Feeling Like a Fraud, But Great at Getting Lost

Dear Felienne!

What a wonderful conversation we had about Hubert Dreyfus’ book “What Computers Still Can’t Do”¹. I loved every second of it. My highlight was when you wanted to get a tattoo with a John Searle quote about the impossibility of representing certain mental capabilities in the computer!

AI-Tattoos and Failing to “Make Things Meaningful”

This made me think of how people nowadays probably get tattoos with quotes provided by AI. I wonder how often Claude, ChatGPT, and Gemini give the same stupid suggestions when asked for a unique, personal tattoo. Some people will likely get hollow tattoos saying things like “Make it meaningful” based on the AI’s recommendation, thereby failing to put the words into action. Others might have tattoos proclaiming “I choose who I become”, while in reality, the AI has chosen this tattoo (and much more about their life) for them. But ok, luckily I am not the tattoo-police, but rather, the co-host of our podcast! So let me get to the actual homework.

Imposter Syndrome Versus True Imposters

I did read the paper “What Can Computers Do Now?” (2024) by Ben Schuering and Thomas Schmid² you assigned, and I did not like it! But then again, I like reading papers that I don’t like, so all is good. In fact, there was a part I actually enjoyed. Namely, they write…

Today’s hybrid AI systems are close to demonstrating all the features shown by Dreyfus, but do so in an almost imposter-type way (Bonnemains, Saurel, and Tessier 2018³).

What they say here (with the help of the authors they cite) is that AI has improved. Nowadays, it can almost do those things computers could not do in Dreyfus’ time. Examples that Schuering and Schmid provide are “ethical and moral reasoning”, which, according to them, are successfully displayed by autonomous vehicles and medical systems. While I don’t exactly agree with this example, I do agree with their remark that AI does things in an impostor-like manner. In fact, I would go even further and argue that these AI systems are imposters! They pretend to be human-like, empathetic, and ethical, when at their core they are not! If this is not the definition of an imposter, I don’t know what is…

Dropping the Syndrom in Imposter Syndrome

The unfair thing is: we humans are the ones with imposter syndrome. We are the ones worried we are a fraud and fearing that someday people will find out the truth about our incompetence. Even now, writing this column, I have this fear that I write something stupid which will finally reveal that all this time, I was just pretending. Everyone will eventually know what I knew all along: that I was just lucky things worked out so well for me for so long. In contrast, AI is not facing such worries. LLMs will just lie about the things they do not know, but they will not lie awake because of it. In fact, AI can’t fear; it has no idea what it means to worry, and has no experience of human struggle. So how could it ever be considered human-like?

Thinking a bit more about imposter syndrome and academia, I am wondering: Maybe AI is changing things up in that regard quite a bit. For a long time, I wished academics would not suffer from imposter syndrome so much. But recently, with more and more researchers using AI to do their job, and Dutch news items like the one about hundreds of Dutch scientists citing non-existent sources⁴, I am thinking: these academics should feel like imposters. They are faking it! So with the rise of AI, maybe feeling like an imposter is increasingly warranted. Why deal with impostor syndrome if you can simply become an imposter instead?

Maybe our imposter feelings are part of what makes AI so attractive to (some of us in) academia. We feel like imposters, and now, finally, reassurance about our greatness is just one click away. I have been there, and yes, I am tempted to ask AI if what I just wrote makes sense. (I did. Gemini assured me: “Yes, this makes perfect sense—and it is a brilliant, slightly cynical, but highly accurate take on the current state of academia.” You can imagine my sigh of relief, followed by a wave of worries for academia’s future!)

Enjoying Mistakes and Being Different

One homework question asked us to think of things computers can’t do now. I’m not sure if computers can’t do it, but I think computers are still quite bad at responding to mistakes.

Of course, there are all kinds of computer science methods where learning from mistakes is a thing (like automata learning and so on). That is not the kind of mistake-making I have in mind. If I think of my own life, I make thousands of tiny mistakes every day. Usually they throw my day around in unpredictable but often good ways. I might forget to buy something at the supermarket and then improvise a nicer meal instead. I don’t think computers respond to or make use of mistakes in such a flexible, open-ended, creative fashion. For instance, consider how a GPS navigation system responds to a person who is taking a wrong turn. It will just tell the person to turn around and fix their mistake. Or, if these instructions are ignored, it will recalculate a route to the original destination. However, it won’t tell the user “hey, you have never been here, that’s cool, look around! You seem to be lost! Do you remember the last time you were lost? What a weird feeling! I will turn myself off now so you get to enjoy this feeling and discover new places.” No, the GPS will stick to the original plan. It’s just so boring.

Yes, I know a creative GPS system could, in theory, be built! Yes, I agree that computers can improvise. (I actually wrote my BSc thesis on improvised computer music a very long time ago). I actually see value in that. I do appreciate computer creativity. I also see value in simulated intelligence. However, I see value in those things because computer intelligence is different from human intelligence and because computer creativity is different from human creativity. I’m not against it. I just don’t want to conflate the two or pretend they are the same when they are not. Humans are not the same as computers. Computers are not the same as humans. This is important to me, because for me, the added value of computers lies in them doing things differently. So if I think about what computers are bad at these days, I think some computers are quite bad at being computers. They are getting worse at it. They are becoming increasingly human-like, and this is not for the best.

What Computers Should Not Do

You also asked us to come up with things computers should not do, independent of whether they could do them. What a beautiful question! It made me think of something that happened a few years ago at our university when ChatGPT was just becoming a thing. Back then, a colleague (who doesn’t work for us anymore!) used an LLM to write a laudatio for a student. I was really shocked by this. It felt so inappropriate and disrespectful towards the student. So here is one answer: writing laudations is definitely a no-go.

Yet, in my opinion, this does not mean that an AI should never be used to write text. Maybe a teacher wants to use AI to produce a text that shows common mistakes Germans make when they learn Dutch. In such a case, I would say, go ahead, have the AI write that text and use it in your teaching! Now, using AI to generate a text would feel appropriate to me. Essentially it comes down to context. There’s not one thing AI should never do. It’s contextual!

This reminds me of a theory we have in privacy research. It’s a theory by Helen Nissenbaum called contextual integrity⁵. And in this theory, Nissenbaum explains that the appropriate flow of information is contextual. If someone shares, for example, a picture, it matters who is in the picture, who shares it, with whom they share it, how they share it and so on. If a doctor, e.g., shares a picture with a colleague to get a second opinion, it’s completely different from them sharing the same picture with their neighbor to make fun of it. We have norms about that. Privacy violations occur when these norms are violated. And I think something similar is going on with the use of AI and computers. We have norms and values in our society, and using AI in some contexts will definitely violate these norms and values.

Maybe AI is not to blame here. Arguably, AI itself lacks the real-world context to make good judgment calls about its use. However, to my big disappointment, humans also display bad judgment and use AI for all kinds of inappropriate things…But then again, humans might learn…I don’t think AI will learn when its use is violating human values and norms. It might be prevented from some actions, just like ChatGPT is prevented from producing pictures of poop. But to bring this back to Dreyfus, AI will never have the common sense needed for this. In any case, I don’t see AIs refusing to write laudations, eulogies, or wedding vows anytime soon. For humans, I still have hope they’ll learn that this is not appropriate.

To sum up, what computers should and should not do ultimately is a matter of adhering to and balancing the various values at play in a certain context. And that’s hard. But we have methods for that. For instance, in my field, there is a framework called Value Sensitive Design (VSD) ⁶, which supports designing technology in accordance with human values. Something similar must surely exist for the use of technology. I will have to look into that! But that’s for another time. Talk to you soon!

Cheers,

Hanna