Psychology of Artificial Intelligence

In your own words, what is artificial intelligence? Give an example

AI = systems built to do things we associate with human intelligence — perceiving, using language, reasoning, learning. Examples: a language model like ChatGPT, an image classifier, a self-driving
car’s vision

What is a large language model (LLM)? Define it and give an example.

A program trained on huge amounts of text to predict the next word; by doing that at scale it can generate fluent language, answer questions, translate, and write code. Examples: ChatGPT, Claude, Gemini.

What is the difference between “narrow” AI and “general ” AI (AGI)? Give an example of narrow AI.

Narrow AI does one kind of task (playing chess, captioning images, filtering spam).

General AI (AGI) would flexibly handle many novel tasks like a person — which does not yet exist. Today’s systems are narrow, even when impressive

Why is “intelligence” hard to define and to measure?

We cannot observe it directly — we can only infer it from behavior — and definitions disagree (reasoning, learning speed, adapting to novelty, problem-solving). Any test can also be confounded by other factors (reading speed, prior exposure). (Something we infer rather than see directly is called a “latent construct,” but you don’t need the term.

What is the “testing principle” (also called the testing effect or retrieval practice), and why does it help learning?

Actively recalling information — e.g. self-testing — strengthens memory much more than re-reading or highlighting. The effort of pulling the answer out of your head is what builds durable memory (and even pre-tests on not-yet-studied material improve later learning).

What is the spacing effect?

Learning is more durable when practice is spread out over time rather than crammed into one session. The same total study time yields better long-term retention if it is distributed across days — which is also why revisiting material for a cumulative exam helps.

What are “desirable difficulties”?

Study conditions that make learning feel harder and slower in the moment — spacing, self-testing, interleaving topics, varying where you study — but that produce stronger, longer-lasting learning. Easy, fluent studying often feels productive yet teaches little.

This course is built on frequent low-stakes quizzes and cumulative exams. Which learning principles does that design reflect?

Retrieval practice (quizzes make you recall), spacing (cumulative exams force you to revisit material over time), and desirable difficulties (effortful recall beats re-reading). The “preview” quiz questions also exploit the finding that pre-testing boosts later learning.

Why does re-reading your notes feel effective but build weak long-term memory?

Re-reading makes material feel familiar and easy (“fluent”), and we mistake that fluency for learning. But recognizing something is far easier than recalling it; only effortful recall (self-testing) builds durable memory

Is sheer practice time (the “10,000-hour rule”) enough to build expertise?

No. What matters is deliberate practice: focused, effortful work at the edge of your ability, with feedback and harder goals — not mere clock hours. Prior knowledge, good instruction, and the domain matter too; 10,000 hours is a rough average, not a threshold

Vocabulary keeps growing into old age, while speed at solving brand-new puzzles declines. Which is crystallized and which is fluid intelligence?

Crystallized = accumulated knowledge and vocabulary (grows or holds with age). Fluid = reasoning and speed on novel problems (declines with age).

What is “general intelligence” (g), and what observation motivates the idea?

The claim that a single general ability contributes to performance across many different mental tasks. It is motivated by the fact that scores on very different cognitive tests tend to be positively correlated — people who do well on one tend to do well on others

Average IQ scores rose steadily across the 1900s (the “Flynn effect”). What is the most likely cause?

Environmental and societal change — better nutrition, more and longer schooling, more abstract reasoning in daily life, and greater test familiarity — not genetic change (far too fast for that)

How can we quantify something we believe exists but cannot measure directly (like intelligence, or hunger)?

Define it through measurable indicators and behaviors, then infer the hidden quantity from them— the way we infer gravity from how fast things fall. (This move is called “operationalizing a latent construct,” but the idea, not the term, is what matters.) Good measures must be reliable and valid.

What does an IQ score actually represent?

It is a relative score: how you performed compared with other people your age, scaled so the average is 100. It is not an absolute “amount” of intelligence, and it captures only some of what we might mean by the word.

What does it mean for a test to be reliable?

It gives consistent, repeatable results — roughly the same score each time you measure the same thing. (A scale that always reads 5 lb too high is reliable: consistent, even though wrong.)

What does it mean for a test to be valid, and what can undermine validity?

