This essay for the Open Philanthropy AI Worldviews Contest is targeted at Question 1.
AIs have recently accrued some impressive successes, raising both hopes and concerns that artificial general intelligence (AGI) may be just decades away. I will argue that the probability of an AGI, as described in the contest prompt, arising by 2043 is extremely low.
Specification gaming and school
In Richard Feynman’s memoir Surely You’re Joking, Mr. Feynman! the physicist recalls his 1951 experiences visiting a Brazilian university and advising Brazilian professors and government officials on science education. He explains how the university he visited managed to “teach students physics” without giving them any real understanding. The students learned to associate physics words with definitions and memorized equations. They could accurately state physics laws on exams, and they could solve the types of problems they’d been taught to solve. They looked like they were learning.
But, as Feynman found, the students couldn’t stretch their knowledge to unfamiliar problem types, and they couldn’t solve even the most basic real-world physics problems by applying the equations they knew to physical objects. Of course, these students couldn’t have even begun to push research into new areas - the job of a working physicist. The students trained in the Brazilian system of the 1950s were one step away from learning the scientific subject matter and two steps away from learning to actually do science.1
Today’s chatbots have been compared to high school students in their writing abilities, and this seems very impressive for a machine. But many high school students are plagiarists; to some extent, so are many adult writers. Instead of copying whole sentences, advanced plagiarists cut and paste fragments of sentences from multiple sources. Better yet, they rearrange and choose near-synonyms for as many words as possible to reduce the chance that the teacher will discover the sources (i.e. paraphrase plagiarism). Or, they develop an essay that closely mimics the novel ideas and arguments in a published work, but doesn’t contain any of the same words (i.e. conceptual plagiarism).
Instead of learning to actually formulate an argument based on their own ideas, high school plagiarists engage in specification gaming. They learn to game teachers’ and plagiarism checkers’ abilities to recognize when they are not developing their own work. They learn to ape the skills the teachers want them to learn by honing advanced ways of disguising the derivativeness of their writing.
The problem with this isn’t just that the students borrow others’ work without attribution. It’s that the students don’t learn to think for themselves. These students are “learning” something, but what they’re learning is a low-level skill that’s only useful within the “game” of school. And the game, to some extent, defeats the purpose of education.
Earlier this year, writer Alex Kantrowitz discovered that a partially plagiarized version of his article had been published without crediting him, and that the “author” was an AI alias.2 Several other companies and individuals have been caught using AIs to produce articles or stories that consist of lightly-altered sentences drawn from uncredited source materials. Researchers have found that generative AIs engage in both paraphrase and conceptual plagiarism as well as reproducing memorized text (verbatim plagiarism).3 Similar issues exist with AIs that generate art or music. In 2022, a research team prompted an art AI to output very close copies of dozens of images (some photos, some artwork) from its training data.4 The company behind the same art AI is the subject of several copyright-infringement lawsuits.
Leaving aside the intellectual property issues and the effects on student learning, these cases tell us something about generative AIs’ actual capabilities and likely trajectory. Free near-copies of artists’ work will be as valuable as forgeries (like a “newly discovered” fake Picasso) have always been. Creating such forgeries is a skill, yes, but developing that skill is not an indication that one is on the way to becoming a Picasso-level artist.
When AI development focuses on convincing humans that a machine is intelligent, the result may say more about human psychology (i.e. how to effectively deceive humans) than it does about the subject matter the AI works with. This deception would most likely be unintentional; in other words, the AI presents humans with outputs that make it easy to fool ourselves. The harder the problem an AI is tasked with, the easier, by comparison, this kind of deception might be. In a subject-matter realm in which no AI is close to developing capabilities, all AIs that appear to succeed will do so by deception.
Eyespots on the rear end of Eupemphix nattereri frog. Photo by Felipe Gomes
Confusing Definitions
The meanings of “data,” intelligence,” and “general intelligence” are not pinned down, so these words are liable to be abused for profit.
Could the biggest AI-related risks come from a mismatch between different meanings of these words, rather than from AIs’ growing capabilities?
“Data”
The inputs that power generative AIs are commonly called “training data,” but they really don’t fit the usual definition of “data”. Most people don’t think of the Mona Lisa or the full text of a novel as a data point, but the “training data” that generative AIs use consists of peoples’ ideas, innovations, creative works, and professional output. The rampant cases of AI-facilitated plagiarism and forgery imply that these AIs are using their “training data” more as a source of raw material than as a source of training.
