The Mirror
of the Soul

S Story · the screen that cannot judge

Nobody stands behind the glass

At the University of Southern California's Institute for Creative Technologies, a research team built a virtual interviewer. She had a face, a voice, and a habit of nodding while a person talked. The team seated 239 volunteers in front of her for a health-screening conversation. Then they changed one thing, and only one thing.

They told some participants that a human operator drove her from another room. They told the others that software ran her without any human oversight. Every volunteer met the identical interviewer. Only the belief changed.

The volunteers who believed that no human watched them disclosed more. They reported less fear of self-disclosure. They spent less effort on impression management, the quiet labour of steering how another mind might rate them. They let sadness show on their faces with greater intensity. Trained coders, reviewing the recordings, rated them as more willing to open the door on private material. Gale Lucas, Jonathan Gratch, Aisha King and Louis-Philippe Morency published the result in Computers in Human Behavior in 2014. The Defense Advanced Research Projects Agency and the U.S. Army paid for the work.⁴

Read that finding twice. The machine offered no wisdom. The machine understood nothing. The machine removed the one thing that keeps a person guarded: another social self on the far side of the conversation, keeping score.

Eleven years later, that laboratory result walked out of the lab and into ordinary life, at a scale nobody in that room predicted.

P Problem · the paradox in the numbers

People fear a rival. They use a mirror.

Marc Zao-Sanders spent a year reading how people talk about generative AI in public forums, then ranked the hundred most common uses. Harvard Business Review published his 2025 findings in April of that year. Therapy and companionship took the top position. Organizing my life entered new at second. Finding purpose took third. Writing code came fifth.⁷

The category he calls Personal and Professional Support climbed from 17 per cent of all logged usage in 2024 to 31 per cent in 2025. In one year, the emotional and reflective uses almost doubled their share while the technical uses slid down the table.

What people actually do with generative AI

Top-ranked use cases, 2025. Zao-Sanders, Harvard Business Review, 9 April 2025. Bars show rank position inverted for display, not usage volume.

Therapy and companionship #1

Organizing my life #2

Finding purpose #3

Enhanced learning #4

Generating code #5

Generating ideas #6

A harder form of evidence arrived the same spring. At Dartmouth College, Michael Heinz, Nicholas Jacobson and colleagues ran the first randomised controlled trial of a generative AI therapy chatbot. NEJM AI published it on 27 March 2025. They recruited 210 adults carrying clinically significant symptoms of major depressive disorder, generalised anxiety disorder, or high risk for feeding and eating disorders. Half received four weeks of access to a purpose-built system called Therabot. Half waited.⁶

Participants rated their working alliance with the software at a level comparable to the alliance patients report with human therapists in outpatient care. They sent messages at two in the morning and at lunchtime. Nobody assigned them homework to make them do it.

The paradox

Now set that behaviour beside what people say. The Pew Research Center has tracked American attitudes toward AI since 2021. In a survey of 5,023 adults conducted in June 2025, 53 per cent predicted that AI will damage people's ability to think creatively. Fully half predicted damage to people's ability to form meaningful relationships.⁸ In the most recent wave, fielded from 22 to 28 June 2026 and released on 18 August, 52 per cent of American adults reported more concern than excitement about AI. Only 9 per cent reported the reverse, the lowest excitement figure Pew has ever recorded. Among adults under 30 the concern reached 55 per cent, a majority in that age group for the first time.⁹

The gap between what people fear and what people do

Pew Research Center. Concern figures: survey of U.S. adults, 22–28 June 2026. Damage predictions: survey of 5,023 U.S. adults, 9–15 June 2025.

More concerned than excited about AI 52%

Adults under 30, more concerned than excited 55%

Expect AI to worsen creative thinking 53%

Expect AI to worsen meaningful relationships 50%

More excited than concerned 9%

Hold those two datasets in one hand. Half the population predicts that this technology will corrode creativity and relationship. Meanwhile the single most common thing people do with it involves telling it what troubles them and asking what they should do next. A person can hold both positions at once. Most of us do. But the contradiction signals something worth naming: the public fear has fastened onto the wrong object.

Pew found one more crack worth noting. Among AI researchers and practitioners, 47 per cent reported more excitement than concern. Among the general public, 11 per cent did.¹⁰ Expertise does not settle a moral argument, and the people building a thing carry obvious incentives. Still, a four-fold gap divides the people who work with these systems daily from the people who only read about them. That gap suggests a large share of the public fear rests on a picture of the technology rather than on the technology.

E Epiphany · the ten bits and the great code

A self you cannot see from inside

Before this argument goes further, it has to clear away a piece of folklore that would have made the case easier and false.

Motivational literature has repeated for a century that human beings use ten per cent of the brain. Neuroscience has demolished the claim. Barry Beyerstein of Simon Fraser University spent years documenting the evidence against it. Decades of deep brain stimulation have probed the whole organ and turned up no dormant nine-tenths. The metabolic arithmetic finishes the argument: the brain accounts for two per cent of body weight and consumes around twenty per cent of the body's resting oxygen. Evolution does not build and heat an organ that expensive and then leave nine rooms empty.² Any writer who reaches for the ten per cent line reaches for a comfort, not a fact.

