Chatbots-Are-AI-Antipatterns | Hacker News Reader

Chatbots-Are-AI-Antipatterns

hello-jp.net

49 pointsjpoersc59 comments

The natural-language chat box (or voice command) is seductive as an interface because it elides the question of how somebody will use a product. It makes a product that uses it seem more potentially capable, because you can imagine asking it anything (we all remember the sense of possibility in the first Zork-style games). If the user was provided with an affordance to understand the range of capabilities of the product, they might realize it's too limited for them, or realize that their use case is served by something else.

The chat/voice interface is like Skinner's Box: we keep "using" because we don't know what will work and what won't, and when something works it seems magical. And it keeps being pushed on customers because it's beneficial for corporations: customers will fruitlessly navigate chat support rather than realizing they should give up and call, businesses will invest in products with open-ended interfaces because they can convince themselves it's flexible enough for their needs.

>And it keeps being pushed on customers because it's beneficial for corporations: customers will fruitlessly navigate chat support rather than realizing they should give up and call

Businesses tend to thrive when consumers choose to keep using them, not when they don't. How does users fruitlessly navigating chat support serve that end?

It reduces costs in the short term by firing all your customer support staff.

There is no long term in modern capitalism.

That’s true in theory, I’m not sure about in practice. It could also be the case that a company says “human support is too expensive, throw in a chat bot to stem the tide”, see the number of calls drop, and call it a success. Never mind how many potential customers are lost (which actually requires competence to measure.)

That's a _long-term_ problem, though. This would be far from the first time that companies prioritise short-term savings over long-term revenue due to perverse incentives.

No-one's getting promoted for "declined to ruin our customer support, thus avoiding loss of X recurring revenue per year over the next decade", but they might get promoted for "reduced cost of customer support by X with AI!!!"

In the long run the invisible hand will dole out rewards and punishments here, but thinking about the short term is just _easier_ than thinking about the long term.

> In the long run the invisible hand will dole out rewards and punishments

Only in competitive markets.

The only way I'm using a command entry if it's the result is deterministic. Fuzzy searching over the choices is nice, but I still want that to be deterministic. And some way to determine the actual choices available.

There's so many studies done on human computer interactions that at this stage, I'd believe it's just pure malice. Corporations want to push the idea that their tools will let you do your task without being trained in it, which have never happened even with physical tools. Even game console and smartphones require some guidance from other people.

It's very funny that this article that I mostly agree with has a chat bot in the bottom left corner that wants to discuss why chatbots are a bad solution

> It seems we're exploring why Jakob and I think chatbots might not always be the best solution! I see you've already been introduced, so let's dive deeper. Given our discussion on chatbots, what are your thoughts on the hybrid interfaces Jakob mentions towards the end of the article?

Chat is a terrible primary interface for working with people, too. It's fine for cursory stuff, but it doesn't suffice for anything that requires high information throughput. You can use it to follow up on a conversation that you had in person or in a video call, you can use it to check in about something that was fleshed out on a real doc, and you can occasionally use it to brainstorm a quick solution to something.

But at some point in real work, you always need to move up to a higher bandwidth medium. If that medium is going to be text then it's got to involve rigorous specification to avoid ambiguity. If it's going to be quick and casual then it has to have other channels to convey meaning than just the words used—real humans use body language, tone of voice, and back channeling to ensure the message is getting across with high fidelity. There's no such mechanism in chat, which makes it frustrating to use with people or with robots.

I suppose those higher fidelity things are why it's so exhausting trying to voice-command a bot like Siri/Cortana/GA— at the end of the day, they're just doing a straight transcription into text, and then operating on that text.

So rather than the machine benefitting from that extra info, it's actually just making it worse by introducing another link in the telephone chain; hence users getting exasperated and giving up to go back to conventional HMI paradigms of keyboard/mouse/touch.

IDK, to me text has higher information density and throughput than a video meeting. Ever listened to a swagger file?

That really depends on the meeting and the nature of the information being shared.

If you’re communicating purely technical information about a subject everyone is already aligned on, text is fine.

But the information density present in facial expressions, vocal inflection, body language, etc. is significant and much is lost without it.

Meh, I go to three types of video meetings. The ones where the speakers have their cameras off, the ones where noone watches the screen, and the ones where its a screenshare.

You are right, though. If all your meetings are about persuasion, video and audio will be best.