It actually measures what it claims to measure. A confound undermines validity — e.g. a “reasoning” test that really rewards reading speed, or an AI benchmark whose items leaked into the training data. A test can be reliable yet not valid, so you need both.

Give one limitation or criticism of IQ testing

Any one: it can reflect cultural background, schooling, or test familiarity rather than pure “ability”; it captures a narrow slice (not creativity, practical or social intelligence); a single number hides very different profiles; and scores can be affected by motivation and stereotype threat

What is a confound when we design an assessment or experiment?

A third factor that varies along with the thing you meant to measure and offers an alternative explanation, so you can’t tell which produced the result — it threatens validity. (E.g. a “reasoning” test that rewards reading speed; for AI, test items leaking into the training data.)

Name three things that make assessing intelligence in AI systems hard

Any three: test items already in the training data (→ recall, not reasoning); prompt sensitivity

(rewording swings the score); teaching-to-the-test (Goodhart’s law ); shortcuts / “Clever Hans”

(right answer, wrong reason); a jagged/alien profile (superhuman on one task, sub-child on a neighbor); no fair human–machine comparison.

Goodhart’s law says: “when a measure becomes a target, it stops being a good measure.” Explain what this means, and why it matters for assessing intelligence in both people and AI.

Once you optimize for a metric — coach to the test, or train on the benchmark — high scores stop reflecting the real ability the metric was meant to track. In people: test-prep can lift an IQ or SAT score without raising the underlying ability. In AI : training on (or leaking) the benchmark inflates scores without real generalization — so the number no longer measures “intelligence.”

What is prompt sensitivity, and why does it complicate claims about an AI’s ability?

The same task, reworded or reformatted, can produce very different AI answers and scores. If a “skill” appears or vanishes with trivial rewording, the ability isn’t robust — so a single high score may not reflect a real, general capacity.

The ARC benchmark was designed to measure intelligence differently from an ordinary skill test. What does it try to measure, and how is that different from raw skill?

It targets skill-acquisition efficiency and generalization — how well a system solves novel problems from just a few examples — rather than skill on one well-practiced task. So a system can’t win just by memorizing; it has to generalize and abstract.

What does it mean to “probe the psychology” of an AI model, rather than trust a benchmark score — and why is it useful?

Treat the AI as an object of study and test it the way psychology tests minds — with controls, comparison conditions, and alternative explanations — instead of trusting a headline number. That’s how you learn what it can really do and how it does it.

What do Bender et al. mean by calling large language models “stochastic parrots”?

An LLM stitches together the form of language from statistical patterns of how words co-occur (stochastic) by echoing and recombining its training text (parrot), with no access to meaning, reference, or intent

What is the “form vs. meaning” distinction at the heart of this argument?

Form is the surface pattern of language (which words tend to follow which). Meaning involves connecting language to the world, to referents, and to communicative intent. Bender et al. argue

LLMs learn form only; understanding requires meaning

Why do people so readily attribute understanding to a fluent chatbot?

Because the text is coherent, we supply the meaning — “coherence is in the eye of the beholder.” We are wired to read intention and mind into fluent language (the old “ELIZA effect”), so fluency easily fools us into assuming comprehension

Give one strong argument against the “stochastic parrot” view — a reason an LLM might capture some genuine meaning.

Anyone: meaning need not be reference — a concept’s meaning can be its web of relations to other internal states, which distributional learning captures; probing recovers structured internal representations (spatial or game-board maps), not just echoed strings; next-word prediction can induce real structure (grammar, world knowledge); ruling understanding out in advance is as unwarranted as granting it

The debate over whether LLMs “understand” is partly definitional. Why does that make it an empirical question?

If we pin down what we mean by understanding (e.g., building a model of the situation, generalizing correctly, tracking who-did-what), we can then test whether a system has that property — turning a word-fight into experiments on the model’s behavior and internal representations.

Give two ways human vision differs from a digital camera.

Any two: resolution is non-uniform (sharp only in the small fovea, blurry periphery); we actively sample with rapid eye movements, not one snapshot; perception is inference (fills gaps, is fooled by illusions); processing is hierarchical and distributed (edges → parts → objects, with feedback); it splits into “what” vs. “where” streams

What are saccades, and why does the visual system use them?

