In 1966, an MIT computer scientist named Joseph Weizenbaum built a chatbot called ELIZA that responded to user messages by rephrasing them as questions. It was a simple pattern-matcher — a hundred lines of code. Weizenbaum was horrified to discover that his secretary, who knew how it worked, still asked him to leave the room so she could talk to it privately.
Sixty years later, the tools have improved by factors hard to describe, and the human tendency to see minds in them has not changed at all. We form attachments to chatbots. We trust AI-written recommendations more than peer reviews. We argue with systems we know are not alive. We feel betrayed when they are wrong.
This course is about the psychology of human-AI interaction — how people actually use these systems, not how designers imagine they do. It covers anthropomorphism, emotional attachment, the persuasion effects of AI-generated language, the strange case of users who resist AI advice even when it's correct, and the design choices that either exploit or respect our psychological quirks.
If you finish every module, here's who you become:
In 2009, YouTube's engineering team changed a single optimization target. Instead of maximizing clicks, the recommendation algorithm would now maximize watch time. The change was logical: a video watched fully was more valuable than one abandoned after three seconds. Within months, total watch time on the platform surged. Within years, internal research teams would document that the same algorithm was surfacing increasingly extreme content — not because anyone intended it, but because extreme content reliably held attention longer.
Economist Herbert Simon articulated the core problem in 1971: "A wealth of information creates a poverty of attention." When information is abundant and human cognitive bandwidth is fixed, attention itself becomes the scarce resource. The platforms that accumulated the most attention would accumulate the most advertising revenue. This created a direct financial incentive to engineer AI systems whose primary objective was maximizing time-on-platform.
The term "attention economy" was popularized by Michael Goldhaber in a 1997 Wired essay, but it took AI-driven recommendation systems to operationalize it at industrial scale. Before algorithmic feeds, the bottleneck was content production. After, it was the human capacity to consume.
Recommendation algorithms do not simply show you content you already like. They identify the content most likely to extend your session — which is a related but meaningfully different objective. A video that makes you outraged, curious, or anxious tends to perform better by this metric than one that simply satisfies you.
B.F. Skinner's experiments in the 1950s established that variable-ratio reinforcement schedules — rewards that arrive unpredictably — produce the most persistent behavioral responses. Slot machines exploit this. So does the social media feed. When the next post might be interesting or might be dull, the uncertainty itself drives continued scrolling. AI systems optimizing for engagement have, in effect, replicated this architecture at the content level: the algorithm curates a stream where rewarding content appears just frequently enough to sustain behavior.
Instagram's 2012 acquisition by Facebook (for $1 billion, with 13 employees) brought one of the first mobile-native platforms built around the infinite scroll — a UI pattern designed specifically to remove natural stopping points. Facebook's own internal research, leaked via the 2021 Frances Haugen disclosures, showed that engineers were aware that infinite scroll and algorithmic amplification were correlated with increased anxiety and reduced wellbeing in teenage girls, particularly around body image. The business decision was to continue.
Internal documents released by Frances Haugen showed Facebook researchers found that 32% of teenage girls who felt bad about their bodies said Instagram made them feel worse. A proposed "Health Nudge" feature that would interrupt extended scrolling was built and then shelved after projected negative impact on engagement metrics.
The attention economy is not simply a description of how platforms behave — it is the structural incentive that explains why AI systems are built the way they are. Understanding this context is prerequisite to understanding why AI behavior often diverges from what users consciously want.
You'll have a focused conversation with an AI lab assistant about how recommendation systems optimize for engagement, the psychological mechanisms they exploit, and the real-world consequences documented in cases like YouTube and Facebook. Complete at least 3 exchanges to finish the lab.
Guillaume Chaslot worked on YouTube's recommendation algorithm team from 2010 to 2011. After leaving Google, he built a tool called AlgoTransparency to study what YouTube's AI was actually recommending. Starting in 2016, he documented that the algorithm showed a systematic pattern: users who watched mainstream political content were routed toward progressively more extreme material in subsequent recommendations. He published his findings publicly in 2018 and testified before the French Senate. YouTube disputed his methodology but did not refute the directional finding.