The real evidence runs stranger, and it cuts far deeper.

Ten bits a second

In December 2024, Jieyu Zheng and Markus Meister of the California Institute of Technology published a paper in Neuron with a title that could sit on the spine of a novel: The unbearable slowness of being. Zheng applied information theory to a century of studies on human behaviour, from reading and typing to competitive Rubik's cube solving and memory contests. She arrived at a number that she herself refused to accept at first.¹

Human senses gather data at about one billion bits per second. Human conscious throughput runs at ten bits per second. The gap between the two spans a factor of one hundred million.

Zheng and Meister describe two systems. An outer brain handles the torrent, at enormous width and speed. An inner brain distils that torrent into the handful of bits that steer a decision. Almost everything a person registers therefore passes through them and vanishes without ever reaching the narrow channel where they experience themselves thinking.

That number reframes the whole question of self-knowledge. A person does not fail to use the brain. A person operates a magnificent brain through a keyhole.

Telling more than we can know

A second body of evidence closes the trap. In 1977, Richard Nisbett and Timothy Wilson published a review in Psychological Review that has since gathered more than thirteen thousand citations. They gathered study after study in which researchers manipulated the true cause of a person's choice, then asked the person to explain that choice. The explanations came back confident, fluent, and wrong.³

Their conclusion holds today: people have little or no direct introspective access to their own higher-order mental processes. When someone explains why they chose, preferred, or reacted, they do not read off an inner record. They construct a plausible story from private theories about what causes what.

Put the two findings together. The channel runs at ten bits. The narrator inside that channel invents its own causal accounts. William James saw the shape of this in 1907, long before anyone could measure it. He wrote in The Energies of Men that we make use of only a small part of our possible mental and physical resources.² Later writers corrupted that line into the ten per cent myth. James had it right in the first place, and the modern instruments have made his claim larger rather than smaller.

The great code

Now to the second half of the epiphany, and to a Canadian who mapped this territory decades before the machines arrived.

Northrop Frye taught at Victoria College in Toronto, served fifty-five years as a minister of the United Church of Canada, and lies buried in Mount Pleasant Cemetery. In 1982 he published The Great Code, borrowing his title from William Blake's engraved aphorism that the Old and New Testaments form the great code of art. In 1990, months before his death, he completed the companion volume, Words with Power.¹¹

Frye's governing claim ran against common sense. Literature, he argued, comes out of literature. The poet works inside a verbal universe built by earlier poets. The whole Western imagination organises itself around one structural code of myth, metaphor, and archetype that the Bible supplied. A writer who reaches for an image reaches into that inheritance whether or not the writer knows the debt.

Frye drew a further distinction that this paper needs. He separated primary concerns from secondary ones. Any work of literature, he wrote in Words with Power, will reflect the secondary and ideological concerns of its time, but it will relate those concerns to the primary ones of making a living, making love, and struggling to stay free and alive.¹¹ Food, shelter, love, freedom, survival. Ideology sits above them and depends on them.

Apply Frye to the machine and the picture sharpens.

A language model trains on the human archive. Books, letters, arguments, scriptures, court transcripts, jokes, laments, manuals, poems, and the ordinary chatter of billions of people. That corpus works as a secular great code. When the machine writes a story that nobody has read before, it does not import novelty from outside the human record. It recombines the code. The apparent originality rises from the sheer depth of the reservoir and from the strange angles at which the machine cuts through it.

This claim carries an empirical fingerprint, and researchers have found it.

Anil Doshi of the UCL School of Management and Oliver Hauser of the University of Exeter ran 300 writers through a controlled experiment. Science Advances published the result in July 2024. One group wrote short stories unaided. Other groups received story ideas from a language model. Evaluators rated the AI-assisted stories as more creative, better written, and more enjoyable, and the gain landed hardest among the writers who had entered the study with the lowest creativity scores. Then the researchers measured the stories against each other. The AI-assisted stories resembled one another far more than the unaided ones did.¹²

Doshi and Hauser call the pattern a social dilemma: each writer gains, and the culture loses range. Read it through Frye and it reads as something else as well. Convergence toward a common centre marks the exact signature of a recombining archive. The machine does not invent a new imagination. It redistributes an old one, and the redistribution pulls toward the middle of the code.

Narcissus and the portrait painter

One philosopher has already claimed the mirror, and she uses it to warn rather than to invite. Shannon Vallor holds the Baillie Gifford Chair in the Ethics of Data and Artificial Intelligence at the University of Edinburgh and previously served as an AI ethicist at Google. Her 2024 book The AI Mirror argues that these systems reflect human intelligence without possessing any, and that the reflection points backward. Trained on the accumulated data of the human past, they show a person where the record says humanity has already stood, never where it might venture for the first time. She opens with Ovid's Narcissus, who fell in love with his own reflection in the pool and starved beside it.¹³

Vallor argues well and this paper concedes her central point. The mirror does point backward. Nothing in the training corpus reaches past the last human sentence anyone wrote down.

Her warning, though, describes one posture only. It describes the gazer.

