Revisiting Generative AI in psycholinguistics

In late April 2023, I wrote a blog post on ChatGPT and its incorporation into academic research. Three years later and things are really basically the same. More or less, the landscape is identical to what it was three years ago, except that ChatGPT has a number of competitors, and OpenAI appears to be losing footing because the systems they have developed are not capable of producing behaviors that solve NP-hard cognitive science problems. Some have pointed out that AI systems are appealing to higher education administrators because their jobs are automatable. Others, particularly researchers in STEM fields, feel entitled to use AI to get the un-fun work out of the way (writing papers) so everyone can spend more time on everyone's favorite part (devising experiments).

An increasingly large community of psycholinguists have betrayed open science practices by their reliance on closed systems whose data sources and behaviors cannot be traced. A person whose work I deeply admire told me that maybe I could consider asking Claude for advice navigating the academic landscape in 2026. The University where I work used ChatGPT to rewrite our faculty webpages so that they could be more appealing to prospective students. Sadly, the resulting pages are mostly hallucinated garbage that is homogeneous, generic, and unappealing to people who are highly adept at identifying AI-generated material. Nothing says "Study Linguistics!" like a page that looks just like "Study Chemistry!"

I have not seen many signs that Generative AI has been widely adopted to create psycholinguistic stimuli, but there are some persistent issues that have not been addressed. Namely, as with industry researchers in tech, many of my colleagues willingly admit to using GenAI to solve problems in their workplaces for which specialized tools exist. Somehow, people still not recognize that the purpose of widespread adoption of commercial, walled-garden AI systems is not to empower people, but to create a subset of individuals who will profit in the short term (say, in the form of publications) until they themselves become disposable, while also actively harming the career prospects of others who choose to not use AI for whatever reason. People in this scenario are playing a kind of classic capitalist game that is liberates no one, creates spare time for no one; AI is a curse. It is a curse to always working faster, more efficiently than before, and every moment you gain is time you'll be doing something else you also don't want to do.

I wish to state that I am still not, as in 2023, opposed to automation across the board. There are many things that are useful to automate, such as things humans are likely to get wrong. However, I think it has become clear to me in the last three years that the mass deployment of artificial intelligence fundamentally undermines the human elements of cognitive science, the human elements of research dissemination and the construction of knowledge. Moreover, the use of Generative AI tools promotes uncritical thinking, deskilling, and actively encourages offloading of cognitive capacity to tool-shaped objects. This is especially salient for writing and science communication more generally, but also applies to learning to code. Speed-ups that one observes with systems that have enormous human and environmental costs will be completely obliterated if these systems are ever taken offline.

In 2023, I wrote a list of concerns I had about ChatGPT specifically as a commercial product that limited (and continue to limit) its effectiveness as a stimulus generation tool. The points below are ones that I am revisiting here because of their consequences for how we live on earth, and how labor is threatened by encroaching tools.

1. Replicability - The systems are not reliable

2. Transparency - Unknown data, "technology is magic"

3. Computation will cost us millions of lives

4. AI is anti-labor

5. AI systems will profit from your labor, and will never pay dividends

Now that my focus has shifted to the anti-labor orientation of academic administrations, and the perpetual do-more-with-less attitude that pervades academic research in the modern era, I wish to revisit these points in the context of the conduct of research in general. So that prospective readers can turn around if they wish, and avoid being frustrated with me and this perspective, I wish to point out that I know I am not the first person to say these things. My goal in this post is to convey my perspective on how Generative AI undermines academic labor, and for that reason I wish for my work to be completely free of generative AI use. If the lack of support for Generative AI prevents me from obtaining, say, tenure, then I suppose the field would only become more frustrating to work in. I am proud that my CV represents my own work and the work of my collaborators, and is not plagiarized from the works of others. I have never automated a single aspect of my research or teaching process using Generative AI systems and never intend to do so.

With that in mind, let us consider these points in turn. Text marked between == indicates writing from my 2023 post.

Replicability - The systems are not reliable

Within the field of psychology, replication has come to play an increasingly large role in establishing whether effects on cognition, behavior, or other mental states are justified by the data we collect.

