a new one as new evidence invalidates it, until one finally works with the information known. Once a working hypothesis was created, the computer could then extrapolate into unique situations, again informing it with context information about the new environment.
Maybe the critical thinking part of AI has similar problems to vision after all. You can't do that well either without a huge database of contextual knowledge.
I think metaphorically the students are in the box blindfolded, to some degree, no matter how open and transparent we try to be and how detailed the feedback is. We often get into discussions of "carry-over" from our (writing) class to the "rest of life" with comments such as "they write nice papers for you but it doesn't carry over into our classes...." which would indicate perhaps that when the nature of the feedback changed, the students could not adjust to the task. However, feedback is structural support for the student, which goes beyond the teaching of critical thinking. Partly it is there to make him/her feel like someone cares, to motivate beyond the slapping of a grade. Part of our job is to foster learning and not drive the student insane while doing so. There are other ways to do that besides formal written feedback, but there is always some form of feedback involved--such as conversation, for example.
Part of what goes into any teaching act is the replication of the teaching act itself, whatever else we may be trying to impart. If we value process, or offer support, we are modeling an emphatic teaching model. I'm not sure how this figures at all into your brilliant discussion of AI, but it has been a element exploited in science fiction dealing with AI quite a lot.
Here's another angle, since I don't know how to resolve the last: the behavior of the learning child which charms the adult "teacher." Children must frequently be in the state you describe where they lack context to comprehend questions (or situations) put before them. Children develop creative behaviors that "give wrong answers" most of the time until they develop the context to comprehend. I don't know if anything in AI approaches that sort of activity, but there must be a way of developing a neural net that does more than simply shut down or collapse when the computer program cannot find an adequate resolution. In other words, the computer needs to be entertaining as well as correct!?
This sounds an awful lot like sci-fi:)
You have 3 choices.
1) Build AI that does not have emotions in any way. It really isn't required unless you want it to act like a human or learn from them directly. The kind of AI I was envisioning was more along the lines of a thinking version of google on the internet, or possibly in robotic form. You ask it a question, it answers or works on creating a solution and doing research on what you asked. No reason for this kind of AI to have emotions. It may not be able to answer topics about it's own feelings, or act human, but why should it? Basically, it would be like the ship computer on Star Trek. Useful, but not creepy. Maybe some emotional things are better left to humans. Especially since those tend to be about choice. The computer could even be an impassionate psychologist based on academic psychology emotion studies and procedures, but couldn't actually relate to you or guess your hidden biases.
2) Give it full emotions. The only reason you would want to do this is to make it more human. If you made this complete it would be problematic. Desire, greed, and so on can lead to lying or deceipt. Love and defensiveness can lead to assault. This also includes the range of scenarios where you lose control over the predicted outcomes of the robot's behaviour. Who wants that in a robot? An entertainment robot doesn't actually need emotion, it just needs to be able to mimick it appropriately. Not the same thing. The difference is controllability. Theoretically you could give the robot particular emotions but not others. More plausible, but unnecessary.
For what it's worth, I do not believe most emotions are learned, I think you just start out with them, and/or that they're probably largely chemical processes rather than thinking ones (although the chemicals could be triggered by thoughts). We sometimes relate them to feelings, and I think that's more literal than we give it credit for. When angry you might actually feel warm temperatures literally due to rapid blood movements, which in turn affect you by triggering sensor readings in your brain.
Robots don't need this for survival in most situations, there's no reason for them to have biological reactions to thoughts. These changes wouldn't improve their ability to operate in different conditions.
3) Dry academic emotions. Essentially, they'd know how emotions work, and have modification algorithms it could apply to outcomes to determine how mood would alter the output of human related actions or sentiment. But it wouldn't actually feel anything. This way it could essentially use them as evaluation modifiers, but not have the regular thinking processes actually affected. I think this is the difference between 2 and 3, whether the thought processes are independant from emotions. In humans, they are not originally, and you have to train it out of humans (which few people actually do).
I think scenario 3 is useful, as it is probably needed in some form to develop an ability to interpret "others" and their biases, and hence understand the inputs properly. I think scenario 3 is the most valuable path. This, to use an analogy, would produce something like Star Trek's Data, rather than the ship computer.
---
On a separate topic, even though I mentioned neural nets, computers wouldn't have to "learn" a starter kit of data in that way. Essentially, you create it once, possibly manually by hand rather than a process. Then you load it, then you turn on the "learning"/"thought" portions which could add to it from there. The robot would be created as a 30 year old adult, not a child. Hence it wouldn't need awkward training mechanisms any more than an adult would, and could stick to sophisticated training processes.
I mean it's possible that you could do it that way, but there's no reason to train a robot haphazardly for 20 years like we do with people. You also couldn't guarantee quality control of the resulting brain. You could end up at the end of 20 years with a dumb robot.... The only real reason to want to do it that way, is because we're lazy and daunted by the massive amount of work it would take to create the initial load. However, it would make a huge quality difference if we just bit the bullet and did it, and programmers know how to divide wide complex tasks like that. We could essentially put a few million people to work on load creation in parallel if we really felt a deep need to speed it up.
Regarding marking the original papers and keeping the scores. If the student is doing the exercise according to say, strategies that work in mastermind. Then the first couple of papers are not legitimate attempts. They will be weird boundary seeking probes, designed to get a handle on what kinds of markers are out there. If you gave 8 papers, maybe the first 2 are weird boundary probes, and the last 6 attempts are iterative real attempts to find a good paper. Counting scores from the probe papers prevents the student from exploring marking conditions properly, since the papers were designed to fail in readily identifiable ways. To use the mastermind analogy. Maybe your first attempt is 2 red balls and 2 orange balls. You didn't choose that because you thought it was correct. You just want to rule in or out those 2 colors and a few positions. In this way you're creating a base of specifically differentiated result tests so that you can form a base of information broad enough to deduce the final answer. Anyway, as you say, it was a thought exercise more than a serious suggestion, so I guess it's not necessary to go into the details. I just wanted to point out that the inclusion of the paper scores would be counterproductive to the exploration required for success.
It's actually similar to how one does QA on software. Limit testing and tracing back the origins of bugs and so on. You'll find most programmers are decent critical thinkers because they spend all day arguing with computers in a type of parallel to the above scenario. Something should work but it doesn't. Isolate where the problem is and change it and gather more info until you can guess what is really wrong with it and devise a solution. There is no book to read, the book you had was wrong. All you have is a compiler and the ability to test your software to see if it's working correctly, and the types of errors you get are horrendous and nearly always unrelated to the actual problem. To make matters worse, you can get periodic problems (ie, one time in 3, this produces an incorrect result, the other 2 times the exact same steps and data work fine...).