How HCI Might Engage with the Easy Access to Statistical Likelihoods of Things

“Everything that needs to be said has already been said. But since no one was listening, everything must be said again.” —André Gide

Together We Are Likely

Context and Creativity

Broken Intuition and Broken Metaphors

The Role of Human-Computer Interaction Research

tl;dr

unintuitive statistical likelihoods of language and vision are now readily available via API; HCI has a vital role in understanding and scaffolding human interaction where our intuitions fail

Together We Are Likely

chatGPT didn’t write this, because if it had I wouldn’t have had to write it. I would have missed out on the iteration, reflection, and wrestling with ideas that writing this manually required me to do over the past six months. I began writing this in November 2022, several decades ago in LLM years.

Lest I set your expectations too high, let me assure you that even 9 or 10 full months of human wrestling still produces prose that is likely. We share too much context, and too many equally capable people are wrestling with the same questions. Oh to be Darwin or Mendel, happening alone together upon a similar context that would lead them to joint paths both statistically likely given priors and also revolutionary.

My hope is thus only that this post is slightly less likely than one written by chatGPT trained on data that is two years old. Soon enough, it will be gobbled up into the training data and chatGPT will assimilate whatever iteration on cleverness this post contains, steadily marching along with humanity into our statistically likely future.

Speaking of several years ago… in 2016, I tweeted:

I thought many of my tweets (toots, now) were clever, but from the macroview of Twitter I was tapping out statistically likely characters, just like everyone else.

Ten years ago, Target used the statistical likelihoods computed over its massive customer data to correctly predict that a teenager was pregnant before she had told anyone, using a “pregnancy prediction score” cobbled together from statistical trends on data like how much unscented lotion a person was buying. It was shocking at the time, in large part because it was surprising that it was possible, and perhaps separately surprising that Target didn’t stop to ask what bad things could happen if they predicted stuff like this.

We are all statistically probable, even Target’s data science team. If they were implicitly predicting pregnancies to target ads, and didn’t think too much about what could go wrong, so too were lots and lots of others predicting things in whatever domains they happened to work in. If a single prediction was surprising and the fallout unintuitive to many, imagine what happens when you train on an unfathomably large corpus of natural language text, which is so core to what we think of as human intelligence, and don’t just make one prediction for a few queries, but rather predict thousands of tokens based on hundreds of words of input.

Human intuition completely breaks down.

I think that is why experts and novices alike can’t quite wrap our heads around what is going on. Each statistically likely token boggles our minds, and there’s hundreds and thousands and millions of them, produced reliably in sequence in response to whatever language we give as input. Is it so surprising that even the experts can’t agree if this is AGI or exponentially-scaled babble, the end of humanity or merely a huge disruption to information work. And, maybe not even that!

Predictably, those arguing for something in the middle don’t get much attention.

If you cross all of this with a healthy dose of humans also being enormously complex and nobody agreeing on what intelligence is, I think you get close to understanding the current state of things. But, this intersection of poorly understood technology and deeply intertwined human interactions is where I think HCI has an enormous opportunity, if we are able to take it on and lead.

Context and Creativity

If all these models are doing is producing statistically probable language, why do they sometimes seem creative? I think it’s a combination of humans driving and curating AI outputs, but also another example of our intuition about creative production breaking down with large data. My intuition is no better on this than everyone else’s, so I’ll give the caveat that maybe there really is some emergent phenomenon happening.

Austin Kleon (who happens to be my cousin, what are the chances?!) has a great series of books starting with “Steal like an Artist”, which has the premise (my version) that oftentimes folks are so obsessed with being completely new, they fail to realize that all human creativity is a collective iteration. We all “steal” from each other, nothing is “new” even when it is new, embrace it.

At the beginning of the pandemic, I made the video below, which got a few thousand retweets. As it spread, at least one person complained that I had copied the idea from someone else.

with @zoom_us 's new video background feature, i can stay engaged and nodding thoughtfully, even when i'm not even in the room -- this is the kind of innovation we need! pic.twitter.com/J4prwOEjxp

— Jeff Bigham 🍂🐿🍁🍁🌻🌻🎃 (@jeffbigham) March 22, 2020

I don’t mean to argue that people are stochastic parrots, or that we shouldn’t fairly compensate the artists that large models learn from (and arguably steal from). But, even as we exercise our agency, we are probable. Distinguishing parrot from probable is difficult and maybe ill-defined, and the difficulty in doing so may also be at the crux of why our intuition breaks down when confronted with large language models.

