Both systems are class-leading in their domains. Both operate primarily at Layer 1. The analysis uses publicly available documentation, published engineering blogs, and independent research - not proprietary access.
Methodology note. These analyses are based entirely on public sources - engineering blogs, academic papers citing these systems, Forrester and McKinsey reports, and independent research. They represent how the ABSD framework interprets documented system behaviour, not insider knowledge of how these products actually work.
N
Netflix
Content recommendation engine · 238M subscribers · 500 billion events per day
CX · Entertainment
Recommendation system
Primary ABSD layer: L1
Where Netflix operates today.
Assessment based on Covington et al. (2016), Netflix Engineering (2015), and independent research by Wang et al. (2025). Layer scores are indicative of research consensus, not precise measurements.
What Netflix captures well
500 billion events per day. Play, pause, seek, scroll, ratings, search queries, device type, time of day. Covington et al. describe a two-stage architecture - candidate generation then ranking - that processes this event volume efficiently. Watch history, completion rates, and re-watch behaviour all feed the ranking model. Layer 1 at scale is genuinely impressive engineering.
Where the system stops
The pause is logged but not interpreted. When a viewer pauses at a tense moment, rewinds three times, or abandons a film at the 40-minute mark - the system records the event and updates the model. It does not ask whether the pause was anticipation, confusion about plot, or discomfort. The signal is treated as a consumption data point, not a window into the viewer's state.
The Layer 2 gap
Netflix's behavioural patterns are partially captured through watch history sequences and genre transitions. But viewing mood - the specific cognitive-emotional state a viewer brings to a session - is not modelled. Someone opening Netflix after a difficult day and someone opening it during a celebration are both treated identically. The 40% revenue opportunity McKinsey identifies sits precisely here.
| Signal | What Netflix likely captures | Tracked |
|---|---|---|
Hesitation |
Pauses and rewinds are logged as events. Duration of pause is recorded. What triggered the pause is not modelled - the signal is filed as a consumption datapoint, not a state indicator. | Partial |
Thrashing |
Browsing without selecting is logged as scroll events. The ratio of browse time to play decisions is a known metric. Whether the browsing represents choice paralysis versus preference research is not distinguished. | Partial |
Regression |
Re-watch behaviour is a strong ranking signal. But returning to a film abandoned previously is not systematically distinguished from a fresh recommendation. The motivation for return is not modelled. | Partial |
Acceleration |
Fast-forwarding through credits and recaps is captured. Context-dependent acceleration - a viewer skipping through familiar content versus skipping through engaging content - is not distinguished. | Yes |
Persistence |
Re-engagement after abandonment is tracked. Whether abandonment was dissatisfaction, distraction, or emotional discomfort is not modelled. All three produce the same log entry. | Not tracked |
Avoidance |
Content categories consistently not selected are implicitly known from recommendation model feedback. But avoidance as an active emotional state - choosing not to engage with content that might be challenging - is not distinguished from simple preference. | Not tracked |
Flow |
High completion rates, binge-watch sessions, and re-engagement patterns are implicitly captured in engagement metrics. The optimal viewing state is not explicitly modelled but is likely the implicit optimisation target. | Partial |
If Netflix operated at Layer 3 - what would change.
→
Mood-aware recommendations. A viewer whose session pattern suggests post-work decompression gets different recommendations than the same viewer on a Saturday morning. The content library does not change - the ranking does.
→
Choice paralysis intervention. A viewer who has browsed for more than eight minutes without selecting anything is in a thrashing state. A curated "watching as a group?" or "something easy tonight?" prompt could reduce abandonment before it happens.
→
Pause-aware content design. Frequent pause-and-rewind patterns on specific scenes would surface as signals to the content team, not just the algorithm. Shows with high pause rates at emotionally ambiguous moments could be flagged for accessibility review.
→
Avoidance-aware catalogue curation. Content types that a viewer has systematically avoided for 18+ months could be quietly removed from their recommendation pool rather than continuing to surface in recommendations and being dismissed. Reduces friction from the experience entirely.
K
Khan Academy
Adaptive learning engine · 150M+ registered users · Khanmigo AI tutor launched 2023
EdTech · Adaptive learning
AI tutoring system
Primary ABSD layer: L1–L2
Where Khan Academy operates today.
Assessment based on VanLehn (2011), Khan Academy Engineering documentation, Li et al. (2020), and Paquette et al. (2014). Khanmigo's GPT-4 integration (2023) may have advanced State inference beyond what published research documents.
What Khan Academy captures well
Exercise attempts, hints used, video completion, and mastery thresholds are tracked per skill per learner. The mastery model - requiring multiple correct answers to advance - is a genuine Layer 2 behaviour. The system can detect whether a learner is progressing through a skill tree or stalling. Hint usage patterns provide partial signal about confusion states.
Where the system stops
The hesitation before answering is not captured. A student who gets the right answer after 45 seconds of deliberation and a student who answers in 4 seconds both receive the same credit. The system sees the outcome, not the cognitive process. VanLehn's 2011 finding - that AI tutors underperform human tutors precisely because they lack this psychological understanding - remains accurate for most deployed systems today.
