Brief notes on the Wikipedia game
Alex Turner introduced an exercise to test subjects’ ability to notice falsehoods: change factual statements in Wikipedia articles, hand the edited articles to subjects and see whether they notice the modifications.
I’ve spent a few hours making such modifications and testing the articles on my friend group. You can find the articles here. I describe my observations and thoughts below. The bottom line: it is hard to come up with good modifications / articles to modify, and this is the biggest crux for me.
The concept
Alex Turner explains the idea well here. The post is short, so I’m just copying it here:
Rationality exercise: Take a set of Wikipedia articles on topics which trainees are somewhat familiar with, and then randomly select a small number of claims to negate (negating the immediate context as well, so that you can’t just syntactically discover which claims were negated).
“By the time they are born, infants can recognize and have a preference for their mother’s voice suggesting some prenatal development of auditory perception.”
-> modified to
“Contrary to early theories, newborn infants are not particularly adept at picking out their mother’s voice from other voices. This suggests the absence of prenatal development of auditory perception.”
Sometimes, trainees will be given a totally unmodified article. For brevity, the articles can be trimmed of irrelevant sections.
Benefits:
Addressing key rationality skills. Noticing confusion; being more confused by fiction than fact; actually checking claims against your models of the world.
If you fail, either the article wasn’t negated skillfully (“5 people died in 2021” → “4 people died in 2021″ is not the right kind of modification), you don’t have good models of the domain, or you didn’t pay enough attention to your confusion.
Either of the last two are good to learn.
Features of good modifications
What does a good modification look like?
Let’s start by exploring some failure modes. Consider the following modifications:
“World War II or the Second World War (1 September 1939 – 2 September 1945) was...” → “World War II or the Second World War (31 August 1939 – 2 September 1945) was...
“In the wake of Axis defeat, Germany, Austria, Japan and Korea were occupied” → “In the wake of Allies defeat, United States, France and Great Britain were occupied”
“Operation Barbarossa was the invasion of the Soviet Union by...” → “Operation Bergenstein was the invasion of the Soviet Union by...”
Needless to say, these are obviously poor changes for more than one reason. Doing something which is not that, one gets at least the following desiderata for a good change:
The modifications shouldn’t be too obvious nor too subtle; both failure and success should be realistic outcomes.
The modification should have implications, rather than being an isolated fact, test of memorization or a mere change of labels.
The “intended solution” is based on general understanding of a topic, rather than memorization.
The change “The world population is 8 billion” → “The world population is 800,000” definitely has implications, and you could indirectly infer that the claim is false, but in practice people would think “I’ve previously read that the world population is 8 billion. This article gives a different number. This article is wrong.” Thus, this is a bad change.
Finally, let me add:
The topic is of general interest and importance.
While the focus is on general rationality skills rather than object-level information, I think you get better examples by having interesting and important topics, rather than something obscure.
Informally, an excellent modification is such that it’d just be very silly to actually believe the false claim made, in the sense of “How on Earth could a reasonable sane person believe that!? You can obviously see how this is false from X, Y or Z”, while still being non-trivial to notice.
Examples
I’ve created eight modified articles. You can find them at the Google Drive folder here.
(Details: For each article you can find the original article stripped for brevity, the edited article, an answer file and Wikipedia auxiliary files. You should download the whole folder and open the html file for the edited article in your browser. Tested on Ubuntu, should work on other systems as well.)
If you only have the energy to look at one article, see the Industrial Revolution one. If you want to look at two more, see the World economy and Price gouging articles. I think the articles steeply drop in quality (more on this below), but I put all articles I created in the folder anyways.
When I play-tested these with my friends, we usually took around 10 minutes per article for reading and writing down thoughts.
Takeaways and thoughts
The biggest challenge is coming up with claims to modify.
I quickly noticed that articles often don’t have lots of very clear-cut factual claims you can invert. To understand what I mean, look at the Wikipedia article on philosophy, for example. It just isn’t amenable to modifications satisfying the criteria laid out above.