If someone is in a technical role and most of their meetings are like this, that's a bad sign.

I'm not contrasting voice with text, I'm contrasting chat with voice and with most other forms of text.

Text is excellent for communication if all involved parties are decently literate.

This, admittedly, means it's very often not excellent.

Text != Chat. Chat is an abbreviated form of text that isn't particularly great for most purposes even among literate people.

When I was 14 I went to live alone in a motel room in Vermont (long story). I had little money left for food, and one weekend had to go without food at all. But I had a bag of sugar, so I reasoned that I could live on sugar alone— it was calories after all.

I discovered that eating nothing but granulated sugar for 48 hours made me quite sick.

Watching companies force chat interfaces into everything is like watching a bunch of kids discover that you can’t live on a diet of sugar. They won’t listen until it makes them good and sick.

This reminds me of the time I had a horrible ant infestation in my dorm room shared kitchen (also had a horrible room mate infestation, but separate story). Lines of ants everywhere through the kitchen, but I had a 5lb bag of sugar.

Went outside, found the ant hill right next to the building window and dumped the bag of sugar on the hill. Half of it was gone in a day! The next day a little more, but there were NO more ants in the house. The next day there was still sugar left and the day after that. The ants started stopped coming out of the hill much, and I think they had filled up their nest with so much sugar they weren't getting any other nutrients. After a week there was still sugar, but they didn't come back into the kitchen for the rest of the term, when I moved out.

I love this story. And I love when people reply to me with another story.

A ui is worth 1,000 tokens (at least).

I agree that chat isn't the end interface for AI powered applications.

That said the ability to describe desires in "fuzzy" ways is a usability value unlock.

My take is that natural language should be use surface functionality, translate between structures, and generate personalize defaults.

Its a hybrid approach and why my co-founder and I started https://tambo.co

It's only useful if the backend is wired to act on the request.

LLMs handle the query understanding part, but it is still up to the backend developer to implement the functionality to fulfill the various possible requests. And the opaque nature of the interface gives the user no clue as to what requests are supported. We all experienced this problem with smart speakers, with their ever-changing features, and resigned to using an increasingly limited set of features.

> That said the ability to describe desires in "fuzzy" ways is a usability value unlock.

Ctl-Shift-P does the job where available. Not entirely sure how AI makes it any better.

I think there’s merit in a hybrid approach, and more people should do this instead of slapping a chatbot everywhere, but your approach seems completely backwards from a usability perspective. Chat interfaces (edit: and similarly, CLIs) are at complete odds with discoverability[1]. In your demo, I’m trying to understand the point of the chatbot when all it does is convert a button, with a pretty decent CTA (“try sending us an email”), to a sentence in the first person? Why not a large “Compose Email” button? And if the chatbot can do more than one thing, how many of these potential “conversation starters” are you going to display?

What we need are designers who can help establish a foundational structure (information architecture) that leads to discoverable and simple UIs to nudge users in the right direction. Once users are at a place where they know what’s possible/available, then perhaps you can allow some fuzziness to help them cross the line to accomplish their task.

[1] https://www.interaction-design.org/literature/topics/discove...

Oh, do I agree…

Don’t trap me in a chat window https://austinhenley.com/blog/trappedchat.html

Natural language is the lazy interface https://austinhenley.com/blog/naturallanguageui.html

> You want to claim a return in an online shop. Which customer service experience do you prefer? An agent (human/AI/hybrid) via chat or phone vs. a well-designed self-service GUI? I asked approximately 20 colleagues and friends across age groups. 18 mentioned they would choose the visual UI.

These friends probably never had to deal with Google support.

"A well-designed self-service GUI" doesn't exist once the company is optimizing for the opposite of what your are trying to do (cancel subscription, return item, file insurance claim, etc.)

Then use the regulatory pen and make them.

Like good grief I know we're insanely politically polarized at the moment but I think every single person in America except for like 200 executives would get behind a law that says "cancelling a subscription must be exactly as easily done as getting one."

The market has already decided. If it was important for people, companies that offered easier cancellation would out-compete with those who don't.

Fuck the market. If that is indeed what it decided, it’s wrong. Markets are wrong all the time. I borrow a line from Rick and Morty, the boos of the market mean nothing to me. I've seen what makes it cheer.

There is no reason at all to require a market to give us permission to make a better society.

> There is no reason at all to require a market to give us permission to make a better society.

