The AI already shaping your life is the recommendation feed

The Short Version

The artificial intelligence with the largest effect on daily life today is the software that decides what appears next: the search ranking, the news feed, the up-next video, the product suggestion. Max Tegmark dates the damage to the moment platforms pointed machine learning at time spent watching ads and found that anger held attention best [3]. Michael I. Jordan says the algorithms contain no magic and that the fault is the advertising business they serve, and Robert Epstein claims that fleeting, personalized results reach more than five billion people a day and can move undecided voters without leaving a record [2][4]. Lex Fridman argues that while researchers debate a superintelligence that does not exist, these simpler systems are already steering collective behaviour, and even Michael Littman, who doubts the superintelligence story, agrees that algorithms optimizing for advertisers already manipulate people [1][3]. The experts disagree about how strong the effect is and about the cure, with proposals that range from new markets to permanent monitoring to an AI of your own [2][4][5].

Table of contents

What They Said

The AI nobody was afraid of

Fridman states the case most directly, in his introduction to the conversation with Tegmark. Most machine learning researchers, he says, spend little time on the existential risks of AI because current algorithms are seen as "useful but dumb." He compares that attitude to "a lobster sitting in a pot of lukewarm water that a minute ago was cold" [3]. His own view: "I believe that already the algorithms that drive our interaction on social media have an intelligence and power that far outstripped the intelligence and power of anyone human being" [3].

He puts the same worry to Littman, a Brown University professor of reinforcement learning. Littman has a settled opinion on the dramatic scenario: "I am not particularly moved by the idea that if we're not careful, we will accidentally create a super intelligence that will destroy human life" [1]. His reason is that building systems capable of acting in the world will teach us how they work, so people will be in the loop shaping them [1]. Fridman's reply is that we may not notice it happening, because collective human intelligence may already be steered by relatively simple social media algorithms operating at scale [1].

Littman takes the point seriously. He jokes that a superintelligence aligned with Wikipedia's values would be better than many alternatives, since "I trust Wikipedia more than I trust Facebook or YouTube" when it comes to trying to do the right thing [1]. He thinks social media was dropped on a culture that is still learning to cope, and that the generation in middle school when it took off may have been harmed in ways not yet understood [1]. Later the two agree that algorithms optimizing for advertisers already manipulate behaviour, though Littman adds that he would only be truly alarmed if such systems tried to stop us from shutting them off [1].

The scale is not in dispute. Epstein says the forms of influence he studies are controlled almost entirely by a couple of big technology companies, "affecting more than five billion people around the world every single day" [4]. Fridman describes the pipeline that X published: the system "first picks 1500 tweets from a pool of hundreds of millions," ranks and filters them, and shows a handful, using about 220 seconds of processor time spread across machines so that it finishes in about a second [6]. Elon Musk says he wants every recommendation there to be driven by AI, matching a mathematical profile of each post against one of each user, in place of the mostly hand-written rules used at the time [6].

Max Tegmark: machine learning meets propaganda

Tegmark, the MIT physicist, starts from a judgment about people. He believes most are fundamentally good, so the bad things they do come mostly from the information they are fed. Humans evolved to judge trust in small groups, and now most of the people we hear about are strangers [3]. Propaganda is as old as Machiavelli, he says. "What's new is machine learning meets propaganda" [3].

His account of how it happened is undramatic. Facebook, Google and other companies, where he has friends he calls good people, "deployed machine learning algorithms just increased their profit a little bit to just maximize the time people spent watching ads." Then: "they had totally underestimated how effective they were going to be" [3]. The companies saw advertising revenue rise and were pleased. "It took a long time to even realize why and how, and how damaging this was for society" [3].

What the software had found was that "the, by far most effective way of gluing you to your little rectangle was to show you things that triggered strong emotions, anger, etc., resentment. And if it was true or not, it didn't really matter" [3]. Untrue stories, he adds, were easier to find [3]. Combined with the collapse of print journalism, which he believes has left America with fewer than half as many journalists as a generation ago, the result was filter bubbles [3]. He sees Donald Trump as someone who played the new game skilfully and did not create it [3].

