MDN can now automatically lie to people seeking technical information | Hacker News Reader

MDN can now automatically lie to people seeking technical information

github.com

181 pointsrobin_reala76 comments

This just looks like the poster child of "hell is made of good intentions".

It may look seemingly harmless at first, even a good idea.

But you'll seed doubt. As that will erode the value of the whole site. Which was its main asset.

Is this being driven by good intentions? It looks like they are jamming the latest AI hotness into a feature without taking much time at all to think it through or implement things well.

I’d bet heavily someone had “ship an AI product” in their APGs

MDN core maintainer confused, to revert: https://github.com/mdn/yari/issues/9208#issuecomment-1615411...

I can't wait for this fad of people trying to use LLMs for things that require factual accuracy goes away (or the tech actually gets there, I guess)

Transformer models are capable of absurd sci-fi miracles when it comes to things that don't require truthfulness but this ain't it

Imagine putting an LLM anywhere health, safety, or regulation are involved. It would be engineering malpractice.

Seeing a lot of it going on in healthcare right now, especially healthcare startups. Wonder how soon regulators will step in and start stomping down on it... it usually takes several years to a decade for any federal rule to be issued so startups may make it through long enough to turn the bag over to someone else.

Don't worry: nobody would be incautious enough to hook a safety-critical system up to a general purpose computer, of all things.

That's not even funny.

A lot of these systems are running actually on Windows, in version that aren't supported since years.

Even if the AI didn't make stuff up, it's unclear to me what value it adds over...

- The existing human edited text already on the page

- The playground to run the code in question

- One of the many existing autocomplete / static analysis tools for HTML / CSS / JS out there. If you want a tooltip explaining what the second argument to a built in JS function in a code snippet does, this is a solved problem!

I've never witnessed a github issue where the comments are unanimous like this one. So far there isn't a single naysayer.

Could probably build fairly successful political party off of this once-in-a-lifetime momentum. :)

One came along some hours after you wrote that. Xyr reasoning was utterly foolish and dopey, and I'm not going to point xem out; merely observe that even there there wasn't unanimity.

Interestingly, there are more negative comments on Hacker News than on the bug discussion. This probably tells something about the audiences of MDN bug commenters and Hacker News readers.

This is not a good look for what is the de facto go-to (from my perspective) reference guide for web development. I’m not too crazy about MDN, but I expected better from them.

I actually think this feature would be fine if implemented better. GPT-4 is perfectly capable of accurately explaining what a short code block does if you prompt it with relevant documentation… which MDN has in abundance, but isn’t using, for some reason.

The mistakes seem to be (1) using GPT 3.5 instead because it’s cost effective, even though GPT 3.5 has severe issues with hallucination, and (2) not including context from the surrounding article into the prompt.

Also, since I assume MDN has a finite number of things that need explaining, why not let users mark an AI explanation as wrong, and submit their own corrected versions?

> Also, since I assume MDN has a finite number of things that need explaining, why not let users mark an AI explanation as wrong, and submit their own corrected versions?

I mean if you implement it well you don't really need the model. If anything it would be more effective if it was always just slightly wrong.

I can't imagine any situation where it would be acceptable. I have personally wasted hundreds of hours over the course of my career due to incorrect documentation. There are millions of programmers in the world. Incorrect documentation has surely caused hundreds of millions of hours of wasted labor, or tens of billions of dollars of economic damage. That's just the damage of wasted development time, not the harm buggy software causes users.

For me, the most obvious problem is that it’s often mixing wrong information in. I found a couple of very simple examples where it accurately explained things but anything longer tended to have multiple incorrect points

On a deeper level, though, I don’t know that it’s actually useful. I haven’t found a single example where it added anything which wasn’t present in the authored text, and the explanations tend to be very pointless mechanical descriptions which miss the larger point of why you’d want to do things. On the popover API page, for example, it had a bunch of regurgitated comments before telling me that you’d need to write a bunch of JavaScript to actually toggle popover visibility when the whole point of that API is not needing to do that. That was also interesting with the incorrect claim that it was a non-standard attribute and you’d need to implement it using a JavaScript framework like Vue.js when you think about how much code someone might write if they relied on that explanation.