A couple of years ago, I was surprised to learn that almost everyone I knew was watching reruns of The Office when I was also watching reruns of The Office. But, if I am doing something, it is very likely that others are doing that same thing because of our increasingly shared context (likely, a confluence of Netflix getting the rights to the show and heavily promoting it, and The Office memes simultaneously spiking on social media.)

People who are surprised that someone else is also watching reruns of The Office on Netflix, despite it being the statistically likely thing given the hidden context we share, will definitely be surprised by an LLM that writes a convincing holiday poem in the style of an Office script.

Stochastic parrots in isolation aren’t so intelligent, but when the parrots are integrated into humanity’s distributed cognition they might be a surprising multiplier. Or not.

Broken Intuition and Broken Metaphors

An LLM can create a holiday poem in the style of The Office for essentially the same reasons as why Target was able to predict that teenager’s pregnancy and for the same reason we were all rewatching The Office (there are lots of holiday poems and writings about the The Office in the LLM’s training corpus). Each token output is statistically likely given the prompt and the token preceding it. Or, maybe that doesn’t entirely capture it, experts don’t exactly agree. Yet, this is much harder to grock because it’s not just one prediction dependent on a small number of describable input parameters, it’s 100s of predictions back to back dependent on inscrutable inputs represented in natural language.

These prediction cascades, where many predictions and statistical likelihoods are brought together token-by-token many times in sequence, don’t even seem like predictions. The size of data completely escapes even the experts’ intuition, and we are not accustomed to seeing machines produce plausible language of this length, fluency, and complexity. In the months since I began writing this, the idea that machines can do this seems much less novel, and perhaps we are starting toward disillusionment in the discrepancy between what we imagined based on what we first saw and the reality sinking in as we’ve seen more.

The best models are also massively fine-tuned and shaped by humans, in scales and methods that are purposefully held opaque. Human input is the secret that led Google to originally win search and maps, even as competitors searched for better algorithms. If OpenAI is talking openly about instruction tuning and RLHF, I’d be intrigued to learn all the human work that they’re not talking about. How far can a statistically scaffolded massive human labeling enterprise get in mimicking intelligence? More than that might be happening, but also I suspect it can (and has) gotten pretty far.

Data at the scale LLMs are trained on is so mysterious, we don’t have metaphors for describing how that data is transformed or not into the model’s outputs. Is it copying its training data or learning from it? Kind of both, and also kind of neither. It seems as though at least some of the amazing power of image diffusion models may come from finding and very slightly modifying images in the training data, but copying alone doesn’t capture the fullness of what is happening. The models are so complex it’s hard to convincingly argue against the idea that there are some emergent properties beyond what we might think of as just next token prediction (it’s also hard to convincingly argue against the idea that it’s just next token prediction at a scale we haven’t seen before).

Our intuition and metaphors have fallen short, which also makes it magnificent and also scary. Lacking effective metaphors, broken metaphors tempt us (e.g., LLMs are sentient, LLMs are AI Hype, LLMs are evil, LLMs have emergent intelligence). We’re so eager to intellectually frame these things that we too readily disregard the obvious problems with each of these framings.

What’s the right metaphor for massive statistics over language, constantly massaged by humans, interpreted by humans, iterated on by humans, and curated by humans into something that is legitimately impressively different than what we’ve seen before, even while limited in predictable and unpredictable ways?  Predictably, I don’t know either.

The Role of Human-Computer Interaction Research

I look at this from the perspective of what human-computer interaction can contribute. How can we help humans thrive in a world in which there is easy access to unintuitive statistically likely things.

I bucket the opportunities into - Benefit, Understand, Protect and Thrive

• Benefit – we need to do the hard work to figure out how to leverage LLMs to benefit humans in the things that they want to do. This is the AI Autumn argument, which I also wrote about in a post called “What will the people be doing?” There’s a rich harvest of generative-AI tools and features that people can benefit from – but, there’s no free lunch, even when LLMs make it trivial to get 90% of the way to a feature, that last 10% is going to take at least as long as it always did to design well, maybe longer.
  
