by Paul W. Bennett 
Originally published on Policy Options
September 5, 2025

by Paul W. Bennett 
Originally published on Policy Options
September 5, 2025

Consider these everyday experiences in today’s digitally dependent world rich with artificial intelligence (AI). A convenience store cashier struggles to make change. Your Uber driver gets lost on his way to your destination. A building contractor tries to calculate the load-bearing capacity of your new floor. An emergency-room nursing assistant guesses at the correct dosage in administering a life-saving heart medication.

All of these are instances of an underlying problem that can be merely an irritant or a matter of life and death. What happens when brains accustomed to backup from phones and devices must go it on their own?  

Increasingly we are relying upon technology to do our thinking for us. Cognitive offloading to calculators, GPS, ChatGPT and digital platforms enables us to do many things without relying on human memory. But that comes with a price.   

Leading cognitive science researchers have begun to connect the dots. In a paper entitled The Memory Paradox, released earlier this year, American cognitive psychologist Barbara Oakley and a team of neuroscience researchers exposed the critical but peculiar irony of the digital era: as AI-powered tools become more capable, our brains may be bowing out of the hard mental lift. This erodes the very memory skills we should be exercising. We are left less capable of using our heads.

Collective loss of memory

Studies show that decades of steadily rising IQ scores from the 1930s to the 1980s — the famed Flynn effect — have levelled off and even begun to reverse in several advanced countries. Recent declines in the United States, Britain, France and Norway cry out for explanation. Oakley and her research team applied neuroscience research to find an answer. Although IQ is undoubtedly influenced by multiple factors, the researchers attribute the decline to two intertwined trends. One is the educational shift away from direct instruction and memorization. The other is a rise in cognitive offloading, that is, people habitually leaning on calculators, smartphones and AI to recall facts and solve problems. 

The AI literacy gap facing Gen Alpha

AI threatens Indigenous data sovereignty and self-determination

Surveying decades of cognitive psychology and neuroscience research, Oakley and her team show how memory works best when it involves more than storage. It’s also about retrieval, integration and pattern recognition. When we repeatedly retrieve information, our brains form durable memory schemata and neural manifolds. These structures are indispensable for intuitive reasoning, error-checking and smooth skill execution. But if we default to “just Google it,” those processes so fundamental for innovation and critical thinking may never fully develop, particularly in the smartphone generation.

A key insight from the paper is the connection between deep learning behaviours in artificial neural networks (consider “grokking” in which patterns suddenly crystallize after extensive machine training) and human learning. Just as machines benefit from structured, repeated exposure before grasping deep patterns, so do humans. Practice, retrieval and timed repetition develop intuition and mastery.

Atrophy of mental exercise

The researchers sound a cautionary note. Purely constructivist or discovery‑based teaching, starting with assumptions that “students know best” and need little guidance, can short‑circuit mental muscle‑building, especially in our AI world. The team found that when students rely too early on AI or calculators, they skip key steps in the cognitive sequence: encoding, retrieval, consolidation and mastery of the brain’s essential building blocks. The result is individuals whose mental processes are more dependent upon guesswork, superficial grasp of critical facts and background knowledge and less flexible thinking.

Even techno skeptics see a role for digital tools. Oakley and her colleagues argue for what they term cognitive complementarity — a marriage of strong internal knowledge and smart external tools. ChatGPT or calculators should enhance — not replace — our deep mental blueprints that let us evaluate, refine and build upon AI output. That’s the real challenge that lies ahead.

The latest cognitive research has profound implications for educational leaders, consultants and classroom teachers. Popular progressive and constructionist approaches, which give students considerable autonomy, may have exacerbated the problem. It’s time to embrace lessons from the new science of learning to turn the situation around in today’s classrooms. This includes reintegrating retrieval practice (automatic recall of information from memory), spaced repetition and step-by-step skills progression in Grades K-12.

Using your head

What are the new and emerging essentials in the AI-dominated world? Oakley and her team deliver some sound recommendations, including:

• Teaching students to limit AI use and delay offloading.

• Training teachers to design AI‑inclusive but memory‑supportive curriculums, demonstrating that effective AI use requires prior knowledge and the ability to distinguish fact from fiction

• Guiding institutions to adopt AI in ways that build upon, not supplant, the human brain, such as editing original prose or mapping data.    