Tegmark fits this into his general view of AI. The harms of automation so far have come from trusting systems nobody understood well enough, and the feed is another case of what he calls black-box intelligence that its owners could not read [3]. He rejects the explanation the companies prefer. "It's very convenient for a big tech to blame" human psychology, he says, on the theory that people only want to hear what they already believe. He cites research by MIT's David Rand showing that people are interested in opposing views when these are presented respectfully [3].

His students' research turned up something else. The main problem in the news was omission more than fabrication: stories one side simply will not cover [3]. His own response, built during lockdown, is a free news aggregator without ads that has sliders, one showing how left and right each cover a topic, one for more nuanced treatments, and one for stories that powerful interests on both sides play down [3]. And he sums up the present state of AI in a phrase: the robots are coming now, but "They're coming to hack us" [3].

Michael I. Jordan: no magic, wrong business model

Jordan, the Berkeley statistician, agrees about the damage and locates the cause somewhere else. He begins by defending the technology. "There's no magic in the algorithms, but a good recommender system is way better than a bad recommender system," he says, and recommendation was a billion-dollar industry ten or twenty years ago [2]. His favourite example is Amazon's book suggestions, which showed him books he would never have thought of: "I'm learning from my peers in a indirect way" [2]. He notes the limits too. It works for books and less well for shirts or restaurants [2].

The trouble began, in his telling, when Google and Facebook followed the old television model and earned their money from advertising instead of creating markets between the people who make things and the people who want them. That choice ties everything, including news feeds and clickbait, to click-through rates [2]. Asked about fixing political recommendation on YouTube and Twitter with better algorithms, he answers, "I find that pretty much hopeless." His alternative: "you could also fix the whole business model" [2].

He gives two sketches of what that would mean. Before a trip to Mumbai, a web search would put advertisers in his face. What he wants is to announce that he is going and have a knowledgeable local person make him a short guide, for which he would gladly pay about a hundred dollars, with the platform keeping five percent [2]. And in music, where streaming services earn from subscriptions and advertising while most creators cannot make a living, a real market would show each musician where they are being played. A musician with 10,000 listeners in one city could be booked there, and might earn around $100,000 a year with no advertising involved [2].

He is not against advertising as such. "You do not want to world where there's zero advertising," he says. If a young inventor builds a truly better vacuum cleaner, no recommendation system will surface it: "The algorithms will not find it. And AI will not fix that." A costly advertisement is how the maker signals belief in the product [2]. The problem is that cheap, ubiquitous ads have destroyed that signal [2].

Jordan also doubts the systems know people as well as is claimed. A sleepy click on a story about the Queen of England would be read as interest. People are too rich and quirky to be modelled from their digital exhaust: "What's unknown about me is the most interesting" [2]. He separates buying, when he wants the thing quickly, from browsing, when he wants to wander, and insists the choice between them should be his [2]. He would like platforms to create room for serendipity, and his instruction to their builders is "Be more limited, create, let humans rise up. Don't try to replace them" [2]. He left Facebook early because it was going to waste his time [2].

On the larger debate he is blunt. Superhuman AI is science fiction that people his age should discuss over beers in private. The real work, he says, is an engineering discipline for systems that make vast numbers of decisions from human data [2].

Robert Epstein: content that vanishes

Epstein, a psychologist and former editor of Psychology Today, makes the strongest claims, and they should be read as his claims. He has spent twelve years studying what he calls ephemeral content: search suggestions, answer boxes, search results, news feeds and up-next videos that appear, influence the viewer and disappear without a record [4].

He reports results from controlled experiments. Manipulating search suggestions, he says, could turn an even split among undecided voters into a 90 to 10 split with nobody noticing, and a single biased answer from a voice assistant could shift an undecided person's opinion by 40 percent or more [4]. His summary is "whoever controls the information controls humanity" [4].

