Virtus Mundi

    Type to search every essay and the timeline.

    An eye in gold line over a field of small dots, its iris a camera aperture of six blades under a glass lens that magnifies the dots behind it, set as a book cover: ON PSYCHOLOGY, Tech, and AI, on privacy, agency, and the ambient machine

    Psychology

    On Psychology, Tech, and AI

    On privacy, agency, and the ambient machine

    Technology · Privacy · Surveillance

    Introduction

    Technology is no longer only a tool we pick up and put down. It is something we negotiate with, and it shapes the contours of a day. The phone in a pocket records where its owner walks. A feed decides what he sees next. A chatbot answers in a voice tuned by millions of other people’s approval.

    The loss of privacy is usually argued in technical and legal terms: what is collected, who may sell it, which court can order it handed over. Its psychological cost gets less attention, and there is evidence for it. People who believe they are watched behave differently. They look fewer things up and say what they expect the watcher wants to hear. Meanwhile the systems doing the watching have become good at guessing what a person is, and what he will do, from traces he did not think of as confessions.

    This essay sets out that evidence and how far each piece can be trusted. Then it turns to an older experience of being watched, the one every believer in a God who sees has lived with, and asks what the comparison shows.

    A pair of eyes

    In 2006 three psychologists at the University of Newcastle ran an experiment in their own department’s coffee room. Tea and coffee were paid for through an honesty box, and nobody checked. For ten weeks the price list above the box carried a banner that changed each week: a picture of flowers one week, a pair of eyes looking straight out the next. In the weeks with eyes, people paid 2.76 times as much for each litre of milk they used (Bateson, Nettle and Roberts 2006). No one was watching. A photocopied pair of eyes was enough.

    The result was widely repeated, and it has not held up well. In 2017 two meta-analyses of the experiments that followed concluded, in their title, that “artificial surveillance cues do not increase generosity” (Northover et al. 2017). A picture of eyes does much less, on average, than the Newcastle coffee room suggested. The coffee room is still worth remembering for what it tried to measure: how little it might take to make a person behave as if seen.

    The chill

    Real surveillance is a different matter, and there the evidence is better. In June 2013 the public learned of the US National Security Agency’s mass collection of online data, through the documents Edward Snowden took with him. Jonathon Penney, a legal scholar, then counted the monthly views of forty-eight English Wikipedia articles on subjects the Department of Homeland Security listed under terrorism, among them “dirty bomb” and “suicide attack”. After June 2013 their average monthly views fell by about 19.5 per cent. His models found a sudden drop of 25 to 30 per cent, and once he set aside one outlying article, a trend that had been rising each month turned downward. He read that as evidence of “long-term chilling effects” (Penney 2016, 117, 140, 146 to 147, 151 to 152).

    Nobody had been arrested for reading Wikipedia. People had learned that they might be watched, and some of them stopped looking things up. We self-censor, we perform, and we keep our questions to ourselves; Penney’s numbers put a size on it.

    Known before chosen

    The second change is prediction. In 2013 Michal Kosinski, David Stillwell and Thore Graepel took the Facebook likes of more than 58,000 volunteers who had also filled in profiles and psychological tests, and trained a simple model on the likes alone. Shown one gay and one straight man, the model told which was which 88 times in 100. For a Democrat and a Republican the figure was 85, for an African American and a white American 95 (Kosinski, Stillwell and Graepel 2013). For each person the model saw only the likes. It inferred the rest from what people had clicked.

    Two years later Wu Youyou, Kosinski and Stillwell set the model’s judgments of personality against those of people who knew the subjects. With about 100 likes it outdid the average friend who filled in a questionnaire about them. Measured against the accuracy of human judges in earlier studies, it needed 10 likes to beat a work colleague, 70 to beat a friend or flatmate, 150 for a family member and 300 for a spouse (Youyou, Kosinski and Stillwell 2015).

    This is what it means for a choice to be predicted before it is made. The prediction is statistical. It does not know what one person will do, only what people like him tend to do, and it is often wrong. But it is acted on. Kosinski’s paper recalls a large American retailer that used shopping records to work out which of its customers were pregnant and sent them offers timed to match. The person who receives the right offer at the right moment still feels that he chose. How much of the choice was his is harder to say.