Does GPT-4 really never hallucinate?

This is the first time I've personally heard someone claiming anything like that, though I don't tend to do anything with LLMs (due to this bullshit factor).

Anyone who says that GPT-4 doesn't hallucinate is either clueless or an AI grifter, though I don't think GP was claiming that, just calling out that GPT-3.5 is really, really bad at it.

Cool, thanks for clarifying. Was hoping they'd managed to crack that nut, but doesn't seem like it yet. :)

Hey ChatGPT, please generate a Firefox Corp-style apology for this snafu... Be sure to use some bullshit blah blah for our original aim of wanting to enhance user experience and listening to the users' feedback, and some more bullshit blah blah about doing better in the future.

Who owns MDN? I thought it was disbanded by Mozilla few years ago.

The team was disbanded a few years ago, which led to the creation of Open Web Docs (I think it is charity, the website: https://openwebdocs.org/ ) where several former editors are still maintaining MDN content. It looks like Mozilla has rebuilt a team of developers though and is trying to extract money from MDN (ads, AI, …), though they don't contribute significantly to the content nowaday.

It is now maintained by this club, I think: https://openwebdocs.org

FWIW, OWD writes open source technical content and is not responsible for the design of MDN nor the blog, ads, or login and membership features of the site. That is done by Mozilla.

The OWD team writes technical documentation on APIs, HTML, and JS. OWD also works on information architecture and browser compatibility data. They contribute mainly to https://github.com/mdn/content/ and https://github.com/mdn/browser-compat-data/.

Mozilla, not OWD, is responsible for Yari, the platform behind MDN Web Docs (https://github.com/mdn/yari). The MDN blog, ads, AI and design, the MDN infrastructure, are fully owned and controlled by Mozilla.

I know some folks who work on this content and TBH it’s the folks we’d want working on this content.

No, these people write only docs. Mozilla seems to have rebuilt a dev team, after they disbanded the MDN team a couple of years ago, with the goal to extract money from the community work.

ChatGPT: I apologize for any inconvenience caused by the recent addition of the "ai assistant" feature on the MDN website. We understand that this feature has generated misleading information, leading to confusion and frustration among users. As an AI model, I acknowledge that the generated text may contain convincing falsehoods or inaccuracies.

At Mozilla, our goal has always been to provide accurate and reliable technical information to the developer community. Unfortunately, in this instance, the introduction of the AI feature did not align with that objective. We deeply regret the impact it has had on users seeking reliable information.

Our intention with the AI assistant was to enhance the user experience by providing additional context and explanations. We wanted to create a trusted companion that would assist developers in understanding complex concepts more easily. However, we recognize that this implementation fell short of expectations and introduced more problems than solutions.

We value the feedback and concerns raised by the community, and we take them seriously. Going forward, we will reassess our approach to ensure that any new features or enhancements undergo thorough review and testing. We will prioritize human oversight, peer-review, and fact-checking to ensure the accuracy and reliability of the information presented on MDN.

Our commitment to providing a reliable resource for developers remains unwavering, and we apologize for any loss of trust this incident may have caused. We appreciate your patience as we work to rectify the situation and improve the overall user experience on MDN.

Thank you for your continued support, and we assure you that we will learn from this mistake and strive to do better in the future.

Sincerely, [Your Name] Mozilla Community Manager

https://chat.openai.com/share/2649a208-7987-4c30-9bed-0738b8...

Given the clear nature of GPT to simply spew things that sound “right”, why is a technically proficient organization such as this one jumping headfirst into integrating this with their products where it is clear that is not yet up to this task?

LLM’s solve all the problems with expert systems (remember those?) by removing expertise.

This is not a criticism.

General intelligence is not represented by the graybeard in the corner office. At best it’s a knucklehead in the next cubicle.

@dang: Title is clickbait. Perhaps replace with something like:

Bug report: MDN's AI Explain feature generates false explanations

not really, it is in fact generating lies automatically

IMO "lying" implies intent, and "MDN" misses that it's specifically an LLM "AI help" assistant that most people won't yet be aware of.

