Isn't the emotion economy just products sold to NPCs?

The verdict up front: No, and the reasons for "no" are more interesting than a "yes" would be. The proposition is wrong twice over: emotions are not hard-coded scripts (they are constructed, which is precisely why they can be sold), and "automatic" does not equal "NPC" (statistically predictable is not individually manipulated). What the emotion economy sells into is the autopilot mode everyone falls into, a real, harvestable layer; but what it mostly does is supply the concepts and triggers out of which your emotions are built and harvest your reactivity, while falling far short of reliably reprogramming your beliefs. "NPC" is not a kind of person, it is a mode; and the genuinely premium product is not sold to that mode, it is the exit from it, except that this exit is small, occasionally backfires, and runs against commercial incentives.

1. Is the premise right: are emotions "hard-coded"?

The first half of the claim, "emotion is hard-coded and thought-free", is half right, and the wrong half is the half that matters most.

• There is a layer that is fast, automatic, and hard to veto in the moment. [Solid] LeDoux's animal work established a thalamus→amygdala "low road" that triggers freezing, startle, and heart-rate spikes before the cortex has identified the stimulus (about 12ms in rats); fast automatic "good/bad" evaluation is solid too (Fazio's evaluative priming). And roughly 40 to 45% of everyday behavior is habit (Wood 2002), programmed responses fired by context cues that can override your current goals. This is the kernel the claim gets right: that jolt, that flare of anger, is real, fast, and does not wait for you to think.
• But felt emotion is not a hard-coded reflex. [Real but Overstated → Flips to the Opposite] The most dramatic finding: Joseph LeDoux, the man who popularized the "automatic fear circuit" story, has personally disowned the idea that the amygdala is a fear center. He splits emotion into two systems: non-conscious "survival circuits" (automatic threat detection, present even in bacteria) and consciously experienced "feelings," which are assembled separately; he has also called for renaming "fear conditioning" to "threat conditioning" (LeDoux & Pine 2016). The human lesion evidence is sharper still: patient S.M., with total bilateral amygdala destruction, felt no fear at all facing snakes, haunted houses, or horror films, yet still panicked under CO₂ inhalation, the "fear center" is neither necessary nor sufficient for the feeling.
• "No fingerprints." [Real but Overstated; Actively Contested] Barrett's theory of constructed emotion: emotions are built on the spot out of core affect (valence × arousal) + learned concepts + context, not triggered by dedicated circuits. Siegel et al. 2018 (202 studies) found no consistent autonomic "fingerprint" for discrete emotions; the "expected" facial expression correlates with actual emotion at only r≈.30 and appears only ~15 to 25% of the time during genuine emotion (of 55 judo champions, only 8 showed a "Duchenne smile"). The other side has real evidence too (Saarimäki 2016: multivariate neural decoding classifies 6 emotions across people), but Barrett's rebuttal is that what gets decoded is a statistical summary, not a biological essence. This debate is genuinely unsettled.
• The two most dramatic pieces of "automaticity" evidence are ground zero of the replication crisis. [Solid] Subliminal advertising hypnotically changing behavior was fabricated (Vicary later admitted he made up his 1957 numbers); Bargh's "elderly words make you walk slower" failed a double-blind replication (Doyen 2012: participants slowed only when the experimenter knew the hypothesis = an expectancy effect); Kahneman publicly retracted his "you have no choice but to believe it." So the strong version, "invisible cues reliably override your will", has been falsified, not established. Even the System 1 vs System 2 dichotomy itself is called "a convenient and seductive myth" by Bargh, the founder of automaticity research (Melnikoff & Bargh 2018).

Bottom line: the accurate statement is "triggers can be hard-coded, feelings are constructed," and the seam between the two is exactly where concepts, framing, and trainable awareness get in.

2. The fatal irony: the critics of the emotion economy are all constructionists

This is the sharpest finding of the round. The entire critical tradition aimed at the emotion economy (Hochschild's "emotional labor," Illouz's "emotional capitalism," Ahmed's "affective economies") is thoroughly social-constructionist: you can only speak of "managing," "commodifying," or "circulating" emotion if emotion is malleable. Hochschild's core mechanism is that sustained emotional labor internalizes "feeling rules" until the "managed heart" runs automatically (and what burns people out is the falseness of surface acting, meta-analytic ρ≈.39.48 with exhaustion; deep acting, by contrast, correlates positively with performance).

So the popular line "the emotion economy = products sold to NPCs" borrows the critics' conclusion (emotion is manufactured) while contradicting the critics' premise (it assumes emotion is an automatic reflex):

If emotion were a hard-coded, thought-free output, there would be nothing malleable to shape and nothing to sell. The commodification thesis requires plasticity, not autopilot.

In other words: the emotion economy works precisely because emotion is constructed. It sells not by loading instructions into a sealed reflex machine but by supplying the concepts and triggers you use to construct your feelings, advertising has never sold the product, it sells a concept in which "owning it = you will feel a certain way." (Caveat: Hochschild is in fact a bio-social hybrid theorist; the direction holds, don't take the details as absolute.)

