1. TD-Gammon (1992) was trained by researcher Gerald Tesauro at IBM using a technique called temporal-difference learning. What does this method update?
✓ Correct. TD-learning bootstraps from intermediate predictions: if the model predicted 60% win probability at move 20 and 45% at move 21, it adjusts the move-20 prediction toward 45% without waiting for the game to end. This allows learning from every position in every game, not just the final result.
✗ Temporal-difference learning updates the network's position evaluations using differences between successive predictions — not waiting for the final game outcome. If the predicted win probability shifts between move 20 and move 21, the move-20 evaluation is corrected toward the move-21 estimate.
2. How can a player model be used to personalize the game experience beyond difficulty adjustment?
✓ Correct. Player models drive many personalization applications: recommending which content to try next, reordering menus to surface preferred modes, timing re-engagement messages, adjusting monetization prompts, and cueing narrative content that matches the player's demonstrated interests. Difficulty is just one dimension.
✗ Player models are used for many personalization applications beyond difficulty — content recommendations, UI personalization, quest suggestions, re-engagement timing, and monetization prompts. They are a general-purpose representation that any game system can act on.
3. AI image generation tools have been documented to over-represent certain demographics in hero roles. What drives this pattern?
✓ Correct. The model learned from training data in which hero characters skew toward certain demographics. It reproduces that statistical correlation — "hero = [demographic X]" — without any explicit rule encoding it.
✗ Algorithmic bias in hero representation stems from training data: if the source material consistently depicts heroes as a particular demographic, the model learns that pattern and replicates it when asked to generate a hero. No explicit rule is needed — the correlation is in the data.
4. The module describes thinking of AI as a "junior team member." What does this metaphor imply about how designers should use AI outputs?
✓ Correct. A junior team member produces drafts that need review, correction, and direction from someone with more experience and context. AI works the same way — useful, energetic, fast, and requiring experienced oversight to produce work that actually meets the design goal.
✗ The junior team member metaphor means: useful for first drafts and high-volume generation, but requiring an experienced designer's direction, quality control, and judgment to turn raw output into work that serves the actual creative goal. The AI is not self-directing.
5. Dark patterns in games are best defined as:
✓ Correct. Dark patterns are deliberate design tricks — artificial scarcity, near-miss engineering, sunk-cost nudges — that exploit psychology to extract spending or behavior that players would not choose if presented with full information.
✗ Dark patterns are UI/UX design techniques that deliberately manipulate psychological biases — hiding unsubscribe options, engineering near-misses in loot animations, manufacturing false scarcity — to push players toward purchases against their own interests.
6. Which of the following best defines a Finite State Machine (FSM) in the context of NPC AI?
✓ Correct. An FSM has a finite, bounded set of states; the NPC is in exactly one at a time; and transitions fire when specific conditions are met.
✗ An FSM is defined by a fixed set of states with one active at a time, and transitions triggered by conditions — not scoring, tree traversal, or planning.
7. FOMO (Fear of Missing Out) mechanics in games typically work by:
✓ Correct. FOMO mechanics manufacture urgency — the battle pass expires, the limited skin leaves tomorrow, the event ends at midnight. The time constraint removes the player's ability to think slowly and deliberately, which is the entire point.
✗ FOMO mechanics work by creating artificial time pressure — limited-time items, seasonal events, expiring passes — that forces players to decide immediately rather than carefully. The urgency is designed, not incidental.
8. Why do some players consider opaque DDA systems to feel "unfair," even when those systems are making the game easier for them?
✓ Correct. The psychological issue is one of earned achievement. If the game secretly became easier, a player's eventual success may feel like the system helped them rather than reflecting their own improvement. Transparency (explicit assist modes, like in Celeste) is one design response — it preserves player agency by making the assist opt-in and visible.
✗ The fairness concern is about earned achievement: if a player suspects the game got easier without being told, their success feels less meaningful. This applies even in single-player contexts. Alternatives like opt-in assist modes (Celeste) address this by making assistance transparent and voluntary.
9. COPPA (Children's Online Privacy Protection Act) requires game developers to:
✓ Correct. COPPA's core requirement is verifiable parental consent before collecting any personal information from children under 13. The FTC's $520 million action against Epic Games in 2022 is the highest-profile enforcement of this rule in gaming history.
✗ COPPA requires verifiable parental consent — not just a checkbox "I am over 13" — before collecting personal data from children under 13. Epic's 2022 FTC fine of $275 million for COPPA violations illustrates how seriously this requirement is enforced.
10. Which of the following is something current AI systems definitively cannot do in game design contexts?
✓ Correct. Game feel — the embodied, intuitive sense that an interaction is or is not satisfying — cannot currently be generated, evaluated, or approximated by AI. It requires a human who has experienced what good game feel feels like.
✗ Game feel — the tacit, embodied sense that a jump arc is right or a combat hit lands satisfyingly — cannot be evaluated by AI. It requires human experience and human embodiment. This is exactly why it becomes more valuable as AI handles more of the production work.