From the title I was thinking along the lines of MDN automatically swapping out an article on encryption to falsely claim a function doesn't exist depending on country to comply with local laws, for example.

You're right. It should be: "MDN LLM assistant generates bullshit to people seeking technical information".

The HN guidelines are to use the title of the source article, which this does - and in this case it seems pretty accurate.

> Otherwise please use the original title, unless it is misleading or linkbait; don't editorialize.

-- https://news.ycombinator.com/newsguidelines.html

Now if only an AI could explain to me those intractable terms and contracts of BigCo's. /s

https://tosdr.org has existed for years before the current LLM wave

that's pretty nice although I really don't like how it puts a value judgement on everything. Characterizing things in a way that's not just "good(-er) and bad(-er)" would be nice.

For instance, on the "pornhub" one they put in red that the service can monitor your "private" messages. I can reverse the value of that: with the problems of sex trafficking, underage exploitation and other types of crime that could happen there, we should be happy the company is dedicated to doing that. All of those red items in fact look like things I've got no problem with there.

Did the AI Explain button replace a more reliable feature?

With placebos, the correct question is: does it displace more efficacious procedures.

I'd say yes. Even if something is a placebo, it's ability to displace better solutions is the problem

I don't use MDN. I'm just asking if they removed something when they added this feature.

Characterizing LLMs as placebos or "designed to produce gibberish" is inaccurate.

They could for example only use LLM's to match query's and use as source for responses like many other -docs as AI- implementations do, wondering why Mozilla wouldn't do exactly the same. as Hallucinations are a normal occurance if not doing so.

I agree it's an odd implementation and seems bandwagon-y in this context. Maybe it's the first step in a larger plan?

I compare them to placebos because they can be effective and they can have zero active ingredient.

That's a fairer comparison that I read it as. I'd say they're more like a pharmacy, usually dispensing medicine. But we don't really know how half the medicine works, and sometimes it hurts people.

Would you shop at such a pharmacy? I wouldn’t and I don’t.

That accurately describes all pharmacies. Medicine is hard.

Oh, please. Have you never read the little insert that's packed in with all prescription medicine? Whether the mechanism is fully understood or not, it's at least been studied and demonstrated to have some benefit, and has _specific_ warnings about the ways it's been seen to harm people and how often.

Unless you're referring to the scam-filled supplements aisle, in which case I agree wholeheartedly with your comparison.

I was thinking more about SSRI's, which make up over half (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8450889/) of US pharmacy orders. While we do know that they work for some people, they only work for 20% of people (https://www.ncbi.nlm.nih.gov/books/NBK361016/). They also come with common side effects that can lower quality of life (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2719451/). And we do not understand how they work in the first place (https://www.discovermagazine.com/health/how-antidepressants-...). Plenty of parallels with the common lamentations over generative AI.

Your own links demonstrate the fundamental difference. The side effects have been studied and can be identified. I'm at this very moment taking a break from a medication because I was able to recognize a severe side effect that came up. I can't recognize when an LLM has generated a nonsensical piece of technical documentation; if I could, I wouldn't need to use it in the first place.

The side effects of incompatible code are similar to the side effects of incompatible medication. Thanks to the identified undesirable effects, you usually identify the problem and resolve the issue with other resources. Occasionally the problem is more severe or persists unidentified. We should definitely be working to minimize these effects and prevent them where possible, but that doesn’t mean we stop trying/utilizing medicine. These are honestly just all the same arguments used against Wikipedia.

- “Incompatible code”? What does this even mean? “Incompatible” with what?

- Have you actually read the examples in the linked github issue? What “undesirable effect” are you able to identify when an LLM tells you that a JS function simply doesn’t exist? Or when something fails subtly enough that the code you wrote based on the LLM’s advice isn’t the obvious culprit? Do you do much debugging?

- Wikipedia requires references, while this LLM directly contradicts the one data source that it’s meant to help explain. Failing to see the comparison.