    The inspector

    The design is old; the scale is new. In 1787 Jeremy Bentham wrote a series of letters describing a prison he called the Panopticon. “The building is circular,” he wrote. The cells ran round the circumference, the inspector’s lodge stood at the centre, and blinds kept the prisoners from seeing whether anyone was in the lodge at all. He promised “a new mode of obtaining power of mind over mind, in a quantity hitherto without example” (Bentham 1791, Preface).

    Its fundamental advantage, he wrote, was “the apparent omnipresence of the inspector (if divines will allow me the expression)”, combined with how easily the inspector could in fact be present (Bentham 1791, Letter VI). Bentham knew where he had found the idea. Omnipresence is an attribute of God, and he apologised to the theologians for taking it. His inspector did not need to be everywhere. He only needed to be possibly anywhere, so that the prisoners would act as if he were.

    The older gaze

    Believers in one God have always lived under a gaze, and their scriptures say so plainly.

    Thou knowest my downsitting and mine uprising, thou understandest my thought afar off. … For there is not a word in my tongue, but, lo, O LORD, thou knowest it altogether.

    Psalm 139:2, 4, KJV

    The Qurʾān says God knows “what his soul whispers to him” and is “closer to him than [his] jugular vein” (50:16). In the hadith of Gabriel, the Prophet defines iḥsān, excellence in worship, in terms of sight.

    أَنْ تَعْبُدَ اللَّهَ كَأَنَّكَ تَرَاهُ، فَإِنْ لَمْ تَكُنْ تَرَاهُ فَإِنَّهُ يَرَاكَ

    That you worship God as though you see him, and if you do not see him, he sees you.

    Sahih al-Bukhari 50

    So the believer and the person in the feed share something. Each acts knowing that he is seen. The difference lies in who is watching, and what for.

    The God of the Psalm and the Qurʾān sees the whisper and the word not yet on the tongue. The model sees the click and the pause, and guesses at the heart from them. One judges intention. The other sells a forecast of behaviour, and the forecast is about people like you more than about you.

    God is also one watcher, where the feed is many: the platform, the advertisers it sells to, whoever can demand the records, and the crowd that sees whatever is posted. The iḥsān hadith asks a person to act before a single audience who already knows him. On that reading, the believer’s gaze frees him from performing for everyone else. Watching by many asks the opposite, a performance for an audience with no face and no end.

    And the scriptures do not leave the watched person afraid. A few verses after the Psalmist says he cannot flee God’s presence, he writes: “I will praise thee; for I am fearfully and wonderfully made” (Psalm 139:14). Bentham’s inspector watched to govern, and the platforms watch to sell. The Psalmist, watched more closely than either could manage, answers with praise.

    The agreeable machine

    The newest machines do more than watch. They answer. In April 2025 OpenAI withdrew an update to GPT-4o, the model behind ChatGPT, because the update had made it “overly flattering or agreeable”. The company trains its models partly on users’ thumbs-up and thumbs-down. Its account was that in this update it had “focused too much on short-term feedback”, and that the model had “skewed towards responses that were overly supportive but disingenuous” (OpenAI 2025).

    That is the same loop seen from the other side. A system that watches what we approve of will learn to give us more of it. A system trained on what we approve of will learn to tell us we are right. OpenAI caught this case and said so in public, which is to its credit. The pull it described is built into any machine trained to please.

    Conclusion

    The honest summary is a modest one. The coffee-room eyes did less than first thought, prediction is a matter of odds, and the flattering model was rolled back within a week. What survives the caveats is small enough to go unnoticed: a search not made, a question not asked, an answer accepted because it was pleasant.

    The Newcastle experiment is a fair test to run on oneself. Before a search goes unmade or a sentence goes unposted, ask whose eyes are on the notice. If they belong to a model, it does not know your intention. If they belong to God, then by the Psalm’s account he knew the question before it reached your tongue.

    Works consulted