In March 2016, Microsoft released an AI-powered chatbot named Tay onto the social media platform Twitter. Designed to “learn” from its interactions with human users, the experiment was intended to showcase the future of machine learning. Within twenty-four hours, the experiment was terminated. Tay had transitioned from a friendly conversationalist to a bigot, spewing racist and sexist rhetoric. This incident is often cited as a cautionary tale of internet culture, but for cognitive scientists, it points to a much deeper question: does a machine “remember” its experiences, and can it ever truly escape the biases of its “prior knowledge”?.
The Machine Brain: Neural Networks and Activation Patterns
To understand how an algorithm “remembers,” we have to look at the architecture of modern AI, which is heavily inspired by human cognitive structures. Modern networks utilize neural networks, which are connectionist and sub-symbolic systems designed as brain analogs. In these systems, knowledge is not stored as a neat list of facts in a digital filing cabinet; instead, it is non-hierarchical and distributed.
In these “modern” networks, meaning is an emergent property. It arises from complex patterns of activation across the network. When the algorithm encounters new information, it isn’t just recording data; it is establishing or strengthening specific pathways. This mirrors the human Spreading Activation Theory, where activating one concept (or node) automatically lights up related concepts in the network. For a machine, “remembering” is essentially the ability to re-trigger a specific pattern of electrical activity that represents a learned association.
The Bottom-Up Bottleneck
Despite these similarities, there is a fundamental complication in the way algorithms process information compared to humans: they are primarily data-driven (bottom-up) processors. In human memory, we rely on a constant, dynamic balance between bottom-up input and top-down influences, such as our goals, perspectives, and existing schemata.
A machine, particularly during its training phase, relies heavily on association learning. It forms links between stimuli based on frequency of co-occurrence and spatiotemporal contiguity. If a dataset frequently pairs certain groups of people with negative descriptors, the algorithm’s neural network will build a strong “link weight” between those concepts. Because the machine lacks the top-down “effort after meaning” that humans use to build coherent mental models, it often fails to distinguish between a meaningful relationship and a random or biased association in its data.
Hidden Assumptions: The “Naive Models” of AI
The failure of Microsoft’s Tay bot reveals that algorithms, like humans, are susceptible to inherent bias in their “background knowledge”. Humans enter the world with “naive models”—initial, often incorrect theories about how things work. We then spend years undergoing conceptual change to restructure those models into valid scientific ones.
AI systems are “born” with a knowledge base derived from their training datasets. If those datasets contain hidden assumptions or societal prejudices, the machine encodes them as foundational truths. These become the algorithm’s version of a “naive model”. When Tay began interacting with malicious users, it wasn’t just learning new facts; it was engaging in accretion, an additive mechanism where it simply filled new “slots” in its existing racist schemata. Because it lacked a higher-order system to question the validity of this input, it “learned” the bias as a functional way to predict and generate conversation.
The Challenge of Restructuring and Inhibition
Can a machine “unlearn” a misconception? In humans, this is a slow and difficult process known as restructuring. Even after we learn the truth, behavioral and neurophysiological evidence shows that our old misconceptions are never truly deleted; they stay in our Long-Term Memory (LTM) forever. When experts solve problems, they actually take longer than beginners because their brains show increased activity in the prefrontal cortex—the area responsible for monitoring and inhibition. The expert must actively “silence” the ghost of the old misconception to find the right answer.
Algorithms currently struggle with this inhibitory control. Once a pattern of activation is strengthened in a neural network, it is difficult to “suppress” it without completely retraining the model. There is no “prefrontal cortex” for an algorithm to decide that a concept or mental model needs to be fundamentally rebuilt. While humans can use Resubsumption—applying an explanatory framework from one domain to another to resolve conflicts—machines often lack the analogical reasoning required for such a large-scale change.
Construction, Integration, and Meaning
To build better “memories” for machines, researchers like Walter Kintsch (1998) have explored how machines might represent knowledge in ways that mimic human text comprehension. Kintsch’s Construction-Integration Model suggests that we first “promiscuously” activate many associations (Construction) and then “prune” the irrelevant ones (Integration) to create a coherent Situation Model.
For a machine to have a “memory” that is more than just a recording of biased data, it must move beyond rote learning and toward meaningful learning. This requires the ability to perform elaborative encoding, where new information is integrated into a larger, coherent structure of “global knowledge”. Without this, the machine remains a “passive record keeper” in its Short-Term and Long-Term stores.
Conclusion: The Future of Algorithmic Mind
So, can an algorithm have a memory? The answer is yes, but it is a memory that currently lacks the top-down oversight and inhibitory flexibility of the human cognitive system. Like the “ghosts in the machine” that haunt our own long-term memories, the biases encoded in an algorithm’s neural network are remarkably persistent.
The challenge for the future is not just feeding AI more data, but developing systems that can build mental models—temporary, fuzzy, but highly coherent representations of a situation. We need algorithms that can not only “remember” what they were told but also possess the cognitive architecture to decide when it is time to demolish and rebuild their internal world. Until then, we must remember that an algorithm’s memory is only as healthy as the world we use to train it.
Leave a Reply