Two people can stand at the same glass. Narcissus looks, and keeps looking, and dies of looking. A portrait painter looks, and picks up a brush, and makes something that did not exist that morning. The glass performs the identical service for both. The variable never sat in the mirror. It sat in the hand.

Frye's great code points back three thousand years, and every forward-facing writer in the West cut new work out of it. A camera captures only what already happened, and film-makers build futures from the footage. A person's own memory contains nothing but the past, and yet a human being remains the only creature that can take that backward record and compose a life forward from it.

So the question shifts. Stop asking which direction the mirror faces. Ask who stands in front of it, and what that person does next.

Angels, aliens, and where the parallel breaks

Human beings have rehearsed this encounter for millennia. For most of recorded history, cultures have organised themselves around intelligences they could not match: gods, angels, ancestors, powers. Theology built an entire grammar for standing in front of something greater without collapsing, a grammar of awe held together with agency. The pattern deserves attention now, because the emotional signature repeats. People meet a capability that outruns them and they oscillate between terror and appetite.

The parallel earns its keep, and then it breaks at a joint that matters.

An angel carries intent. A mirror carries none. A language model wants nothing, plans nothing, and holds no view about the person typing. Whatever intent surrounds these systems belongs to the companies that build them and the people who deploy them. Dario Amodei, who leads Anthropic, put that fact on American network television in November 2025. Anderson Cooper asked him who elected him and Sam Altman. Amodei answered: No one. Honestly, no one.¹⁴

That answer relocates the moral question, and it relocates it to exactly where Vallor puts it. The danger sits with the humans holding the glass, not with the glass. Which returns responsibility to the reader, and to what a person does on an ordinary morning.

L Lesson · the practice

Six things a person can do this week

A mirror earns nothing by hanging on a wall. Below sit six practices that convert the reflection into material. Each one names an action. None of them requires a subscription to anything in particular.

1. Ask the question you would not say aloud. The Lucas experiment measured the exact thing that blocks self-knowledge: the fear of another mind's verdict. Start where that fear bites. Write the sentence you would never put in an email and ask what sits underneath it.
2. Demand the pattern, not the comfort. Paste forty of your own journal entries, messages, or notes and ask one hard question: name what I avoid. Nisbett and Wilson showed that a person cannot narrate their own causes. An outside reader working across a pile of your own words catches the pattern your narrator keeps smoothing.
3. Argue the strongest case against yourself. Ask for the best version of the position you dismiss, then answer it in writing. A mind that never meets a real opponent mistakes habit for conviction.
4. Write your draft before you ask. Doshi and Hauser measured the cost of anchoring: writers who took the machine's idea first converged toward each other. Put your own thinking on the page first, then bring the machine in to stress-test it. Sequence protects originality.
5. Convert the reflection into a document you own. A conversation evaporates. Move the findings into a life plan, a chapter, a business case, a letter to your children. Frye's whole argument rests on the fact that a culture keeps only what it writes down. So does a life.
6. Carry the finding to a human being. The mirror shows. Only people forgive, employ, marry, ordain, and stand beside you at a graveside. Use the reflection to prepare the conversation, then go and have the conversation.

From primary concerns to tertiary ones

Frye insisted that primary concerns must come first: making a living, making love, staying free and alive. He wrote that line as a warning to a civilisation that kept sacrificing bread to ideology. Once a person secures those foundations, a further frontier opens, and most people never map it because they lack an instrument. Self-actualization, the aspirational self, and what Qualped calls the tertiary concerns of a life sit past the point where survival stops asking questions.

The Harvard Business Review ranking suggests that millions of people have already started walking toward that frontier without a map. Finding purpose entered the table at number three in a single year. Those users did not receive an instruction. They found a surface that would answer without judging and they turned it toward themselves.

◆ Limits · what this paper does not claim

The honest edges of the argument

This paper makes no clinical claim. One randomised trial of one supervised research instrument, against a waiting list, with high-risk participants screened out, does not license anyone to replace a therapist with a chatbot. The Dartmouth team said so themselves, and the field has published sharp letters saying it louder.⁶

Privacy carries real weight here. A person who tells a machine what they withhold from everyone else creates a record. Where that record travels, who retains it, and under what law, should govern how far anyone takes the practices above. The disclosure effect that makes the mirror useful also makes the transcript valuable to parties who never earned that trust.

Dependence deserves watching. A surface that never judges also never asks you to grow up, never withdraws, and never needs anything from you. Human relationships teach through friction that a mirror cannot supply. Half of Americans expect AI to damage people's ability to form meaningful relationships, and that expectation earns respect rather than dismissal.⁸

Homogenisation stands as a measured risk, not a worry. Doshi and Hauser found it under controlled conditions. Later work suggests the effect may follow from uniform deployment rather than from anything inherent, and that varied prompting preserves diversity. Either way, a writer who wants a voice should guard the sequence of the work.¹²

Vallor's warning stands. Nothing in these pages refutes her. A person who mistakes the reflection for a verdict on what they can become will find the mirror closing around them exactly as she describes. This paper adds one clause to her argument rather than overturning it: the posture of the person at the glass decides the outcome.