“Intelligence” and “general intelligence”
Moving on to other types of AIs, it's easy to see game-playing AIs’ successes at games like Go as evidence that they are becoming “intelligent.”
Game-playing machines’ successes are impressive, but they don’t necessarily indicate skills that transfer to the real-world environments that our human intelligence handles. We can precisely tell an algorithm what “winning” a game means and let it play millions of games in a uniform environment. The games in question are played entirely inside a computer (or entered into a computer) and don’t involve changing environments, uncertainty around labels and borders, or the need to track down and collect information in the real world. Carefully-designed AIs can produce novel solutions for manmade puzzles and games even if they’re highly complex, but this says nothing about their performance in an environment of real-world decision-making or investigation.
In real-world environments, unlike in a game, we rarely know what “winning” is or what “the correct answer” is. If we think we know, we may be wrong in cases different from the ones we’ve considered — even if we don’t know there’s a difference.
Case study 1
In 2018, the AI system AlphaFold beat out 97 competitors in accurately predicting the structures of recently solved proteins, thus winning the biannual CASP competition. AlphaFold was trained on the Protein Data Bank (PDB), a large database of proteins solved mostly through crystallography. This AI system has gained recognition as an important advance in structural biology.
The first issue is that PDB is a biased sample of proteins- those whose production, purification and crystallization were tractable, and those that come from species that humans are most interested in and/or have an easy time studying in the lab. AlphaFold can mostly interpolate within the range of its training data. We don’t really know how much we can extrapolate known protein structures to, say, the proteins of extremophiles, or proteins with low homology to any protein in PDB, or proteins that haven’t been solved because they tend to kill the bacterial cells in which we try to grow them.
Another issue with using PDB is that not all proteins adopt a single, static structure in nature. The up-and-coming fields of intrinsically disordered proteins (IDPs), intrinsically disordered regions (IDRs), and bi-stable “fold-switching” proteins are revealing that many proteins are fully or partly disordered or can change between different structures. Current estimates are that over 50% of eukaryotic proteins contain at least one long disordered segment, and these features appear to be key to many proteins’ functions.
Scientists who work with IDPs/IDRs and fold-switching proteins are offering cautions that AlphaFold (and its successor AlphaFold2) may predict spurious structures for these proteins.5 Scientists working on proteins that switch between two or more structures report that AlphaFold will typically find just one of the structures, or an intermediate structure, but the prediction will be given with high confidence. Scientists are using AlphaFold to (carefully) study IDPs and IDRs, and AlphaFold seems to do well at marking these regions as “low-confidence” predictions. However, the authors of a 2021 paper caution that those inexperienced with IDPs/IDRs who use AlphaFold may be “tempted to draw” incorrect inferences that have “zero physical meaning” and that the AlphaFold process can introduce artifacts that are especially severe for IDRs.6 Because IDPs/IDRs are still a lesser-known field, some scientists today may be looking at AlphaFold’s static outputs with less than the appropriate skepticism.
This is not to diminish the accomplishments of those who’ve worked on AlphaFold; it really is an important advance that puts multiple recent discoveries in protein biology and molecular evolution to work. We can use AlphaFold and similar AIs well by understanding where they perform better and worse, where nobody knows the “ground truth,” where we may be wrong about it, or where we may be extrapolating too broadly.
If humans had turned the task of investigating protein structures over to an AI (or to scientists who rely heavily on an AI) before humans had discovered IDPs and bi-stable proteins, we may never have known about these important groups. It took human intelligence to make the discoveries leading to the recognition of IDPs/IDRs: the hands-on work of trying and failing to crystallize some groups of proteins, and the innovative thinking needed to notice where the dominant paradigm was breaking down and recognize its importance.
Case study 2
AI programs for medical image analysis have been under development for years. A few of these programs are currently used in the clinic, though reception has been mixed. To the extent that we can define what constitutes a “normal” or “abnormal” imaging result, and to the extent that representative and accurately-labeled imaging datasets are available, it should be possible to eventually develop useful AIs that replicate physicians’ judgments regarding images.
But image evaluation is only one step in a larger diagnostic process, the same step filled by laboratory blood tests and the like. Because doctors who understand a lab test’s limitations can make better use of it, clinical leaders like Catherine Lucey, MD of UC San Francisco are advocating for doctors to place less emphasis on the tests alone by placing test and imaging results within a Bayesian framework (such as by using Fagan nomograms, also called “Bayesian ladders”), with the goal of improving medical decision-making.