If a medication is incompatible with a person's specific biology, they don’t see improvement (or their health declines). If a function is described incorrectly, you won’t see improved functionality (or you’ll see bugs). It was incompatible with the syntax/compiler/the rest of the code.

Though this one won't, LLMs will cite references if asked. They’ll also provide step-by-step rationalization (though they don’t think in that way). Just like Wikipedia, you still have to check those references to verify.

> If a medication is incompatible with a person's specific biology, they don’t see improvement (or their health declines).

And the specific ways that their health might decline have been studied, a list of which is included with every package of the medicine. Maybe you even get regular bloodwork to check for the specific side effects this medicine’s been seen to cause.

> If a function is described incorrectly, you won’t see improved functionality (or you’ll see bugs).

Hopefully! Again, how much development/debugging do you do on a daily basis? Because real developers don't accept the notion that it's ok for bugs to occur as a result of following reference documentation.

Relevant username.

But I thought it was supposed to be my Trusted Companion… That phrase had my eyes rolling

This is a heavily editorialized title.

It’s just a feature that explains a web platform doc using LLMs.

They can be inaccurate sometimes, but as long as that’s clear to the end user (in this case, developers) I don’t think that’s a problem. Heck, many of us do this day-to-day with ChatGPT.

In this case, the sin was not making the potential for inaccuracies clear enough.

I cannot see anywhere that MDN say the explanation might be wrong.

Im pretty sure they use a vector storage database against their docs/resources so when you query they can just parse those sources and produce a free hallucination response to it.

No need to make things up if you have the sources anyways.

“Hallucination free” or “Free hallucination”?

The latter would be very different interpretation and some here would probably agree with that accusation

"Hallucination free"? No, no, no, it's "Hallucination? Free!!".

https://youtu.be/5yuL6PcgSgM

The linked bug report has four examples of it being wrong. And the bug reporter mentions there's no workflow to submit corrections.

Also hallutination was probably the wrong word to use for LLMs being incorrect since it might imply they could be correct if they would only just focus on the right input data rather than, say, simply not understanding complex technical concents, but I guess that ship has sailed.

stand corrected, didnt read TFA

> In this case, the sin was not making the potential for inaccuracies clear enough.

I completely agree and the tone of some of the comments in the issue are disproportionate to the problem.

I don’t really understand viewpoints like these.

The submitter knows the AI isn’t lying and the submitter is being purposefully obtuse in order to drive controversy.

Hallucinating, lying. What difference does it make when it's giving you an inaccurate, possibly made up explanation of an API?

> The submitter knows the AI isn’t lying

You've somehow internalized something as "obvious to everyone" that's actually very wrong and dangerous about AI. Of course the AI is lying. Or, perhaps a better term would be "bullshitting", defined as "sounding convincing while having no regard for the truth". By that definition, bullshitting is all LLMs ever do.

That's an ungenerous way of describing direct access to semantic space. gAI is creative. Rules-based AI is logical. We need them both.

It may be "ungenerous" but it's accurate. Unlike the output of LLMs.

This is a good point. The problem is that most people simply don’t care about these nuances. If the AI hallucinates or is wrong, then in their eyes it is lying.

It doesn’t matter if that opinion is wrong because its still the opinion that matters.

But MDN's "AI Explain" is just straight LLM output, right? It's neither generative (LLMs are transformers, not GANNs, right?) nor rules-based: it's predictive. We need that too. But not for technical documentation! Not on its own!

Aye, I agree this sounds like a strange and ill-fitting implementation. And it bums me out, because the more LLMs are used incorrectly, the harder it is to get people to use them at all.

You’re worried about people not using LLM’s _enough_???

On the user side, yes. It's an uphill battle getting most of the people I know to understand and use ChatGPT.

The accusation is not that AI is lying it is that MDN is, especially given they are showing it at the same level as curated source with reputation for high degree of quality without full disclosure at error rates to expect

We don’t need more garbage on our internet. MDN has been a high quality source of information forever now. This is a large step backwards. That’s the viewpoint.

Reply on news.ycombinator.com