==With Generative AI, the replicability question is altogether different — in that it is not always possible to get the service to give you the same answer. Randomness is built into the architecture for generative models to prevent the model from always producing the highest probability linguistic sequence. Typically, a model then selects the continuation or response that has the highest probability. ==

Generative AI systems are approximators, and do not solve tasks in the same way every time. In this respect, these systems are like people. However, in the context of a research and analysis pipeline, the growing use of Generative AI jeopardizes our ability to conduct replicable research. Automatically extracted annotations can vary from run to run, change systematically---and unsystematically---depending on the prompt, and also differ substantially across different versions of models by the same provider, across different providers hosting the same model, and across different computational systems (e.g., the Crab versus Chat).

From the point of view of the encroachment of Generative AI systems on everyday labor tasks, it is clear that the management class is less interested in the quality of the output and more interested in finding ways to eliminate jobs, make them more dependent on their employers, and line their own pockets. In the context of the University, the appeal is to eliminate instructors, to eliminate advisors. The "advisor clone" idea that appears to appeal to many administrators ultimately means you can hire half as many advisors. If you make more copies, you can fire even more. Administration simultaneously advertises the University as a "place-based institution" and yet treats undergraduate and graduate education as a series of micro-transactions more or less accomplished through self-checkout.

Transparency - Unknown data, "technology is magic"

With this in mind, you have no direct control over the generations produced by ChatGPT. This becomes even clearer when we might want to generate outputs with even more constraints. Unlike GPT-whatever’s predecessors, you cannot inspect the working components. That means no embeddings, no logits, no indices or ghosts of the representations that are being calculated. While we assume that most systems behave like its large language model predecessors and contemporaries (e.g., GPT-3, T5, etc.), we cannot know what computations it is doing at any one point in time, and are never able to retrace its steps.

Commercial generative AI systems are fundamentally based on heavily exploited human labor. In all of these systems, the human is doing a lot for their human-in-the-loop reinforcement learning procedure; workers experience traumatic and intense challenges every day to identify hate speech, moderate posts, flag individuals who are experiencing suicidal ideation. These professionals are not the only people training the model. Each model is updated in response to the inputs that we give it, both in the short term (while we are interacting with it) as well as the long term (the systems these corporations release). Some systems are purportedly "private" and "HIPAA compliant" and "FERPA compliant" but how much that is really the case is anyone's guess in an industry that is basically unregulated.

There are also a lot of hidden things under the hood, similar to how image generation models that relied on prompting (e.g., DALL-E) embedded secret rules to make it less racist. While this is a good thing for the end user, this front-end filtering it not disclosed to us, and is considered proprietary information. Any number of layers and layers of constraints may be being applied to the output that have nothing to do with the generative capacity of this model, and instead are meant to keep the model from producing racist, sexist, ableist, and other -ist language. We know some because of recent leaks of proprietary code. Claude includes vibe-coded regular expressions meant to trigger the models to treat individuals who yell at it differently.

From a labor perspective, I am concerned that the systems induce a kind of magical thinking on behalf of the user. Sam Altman even invokes terms like "magical" to describe new products that they are developing at OpenAI. Advertising copy everywhere, particularly in big cities with a tech presence, presumes that Generative AI systems are inevitable. But if the system cannot tell us what it is doing, and if we cannot inspect it, the system distances us from the actual products of our use of it. We have no way of understanding how it works. While I know a lot of people would prefer to not know how their cars or bicycles or telephone function, knowing at least some of the basic components of how our possessions function means we can at least explain them to others when they are broken. Generative AI provides no such autonomy.

Computation will cost us millions of lives

Which of course brings us to the question of what your work is doing for OpenAI. OpenAI is funded by a number of investors, most prominently Microsoft, whose computers are an array of precious metals held in heavily refrigerated storage centers to keep from overheating. This is typically in the form of chilled water. This water can either be refrigerated, which requires electricity (on top of the electricity to run the computations) or it comes from snowmelt as it comes down from the North Cascades and various large rivers in the Pacific Northwest and elsewhere. Each computation eats up some of this cold water, but also other water, which fuels hydroelectric dams. Water drops from great heights down artificial precipices, to leverage gravity to create “clean” energy. These dams were built on some of the largest rivers in the United States and used to host massive salmon runs that are now more or less salmon ladder-jumps, for those that don’t die trying to swim upstream to spawn. Anyway, that water is one of the more renewable sources of electricity in the United States, and Generative AI systems would have you believe that the electricity is being put to good use. In reality, querying a generative AI system consumes massive amounts of electricity.