Using our heads and tapping into our memory banks must not become obsolete. They are essential mental activities. Access to instant information can and does foster lazy habits of mind. British education researcher Carl Hendrick put it this way: “The most advanced AI can simulate intelligence, but it cannot think for you. That task remains, stubbornly and magnificently, human.”

The most important form of memory is still the one inside our heads.

*Composed in a fierce dialectical encounter with ChatGPT.

This <a target=”_blank” href=”https://policyoptions.irpp.org/2025/09/ai-memory/”>article</a> first appeared on <a target=”_blank” href=”https://policyoptions.irpp.org”>Policy Options</a> and is republished here under a Creative Commons license.<img src=”https://policyoptions.irpp.org/wp-content/uploads/2025/08/po_favicon-150×150.png” style=”width:1em;height:1em;margin-left:10px;”><img id=”republication-tracker-tool-source” src=”https://policyoptions.irpp.org/?republication-pixel=true&post=295565″ style=”width:1px;height:1px;”>

Artificial intelligence is built on the thinking of intelligent humans, including data labellers who are paid as little as US$1.32 per hour. Zena Assaad, an expert in human-machine relationships, examines the price we’re willing to pay for this technology. This article was originally published in the Cosmos Print Magazine in December 2024.

From Blade Runner to The Matrix, science fiction depicts artificial intelligence as a mirror of human intelligence. It’s portrayed as holding a capacity to evolve and advance with a mind of its own. The reality is very different.

The original conceptions of AI, which hailed from the earliest days of computer science, defined it as the replication of human intelligence in machines. This definition invites debate on the semantics of the notion of intelligence.

Can human intelligence be replicated?

The idea of intelligence is not contained within one neat definition. Some view intelligence as an ability to remember information, others see it as good decision making, and some see it in the nuances of emotions and our treatment of others.

As such, human intelligence is an open and subjective concept. Replicating this amorphous notion in a machine is very difficult.

Software is the foundation of AI, and software is binary in its construct; something made of two things or parts. In software, numbers and values are expressed as 1 or 0, true or false. This dichotomous design does not reflect the many shades of grey of human thinking and decision making.

Not everything is simply yes or no. Part of that nuance comes from intent and reasoning, which are distinctly human qualities.

To have intent is to pursue something with an end or purpose in mind. AI systems can be thought to have goals, in the form of functions within the software, but this is not the same as intent.
The main difference is goals are specific and measurable objectives whereas intentions are the underlying purpose and motivation behind those actions.

You might define the goals as ‘what’, and intent as ‘why’.

To have reasoning is to consider something with logic and sensibility, drawing conclusions from old and new information and experiences. It is based on understanding rather than pattern recognition. AI does not have the capacity for intent and reasoning and this challenges the feasibility of replicating human intelligence in a machine.

There is a cornucopia of principles and frameworks that attempts to address how we design and develop ethical machines. But if AI is not truly a replication of human intelligence, how can we hold these machines to human ethical standards?

Can machines be ethical?

Ethics is a study of morality: right and wrong, good and bad. Imparting ethics on a machine, which is distinctly not human, seems redundant. How can we expect a binary construct, which cannot reason, to behave ethically?

Similar to the semantic debate around intelligence, defining ethics is its own Pandora’s box. Ethics is amorphous, changing across time and place. What is ethical to one person may not be to another. What was ethical 5 years ago may not be considered appropriate today.

These changes are based on many things; culture, religion, economic climates, social demographics, and more. The idea of machines embodying these very human notions is improbable, and so it follows that machines cannot be held to ethical standards. However, what can and should be held to ethical standards are the people who make decisions for AI.

Contrary to popular belief, technology of any form does not develop of its own accord. The reality is their evolution has been puppeteered by humans. Human beings are the ones designing, developing, manufacturing, deploying and using these systems.

If an AI system produces an incorrect or inappropriate output, it is because of a flaw in the design, not because the machine is unethical.

The concept of ethics is fundamentally human. To apply this term to AI, or any other form of technology, anthropomorphises these systems. Attributing human characteristics and behaviours to a piece of technology creates misleading interpretations of what that technology is and is not capable of.

Decades long messaging about synthetic humans and killer robots have shaped how we conceptualise the advancement of technology, in particular, technology which claims to replicate human intelligence.
AI applications have scaled exponentially in recent years, with many AI tools being made freely available to the general public. But freely accessible AI tools come at a cost. In this case, the cost is ironically in the value of human intelligence.