11. Mass Effect's "dialogue wheel" UI is associated with which studio?
✓ Correct. BioWare introduced the radial dialogue selection wheel in Mass Effect, replacing the numbered list dialogue menus common in earlier RPGs.
✗ BioWare created the Mass Effect dialogue wheel — a radial UI letting players select tone and direction of NPC conversation, which became a widely copied design pattern.
12. Left 4 Dead's AI Director dynamically controls which game elements based on measured player stress levels?
✓ Correct. The AI Director tracks a "stress" or intensity model from player health, progress, and combat frequency. After intense combat it reduces enemy pressure and spawns healing items; when things are too easy it escalates spawns and reduces resources — all to maintain a paced, dramatic experience.
✗ The AI Director controls enemy spawns, item placement, and music — not graphics, story routing, or matchmaking. Its core function is pacing: it reads player stress and responds by modulating intensity up or down to maintain a dramatic experience.
13. A game company uses k-means clustering on player behavioral data to group players into segments and then personalizes content for each segment. What is one key limitation of this approach for long-term players?
✓ Correct. Static clustering is a snapshot. A cautious new player classified as a "Pacifist Explorer" after their first five sessions may behave like a "Veteran" after 50 hours. Systems that do not re-cluster or use sequence models will keep personalizing for an outdated player type, degrading relevance over time.
✗ The limitation is temporal: k-means clustering is a snapshot, not a continuous model. Players evolve — a new player's cluster may be entirely different from their cluster at 50 hours. Solutions include rolling-window re-clustering or sequence models (HMMs, RNNs) that track behavioral trajectories over time.
14. In a Behavior Tree, what does a Selector node do when it encounters its first child that returns Success?
✓ Correct. A Selector acts as logical OR — it returns Success the moment any child succeeds, without evaluating remaining children.
✗ A Selector (Fallback) node is logical OR: the first child that returns Success causes the Selector to immediately return Success and stop evaluating further children.
15. What is "seeded randomness" and why is it useful in procedural content generation?
✓ Correct. A seeded PRNG produces a deterministic sequence — the same seed always generates the same content. This enables Minecraft's shareable world seeds, Spelunky's daily challenge fairness (all players get the same layout), and Dwarf Fortress's entire civilization histories encoded in a 64-bit integer.
✗ Seeded randomness means using a pseudo-random number generator initialized with a fixed value (the seed). The same seed always produces the same sequence — making content reproducible, shareable, and verifiable. Minecraft world seeds, Spelunky daily challenges, and Elite's galaxies all depend on this deterministic property.
16. DALL-E 3 is developed by OpenAI and is notable for which integration that makes it accessible to game writers and designers without a separate account?
✓ Correct. DALL-E 3 is integrated into ChatGPT, making it accessible without a separate API key or account — users can generate images directly within a ChatGPT conversation. It is particularly strong at following detailed text descriptions.
✗ DALL-E 3 is integrated into ChatGPT — not Photoshop, Unreal Engine, or Midjourney. This makes it uniquely accessible within an existing writing and ideation workflow for designers who are already using ChatGPT.
17. In Drachen et al.'s 2012 Tomb Raider study, which algorithm was used to identify the four player archetypes from behavioral telemetry?
✓ Correct. Drachen's team applied k-means clustering to behavioral telemetry from 10,912 Tomb Raider: Underworld sessions — no surveys, no labels, no designer rules. The four archetypes (Veterans, Solvers, Pacifists, Runners) emerged purely from the statistical structure of behavioral data.
✗ Drachen used k-means clustering — an unsupervised algorithm — on behavioral data from nearly 11,000 sessions. No player surveys, no neural networks, no designer-defined rules were involved. The archetypes emerged from statistical patterns in the data alone.
18. Under GDPR, what is the "right to erasure" and why does it matter for games that use AI behavioral profiling?
✓ Correct. GDPR's right to erasure (Article 17) means any AI behavioral profiling system must be able to completely delete a player's data on request — a technical requirement that must be built in from the start, not retrofitted.
✗ The right to erasure under GDPR (Article 17) means players can demand deletion of their personal data, including behavioral profiles. AI systems that build player profiles must be designed from the start with a complete deletion pathway — it cannot be an afterthought.
19. Retrieval-Augmented Generation (RAG) applied to NPC dialogue means:
✓ Correct. RAG stores memories and lore externally (in a vector database or structured store), then retrieves the most relevant facts for the current exchange — giving the NPC access to a much larger knowledge base than the context window alone allows.
✗ RAG retrieves relevant memories or facts from an external store and injects them into the current context. The NPC draws on this retrieved material when responding, extending its effective memory well beyond what fits in a single context window.
20. Regarding IP and copyright for AI-generated art, which statement most accurately reflects the current legal position in the United States as of 2023?
✓ Correct. The US Copyright Office's February 2023 ruling established that AI-generated images without meaningful human creative input lack the human authorship required for copyright protection. Studios need documented human modification to protect their AI-assisted assets.
✗ The Copyright Office ruled in February 2023 that purely AI-generated images are not copyrightable — neither the AI tool company nor the prompter automatically holds copyright. Meaningful human creative selection and modification is required for any protection.