3. The "NPC" frame: a category error that points at a real problem

• "The mind is mostly automatic, therefore people are NPCs with no free will" is a non sequitur. [Solid] It slides from a subpersonal fact (neural processes are causally determined and often unconscious) to a personal-level conclusion (the agent is empty). Compatibilism (the majority position in the field, around 59%) denies that "determinism = unfreedom": freedom is acting on your own reasons, which is fully compatible with those reasons being neurally caused; the opposite of freedom is external coercion, not causation. And the hardest pillar of the "no free will" case, Libet's "your brain decides before you do", was reinterpreted by Schurger 2012 as an averaging artifact of a noisy accumulator (and it only ever involved meaningless finger twitches).
• The ceiling on predictability measures exactly the distance between "exploitable" and "hard-wired." [Real but Overstated; Direction Solid] The largest study of predicting individual life outcomes (Fragile Families: 160 teams, ~13,000 variables) topped out at R²≈0.2; algorithms inferring personality from Likes reach r≈0.56 (beating spouses) yet still explain only about a third of the variance; psychologically targeted ads do lift purchases by +50%, but that runs on a ~0.01% conversion rate (3.12 million people yielding 390 orders), and the study was later criticized for not actually randomizing, with Cambridge Analytica-style "mind control" widely regarded as inflated. Human mobility is ~93% predictable, but that mostly says "you aren't moving most of the time." Statistically predictable ≠ hard-wired.
• Treating other people as "NPCs" is, psychologically, the textbook mechanism of dehumanization [Judgment], denying others a mind is exactly what it means to see people as automata lacking warmth, emotion, and agency (Haslam). Variation in inner experience is genuinely large (inner speech ranges from 0% to 75% across individuals, with 5 of 30 people reporting none at all; aphantasia at ~0.8%), but that is a continuously varying degree of reflectiveness within every person, never a binary split into a "soulless species."
• The honest reconstruction: the emotion economy sells into the NPC-mode everyone falls into, not to "NPC-people." And meta-awareness (noticing that you are on autopilot) is a measurable, trainable lever, precisely what an "NPC" would by definition lack.

4. The exploitation is real, but mythologized

The "selling" half of the claim: the engagement economy does harvest those automatic moral-emotional triggers, but the jump from "harvesting reactions" to "controlling you" falls apart under the largest causal test we have.

What's real [Real but Overstated]: engagement systems reward out-group / moral-emotional language (each additional out-group word raises share odds by +67%, Rathje 2021, 2.73 million posts), but that is an observational text regression, describing "what spreads," not showing "the platform manufactured the impulse" (Burton 2021 got equally good predictions from a nonsense-word dictionary). Negativity driving clicks has clean causal evidence (Upworthy's 100,000-headline A/B corpus: each additional negative word raises click-through by +2.3%), but it measures clicks, not shares, and the per-word effect is tiny. Dark patterns pushed forced acceptance from 11.3% to 41.9% (Luguri 2021), but that came from a hypothetical online experiment with no real money, never replicated in the field.

What's mythologized [Solid]:

• Subliminal advertising is a founding hoax (Vicary fabricated it); subliminal priming can only push a goal you already have (the Lipton prime works only if you are thirsty) and has no significant effect on real purchase choices.
• The "hypodermic needle / magic bullet" model is a retrospective straw man that mainstream effects research abandoned 60 years ago (Klapper 1960: media mainly reinforce existing attitudes rather than convert them).
• The largest causal test flatly contradicts "reprogramming": Guess 2023 (Science, in collaboration with Meta) randomly replaced the algorithmic feed with a reverse-chronological one for 3 months, which massively changed what people saw yet had almost zero effect on affective polarization or issue attitudes, "monetizing reactivity ≠ reprogramming people."
• Advertising ROI is often about zero or even negative (eBay's randomized shutdown of paid search: zero return on branded keywords, negative on non-branded); the best estimate of the average persuasive effect of political campaigning in general elections is essentially zero (Kalla & Broockman 2018); and observational methods systematically overstate ad effects by 2 to 13× (Gordon 2019).
• Emotion AI that reads faces largely fails on its own scientific terms: Barrett et al. 2019 concluded that facial configurations cannot diagnose internal emotion; the EU AI Act outright bans emotion recognition in workplaces and schools, citing in black and white "limited reliability, lack of specificity and limited generalisability," with penalties at the top tier.

The one solid exception: fear appeals reliably change behavior under bounded conditions (Tannenbaum 2015, d≈0.3). Net ledger: this machine reliably harvests reactivity, but cannot reliably reprogram people.