Even if you have factual statements, it is often of the form of memorized facts (“The capital of France is Paris”) that are in themselves inconsequential. Or, if you have some very consequential facts, like “The population of Earth is 8 billion”, as opposed to 800,000 or 80 trillion, you have memorized those anyways.
Finding consequential statements that people haven’t memorized beforehand, or don’t immediately recognize as true or false, is not easy. I felt like many examples I came up with just weren’t good enough.
(Other source materials – scientific publications, perhaps – could be better in this regard than Wikipedia, though then one runs into issues of technicality and narrowness. I also thought of modifying the overall lean or bias of an article by selective focusing, withdrawal of information and “lies of omission”. Noticing such changes is a skill that clearly has real life applications, but more work is needed to produce articles.)
In any case, I haven’t spent that much effort on this, and feel like I lack imagination on the sorts of topics one could use.[1] Plausibly if you had more people spending more effort on this, you would come up with lots of other good examples and discover heuristics for finding good articles, types of modifications to make etc.
This remains the biggest crux for me: how difficult is it to come up with good modifications? If others find it as hard as I did, then it’s hard to get enough supply of articles. If others are more capable than I am, I’m much more optimistic: the idea is extremely scalable, allows for iteration on quantitative metrics, and one can filter for high-quality examples. So if you like the idea, consider taking some time to think about it and share your thoughts.
- ^
I did try to prompt GPT and Claude models to suggest me ideas, but their article suggestions were mediocre, and suggested modifications were rubbish. (I didn’t spend a lot of time optimizing the prompts, however.) At one point I did use Claude to write a fake paragraph in Wikipedia style for one of my modifications, and was happy with the result.
I tried them all. My notes are full of spoilers.
Industrial Revolution (_2):
I didn’t catch the deliberately-introduced error; instead I thought the bit about a recession in the 1830s was likely false. I took “the Industrial Revolution” to include the “Second Industrial Revolution” (as, indeed, the article seems to) and it seems to me that electricity was important for the progress of that, so I’m not entirely convinced that what was introduced to the article was exactly a falsehood.
Price gouging:
It seemed utterly unbelievable to me that a gathering of economists would disapprove wholeheartedly of “price gouging” and, lo, I was correct. (I feel like this one was pretty easy.)
World economy:
A few things seemed dubious and nothing seemed obviously wrong. There’s a statement that economists usually use PPP to translate currencies and not official exchange rates, but then a table that (if I understand it right) uses both of those. (But this is kinda reconcilable: usually PPP is used for comparisons, and it’s much better, but no harm in including nominal figures as well.) There’s a claim that it’s usual to consider only “human economic activity” and I don’t know what that means—what’s the alternative? E.g., if it doesn’t include making things with automatic machines then this is obviously false; if it’s contrasting with aliens or something then it’s so obviously true as not to be worth saying; probably it means something else but I don’t know what, and looking it up feels a bit cheaty. This feels odd but my epistemic state is more “I don’t understand something here” than “I know what this means and it’s wrong”. (Oh, I guess maybe it means “human economic activity as opposed to other human activities not classified as economic, like parents taking care of their own children”, in which case it’s not content-free and is probably correct.) There’s a claim that before the Industrial Revolution “global output was dominated by China and India” and that India was the largest economy circa 1CE which seems iffy—bigger than the Roman Empire? -- but, I dunno, those both had large populations and China at least was pretty advanced in some ways. There’s a claim that the African Union has an economy bigger than $2T/year; I’m not sure just who’s in the African Union but most of Africa is pretty poor and I’d have thought that if e.g. Nigeria and/or South Africa is at ~$1T then there ought to be a bigger list of $2T individual countries. The first of these feels like a nitpick; for the second, it feels like maybe I just don’t understand a technical term; the third and fourth feel like memorization more than understanding. The thing that seems most likely wrong is the claim about China and India, though I wouldn’t be astonished to find it’s completely true. … The actual fabrication turns out to be none of those things. I did wonder about errors introduced in the table that was falsified, but didn’t consider the thing that was actually done. It’s a fair cop.