Saccades are rapid eye movements (several per second) that jump the high-resolution fovea from one point of interest to the next. Because only the fovea sees fine detail, vision must actively sample the scene rather than take one uniform picture

What does it mean to say “perception is inference”? Give an example

The flat retinal image is underdetermined — many scenes could produce it — so the brain reconstructs the most likely scene using prior knowledge, filling in what’s missing. Illusions show it: the Kanizsa triangle (edges you see but that aren’t drawn), or checker-shadow (two squares look different but are the same gray).

What does it mean that human (and machine) vision is “hierarchical ”?

Simple features are detected first and combined into progressively more complex ones — edges → parts → whole objects — across a series of stages (V1 up to higher areas), with each stage “seeing” a larger region of the image

What is a neuron’s receptive field in the visual system, and how do receptive fields change as you

go up the hierarchy (V1 → IT)?

A receptive field is the region of visual space (and the feature) a neuron responds to. Early on (V1), receptive fields are small and tuned to simple features (edges, orientations); as you ascend the ventral (“what”) stream (V2 → V4 → IT) they grow larger and respond to complex patterns — shapes, parts, whole objects and faces. (This growing-receptive-field hierarchy is exactly what a CNN copies.)

In V1, what is the difference between a simple cell and a complex cell

A simple cell responds to an edge or bar of a particular orientation at a particular position in its small receptive field. A complex cell also prefers an orientation but responds regardless of the exact position (some position invariance). These two cell types inspired a CNN’s two tricks —convolution (like simple cells) and pooling (like complex cells)

What are the visual “what” and “where” streams?

Two cortical pathways: the ventral (“what”) stream supports object identity / recognition; the dorsal (“where”/“how”) stream supports spatial location and guiding action.

Human vision is great at some things and poor at others. Give one of each.

Great at: grasping the gist of a scene in a fraction of a second; recognizing objects across viewpoint, size, and lighting; huge visual memory. Poor at: noticing unexpected events or changes

(inattentional / change blindness); resisting illusions; judging precise sizes, counts, or angles.

What two brain-inspired tricks turn an ordinary neural network into a CNN, and what biology does each echo?

Local receptive fields + shared weights — one feature detector slid across the whole image, like V1’s edge detectors — and pooling (tolerating small shifts). Stacking them builds an edges-to-objects hierarchy, mirroring V1 → IT

What is a receptive field, and what happens to receptive fields in deeper and deeper layers of a CNN?

A unit’s receptive field is the region of the image its response depends on. Early units have tiny fields (they detect edges); deeper units pool many units below them, so their fields grow — letting them detect parts, then whole objects

Why is a CNN called “convolutional ”? What does the shared filter do?

It slides one small filter (a little pattern) across the whole image and, at each spot, reports how well the image matches the pattern — producing a “feature map.” Reusing the same filter everywhere (shared weights) is the convolution, and it mirrors V1’s detectors applied all over the visual field.

Give one concrete way deep nets “see” differently from people.

Any one: adversarial examples — a tiny, invisible change flips a confident label (humans are immune); texture- vs. shape-bias — CNNs judge mostly by texture, people by shape; brittleness to blur, noise, or a new camera

Give one point for and one against treating CNNs as models of the brain’s ventral stream

For : a net trained only to recognize objects turns out to predict neural responses in monkey V4 and IT better than any earlier model, with early-to-late layers tracking the visual hierarchy. Against: nets reach human accuracy by non-human means (adversarial fragility, texture bias, odd cues), so the best predictor of brain data isn’t necessarily a good model — prediction isn’t explanation.

Why can a system be superhuman at labeling images yet useless at guiding a robot?

Recognizing static, labeled images is not active, embodied vision — real-time, 3D, novel and cluttered scenes, in a closed loop with action. (Moravec’s paradox: what’s easy for us — seeing, moving— is hard for AI, and vice versa.)

What does it mean that CNNs are “data-hungry” compared with people?

They need millions of labeled examples to learn a category, whereas a child can learn a new object from a handful of views (few-shot learning). It’s a sign the two systems learn very differently.

What are the two main types of long-term memory?

Declarative (explicit) memory — facts and events you can consciously report — and nondeclarative

(implicit) memory — knowledge that shows up in behavior without conscious recollection.