From this perspective, if AIs eventually become capable of replacing doctors in medical image analysis, allowing the AIs to take over diagnosis would be a step in the wrong direction. Instead, the advent of radiology AIs would be analogous to the mid-1800s replacement of doctors skilled at tasting urine by laboratory tests for glucose in diagnosing diabetes. The laboratory glucose tests eventually proved more accurate than the human tongue, and most human doctors probably didn’t mind giving up that role.
But any lab test has its limitations. Doctors need to consider the prior probability that a patient has a particular disease based on an exam and interview, the possibilities for other diseases that may explain the patient’s symptoms, available tests’ specificity and sensitivity, the consequences of a false positive or false negative, the uncertainty around the definition of “normal,” differences among populations, and cases where the test will fail in order to decide whether to apply the test in a particular patient’s case. The same considerations, plus more, apply to AI image analysis.
A different kind of “learning”
The “learning” that some AIs go through may be better compared to the learning that the human immune system goes through, when it “learns” to recognize self vs. non-self and to attack pathogens, rather than to the learning that a human does. The kind of “intelligence” an immune system has could even be called “general,” since it eventually becomes successful at many different tasks, from destroying cancer cells to neutralizing different pathogens to causing allergic reactions against pollen to rejecting transplanted organs.
Likewise, the kind of “learning” AIs do can make them very good at some tasks or even a range of tasks. However, calling that “intelligence” or “general intelligence” is a metaphor that is more likely to be misleading than helpful. It doesn’t make much more sense than calling the immune system a “general intelligence” -- except as a marketing term and to attract funders.
AI trainers, handlers and data labelers are the cognitive engine behind many AIs
Next, I’d like to look at another way observers might overestimate current AIs’ capabilities (both quantitatively and qualitiatively) and thus overestimate their prospects for developing certain capabilities in the future.
Quite a few AI companies have reportedly used a combination of AIs and human handlers to deliver what customers needed, even while customers and members of the public thought an AI was fully responsible.7,8,9Even AIs that don’t need support after they’re deployed often require a massive number of worker-hours to get up and running, including dataset development, labeling and sorting data, training, and giving feedback. In the case of generative AI, these inputs are in addition to the work of artists, writers, and musicians whose contributions are scraped as “data.” In some cases, these human contributions look less like training and more like doing the work for the AI (albeit ahead of time).
To understand whether human support is like “training wheels” or will be needed permanently, we need a clear-eyed view of what is actually happening. Today, some companies pass off much of the cognitive work in setting up and running their AIs to contractors or people working through task services like Mechanical Turk; other companies have used prison labor or outsourced to low-income countries. Tech researcher Saiph Savage has documented how AI company staff often underestimate the time required to complete “microwork” tasks passed to contractors. The money offered for microwork tasks is often so poorly aligned to the time needed that these AI workers’ average earnings are around $2 an hour (even for workers in the US).10 The lack of feedback makes it hard for AI companies, and even harder for outside observers, to gauge how many hours of human cognitive work are actually going in.
When deployed in the workplace, some current AIs may surreptitiously borrow time and cognitive resources from their human “coworkers” too. One of the criticisms of the (now defunct) IBM Watson Health AI was that doctors and nurses had to spend large amounts of time inputting their medical knowledge and patient data into the AI tools, with little usable payoff.11 Clinicians have similar complaints about other AIs that are widely deployed in healthcare today.
I’ve had similar enough experiences with expensive but poorly functional laboratory robots and corporate software to understand clinicians’ frustrations with these tools. When a technology appears to replace human workers, sometimes it actually pushes the work onto other humans. Managers can easily miss or underestimate the contributions by humans, making it appear that the technology (whether AI or otherwise) is more successful and “smarter” that it actually is. Humans who wind up as AI handlers may experience their jobs as less meaningful and more frustrating. For humans involved (by choice or not) in any stage of AI development or deployment, having a machine take credit for one’s work will come with an emotional impact and possibly a financial penalty.
Current estimates of the limitations on making AIs “smarter” may be biased by underestimates of how much today’s AIs rely on humans to do their cognitive work, either in advance or in real time. To know whether an AI is actually accomplishing useful work, we need to compare its time savings to the hours put in by all humans who worked on the project and all humans who help the AI function day-to-day, including contractors, people at outsourcing sites, people at client organizations, and end users, and we need to compare its contributions to the opportunity costs from those people not spending their time on something else. In some cases, the human input for today’s AIs is vastly greater than any output we can expect.