Cell (biology):
Surely 1-5nm is waaaay too small for even a small prokaryotic cell. The 100nm given as biggest typical size for a eukaryotic cell feels way too small too. Are these really microns rather than nanometres, perhaps? Of course some eukaryotic cells get much larger than 100um—consider a hen’s egg, for instance. But 100um feels like a plausible “seldom much bigger than this” figure. (Yup, this is spot on, and indeed these were originally microns.)
Fundamental theorems of welfare economics:
This is not a field I’m aware of knowing anything about, but the two theorems sound familiar (though I hadn’t associated them with the specific title here). The only thing that seems obviously fishy is that the article says the theorems don’t ensure existence of the equilibria, but also says that the second theorem guarantees that “any Pareto optimum can be supported as a competitive equilibrium for some initial set of endowments”, which sure sounds like ensuring the existence of equilibria to me. There are a couple of other things that look strange but I think they’re more likely just badly-written-article strange: “The production economy is more general and entails additional assumptions” (surely to be more general is to require fewer assumptions, but maybe what’s meant is that in the broader setting you need extra assumptions that automatically hold in the narrower setting or something), and “The assumptions are all based on the standard graduate microeconomics textbook” which is just a really weird sentence. There are lots of technicalities wherein errors could be hiding, but I’m no economist and wouldn’t notice. So, my guess is that it’s the “don’t ensure existence” thing. -- Nope, all those are apparently correct, and actually there are some further technical conditions that I didn’t know about. This seems very much in the “memorization” rather than “understanding” category to me; e.g., for all I know “perfect information” is implied by “rationality” as those terms are used in economics. Maybe given infinite intelligence I could have figured out that there needed to be some sort of condition about externalities, but: bah.
List of causes of death by rate:
This says that neurological disorders are the biggest single cause of death. Ahead of both cardiovascular and respiratory problems? Really? Can’t possibly be right. And, indeed, it isn’t. On the one hand, this is “memorization”; on the other, surely this is pretty well known.
Natural selection:
Surely that description of “microevolution” and “macroevolution” can’t be right; the distinction is about the scale of the changes, not the nature of the environmental pressures that cause them. … Huh, apparently that isn’t the issue; the actual error is one I should have spotted if I’d been reading more carefully and hadn’t seen something else I was sure was wrong. I have to say I still think those sentences about microevolution and macroevolution are wrong. Aren’t they? (I tend to avoid those terms, which I mostly associate with creationists who argue that one is real and the other not, so maybe I’m misunderstanding them, but e.g. the Wikipedia pages for them don’t seem to match what’s claimed in this article.)
Working time:
Surely the decreases in working hours shown in that table are too dramatic. Unless, I guess, they’re the result of e.g. increasing numbers of part-time workers, rather than changes to typical full-time hours or something? (If so, then right next to the table there should be a load of caveats about this.) But this claims that, for instance, in France in 1950 the average worker worked nearly 3000 hours per year, versus about half that now. And can it really be true that Greece has longer-working workers than any other EU country? It’s not that any of the individual figures is completely unbelievable, but in the aggregate something seems really off. I’m sure there has been a decrease in working hours, but it surely can’t be this big. (I may have been encouraged toward this conclusion by noticing that the directory containing this example had what looked like data-processing code, suggesting that something quantitative has been changed in a nontrivial way.) Anyway, cheaty or not, I was spot on with this.
So, I got exactly half of them right (and alternated between right and wrong!). For two of the wrong ones I’m prepared to argue that the thing I thought I’d spotted is wrong even though it wasn’t deliberately introduced as a test; for another, I’m maybe prepared to argue that the deliberate error introduced isn’t exactly an error.
Thanks for writing this, it was interesting to read a participant’s thoughts!