Is procedural knowledge (e.g. how to ride a bike) declarative/explicit or nondeclarative/implicit memory?

Nondeclarative (implicit). It is a motor skill that shows in performance — you can ride without being able to state the rules — which is why H.M. could learn new skills despite dense declarative amnesia.

Is episodic memory (e.g. recalling your last birthday) declarative/explicit or nondeclarative/implicit?

Declarative (explicit) — a consciously retrievable memory for a specific event you experienced, tied to a particular time and place

Is semantic memory (e.g. knowing Paris is in France) declarative/explicit or nondeclarative/implicit?

Declarative (explicit) — general facts and world knowledge you can state, independent of when or where you learned them.

What is the classic “modal model ” of memory?

Information flows through stores that differ in capacity and duration: sensory memory (brief, large) → short-term/working memory (seconds, small) → long-term memory (vast, durable), with attention, rehearsal, encoding, and retrieval moving information between them

What is the difference between short-term memory and working memory?

Short-term memory is passive storage (holding a few items for seconds). Working memory is an active workspace that holds and manipulates information — e.g. doing mental arithmetic or following a plan

Distinguish declarative from nondeclarative memory, with an example of each.

Declarative (explicit, “knowing that”) — reportable facts and events: episodic (your last birthday) and semantic (Paris is in France). Nondeclarative (implicit, “knowing how”) — shows in behavior: procedural skills (riding a bike), priming, conditioning

What did patient H.M. teach us about how memory is organized

After bilateral removal of the medial temporal lobes (incl. the hippocampus), he could not form new declarative memories (dense amnesia for facts/events) yet could still learn new procedural skills (mirror-drawing) without recalling the practice. So declarative and nondeclarative memory are separate systems, and the hippocampus is needed to form new declarative memories.

What is memory consolidation?

The process by which new memories become stable over time and gradually come to depend less on the hippocampus and more on the cortex — which is why H.M.’s old memories survived even though he could form no new ones. It unfolds over nights, weeks, and years (with a role for sleep).

What is catastrophic interference in neural networks, and how do brains seem to avoid it?

Train a network on task A then task B and it overwrites A. Brains avoid this with two complementary systems — a fast hippocampus and a slow cortex — and by interleaving old with new via replay (notably in sleep), consolidating without wiping out old memories.

What is context-dependent memory (encoding specificity), and what’s one study tip it implies?

Retrieval is better when the conditions match those at encoding — divers recalled a word list better in the same place (land vs. underwater) where they learned it. Tips: build many cues by varying where you study, and make study conditions resemble the exam (and test yourself).

Give one way memory is malleable or reconstructive

Any one: false memories (recalling “sleep” from a list that never contained it; a whole invented “lost in a mall” event); suggestibility (a leading question — “smashed” vs. “hit” — reshapes the memory and adds false detail); flashbulb memories feel like exact snapshots but drift like ordinary memories — only our confidence stays high.

Where does an LLM’s “memory” live, and how does each part map onto human memory?

Weights (learned during training, then frozen) ≈ long-term/semantic memory — facts and skills baked in. The context window ≈ working memory — a small active workspace, gone when the chat ends. Retrieval / tools ≈ an external store — the world as memory.

Compared with people, LLMs are rich in one kind of memory but poor in another. Which, and why does it matter

They have lots of semantic memory (general facts, in the weights) but almost no episodic memory— no record of your specific past interactions — unless it is explicitly stored. That’s why a fresh chat doesn’t “remember you.”

What is retrieval-augmented generation (RAG), and what human strategy does it resemble?

RAG retrieves relevant documents and adds them to the model’s context before it answers, instead of relying only on the weights. It resembles using the world as memory — looking things up in notes/books/search and bringing them into the workspace

How is retrieval “reconstructive” in people, and how does that connect to AI “hallucination”?

Human memory rebuilds an event from partial traces plus expectations, filling gaps with plausible detail (so we get confident false memories). An LLM does something similar when it hallucinates — generating fluent, plausible content that isn’t true — because it, too, fills gaps by what should come next.

