I’ve seen some talk recently about the virtues of hallucination, implying that it’s a “good” thing because it’s “creative” to make things up or simply what LMs are “trained to do.” I see my own work sometimes come up in these discussions since I’ve written a lot on creativity and open-endedness. So I want to give my two cents: An important aspect of creativity is the ability to know what already exists, which is how you can push towards novelty. If you are confused about what is real and not real, then it’s hard to know what’s novel. So the less-discussed flipside of hallucination is the tendency to present old ideas as novel, fundamentally misunderstanding what it means to be creative. And indeed we see LMs often do just that when pushed to exhibit genuine innovation. Ask e.g. for innovative new recipes, ideas for new genres of music, or inventions to help with some existing problem, and often what you get back is something that either already exists or already has been proposed! (Note I am not saying that’s always what you get back, but neither do you always get hallucinations when asking questions of fact. But it happens too often.) So false creativity and hallucination are two sides of the same coin. If you have one, you would tend to expect to see the other. In this sense, these are not good things, and the argument that hallucination is an asset “because creativity” is a misunderstanding. A firm grasp of reality is an ally of both authority and creativity, and indeed human authorities on subjects are also often creative. But this point is not just to be cynical: it means that if we genuinely “solve” hallucination, creativity will come along for the ride. It also surfaces an interesting hypothesis about which approaches to solving hallucination are most deeply illuminating for progress in the field: if a “solution” only addresses one side of this two-sided coin (e.g. lowers hallucination without increasing creative quality) it is likely ultimately relatively superficial and not the genuinely transformational path that will lead to something spectacular in the field.

Dec 9, 2023 · 7:36 PM UTC

Woops that was the first time I edited a post and I should have read the warning more closely - didn’t realize a small edit would cut off all the quotes and comments already there! Sorry to @MinqiJiang @MLStreetTalk @c_valenzuelab @tobias_rees and others whose thoughtful comments got detached. Wish I could re-add them.
By the way, regarding the argument that by hallucinating LMs are simply doing what they were “trained to do”: they may or may not be trained to do it, but even if we assume they are, being trained to do something that is pathological does not make it good. If we are training them to hallucinate and (hand in hand with that) be falsely creative, all it implies is that we need to find a better way to train them. The only reason we train them to do something problematic is that we don’t yet know a better alternative, not that it’s some kind of clever, intentional sleight of hand.
Replying to @kenneth0stanley
Interesting. When reducing hallucination are we bound to be stuck with explorative creativity as opposed to transformative creativity? The way out this paradox is think deeper about how to evaluate what is good, and as that is contextual, how to give control back to the user.
It’s an interesting question, but I don’t think reducing hallucination precludes Boden’s transformational creativity. After all, reducing hallucination should merely reduce effort misspent on regurgitating ideas that already exist. Such a reduction increases space for both exploratory and transformational creativity (as well as combinatorial creativity).
Replying to @kenneth0stanley
Rediscoveries are fairly frequent. Good ideas (like backpropagation) get independently rediscovered dozens of times. I don't think this is necessarily a drawback (at least, for non-omniscient systems)...
Yes but the reason the LM gives you back a recipe that already exists when you ask for a novel recipe is not that it “rediscovered” it. Rather, it is simply echoing the presence of the preexisting recipe in the distribution of its training data because it doesn’t understand that doing so is not novel (just as, on the flipside, it doesn’t understand that saying something not in its training data is likely false). This kind of error is not at all related the virtuous kind of reinventing the wheel that might arise for a human out of ignorance. Rather, it is delusionally thinking you are inventing the wheel when someone just told you about wheels yesterday. Maybe we should call it “creative delusion.”
Replying to @kenneth0stanley
In art class, I always welcomed anyone 'reinventing' a work or style, because no-one ever follows the same path. There's often a *lot* of creativity in the way they got there. This may not be very different for LLMs.
As I mentioned in a previous response, there’s a big difference between reinvention in the sense of coming up with an idea someone else had before versus reinvention in the sense that someone told you yesterday exactly the same “idea” you think you had today. The former is perfectly delightful, the latter is delusional.
Replying to @kenneth0stanley
not sure "false creativity" is a coherent concept. the test of creativity is, if it actually moves the needle, not whether it's new in some absolute sense. the true solution probably is not to "solve" hallucination but to make new hallucinations truer
Not so easy to “move the needle”creatively by simply outputting something you already saw in your training data.