Better societies but without good market couldn't compete technologically and militarily, so ultimately they lost. Borrowing a line from Rick and Morty: that was always allowed.

A lot of people don't get that the market isn't a reflection of reality. It is reality - the only reality. That's the point of market realism.

That's... a pretty amazing misunderstanding.

Organized human societies predate markets by millions of years. Markets are reality now, kind of, because of the utter failure on the part of liberal governments to keep them regulated and in their place. This is not a desirable nor permanent nor sustainable situation.

Capitalism subsumes all critiques. There will never be anything other than the permanent market.

Tell that to the coal miners who parked trains on bridges and then blew them the fuck up, and got us weekends and 40 hour workweeks.

If people are too rich, dumb, lazy or some combination of all three to not cancel a subscription because of the "barrier" of several web-pages and reading then they are extremely unlikely to use their money for any productive efforts.

I trust Amazon to make society a better place than lead-brained boomers/millennials who can't figure out how to cancel a subscription

> If it was important for people, companies that offered easier cancellation would out-compete with those who don't.

That's not how the free market works. The free market works companies prioritizing profits over people. So what's "important" for people is only visible in the market when it's also profitable.

No it hasn't. Even the most zealous believers in the free market know this isn't true (assuming they're not wildly naive, too).

Cancellation policies fall into two well-known failure modes of markets to actually indicate what they want: information asymmetry and inconsistencies in time horizons between counterparties.

You have no ability whatsoever to assign market preference to this without at the very least ensuring consumers are actually aware at time of purchase of the friction they'd face at time of cancellation AND the chances they'll want to cancel.

That's all leaving aside that a person's preference for Product A over Product B obviously does not mean that a person prefers every dimension of Product A over Product B.

You forgot to add /s.

1. Customers don't know what cancelling is like before they buy a service. This isn't advertised, obviously.

2. Competing on being easy to cancel is a bad thing, actually. However wins that competition is losing money. That's why you see the competition go in the other direction. As in, who can make the shittiest cancellation interface (Planet Fitness, btw).

3. People who cancel a service are cancelling BECAUSE of the service. Why would they go back to a service they disliked enough to cancel, if the service stays exactly the same? They wouldn't. Meaning, "good cancellation" doesn't count for anything.

So it doesn't help you retain customers, it doesn't help you get back customers you lost, and it doesn't help you get new customers.

> every single person in America except for like 200 executives

Oof sounds like the base is pretty divided in this one. Better send it to committee where it can be studied forever.

I laugh because the alternative is to scream.

I've been thinking for a while, that since these chatbots will be optimized for profit (if accounting and legal can be profit centers, why not customer service), we'll see all the same anti-patterns. You won't be able to cancel the monthly payment, or return the part, or fix the excess billing fee. Instead they will be highly trained in giving you the polite frustration and screaming anger without paying a human to absorb it. Likely, they will also gather any additional saleable personal information and legal justification for doing so.

Would you like to sign up for our new personalized support for only $4/mo? The first month is free with credit card - I mean the Nigerian scam works because of self-selection, amiright?

A well designed chatbot is better than a well designed UI for this use case. A “well designed chatbot” could absolutely do something as common as claiming a return in an online shop.

We gotta compare apples to apples.

The issue, as I see it, is that most of these companies do not have well designed system to begin with. They then attempt to slap on a AI chatbot to deal with their total lack of a coherent systems design.

If these companies had good design, good processes and clearly defined processes, they wouldn't need a chatbot. They could add one to make querying better/easier.

Let's say you go to your ISPs website, because your internet connection is down. Ideally their frontpage would have a direct link to "operational status", but let's be real, it doesn't. You eventually find the status page, it doesn't mention anything about outages. So you call or ask the chatbot, in neither case will you get an answer, because neither the person on the phone, nor the bot has ANY clue that you're internet connection is down. You then switch to a different ISP, who has procedures and systems in place and a link to their status page and they update it before you even think about contacting them. No amount of AI chatbots will fix the problems for the first ISP, because their internal systems suck and can not provide the bot with the required information anyway.

Most interactions with ChatBots and customer service could, and should, be solved with better processes and better systems. We're not even at the point where we can compare the well designed UI to the well designed chatbot, because most businesses can't provide the data and processes for either.

In my experience dealing with Google support across ads, workspace and cloud, the self-service GUI has been preferable because the self-service GUI typically provides zero utility.