  If we don’t design these solutions well, then we’re likely to introduce other problems. One of the earliest models for thinking about human-AI interaction is the utility model from Eric Horvitz. Simply, if the cost of using the thing is more than the benefit you get from it, then people won’t benefit from using it. The original formulation thought about this in terms of error – if the cost of recovering from an error was greater than the benefit from when a prediction was right, it’s a net negative. Even amazing LLMs require fact checking, and errors are just the surface of the issue. For instance, we already know that even slight changes in the sentiment of predictions can lead to changes in the sentiment of what people write, and so we need to make sure we’re measuring utility (or whatever we choose to maximize correctly, which is likely a whole big unwieldy problem unto itself)
  
  I’m not sure that work designing specific prompts to work with today’s specific LLMs will be super impactful long term because the technology seems too likely to change out from under any particular design, but this work will likely be super valuable in the short term in exploiting this powerful new way of accessing statistical likelihoods. Prompt engineering betrays the secret of working with LLMs – they still require a programmer’s mindset to get right. What will the toolkits of the future look like to leverage every-changing LLMs without having to start from scratch with your prompt engineering and design work each time? I’m reminded of the endless complex “crowd algorithms” we designed 10 years ago during the heart of the crowdsourcing era – most of them didn’t matter, but some ideas survived in human-AI interaction. What will be the lasting architectures that involve LLMs that will continue to be useful and important?
  
  Each well-scoped feature we develop with LLMs requires intense one-off design work to get right. I personally think the new AI framings around Safety and Alignment miss the mark here, when what they are really talking about is something that encompasses all of these – design of generative AI. There is ample room for early- and late-stage design tools to help make higher-fidelity prototypes (higher fidelity than the chatGPT web interface) and democratize who is doing the ideation and prototype making.
  
  It’s easy to point out LLMs errors, but tools like CoPilot are super useful… how do we make more interactions with LLMs feel like this and work like this, while still guarding against or designing around the sneaky errors that such features can introduce? I don’t think we’re there yet with most generative AI tools. What will it take so more of what I want to do is preserved and less of what I don’t want to do is required? Success in coding might be a red herring given how formal coding languages are and how accustomed to working with programming systems programmers are.
  
  Current approaches are limited by where we have data – we have tons of text data from the web, we have tons of code and images. What fraction of the human experience does this data capture and what is missing? HCI researchers will need to define where and how to collect and curate new sources of data where access to statistical likelihoods will benefit people. Data will determine where easy access to statistical likelihoods is possible, and thus will scope what innovations in interaction are possible.
  
  
• Understanding – LLMs invite us to examine what we know about our interactions with computers and how we might reshape those interactions given our new access to statistical likelihoods. We could all likely benefit from understanding what part of what we do is statistically likely, how much of that needs to be or should be statistically likely, and what can be learned about creativity or innovation in the differentiation.
  
  Most of what I do is statistically likely – that doesn’t mean it’s not meaningful. In fact, if your intent is to communicate so people can understand, you might want to produce text that is likely given the message you intend to send. When I write a recommendation letter, some of the value of that letter is in the words, and a lot of it is the fact that I wrote them and claimed them as mine. Most of what I did in writing this essay was statistically likely, although maybe not all of it. How much of the writing work did I need to do to benefit from writing it, and what could I have avoided? How do we structure writing as a collaborative iterative endeavor with LLMs?
  
  I want tools that help me leverage the statistical likelihood of things, and I think that means we need to understand the statistical likelihoods embedded in everyday interactions. If we can differentiate likely from creative, rote from reflection, we can empower humans even as they benefit from the new access to statistical likelihoods embedded in what we produce.
  
• Protect, What does it mean to Live in a World of Easy Generation – many people have posited that we’re soon likely to be flooded with generated text, images, voice, video. How do we use LLMs to help us sift through the noise? Maybe it’s the same as the last bullet in recognizing what is or isn’t statistically likely, or maybe it’s about leveraging identities, expertise, and reputation. Maybe I don’t care if something I read is generated text as long as someone I trust has backed it with their reputation.
  
  How do we build tools that help us make sense of credibility? My guess is that humans and human reputations will continue to be the primary way that people make useful shortcuts to gauge credibility. We’ve already seen how social networks can upend our ability to appropriately gauge credibility – when people reported on tweet counts to gauge interest or opinions, credibility was easily manipulated and obfuscated. The surface level problem was the existence of all that disinformation, but the underlying problem was the social mechanics of information spreads – how were (so-called) bot accounts allowed to have influence beyond the people who ran them? What mechanisms should we put in place to help people make sense of it all, even when many human and machine actors may have other motives? That has been and remains a core social computing challenge for the HCI community. The easy existence of fake data might change that in some ways, but it’s a riff on the same fundamental problem we’ve been wrestling with for much longer.
  
