We currently live in the age of “artificial intelligence” – but not as the companies selling “AI” would have you believe. According to Silicon Valley, machines are rapidly outperforming humans in a variety of tasks ranging from mundane, but well-defined and useful tasks like machine transcription to much more vague skills like “reading comprehension” and “visual comprehension.” Some say these skills even represent rapid progress toward “artificial general intelligence,” or systems that can learn new skills on their own.
Given these grandiose and ultimately false claims, we need media coverage that holds tech companies to account. Far too often, what we get instead is breathless reporting, even in venerable publications like The New York Times.
If the media helped us break the hype, what would we see? We would see that so-called “AI” are actually pattern recognition systems that process unfathomable amounts of data using huge amounts of computing resources. These systems then probabilistically reproduce the patterns they observe, to varying degrees of reliability and usefulness, but always guided by the training data. For automatic transcription of several varieties of English, the machine can map waveforms to spelling, but it will be blocked by newly prominent names of products, people, or places. In translate from turkish to englishthe machines will map the gender-neutral Turkish pronoun ‘o’ to ‘he’ if the predicate ‘is a doctor’ and ‘she’ if it’s ‘a nurse’, as these are the most important patterns in the training data.
In both machine transcription and machine translation, pattern matching is at least close to what we want, if we are careful to understand and account for failure modes when using the technology. Bigger problems arise when people design systems that claim to do things like infer mental health diagnoses from voices or “crime” from pictures people’s faces: these things are not possible.
However, it is still possible to create a computer system that gives the expected type of output (mental health diagnosis, crime score) for an input (voice recording, photo). The system will not always be wrong. Sometimes we may have independent information that allows us to decide that it is true, other times it will give a plausible or even unverifiable result. But even when the answers look correct for most test cases, that doesn’t mean the system is actually doing the impossible. It may provide answers that we deem “correct” by chance, based on spurious correlations in the dataset, or because we are too generous in our interpretation of its outputs.
It is important to note that if the people deploying a system believe it performs the task (however ill-defined it may be), the results of the “AI” systems will be used to make decisions that affect the lives of real people.
Why are journalists and others so ready to believe claims of magical “AI” systems? I believe an important factor is show-pony systems like OpenAI’s GPT-3, which use pattern recognition to “write” seemingly coherent text by repeatedly “predicting” which word comes next in a sequence , offering an impressive illusion of intelligence. But the only intelligence involved is that of the humans reading the text. We do all the work, intuitively using our communication skills as we do with other people and imagining a spirit behind the language, even if it isn’t there.
While it may not seem like a big deal that a journalist is won over by the GPT-3, every bit of puff that flatters its supposed “intelligence” lends credence to other “AI” applications – those that self-classify. -calling people (as criminals, as having mental disorders). disease, etc.) and allow their operators to claim that because a computer does the work, it must be objective and factual.
Instead, we should demand journalism that refuses to be dazzled by claims of “artificial intelligence” and looks behind the curtain. We need journalism that asks key questions such as: what patterns in training data will lead systems to reproduce and perpetuate past harms against marginalized groups? What will happen to the people subject to the decisions of the system, if the operators of the system believe that they are correct? Who benefits from transferring these decisions to a supposedly objective computer? How would this system further concentrate power and what systems of governance should we demand to oppose it?
It behooves us all to remember that computers are just tools. They can be beneficial if we give them appropriately sized tasks that match their abilities well and maintain human judgment on what to do with the outcome. But if we mistake our ability to make sense of computer-generated language and images for computers to be “thinking” entities, we risk ceding power – not to computers, but to those who would hide behind the curtain.