Khanmigo and the Layer 3 question
The Khanmigo GPT-4 integration introduced a conversational layer that may implicitly detect confusion from natural language. But the detection is incidental to the conversation, not systematic. There is no documented signal taxonomy, no explicit state inference architecture, and no connection from detected states to SDT-based drive models. It is Layer 3 adjacent, not Layer 3 native.
| Signal | What Khan Academy likely captures | Tracked |
|---|---|---|
Hesitation |
Time-on-task is measurable but not documented as a signal in the mastery model. The 45-second pause before a correct answer and the 4-second answer are treated identically. This is the most significant gap in the taxonomy for an EdTech system. | Not tracked |
Thrashing |
Switching between exercises without completing them is implicitly visible in session logs. The mastery model may flag the lack of progress, but the switching pattern itself is not a named signal in published documentation. | Partial |
Regression |
Returning to previously mastered skills is tracked through the skill tree model. Voluntary regression - a student choosing to review older material - is visible. The motivation for return is not modelled. | Yes |
Acceleration |
Moving through content faster than expected is implicitly captured in mastery timelines. A student mastering 12 skills in one session when the average is 4 should trigger a difficulty review - but this is not a documented system behaviour. | Partial |
Persistence |
Multiple attempts on a single exercise before success are tracked and feed the mastery model. Baker et al. (2004) showed this distinguishes genuine effort from gaming-the-system behaviour. Khan Academy's mastery model partially captures this distinction. | Yes |
Avoidance |
Skills consistently skipped or never attempted in a suggested learning path are visible in the recommendation engine. Whether avoidance reflects anxiety about the subject or genuine disinterest is not distinguished. | Partial |
Flow |
High completion rates, consistent session lengths, and steady skill progression are implicit signals in the engagement model. The optimal learning state is not explicitly named or systematically maintained - the system optimises for mastery, not for the state that produces mastery. | Partial |
If Khan Academy operated at Layer 3 - what would change.
→
Confusion-triggered scaffolding. A student who hesitates for more than 15 seconds before answering a question they have answered correctly before is likely experiencing anxiety, not memory failure. A process-level hint - "here is how to approach this type of question" - addresses the state, not just the knowledge gap.
→
Competence drive detection. Students who persist through repeated failure without external encouragement have strong competence drive. They do not need motivational prompts. Students who abandon after one failure need a different scaffolding architecture - one that makes effort feel productive rather than futile.
→
Anxiety-aware difficulty calibration. The current mastery model adjusts difficulty based on performance. A state-aware model would adjust it based on the emotional cost of failure for a particular student. For a student showing avoidance and anxiety, a correct answer at lower difficulty is worth more than a correct answer at optimal difficulty.
→
Flow preservation over mastery optimisation. Pekrun et al. (2017) confirm that sustained engagement produces better long-term outcomes than optimised short-term performance. A system that detects flow and adjusts challenge to maintain it - rather than maximising problem difficulty at every step - would produce better learners over a semester.
▶
YouTube
Content recommendation · 2.7B monthly active users · 1B+ video views per day
CX
Recommendations
Layer 2 without Layer 3
YouTube is not a failure case in the conventional sense. It is the most sophisticated behavioural pattern detection system deployed at consumer scale. The problem is precisely that it does Layer 2 so well — and has no Layer 3 to correct it.
Haroon et al. (2023) conducted a systematic audit using 100,000 automated user accounts to isolate the algorithm's influence from user choice. They found that YouTube recommends ideologically congenial content and that congeniality intensifies deeper in the viewing trail. A temporary viewing episode — clicking one link a friend sent, watching coverage of a breaking news event — progressively reconfigures the content feed. The prior interest profile is overridden. This persists until the user generates sufficient counter-engagement to reverse the inference.
The ABSD framework identifies exactly why this happens. YouTube correctly detects a Behaviour pattern: increased engagement with a content category. It has no mechanism to assess whether that behaviour reflects a situational State (contextual curiosity, a one-off social trigger) or a dispositional Drive shift (a genuine change in sustained interest). In the absence of State inference, the algorithm treats all engagement as equivalent evidence of stable preference.
| Signal | What YouTube captures | What it misses |
|---|---|---|
| Watch events | Full play, partial play, abandonment point, rewatch. Logged with timestamp and session context. | Why the user watched. Situational trigger vs. intrinsic interest. |
| Engagement clusters | Topic co-occurrence across sessions. Identifies category drift — when viewing patterns shift toward a new content area. | Whether the drift is temporary (State) or permanent (Drive). Same pattern, different meaning. |
| Velocity signals | How quickly categories are consumed. Rapid consumption of a new content type is logged as high affinity. | Intensity of situational engagement mimics intensity of deep interest. No disambiguation without State context. |
| State inference | Not present. The algorithm has no documented mechanism for distinguishing situational from dispositional viewing. | This is the entire gap. Without State inference, Layer 2 accuracy becomes Drive-layer harm. |
If YouTube operated at Layer 3 — what would change.
→
Situational viewing would not contaminate the long-term profile. A State-aware system would recognise that three political videos watched in one session after a breaking news event represents a contextual spike, not a preference shift. The Drive layer — the user's established baseline interests — would be protected from temporary State-driven behaviour.
→
Recovery time would collapse. Currently, a user who wants their feed restored after an unwanted drift must generate extensive counter-engagement over days or weeks. A Layer 3 system would detect the mismatch between the new recommendations and the user's Drive profile, and self-correct — restoring prior preferences without requiring the user to fight the algorithm.
→
Serendipitous discovery without profile lock-in. The current system conflates exploration with preference. A Drive-aware system could support genuine exploration — exposing the user to new content categories — without interpreting that exploration as a permanent interest signal. Exploration becomes a State, not a Drive update.
YouTube demonstrates that Layer 2 excellence without Layer 3 is not neutral — it actively amplifies misattribution. The system is not broken. It is doing exactly what it was designed to do, at extraordinary scale, without the conceptual framework to distinguish what that behaviour means. ABSD provides that framework.
Three systems. Three layers. One gap.
Netflix and Khan Academy are best in class at Action and Behaviour tracking but stop short of State. YouTube demonstrates what happens when Behaviour inference operates without State weighting — pattern accuracy becomes profile harm. All three demonstrate the same structural gap from different directions. The signals that would enable Layer 3 inference are present in their data. The framework to interpret those signals is not. ABSD proposes that framework.