What we have seen in the last three years is an explosion of non-renewable energy to power AI systems. Elon Musk created a massive power system outside of the city of Memphis that burned methane (liquified natural gas) and had such lax controls over leakage that people started getting sick. Data centers are sapping desert aquifers. Between this and the byproducts of non-renewable energy sources -- for systems that people mostly oppose, and have no interest in using -- and their associated costs to the local environment, which exacerbate global warming, the people have led widespread opposition movements to block the construction of data centers. This is encouraging, but technologists see this as worth it. I find it hard to believe that people who claim to be pro-environment through their individual actions while using AI tools recognize that the tools they are using are intended to cause the destruction of human civilization. They are intended to make places uninhabitable, because they're replacing humans anyway.

AI is anti-labor

A concrete concern for researchers hoping to use prompting models for psycholinguistics is replacing human beings in order to obtain savings on human labor in the form of teaching and research assistants in our labs. Labor savings may certainly be found for narrow tasks — RAs probably need sleep consolidation while ChatGPT does not, and people are highly likely to make the kinds of mistakes that can lead to items being eliminated from analyses entirely due to errors we only detect after data have been collected. Faculty and staff now use AI tools to create vibe-coded websites and analysis pipelines and this approach is relatively commonplace among technologists online.

The application of AI as a research assistant rejects the root word "labor" in "laboratory." An artificial intelligence system does not labor but rather computes. The consequence is that we have come to think of research assistants as an annoyance or a cost to running a lab, when there are free systems (or inexpensive ones, relative to stipends). We have decided, somewhere along the line, that teaching people to conduct research by hand is too time consuming. In the end, you might get something you like, but you might not know how it works. We are effectively cutting off the entire future ecosystem of trainees by offloading our research pipeline to a technological ecosystem. We still need young researchers to know and recognize parameters that are critical for generating sentence stimuli. Automation risks making researchers less effective in the long-run.

Recently, I have seen faculty jokingly, or half-seriously, describing one particular AI system as their "graduate student" or "postdoc" that is capable of automatically producing high-quality research outputs. Nevermind that increasingly graduate students are relying on AI systems to conduct their work in ways that necessitate college- and university-level policies to delineate acceptable and unacceptable use cases. Replacing graduate students with stochastic text generation machines only serves to hasten University policies that will cut teaching assistant and research assistant positions, and be used by grant agencies to demand cuts to staffing. AI is a jobs killer, and the use of AI to replace human labor is a way to guarantee that people end up without the means to support themselves in an economy that does not provide any kind of meaningful social safety net, and the pieces that do exist are eroding.

AI systems will profit from your labor, and will never pay dividends

Every single AI tool was created by a business whose business model is getting people to run their models at prices that have been heavily subsidized by venture capitalists. The VC funding racket, as seen by the recent collapse of Silicon Valley Bank, is largely speculative and not grounded in evidence that systems work. For example, OpenAI is subsidized by Microsoft, who immediately integrated ChatGPT into Bing search. These corporations are invested in folks’ dependence on models like ChatGPT and are, in the long run, aiming to crowd out smaller systems that may perform less well on certain tasks. Certainly the approach here is unscientific — the work is not intended for consumption by scientists, is not a suitable object of scientific inquiry, and more importantly the use of this “free” model makes all of us poorer in the knowledge we could have gained about language, statistics, and networks of language communities.

Every one of these systems stands to make a profit from you even if you are not giving them money because your use of the service changes how the service works. Every query you make could be used for an update to the system. Every AI player will make money off ads that better target users querying the search engine, which makes them richer. It’s a trope but we are obviously the product being sold in this scenario. This is at odds with the kind of work scientists do — funded by public dollars, for public sharing of scientific knowledge. If we have to query a private system for private profit, we are undermining the principle value of our work as scientists to society.

In summary

You're better off doing the work yourself. When we use generative AI systems, we all lose. The horizon is approaching, and you have every obligation to prevent the cannibalization of your future self.