Why don’t frozen LLMs obviously show catastrophic forgetting, and when can it return

They are trained once on a huge, interleaved corpus and then frozen — no weight updates while you use them — and “new” information lives in the context window or is retrieved, not written into the weights. Repeated fine-tuning on new data can bring forgetting back.

What is “using the world as memory” (cognitive offloading), and what is one downside?

We store information outside our heads — in notebooks, calendars, phones, and search — and just remember where to find it. The downside: leaning on lookups can leave us with facts at our fingertips but shallower understanding (“Google-knowing”). RAG and tool use are the AI version.

Our senses take in far more than we can fully process. In one sentence, what is the core function

of attention, and name one cost of this design

Attention selects a subset of information for deeper processing, because our capacity is limited —

we can’t process everything at once. The cost: unattended information is poorly processed or

missed entirely (we don’t perceive everything, only what we select)

In the “invisible gorilla” study, about half of viewers counting basketball passes never notice a

person in a gorilla suit. What does this show about attention and awareness?

Attention is the gateway to awareness: when attention is fully engaged elsewhere, even a large,

moving, unexpected event can go unnoticed — inattentional blindness. We consciously see what

we attend to, not everything in the visual field.

Change-blindness experiments show people miss large changes to a scene across a brief interruption. What does this suggest about how much of the world we store in our heads — and what do we rely on instead?

We don’t build a rich, detailed internal copy of the scene; we encode only what we attend to.

Because the world is stable and always available, we re-look when we need a detail — using the

world itself as external memory — so changes outside our current focus slip by

At a noisy party you can follow one conversation and tune the others out — yet your name across

the room can still grab you (Cherry’s “cocktail-party” effect). What two things does this show about

selective attention

(1) Attention is selective — we can focus on one stream and filter the rest. (2) But the filter is

incomplete — unattended input is still processed enough for significant items (your name) to break

through. Selection attenuates; it doesn’t fully block

Attention can be steered “top-down” or “bottom-up.” Give one everyday example of each.

Top-down (goal-directed): deliberately scanning a shelf for your brand of cereal, driven by your

current goal. Bottom-up (stimulus-driven): a sudden flash or a loud bang automatically grabbing

your attention

What does “attention” do inside a transformer (an LLM)

For the word it’s processing, attention weighs every other word by relevance and pulls in information from the ones that matter — a learned, weighted “lookup” over the whole context — to help predict the next word

Why is a transformer’s “attention” a false cognate for human attention — i.e., the same word for

two quite different things?

Human attention is selection under a hard capacity limit, tied to awareness — we can attend to only a little at once. Machine “attention” is a weighted lookup over all tokens at once — no capacity

limit and no awareness. Same name, different mechanism

Older sequence models (RNNs) read a sentence one word at a time; a transformer uses attention

so every word can look at every other word at once. Give one advantage of this “all-at-once” design

Either: long-range dependencies — any word can draw on any other regardless of distance (the

signal needn’t survive many sequential steps, where it fades); or parallelism — all positions compute at once, so training parallelizes, making far larger models practical

Do a transformer’s attention weights tell us why the model produced its answer (are they a faithful

explanation)

Not reliably. Attention shows where the model looked, but studies (Jain & Wallace) find the weights

are not a faithful account of the decision — different attention patterns can yield the same output.

Looking is not the same as explaining.

Compared with an LLM, does a child learn language from far less or far more input — and name

one way a child’s input differs in kind ?

Far less — a child reaches fluency on a tiny fraction of an LLM’s training data (the “data gap”). And

the input differs in kind: it’s grounded/multisensory, social and interactive, and messy/unsegmented— not clean, pre-tokenized text.

In one sentence, what is the basic thing an LLM does, over and over, to produce a sentence?

It predicts the next word — the most likely word given the words so far — then adds it and repeats,

word by word

How can a model learn that “cat” and “dog” are similar in meaning with no dictionary — just

from text?

The distributional hypothesis — “you shall know a word by the company it keeps.” Words in similar

contexts (near feed, pet, vet) get similar vectors — learned from co-occurrence counts (LSA/HAL)

or by training a network to predict a word’s neighbors (word2vec).

A base (pretrained) LLM just continues text — ask it a question and it might reply with more

questions. In concept, what turns it into a chatbot that answers you?