Responses, spoilered:
Industrial revolution: I think if you re-read the article and look at all the modifications made, you will agree that there definitely are false claims. (The original answer sheet referred to a version that had fewer modifications than the final edited article; I have fixed this.)
Price gouging: I do think this is pretty clear if one understand economics, but indeed, the public has very different views from economists here, so I thought it makes for a good change. (This wasn’t obvious to all of my friends, at least.)
World economy: I received some (fair) complaints about there being too much stuff in the article. In any case, it might be good for one to seriously think about what the world economy looks like.
Cell: Yep, one of the easier ones I’d say.
Fundamental theorems of welfare economics: Yeah, I don’t think modification was successful. (But let me defend myself: I wanted to try some “lies of omission”, and something where an omission is unarguably wrong is in the context of mathematical theorems with missing assumptions. Well, I thought for math it’s difficult to find an example that is not too easy nor too hard, and decided to go for economics instead. And asymmetric information and externalities really are of practical importance.)
List of causes by death rate: Leans towards memorization, yes. I’d guess it’s not obvious to everyone, though, and I think one can do non-zero inference here.
Natural selection: (I think this is one of the weaker modifications.)
Working time: Deleted the spoiler files, thanks.
Let’s suppose it’s true, as Olli seems to find, that most not-inconsequential things in Wikipedia are more “brute facts” than things one could reasonably deduce from other things. Does this tell us anything interesting about the world?
For instance: maybe it suggests that reasoning is less important than we might think, that in practice most things we care about we have to remember rather than working out. It certainly seems plausible that that’s true, though “reasonining is less important than we might think” feels like a slightly tendentious way of putting it. (I suggest: Reasoning is very important on those occasions when you actually need to do it, but those occasions are rarer than those of us who are good at reasoning might like to think.)
I think there’s more value to just remembering/knowing a lot of things than I have previously thought. One example is that one way LLMs are useful is by aggregating a lot of knowledge from basically anything even remotely common or popular. (At the same time this shifts the balance towards outsourcing, but that’s beside the point.)
I still wouldn’t update much on this. Wikipedia articles, and especially the articles you want to use for this exercise, are largely about established knowledge. But of course there are a lot of questions whose answers are not commonly agreed upon, or which we really don’t have good answers to, and which we really want answers to. Think of e.g. basically all of the research humanity is doing.
The eleventh virtue is scholarship, but don’t forget about the others.
I used Alex Turners entire shortform for my prompt as context for gpt-4 which worked well enough to make the task difficult for me but maybe I just suck at this task.
Thanks for the link, I wasn’t aware of this.
I find your example to be better than my median modification, so that’s great. My gut reaction was that the example statements are too isolated facts, but on reflection I think they are actually decent. Developmental psychology is not a bad article choice for the exercise.
(I also find the examples hard, so it’s not just you. I also felt like I on average underestimated the difficulty of spotting the modifications I had made, in that my friends were less accurate than I unconsciously expected. Textbook example of hindsight bias.)
Ultimately, though, I would like this exercise to go beyond standard calibration training (“here’s a binary statement, assign a probability from 0% to 100%”), since there are so many tools for that already and the exercise has potential for so much more. I’m just not yet sure how to unleash that potential.
There appear to be two edited versions of the Industrial Revolution article. Which one is recommended? (My guess: the _2 one because it’s more recent.)
Thanks for spotting this; yes, _2 was the correct one. I removed the old one and renamed the new one.
This was interesting. I tried the Industrial Revolution one.
I initially thought it was strange that the textile industry was first (my history is patchy at best). I remembered that industrial looms were an important invention but it seemed to me that something earlier in the production chain should be bigger, like coal extraction or rail, steam engines, or agriculture. I noticed that electricity was not so significant until after the industrial revolution. I think my error sensors were over active though—I flagged a lot of stuff as false and only some of it was. Here’s the summary:
Correctly spotted: Electricity too early, Telephone too early, lack of much mention of steam power + coal.
Incorrectly spotted: Textile industry seemed too dominant, Did not know about/expect there would be a recession in the 1830s