How We Talk About AI Is Important: A Case Study

How we talk about AI is important—and there has been a lot of such talk lately—due to its potential impact on policy makers and decision-makers. How does one separate hype from reality? Such discourse needs to be grounded in the state of technology—what is unlikely or possible and what has already been realized. We also need to watch our language when discussing AI that has verbal capabilities.

The Person-in-the-Machine Effect With Language-Based AI

AI has been around for over six decades but it only captured the popular imagination lately when it mimicked the faculty that makes us human—language. Humans have a natural ability to produce grammatically acceptable sentences in their native tongues. For decades, linguists struggled with this syntax problem: how to create a model that replicates this ability.

The theoretical models were terribly leaky and created both ungrammatical and grammatical sentences. To skirt around this theoretical problem, large language models, based on the vast data on the internet and elsewhere, produce sentences by predicting the next word.

Why, then, ask whether next-word predictors “lie” or “hallucinate”? We endow machines with human qualities when we use such words. Yet, we cannot help ourselves. When we see or hear language, we instinctively imagine a person behind it. After an early mistake by BMW, car manufacturers became mindful that consumers disfavor certain voices in navigation systems. Consumers react strongly to age, gender, ethnic and other stereotypes—even though there really is no person inside the machine.

We instinctively react to human voice because we need to guess a speaker’s age, gender, height, and size to match the person’s speech signal to the intended word. The physical output is highly variable across speakers, which is why the brain has to make these guesstimates to figure out what was said.

The-Expert-in-the-Machine with Autonomous AI

On the one hand, we may be crediting some forms of AI with more capabilities than they actually possess. On the other, we may not fully recognize the actual capabilities of others.

Consider autonomous AI used in the diagnosis and therapy of neurodevelopmental disorders, functions previously reserved for human experts.

AI is being deployed in this area because it overcomes three major obstacles in traditional practice. The first is the complexity of the problem in the case of a disorder such as dyslexia. A person with dyslexia has difficulty learning to read because of deficits in processing language in the brain.

The linguistic system is immensely complex. It is really a conglomeration of many interlocking systems, each with its components, sub-components and sub-sub-components, all the way down to its atomic elements.

Deficits may occur in any part of this complex system. Given wide individual variation, how can we locate the linguistic deficits of each person with dyslexia? Covering the whole linguistic system to find these deficits would take too long. Even covering just one component fully, such as the phonological (sound) component often implicated in dyslexia research, is not feasible.

Traditional evaluations can only do sample testing. But sampling is just as likely to miss a person’s linguistic deficits as find them.

The AI system for dyslexia neither covers all ground mindlessly nor tests the same sample items. Instead, it proceeds intelligently by using a person’s past and current responses to determine the optimal path forward.

To do so, the AI system cross-references billions of data points in its databases. It synthesizes key information to decide on the next step. AI thus overcomes the second obstacle in traditional practice: capacity to hold a vast volume of information to use in analysis. By comparison, our working memory, the mental space for manipulation of ideas and information, is pitifully limited.

Artificial Intelligence Essential Reads

The LLM Prompt Is Dead. Long Live the Prompt!

The Evolution of LLMs Through Real-Time Learning

Compare the AI expert system to a human expert. Take a task commonly given to struggling readers: break up a word into its single sounds (word segmentation). This is directly tied to a student’s ability to learn an alphabetic system like English.

Say the student is asked to break up the word “speak” into its sounds (phonetically [spik]). If the student just gives the whole word back as one indivisible unit, the AI system may explore further whether she can operate with smaller units than the word, such as the syllable. If she breaks up the word into two, the AI system can go on to determine whether she can operate with sub-syllabic units such as onset ([sp]) and rime ([ik]), and so forth.

The AI system weighs all these options in real time to decide what to deliver next during its interaction online with this user. Natural language is processed super fast in the brain. Orthographic processing (or spelling) occurs in 100-200 miliseconds; phonological in 200-500 milliseconds. Thus the third obstacle to overcome is speed.

During evaluation and intervention, the AI system can only match and monitor the user’s language processing in real time if it operates autonomously without human input. Thus, such a system has to use autonomous AI.

Real-World Consequences of Ignoring the State of Technology

Autonomous AI was developed to address dyslexia because of the size of the problem. Affecting one in five people, dyslexia costs US taxpayers around $100b a year in special services, and more to support a disproportionate number of prison inmates with this learning disability.

New York, in following the footsteps of other states, is currently considering legislation to mandate dyslexia intervention. But the language of the bill inadvertently limits the provision of intervention services to human specialists.

Whether in legislation or policy, the language used has to be carefully crafted so that it is forward-looking—because the future is already here.