  The HCI community also has an important role to play in helping to support the design of technology that uses LLMs. How do we help people creating these tools better appreciate the risks that might come with them? Until we see actual use cases of LLMs, it’s hard to say what forms these might take or what will be most meaningful. I’m not sure if jailbreaking LLMs is all that meaningful, and safety and alignment seem quite abstract without being contextualized within a use case. I think relatively little of our interaction with LLMs is ultimately going to be via general-purpose chat interfaces, and that’s good because open-ended chat is the most difficult application to introduce effective safeguards into. Fortunately, there are so many other potentially more useful applications of LLMs; unfortunately, the straightforward way this plays out is that everyone designing features on top of LLMs will need to do substantial work to produce safe and responsible models. I believe there will be many who do this, but if the easiest path is to not do this then we should expect that – how do we build tools and safeguards to protect users from the harmful use cases that will exist, recognizing the incentives may not be in place for many feature/product owners to do so?

• Thrive, Interaction Beyond the next Token – statistically likely language is powerful and the performance of state-of-the-art models is surprising, but I don’t think predicting next tokens will necessarily lead to fundamental changes to interaction in ways we expect. It will probably be useful to generate text in the context of applications we care about. Generating scripts to automate interactions may be useful in some circumstances, although three decades of work on macros and end user programming makes me cautious because so much of getting that right depends on these automations being predictable, accurate, and transparent, which nobody has really figured out how to do especially well.
  
  Many seem to be betting on chat-like interactions taking a more prominent role between users and getting things done, and that's likely true to some extent -- but the focus on chat over GUIs might have has much to do with the fact that they only control the chat interface and thus naturally think in terms of things being brought into it. It’s no surprise companies like OpenAI are pursuing and announcing endless partnerships. The sticky transformative use cases are likely yet to be found.
  
  I don’t think we should throw away the graphical user interface, and am skeptical of search or automation as the framing of how to integrate with it. Do I really want to have a natural language conversation to get things done? Do I want to cede my agency to a generative model that tries to do the whole task for me? There are reasons why I prefer to book my own travel using a GUI rather than talk to even a highly skilled travel agent, and reasons why I wrote this essay instead of typing, “write an essay on how HCI might engage with the new easy access to unintuitive statistical likelihoods” into chatGPT. Creativity is collective iteration. We will iterate collectively on familiar interaction patterns combined with new capabilities, rather than anything truly different supplanting interaction paradigms entirely.
  
  How do we build for human thriving? Part of the answer letting ourselves be guided by what we want to be doing in the future.
  
  

Humans have never before had such easy access to statistical likelihoods of things, and interestingly the least predictable part about all of this is how it might upend the many things that humans do and care about (or not). My personal bet is that this all will be both more transformative than we think, and also less. In ten years, I expect we’ll be using computer interfaces that would be familiar to someone 2023, yet many of the interactions that don’t quite work fluidly today will work much better – LLMs etc will be the lubricant that helps remove friction, and we’ll get to spend less time fighting with our computers and more time on creative human production.

I believe most of the directions above hold even if we somehow qualitatively move beyond statistical likelihoods and closer to something like AGI … we interact with intelligent humans all the time, and yet nobody would describe that as simple, let alone solved. We’ll still need innovations in HCI to allow humans to benefit, understand, protect and thrive. In a world in which we’re interacting regularly with “AGI”, we’ll need even more innovation in HCI.

Yet, futures in which humans and human concerns are at the center are not the only possible outcome – we will need to make it this way! HCI as a field is core to not only understanding but shaping how this develops, because we are the only field positioned at the intersection of what matters – our work understands people and technology, designs futures that reimagine how people and technology will work together, critically reshapes narratives around benefit and harm, and ultimately builds interactions that enable people to thrive.

chatGPT felt almost like the end when it first came out – “we’ve done it!”  But, we haven’t done it at all. We’re still very much at the beginning of what interaction will look like in the era of easy access to statistical likelihoods. HCI needs to lead the way.

This page and contents are copyright Jeffrey P. Bigham except where noted.
Blog posts are not intended to be final products, but rather a reflection of current thinking and/or catalysts for discussion, like tweets but longer.