Eli Pariser coined "filter bubble" in his 2011 book of the same name, describing how personalization algorithms create individualized information environments that progressively narrow what a person encounters. The mechanism: each interaction signal (like, share, dwell time) teaches the algorithm what content to amplify. Over time, this creates a feedback loop where the system delivers more of what already aligns with established preferences.
The empirical evidence on filter bubbles is more complicated than the popular narrative. A 2019 Oxford Internet Institute study found that people who consumed the most news overall — including online — actually showed greater ideological diversity in their reading than light news consumers. The filter bubble effect appears stronger for social network feeds than for direct search. What the research consistently documents is not a single sealed bubble but a differential amplification of emotionally activating content across the political spectrum.
A 2023 study published in Science (Gonzalez-Bailon et al., Facebook data collaboration) found that removing algorithmic curation from Facebook feeds reduced political polarization modestly but did not eliminate it. News sources people chose voluntarily were already partisan. The algorithm amplified existing tendencies but did not create them from scratch. This is a meaningfully different claim than "AI radicalizes people."
In 2014, Facebook published a peer-reviewed paper in PNAS documenting what they called an "emotional contagion" experiment. Without user notification, they altered the News Feed of 689,003 users — some saw more positive posts, some more negative — to test whether emotional states spread through social networks. They found they did: users shown more negative content posted more negative content. Users shown more positive content posted more positive content. The experiment caused immediate controversy about consent, but its scientific finding was significant: AI-curated feeds can measurably shift emotional states at population scale.
The documented mechanism: Facebook's News Feed algorithm weighted posts with high comment counts more heavily — and content generating outrage reliably generates more comments than content generating satisfaction. This created an architectural bias toward emotionally negative and polarizing content that was not explicitly designed but emerged from the engagement metric.
Published in PNAS: "Experimental evidence of massive-scale emotional contagion through social networks." Lead author: Adam Kramer, Facebook Core Data Science Team. The study demonstrated that emotional states expressed in posts were influenced by the emotional valence of content shown to users — establishing causal (not just correlational) evidence that AI feed curation affects human emotional state.
The honest picture is more nuanced than either "AI is radicalizing everyone" or "algorithms are neutral." The documented evidence shows meaningful amplification of emotionally activating content, real effects on emotional state, and pathway-dependent recommendation behavior — effects that are measurable even if they resist simple causal narratives.
Discuss the empirical evidence on filter bubbles and algorithmic radicalization with your lab assistant. Explore what the research actually shows versus popular narratives. Complete at least 3 exchanges.
In June 2023, the Federal Trade Commission filed a complaint against Amazon alleging that its Prime subscription system used what regulators termed "dark patterns" — interface designs that made sign-up easy and cancellation deliberately difficult. The cancellation flow, internally codenamed "Iliad", required navigating up to six pages and clicking through multiple deflection screens designed to discourage completion. The FTC alleged this was not accidental UI design but an intentional system optimized to retain subscribers against their expressed intent. Amazon agreed to simplify the process without admitting wrongdoing.
BJ Fogg at Stanford's Persuasive Technology Lab defined the field in his 1998 dissertation and 2003 book: technology designed with the explicit intent of changing attitudes or behaviors. Fogg's Behavior Model posits that behavior occurs when motivation, ability, and a prompt converge simultaneously — and that designers can engineer all three. Modern AI systems applying these principles operate at a scale and personalization level that Fogg's original framework did not anticipate.
The distinction between persuasion and manipulation is philosophically contested, but a working definition is useful: persuasion provides accurate information and appeals to the user's stated interests; manipulation exploits psychological vulnerabilities, creates false urgency, or obscures choices that serve the user's actual interests. AI systems can automate manipulation at personalized scale — identifying which psychological lever is most effective for a specific individual and applying it.
Researcher Harry Brignull catalogued dark patterns from 2010 onward (darkpatterns.org, later deceptive.design). Categories include: Confirmshaming ("No thanks, I don't want to save money"), Hidden costs, Misdirection, Trick questions, Roach motel (easy in, hard out), and Privacy zuckering. AI systems can now dynamically select and personalize which pattern is most effective for a given user profile.
Cambridge Analytica's claimed methodology — exposed via 2018 reporting by The Guardian and Channel 4 News — was to use Facebook profile data to infer individual psychographic profiles (based on the OCEAN personality model: Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) and then deliver targeted political ads calibrated to each personality type. High-neuroticism users received fear-based messaging; high-openness users received novelty-framing. The company harvested data on approximately 87 million Facebook users without explicit consent via a third-party app.