The hidden labour behind AI

At a basic level, artificial intelligence works by finding patterns in data, which involves more human labour than you might think.

ChatGPT is one example of AI, referred to as a large language model (LLM). ChatGPT is trained on carefully labelled data which adds context, in the form of annotations and categories, to what is otherwise a lot of noise.

Using labelled data to train an AI model is referred to as supervised learning. Labelling an apple as “apple”, a spoon as “spoon”, a dog as “dog”, helps to contextualise these pieces of data into useful information.

When you enter a prompt into ChatGPT, it scours the data it has been trained on to find patterns matching those within your prompt. The more detailed the data labels, the more accurate the matches. Labels such as “pet” and “animal” alongside the label “dog” provide more detail, creating more opportunities for patterns to be exposed.

Data is made up of an amalgam of content (images, words, numbers, etc.) and it requires this context to become useful information that can be interpreted and used.

As the AI industry continues to grow, there is a greater demand for developing more accurate products. One of the main ways for achieving this is through more detailed and granular labels on training data.
Data labelling is a time consuming and labour intensive process. In absence of this work, data is not usable or understandable by an AI model that operates through supervised learning.

Despite the task being essential to the development of AI models and tools, the work of data labellers often goes entirely unnoticed and unrecognised.

Data labelling is done by human experts and these people are most commonly from the Global South – Kenya, India and the Philippines. This is because data labelling is labour intensive work and labour is cheaper in the Global South.

Data labellers are forced to work under stressful conditions, reviewing content depicting violence, self-harm, murder, rape, necrophilia, child abuse, bestiality and incest.

Data labellers are pressured to meet high demands within short timeframes. For this, they earn as little as US$1.32 per hour, according to TIME magazine’s 2023 reporting, based on an OpenAI contract with data labelling company Sama.

Countries such as Kenya, India and the Philippines incur less legal and regulatory oversight of worker rights and working conditions.

Similar to the fast fashion industry, cheap labour enables cheaply accessible products, or in the case of AI, it’s often a free product.

AI tools are commonly free or cheap to access and use because costs are being cut around the hidden labour that most people are unaware of.

When thinking about the ethics of AI, cracks in the supply chain of development rarely come to the surface of these discussions. People are more focused on the machine itself, rather than how it was created. How a product is developed, be it an item of clothing, a TV, furniture or an AI-enabled capability, has societal and ethical impacts that are far reaching.

A numbers game

In today’s digital world, organisational incentives have shifted beyond revenue and now include metrics around the number of users.

Releasing free tools for the public to use exponentially scales the number of users and opens pathways for alternate revenue streams.

That means we now have a greater level of access to technology tools at a fraction of the cost, or even at no monetary cost at all. This is a recent and rapid change in the way technology reaches consumers.
In 2011, 35% of Americans owned a mobile phone. By 2024 this statistic increased to a whopping 97%. In 1973, a new TV retailed for $379.95 USD, equivalent to $2,694.32 USD today. Today, a new TV can be purchased for much less than that.

Increased manufacturing has historically been accompanied by cost cutting in both labour and quality. We accept poorer quality products because our expectations around consumption have changed. Instead of buying things to last, we now buy things with the expectation of replacing them.

The fast fashion industry is an example of hidden labour and its ease of acceptance in consumers. Between 1970 and 2020, the average British household decreased their annual spending on clothing despite the average consumer buying 60% more pieces of clothing.

The allure of cheap or free products seems to dispel ethical concerns around labour conditions. Similarly, the allure of intelligent machines has created a facade around how these tools are actually developed.

Achieving ethical AI

Artificial intelligence technology cannot embody ethics; however, the manner in which AI is designed, developed and deployed can.

In 2021, UNESCO released a set of recommendations on the ethics of AI, which focus on the impacts of the implementation and use of AI. The recommendations do not address the hidden labour behind the development of AI.

Misinterpretations of AI, particularly those which encourage the idea of AI developing with a mind of its own, isolate the technology from the people designing, building and deploying that technology. These are the people making decisions around what labour conditions are and are not acceptable within their supply chain, what remuneration is and isn’t appropriate for the skills and expertise required for data labelling.

If we want to achieve ethical AI, we need to embed ethical decision making across the AI supply chain; from the data labellers who carefully and laboriously annotate and categorise an abundance of data through to the consumers who don’t want to pay for a service they have been accustomed to thinking should be free.

Everything comes at a cost, and ethics is about what costs we are and are not willing to pay.