5. The exit = your product territory, but it is small

• Cognitive reappraisal can interrupt automatic reactions, but the effect is only small-to-medium, and it can backfire. [Real but Overstated] Reappraisal d+≈0.36, perspective-taking 0.45 (Webb 2012), but the whole class of attentional-deployment strategies averages d=0.00, and "suppressing the experience" comes in at d=−0.04. Worse: reappraising a problem you could actually have solved is associated with higher depression (Troy 2013), the mirror turns into an excuse not to act.
• Mindfulness / meta-awareness is systematically oversold. [Real but Overstated] Against active controls it is no better on most outcomes; about 1/4 of regular meditators report distinctly unpleasant meditation experiences. Even the famous line "between stimulus and response there is a space" is folklore misattributed to Frankl. Teaching people to spot manipulation (prebunking): d≈0.3 to 0.7 in the lab, but only h=0.09 in the field, with no measurable effect on actual behavior, decaying within weeks.
• Ethically, the line between manipulation and persuasion is drawn at "covertness," not at "does it touch automatic emotion." [Real but Overstated] (Susser/Nissenbaum): manipulation = covertly exploiting your decision-making vulnerabilities so that you act for reasons you cannot identify. So mobilizing hard-coded affect transparently, with your informed and reflectively endorsed consent, is not manipulation; doing it covertly and against your own goals is. Which means "people have rich inner lives" is no rebuttal to the charge.
• "Selling agency rather than exploiting automaticity" has a proper academic name: boosting / self-nudging (Hertwig & Grüne-Yanoff 2017). It builds the person's own competence, is transparent, requires their active cooperation, and its effects persist after the intervention ends; it is the architectural opposite of nudging (which borrows your biases without telling you), and it is the academic skeleton of a "mirror, not oracle" product philosophy. The honest tail: "selling the exit" is a justified ethical position but an unproven business model, and the incentive gradient still tilts toward "extraction."

Synthesis: answering "isn't the emotion economy just products sold to NPCs?"

No, but it grazes a more precise truth. Taken apart:

Three notes for a "mirror, not oracle" product

1. What you sell is not "the person who has left autopilot," it is "the moment of leaving it." The emotion economy sells to the NPC-mode everyone falls into; your differentiation is selling the exit (the constructed layer is where trainable awareness and agency live). The positioning is right.
2. You are natively on the ethically right side, because you are transparent, require the user's active cooperation, and leave the capability with the user (the definition of boosting). What the emotion economy gets condemned for is never "touching automatic emotion," it is doing so "covertly and against the user's own goals." You do the reverse.
3. But do not overpromise what the exit can do. Reappraisal d≈0.36, mindfulness no better than active controls, prebunking near zero in the field, and it can backfire. The honest product claim is: give people a lever for seeing clearly and choosing for themselves, not "see clearly and you are free"; and stay clear-eyed that "selling the exit" is a justified ethic and an unproven business, running against the incumbents' incentive gradient.

In one line: the emotion economy is not "sold to NPCs," it is sold to everyone's NPC-moments; and it can be sold because emotion is constructed and can be supplied with concepts and triggers, not because people are empty. What you should build is not one more product stuffed into that moment, but handing back to the person the capability to see the moment and be able to choose to leave it. This path is right, and it is clean; just don't forget that it is small, and that it runs into the wind.

Primary sources (all verified): automaticity/construction, LeDoux & Pine 2016, LeDoux & Brown 2017, Barrett 2017 and Barrett et al. 2019 (PSPI), Siegel et al. 2018 (202 studies), Saarimäki 2016, patient S.M.; habit, Wood/Quinn/Kashy 2002; replication crisis, Doyen 2012, OSC 2015, Kahneman's retraction; dual-process, Melnikoff & Bargh 2018. NPC/agency, Libet 1983 and Schurger 2012, compatibilism (SEP), Mele 2009; predictability ceilings, Salganik et al. 2020 (Fragile Families), Youyou/Kosinski/Stillwell 2015, Song/Barabási 2010, Matz 2017 and Eckles et al. 2018; dehumanization, Haslam. Emotion economy, Hochschild, The Managed Heart (1983) and Hülsheger & Schewe 2011, Illouz, Ahmed 2004; engagement, Brady 2017, Rathje 2021, Burton 2021, Robertson 2023, Guess et al. 2023, Orben & Przybylski 2019; emotion AI, Barrett et al. 2019 and the EU AI Act; behavioral economics (the Vicary hoax, Karremans 2006, Luguri 2021, Gal & Rucker 2018; advertising effects) Blake/Nosko/Tadelis 2015, Kalla & Broockman 2018, Gordon et al. 2019, Lewis & Rao 2015. Exits/ethics, Webb 2012, Troy et al. 2013, Goyal et al. 2014, Schlosser et al. 2019, Roozenbeek et al. 2022; the ethics of manipulation, Susser/Roessler/Nissenbaum 2019 and SEP; Hertwig & Grüne-Yanoff 2017, Reijula & Hertwig 2022. 60+ load-bearing claims in total, 40 of them run through adversarial verification; tiers and effect sizes are labeled throughout.