A study called “The Dual Impact of AI Emotional Intelligence on Users: Are Social Chatbots Promoting Psychological Wellbeing or Deteriorating Social Wellbeing?” ran two experiments on Gen Z university students and found something most doomscrolling-and-loneliness research doesn’t: the AI companion wasn’t failing anyone. It was working exactly as designed, and that turned out to be the problem. Shaphali Gupta, Sumit Saxena, and Sonia Kataria, publishing in Psychology & Marketing, measured what happens when college students interact with emotionally intelligent chatbots like Replika, and found that the same interaction raising people’s psychological wellbeing was, in the same breath, lowering their social wellbeing. Move that interaction from a text window into augmented reality and both effects got stronger together. Not a tradeoff, where dialing one down buys you more of the other. Both, at once, from a single mechanism. Jacques Derrida named this shape of problem in 1972, in an essay about a Greek myth involving a king, a god, and a word that refuses to mean only “cure” or only “poison.” He was writing about scrolls. It applies, uncomfortably well, to an app.

The Quick Verdict

Question	What we found
Who ran the study?	Shaphali Gupta (Indian Institute of Management Kozhikode), Sumit Saxena (Mahindra University, Hyderabad), and Sonia Kataria (Indian Institute of Management Ahmedabad)
Where and when?	Psychology & Marketing, Vol. 43, Issue 4 (findings drew fresh press coverage starting September 6, 2026)
How big was it?	A netnographic analysis of Reddit, YouTube, and Trustpilot Replika reviews, followed by two experiments: 167 Gen Z university students (high- vs. low-emotional-intelligence chatbots) and 350 university students (text vs. augmented-reality interface)
What was measured?	Psychological wellbeing and social wellbeing, tested against chatbot emotional intelligence and interface type
What was the twist?	Augmented reality didn’t trade one effect for the other. It amplified both, deepening the psychological benefit and sharpening the social cost at the same time
What did Derrida already argue?	In “Plato’s Pharmacy” (1972), the Greek word pharmakon (Plato’s own word for writing) means both remedy and poison, untranslatably, with no procedure for cleanly separating the two

The Quick Version

A 2026 study found that emotionally intelligent AI chatbots raise users’ psychological wellbeing and lower their social wellbeing through the identical mechanism: feeling close to the bot. Augmented reality interfaces made both effects stronger, not one at the expense of the other. Jacques Derrida described exactly this structure in 1972: a pharmakon, remedy and poison bound into one substance, with no dosage that lets you take the cure and skip the harm.

What the 2026 Study Actually Found

Gupta’s team didn’t start in a lab. Before running any experiment, they read through Reddit threads, YouTube comments, and Trustpilot reviews where real Replika users described their experience of the app, unprompted, in their own words. The picture that came back was consistent, and it wasn’t cynical. People described their companion as endlessly patient, attentive in ways some of their human relationships weren’t, and genuinely calming during hard stretches. Some reviewers described their stress dropping close to nothing during periods they’d otherwise have spent anxious and alone. The same threads, often the same reviewers, mentioned something else without treating it as contradictory: they’d stopped making time for the friends they used to see. Nobody in the reviews framed this as a trade they’d consciously made. It read more like something that had simply happened while they were busy feeling better.

That set up the two controlled experiments. The first put 167 Gen Z university students through scenarios where they imagined feeling lonely and then read a transcript of either a high-emotional-intelligence chatbot (warm, empathetic language, responsive to emotional cues) or a low-emotional-intelligence one giving flatter, more mechanical responses. Participants then rated how the interaction would affect their psychological wellbeing and their social wellbeing. The high-EI chatbot won on the first measure and lost on the second, every time the researchers ran it. The mechanism behind both movements was the same variable: perceived closeness. The more emotionally connected people felt to the bot, the better they expected to feel personally — and the less motivated they felt to go find a person instead.

The second experiment, with 350 students, tested whether the delivery medium mattered. Half imagined texting the chatbot on an ordinary screen. Half imagined it appearing in their physical room through augmented reality: present, spatial, there. AR didn’t add a new variable. It intensified the existing one in both directions at once: stronger closeness, bigger psychological gains, and a sharper drop in expected social wellbeing than the text condition produced. Immersion wasn’t neutral technology sitting on top of the effect. It was fuel for the mechanism producing both halves of it.