Chatbots are nice for fooling around and that’s what most people use them currently, like I want chatbot to entertain me chat interface is perfect.

GUI is great when I have to do something and I need guidance to do it correctly because in GUI I can see what is expected from me in a glance, I can see required parts and parts I can gloss over.

Let’s not forget about CLI which is perfect when I exactly know what I want to do and I do know how to do it in least amount of motions.

People who think chatbots are great think that natural language can be efficient like CLI but it is not.

> Expecting our users to write a pitch-perfect prompt is like asking the average person to control their computer via the command line.

I admit I lean more on the "command line pilled" side of things, but to me this analogy implies that the chat interface, specifically prompting, are fundamentally intractable as an interface to AI.

Every operating system is necessarily built on compiled/interpreted text, and therefore has the ability to "fall back" to a command line interface to have finer grained control over basically everything.

Even mobile devices, completely 100% GUI, where native terminal emulators are unusable (except for the truly dedicated) are wired this way as soon as you scratch the surface of app development.

I think this analogy has to hold for AI, no? The entire current ChatGPT renaissance is built on the concept of prompting. There are no examples of breaking the mold in this article that don't rely on an underlying "chat" paradigm. Voice agents? Just turning a vocal interface (something predating GPT3) into... prompts for an LLM. Gemini summarizing documents? A button that has a prompt pre-written and ready to go.

Carrying forward the analogy, there will be a "GUI" where your AI agent will provide a streamlined experience that is extremely powerful in matching your context most of the time, but then if you want to tweak anything yourself one of the most powerful ways to do this is to get down into the "command line" of what chat interfaces control what or how these underlying prompts are structured/ordered.

The hate against chat interfaces is missing the real opportunity here. Consider the calendar interface the article describes: that we've managed to actually make it function within chat means we've roughly automated pretty much everything except the human interaction that's left. That's pretty amazing!

The article goes on to claim that hybrid UI's are the future. I disagree, the future is figuring out how to automate prompt generation so I don't need to even interface with the UI to begin with. Hybrid UI's are a patch over something that can't be efficiently automated yet.

The opportunity is the "yet".

What does this even mean? What do these magical AIs with no interface do?

Infer what you're doing already based on your actions.

To take a concrete example from the article, the Cursor screenshot includes a prompt "Please refactor these components ..."

When was the last time I told an engineer on my team to do that? They automatically do that because they feel like it's a good idea. That's a no-interface magical AI.

An AI that refactors with no human interaction sounds like an absolutely terrible idea. I don't want to be anywhere near a codebase where that's enabled

There's a bit of irony in this article having a huge distracting button asking me to interact with the author's chatbot.

What about a chatbot that can render HTML for you, including forms of which it can read the results when you hit SUBMIT, and a text window where you can add extra clarification that it will read?

This contrived post sets up a false dichotomy singular communication channel. There's a difference between command input and data output. You can whatever mode is suitable for each.

At a previous employer, I pitched tooltips as a potential way to surface AI interactions.

I may still try it in an app I'm building.

Docs -> RAG -> tooltips / onboarding sounds like it might work out ok.

Tooltips popping up for user feels like Clippy ;)

^^ I have vague memories of Clippy, but I remember it as obnoxious, often consuming the precious screen real estate on the low res monitors of the day, without offering anything valuable. But tooltips on the web with CSS and libraries like floating-ui can be much more compact, agile and barely noticeable.

I have tried showing help text in an internal app using tooltips when the user would hover over the target element (or show a small icon on touch devices), and the feedback was good as the tooltips were never in the way but easily available for help (accessibility for keyboard users needed some thinking, but for the limited audience for that app, it was not a problem). And while you're at it, may be make it more engaging than a simple text only tooltip (which can be done without any intelligence), and let the host to customize and offer complex workflows.

its because people integrated llm based ai into their existing guis as quickly and thoughtlessly as possible. ie. in the form of an annoying hovering element with a speech bubble protruding from it, and once you begin talking to it, you realize the range of topics it will respond to is very shallow. If at the end of your blog post you could type in "what about the stupid chat thing in your own blog" and it would respond intelligently, not gaslight you etc. that would perhaps be something.

The point is OK (ie, a car with no steering wheel just chat) but pretty obvious and seems like the author is a hypocrite in this regard

Reply on news.ycombinator.com