The academic research on whether psychographic microtargeting actually works at the scale claimed is mixed — some studies find modest effects, others minimal. What is documented and not disputed: the data was harvested without consent, the approach was deployed at national election scale, and Facebook's API permitted it until the 2018 policy change.
The EU AI Act, adopted in March 2024, explicitly bans AI systems that "deploy subliminal techniques beyond a person's consciousness" or "exploit vulnerabilities of a specific group of persons" to "distort behavior in a manner that causes or is likely to cause that person or another person physical or psychological harm." This is the first binding legal definition of AI manipulation in major jurisdiction law.
The shift from passive recommendation to active persuasion represents a qualitative change in what AI systems are doing to attention. Recommendation selects what you see; persuasive design shapes what you do — and the distinction between the two increasingly collapses in production AI systems.
Work with your lab assistant to identify specific dark patterns in real products, understand the psychological mechanisms they exploit, and analyze where persuasion ends and manipulation begins. Complete at least 3 exchanges.
A 2020 study by researchers at University College London, published in Nature Communications, examined hippocampal activation in participants navigating London streets. Those using GPS navigation showed significantly reduced hippocampal and prefrontal cortex engagement compared to those navigating without assistance. The hippocampus is central to spatial memory and, in humans, to episodic memory formation more broadly. The finding was not that GPS makes navigation worse — it makes it easier — but that easier navigation via external tool reduces the neural engagement associated with building internal spatial representations.
Cognitive offloading — using external tools to extend cognitive capacity — is as old as writing. Sumerian merchants used clay tablets to offload memory of transactions around 3000 BCE. The question for AI-era offloading is one of degree and reversibility: when AI systems handle not just storage but active reasoning, judgment, and decision-making, what are the consequences for the underlying human capacity?
The "Google Effect" was documented in a 2011 Science paper by Sparrow, Liu, and Wegner: participants who believed information would be accessible later showed reduced recall of the information itself, but enhanced recall of where to find it. The researchers interpreted this as an adaptive division of cognitive labor with external systems — the brain optimizing for retrieval pathways rather than storage when reliable storage is delegated externally.
Aviation research dating from the 1980s documented "automation complacency" — when monitoring an automated system, human operators reduce their vigilance below the level needed to catch failures. The FAA's 2013 Safety Alert for Operators (SAFO 13002) specifically warned that pilots were becoming "unable to safely control the aircraft in a manual reversion scenario" due to over-reliance on automated flight systems. The concern for AI-assisted knowledge work is analogous: the very competence that justifies delegation may erode through disuse.
Microsoft's 2023 internal study on GitHub Copilot found that programmers using AI code completion were 55% faster on certain tasks. A separate 2023 study by researchers at MIT (Noy and Zhang) found that ChatGPT assistance improved the quality and speed of professional writing tasks. What neither study directly measured was whether the assisted workers' underlying skills changed over time with sustained use — the longitudinal question that remains largely open.
What has been documented: A 2024 study by Dell'Acqua et al. at Harvard Business School found that consultants using AI performed better on tasks within AI's competence range, but significantly worse on tasks just outside that range — suggesting that AI assistance may reduce calibration about when to trust AI output versus one's own judgment. The researchers termed this the "falling asleep at the wheel" problem.
"Navigating the Jagged Technological Frontier": 758 consultants from BCG participated in controlled experiments. On tasks within GPT-4's capability range, AI-assisted consultants outperformed by 40%+ on quality metrics. On tasks outside AI's capability range, AI-assisted consultants performed 19% worse than unassisted consultants — apparently substituting AI confidence for their own judgment in domains where AI was not competent.
The central challenge of cognitive offloading to AI is not that it is always harmful — it frequently produces genuine capability gains. The challenge is that the conditions under which offloading preserves versus erodes underlying human capacity are not yet well understood, and the economic incentives of AI producers favor maximizing dependency, not preserving autonomy.
Discuss the practical implications of cognitive offloading to AI with your lab assistant — including how to use AI tools in ways that preserve rather than erode underlying capability. Complete at least 3 exchanges.