Extra training after pretraining: it’s shown examples of good question-and-answer responses and

tuned using human feedback, so it learns to respond helpfully instead of merely continuing text (its

input is also wrapped so it’s completing an “assistant” reply)

Development is “orderly and early.” Give two facts about how children acquire language that are

hard for a text-trained LLM to match.

Any two: learning starts before birth (newborns prefer their mother’s voice and native rhythm);

word boundaries are found from statistics by ∼8 months; children extract rules (“goed,” “wugs”),

not just memorize; the milestones hold across very different languages and even sign — all on

grounded, social input.

How does an LLM learn (not how it runs)? Describe pretraining in a sentence

By self-supervised next-word prediction at huge scale: hide the next word in a vast pile of text,

let the model guess, compare to the real word, and nudge its weights to be a bit more right —

repeated billions of times (the text is its own answer key)

What is the “poverty of the stimulus” argument

Children end up with rich, abstract grammar, yet their input is limited, noisy, and uncorrected.

Chomsky inferred that the data underdetermine the grammar, so children must bring innate knowl-

edge (Universal Grammar). This made nativism dominant for decades.

What did Saffran, Aslin & Newport (1996) show, and why did it matter for the poverty-of-stimulus

debate?

8-month-olds, after just 2 minutes of continuous nonsense speech, could tell “words” from “part-

words” — segmenting the stream using only statistics (how likely one syllable follows another). It

showed infants are powerful statistical learners, reviving learning-based theories of language

Do LLMs settle the poverty-of-stimulus debate (showing grammar is learnable, so nothing need

be innate)?

They recast it, not settle it. LLMs do learn complex syntax from data alone — strong evidence much is learnable from distribution — but they use ∼10,000× a child’s data, text-only and ungrounded. “Learnable from data” isn’t “learnable from a child’s data,” so the question stays open

Chomsky et al. (2023) called ChatGPT a “false promise.” Give one of their criticisms — and one

weakness of their argument

Criticism (any): it’s a “lumbering statistical engine”; it would learn humanly impossible languages

as easily as possible ones, so it isn’t a theory of the human language faculty; it describes/predicts

but doesn’t explain. Weakness (any): their one concrete claim was asserted, not tested (people

quickly showed ChatGPT handles it); the “impossible languages” claim is testable and turns out

false (models find impossible languages harder ); and they take the model’s self-reports at face value (an LLM has no real introspection).

What is the brain’s “language network ”?

A set of (mostly left frontal and temporal) regions that respond selectively to language. Crucially

it is dissociable from networks for reasoning, math, and theory of mind — so language and thought

rely on different brain systems (people with severe aphasia can still reason)

Schrimpf et al. (2021) compared trained language models with human brain data. What was the

striking finding?

The internal representations of trained LLMs predict human brain responses in the language net-

work surprisingly well — and, strikingly, the better a model is at predicting the next word, the better

it predicts brain activity. Next-word prediction seems to capture something the brain’s language

system also does.

If an LLM’s activity predicts brain responses, does that mean the LLM explains how the brain

does language? Why or why not?

Not by itself. Prediction is not explanation (Guest & Martin): a model can fit brain data well

while working by very different means, so a good predictor isn’t automatically a good theory. You

need controlled tests of mechanism, not just a high alignment score — the same “behavioral match̸= shared mechanism” caution as elsewhere in the course.

Why is the speech signal so hard to turn into words?

It’s continuous with no spaces (no marked word or phoneme boundaries); it’s coarticulated (neigh-

boring sounds overlap — in “bag,” /b/, /æ/, /g/ are produced together); and there’s no invariance

(the same word is a different signal every time — talker, rate, context).

What is EARSHOT, and what is it a model of ?

A minimal neural-network model of incremental human speech recognition (Magnuson et al., 2020). It takes real, continuous speech and recognizes words as they unfold over time, reproducing human-like dynamics — a cognitive model whose goal is to explain how people perceive speech.

What is Whisper, and how does its goal differ from EARSHOT’s?

Whisper (Radford et al., 2023) is a large-scale engineering speech-recognition system trained on

∼680,000 hours of audio to transcribe accurately and robustly. Its goal is accurate transcription —

not to model how humans hear — so it’s a tool, not a theory of perception.