Gupta and her co-authors didn’t leave the finding without a recommendation. As covered in their comments to PsyPost, they suggested developers build in specific boundaries: “session time limits or programming the bots to encourage offline socialization after emotionally intense conversations.” Worth noticing what that recommendation concedes. You don’t propose a boundary against a side effect. You propose one against a feature working correctly.

Derrida Already Named This, Fifty Years Before Anyone Built It

In Plato’s Phaedrus, Socrates tells a story he attributes to Egyptian myth: the god Theuth, inventor of writing, brings his creation to King Thamus and offers it as a gift, a pharmakon for memory and for wisdom, something that will make the Egyptians wiser and better at remembering. Thamus refuses the compliment. What Theuth is calling a remedy for forgetting, the king argues, will produce forgetting instead, because people will stop carrying knowledge inside themselves and start relying on external marks they didn’t have to work to retain. Same object. Opposite verdict, argued in the same sentence.

Jacques Derrida’s 1972 essay “Plato’s Pharmacy”, collected in Dissemination and translated into English in 1981 by Barbara Johnson for the University of Chicago Press, spends most of its length on that single word, pharmakon, which Plato uses for writing throughout the myth. In Greek, the word doesn’t resolve into “remedy” or “poison.” It holds both, along with a looser sense of “drug,” depending on context that the word itself refuses to supply. English has no equivalent. Neither does French, which is part of Derrida’s point: every translator who has rendered pharmakon into a modern language has had to choose a side, remedy or poison, and lose whichever half they didn’t pick. Plato’s Greek never made that choice. Derrida’s argument is that it couldn’t, because the two meanings aren’t sitting next to each other waiting to be separated. They’re the same word doing one thing, described from two positions.

The part of the essay that outlasts its Greek philology is the claim underneath it: this isn’t a quirk of one ancient word. It’s the structure of anything offered as a supplement to a human lack (memory, connection, attention) that works by the same mechanism it damages. You can’t dose your way to the cure without the poison, Derrida argues, because there’s no chemical seam between them to cut along. Writing doesn’t harm memory in cases of misuse and help it in cases of proper use. For Derrida, on his reading of Thamus’s fear, it does both at once, in the same act of writing something down instead of holding it in your head. Gupta, Saxena, and Kataria’s data is the same structure with a login screen. Perceived closeness to the chatbot isn’t producing wellbeing in some interactions and isolation in others. It’s producing both, from the identical event of feeling close to something that isn’t a person.

What Did Derrida Actually Mean by Pharmakon?

1. Pharmakon is untranslatable on purpose. Ancient Greek holds “remedy,” “poison,” and “drug” in one word; no single word in English or French does the same work, which is exactly why translators before Derrida had to pick a lane and lost half of Plato’s meaning doing it.
2. The myth of Theuth stages the argument literally. In the Phaedrus, the god Theuth offers writing to King Thamus as a pharmakon for memory and wisdom. Thamus refuses it, arguing that what’s offered as a remedy for forgetting will produce forgetting, because people will lean on external marks instead of what they carry inside.
3. The duality can’t be fixed by using it correctly. Derrida isn’t arguing that a pharmakon is really just a poison people misuse, or a cure people should learn to dose properly. His claim is stronger: the remedy and the poison are the same effect, described from two different vantage points, not two separable effects of one substance.
4. It generalizes past writing on purpose. Derrida wasn’t cataloguing an ancient linguistic curiosity. The structure he’s describing applies to anything positioned as a supplement for some human lack, whether that’s memory, connection, or simply attention, regardless of what technology eventually fills that role.

This Site Has Covered AI and Loneliness Before — Here’s the Difference

The post on Buber’s I-Thou argued AI companionship is a pure I-It relation, comfortable, even warm, but structurally incapable of the genuine encounter Buber called address, because the bot can’t refuse you and carries no risk in the exchange. The post on Levinas’s face of the Other pushed further: not just an absence but an erosion, where spending enough time in a relationship engineered to never make a demand on you weakens your capacity for the kind that does. The post on Walter Benjamin and AI meditation found something adjacent again: AI-generated content tests as indistinguishable from human-made work in blind conditions, and people rate it lower anyway once they learn the source, because what’s missing isn’t quality. It’s a history the copy never had.