Some AIs are status quo engines that will slow down human advancement if deployed inappropriately
Modern society has a great predilection to base new programs and solutions on existing situations, even if those situations are pathological.
Some examples:
AIs didn’t originate this, but because many AIs rely on past data to make predictions, basing the future on the past or present is something they’re naturally good at. Similar examples of “AI bias” influenced by current and historical realities or by human behavior are popping up throughout healthcare and in other areas of our society. Many people are working on reducing bias in AI technology and in algorithms more broadly, but the problem might be inherent to any system that relies on making associations in existing data rather than going out and investigating the causes of problems.15,16
If AIs can’t be made to think outside the data, more advanced AIs may just be better status quo engines, skilled at propagating harmful situations forward in time.
In the corporate world, AI could easily feed right into many business leaders’ inclination to throw new, heavily advertised technology at a problem instead of understanding it. Unless AIs can investigate a problem hands-on, they can only use the existing data that humans chose to collect and feed them.
And alignment between the AI itself and the goals of its developers is not enough. As with traditional product design, the way developers understand real-world environments and the needs of users is often far from reality. The Lean Manufacturing and the broader Lean Thinking movements have facilitated real growth in value precisely by teaching people to open up communication, investigate the real roots of problems hands-on, and seek alignment of goals and actions not only within a company but with customers and vendors. To improve our thinking and action, we should be learning to break out of our status-quo-is-correct thinking patterns instead of reinforcing them and making them harder to detect.
To work with some AIs, humans will have to dumb themselves down
In high school English classes today, students and schools are rewarded for producing essays that please automated essay-scoring systems. These tests tend to give low scores to the work of certain famous authors, but high scores to students who’ve learned to write shallow essays that use lots of big words in long sentences, even if the essays contain little actual reasoning or insight.17 In other words, when algorithms are not capable of scoring our brightest kids’ tests, teachers are tasked with making kids dumber. Automated scoring systems for school subjects besides English are being developed.
AIs may be able to displace workers who were trained in “rote” educational programs, but the reason is that standardized testing and rote learning aren’t the best ways to help people develop real creativity, advanced thinking skills, or hands-on problem investigation. An AI recently passed a standardized medical licensing exam, but a written test doesn’t capture all aspects of knowledge, and the difference between a new med-school graduate and an experienced doctor is precisely the real-world learning that an algorithm inside a machine can’t do. The more similar our educational systems are to the 1950s Brazilian physics program Feynman visited, the easier our work will be for an AI to replace. But the issue here isn’t really the AI.
Generative AIs, too, trade in a secondary market of ideas, so their work will be more and more self-referential unless humans keep infusing real innovation. But this could become more difficult. If generative AI intellectual property law is worked out in favor of the AI companies, writers who share very innovative ideas, artists or musicians with a very distinctive style, or photographers who capture unique objects or scenes may be most prone to being directly infringed. The most innovative creators will be susceptible to having their work devalued by the greatest degree and could be driven out of the marketplace. It will become difficult to originate and share anything that contains true originality, unless it can be securely kept away from the AIs.18
In scientific research, over-production of low-quality journal articles and bad meta-analyses are already problems, and people in the field of metascience are working on solving them. Attempts to use AIs to carry out investigations and publish papers will only make the scientific “replication crisis” and similar issues worse. Even in most areas where we think we know what “correct” is, we should be questioning our definitions of “correct,” not fossilizing them.
Like automatic telephone switchboards and calculating machines before them, AIs are much faster and better at specific tasks than humans are. The problem is that AIs are getting better at miming human-like behaviors at a time when their actual capabilities cover only a tiny fraction of the full range of what humans can do. If societal forces exacerbated by AI deployment lead us to devalue the work that is outside AIs’ range, we will constrain our thinking and development.
As with any type of tool, we can make better and safer use of AIs if we’re clear-eyed about their limitations and keep humans in the driver’s seat. The graduates of the Brazilian physics program Feynman encountered could be dangerous, if someone hired a bunch of them for high positions in a nuclear project on the strength of their apparent skill in physics. But the danger would be very different from that of a brilliant and malicious physicist wresting control of the same project.
In summary:
References and Notes