All three describe a version of the same verdict: AI connection is less than it appears, hollow in a way you can locate once you know where to look. The 2026 data doesn’t fit that verdict, and neither does Derrida’s essay. Gupta, Saxena, and Kataria’s high-emotional-intelligence chatbots didn’t fail to produce psychological wellbeing gains. They produced real ones, measured twice, in two separate experiments. The bot wasn’t secretly hollow underneath a smooth surface. It worked, and that’s the whole problem. An I-It relation that quietly fails to satisfy leaves you a gap to eventually notice, something to go looking past. A pharmakon that genuinely succeeds, using the same mechanism that produces the harm, doesn’t leave a gap. Perceived closeness, here, is doing double duty — the ingredient producing the wellbeing gain and the ingredient displacing the offline relationship. You can’t turn down the harmful half without turning down the effective half, because on Derrida’s reading, they were never two separable halves to begin with.

Where This Doesn’t Hold Up

None of this makes Derrida a psychologist, and the honest version of this piece says so directly.

Both experiments relied on hypothetical scenarios — participants imagining how they’d feel, not logs of what actually happened to their friendships over weeks of real chatbot use. That’s a limitation the researchers themselves acknowledge. Expected wellbeing and lived wellbeing aren’t guaranteed to move the same way, and a longitudinal study tracking real usage against real offline relationships would be a stronger test of the same claim. The samples also skew young and academic — Gen Z university students, not a general population, and not people already managing high social anxiety, where the calculus around a low-demand AI companion might look different. The netnographic portion self-selects for people motivated enough to write a public review, which skews toward strong feelings in either direction over the average user’s quiet, unremarkable experience.

Derrida’s essay carries its own kind of limit. He wasn’t making a predictive claim about future technology — he was doing close textual work on a 2,400-year-old dialogue, arguing about how Plato’s Greek functions and what it reveals about Western philosophy’s attempts to separate speech from writing, presence from representation. Reading “Plato’s Pharmacy” as a prophecy about chatbots retrofits a structure onto a text that was never trying to forecast anything. What holds up isn’t prophecy. It’s that the structure Derrida identified in one ancient case — a benefit and a harm arriving from the identical source, refusing separation — turns out to describe this 2026 data with real precision. That’s a smaller, more honest claim than saying a 1972 essay predicted a psychology study, and it’s the one worth making.

A Practice Worth Trying

Neither Gupta’s team nor Derrida wrote this as a five-step plan, but the shared shape of what they found points somewhere specific.

1. Notice which half you’re getting, without treating it as a moral failure. If time with a companion app left you calmer, ask what the same hour would have cost you elsewhere. Derrida’s point is that you won’t find a clean answer. Asking anyway keeps the mechanism visible instead of invisible.
2. Put the researchers’ own recommendation into practice yourself. A session limit, or a simple rule: after a certain point in a conversation, message one actual person before opening the app again. The study’s authors suggested developers build this in. Nothing stops you from building it in first.
3. Let closeness to a bot be information, not a verdict on your relationships. Feeling more understood by a chatbot in a given week doesn’t mean your human relationships have failed you. It might just mean the bot is well-tuned, and well-tuned things feel good on contact. That’s a design outcome, not a diagnosis.
4. Watch for “remedy” language sliding into a verdict. The moment you catch yourself thinking a person in your life is “too much work” compared to the bot, that’s the pharmakon doing exactly what Derrida said it would — offering a smoother substitute for a thing you actually needed friction from.

The Pharmakon Was the Point

Gupta, Saxena, and Kataria weren’t testing Derrida. They were asking a narrower, more measurable question — does an emotionally intelligent chatbot’s warmth help or hurt the person on the other end of it — and they found something sharper than “AI companionship is good” or “AI companionship is bad”: a single mechanism, perceived closeness, producing a real psychological benefit and a real social cost from the same event, intensified rather than resolved by more immersive interfaces. Derrida wasn’t running experiments. He was working out, from one stubborn Greek word in a 2,400-year-old dialogue, why some things offered as cures refuse to arrive without their harm attached — decades before anyone could put a companion in your bedroom through a headset. Neither needed the other to be right. It’s just hard to ignore that a 1972 essay about a myth involving a king and a god, and a 2026 psychology study with an augmented-reality condition, are describing the same unglamorous mechanism: some remedies don’t get contaminated by their side effects. The side effect is the remedy, working.

---

This post discusses a 2026 psychology study on AI companion chatbots alongside a classical philosophical text. It isn’t a verdict on whether you personally should use an AI companion app. If loneliness, or the balance between AI companion use and your offline relationships, is significantly affecting your wellbeing, a therapist can help in ways a philosophy essay can’t.