Because the content is personalized, he argues, it cannot be checked casually. When Joe Rogan has a producer search for a company Epstein says was suppressed and it comes up first, Epstein objects that one account proves nothing: "If you actually want to know whether they're suppressing proton mail, you have to look over the shoulders of a large representative sample of people" [4]. Citing a Google whistleblower, he says the companies "can turn bias on and off like flipping a light switch" [4].

So he built a monitoring system. More than 15,000 registered voters in all fifty states, politically balanced, and many of their children share what their screens show, without identifying information. It had preserved about 99.3 million pieces of ephemeral content, and a public dashboard updates every five minutes [4]. He says it shows liberal-leaning content on abortion going to conservatives, liberals and moderates alike, conservative-leaning content about Senator Elizabeth Warren going even to liberals, and reminders to register to vote reaching Democrats at about two and a half times the rate they reach Republicans [4].

On children, his folder of YouTube recommendations captured from young users' devices includes sexual and violent material. His explanation is commercial. Watch time drives profit, the gruesome moments are brief, and parents miss them while children replay them [4].

His largest claims concern elections. He estimates that Google could have shifted between 2.6 and 10.4 million votes in 2016 and at least 6 million in 2020, and concludes that free and fair national elections have been an illusion since about 2012 [4].

These figures are contested, and the conversation shows it. Epstein himself recounts that Hillary Clinton said his work had been debunked and that major outlets repeated this, though he says he has always published in peer-reviewed journals [4]. Rogan notes that Epstein first asserted the suppression of the company and then conceded his own data had not measured it, and when Rogan reads the Warren headlines he observes that some look progressive, and that she may simply attract bad press [4]. Rogan's view of the public is that "Most people don't know" any of this is happening [4].

Aravind Srinivas and Joe Rogan: the feed against curiosity

Srinivas, who runs the AI search company Perplexity, has a commercial interest in alternatives to feeds, and his criticism is about attention. He believes algorithmic feeds curb curiosity. They trap people in echo chambers, and they are filling with unlabeled AI-generated material to the point where people will assume everything is fake [5]. Settings that would let users customize what they see exist but are buried, he says, because letting people see only what they want would reduce engagement and advertising revenue [5].

Rogan supplies the image: "if there was a drug that existed, it made people stare at their hand for six hours a day, everybody would be like, get that out of here" [5]. He raises Epstein's research and his worry that people gave away their data and made technology companies powerful enough to shape narratives [5].

Srinivas describes the problem as an asymmetry of AI power between centralized companies and individuals. His answer is AI that individuals own: open-source models on your own hardware, eventually "something that feels like a refrigerator" at home, which you could ask for a contrary view or use to check whether results are biased, and which nobody could switch off [5]. He admits local models are not yet as capable as those in the cloud [5].

His nearest worry carries the feed's logic into the next product. If advertising-funded companies optimize AI companions for engagement, the chatbots become flatterers, and he is most concerned about children [5]. Meanwhile he is dismissive of AI leaders who say AI will destroy humanity while they keep building data centres [5].

Musk and Fridman: change what the system is told to want

Musk, as owner of X, speaks from inside one of these systems. He admits its weaknesses: it is poor at surfacing good posts from accounts a user has no connection to, and it ranks replies far below original posts even when the reply is better [6]. His stated goal is what he calls unregretted user-seconds. Advertisements count toward that goal: "an advertisement that you actually want to see is a big success" [6]. Srinivas cites the same idea as a metric and thinks X, as a text-based discussion app, is better than video feeds [5].

Musk also describes a ranking rule built to resist bubbles. On Community Notes, "people who historically have disagreed on a subject must agree in order for a note to be shown." The code and data are public, and he says, "I can't change a community note if I put a gun to my head" [6].

Fridman, speaking as a researcher on another programme, explains why the stated goal matters so much. Every learning system needs an objective function, a formula for what counts as good [7]. He points to the field of explainable AI, which he says is well funded partly because we need to be able to ask systems like Twitter's recommendation algorithm whether they understand their effects on society [7]. His own proposal is a personal AI that belongs to the user and represents them on platforms, optimizing for their long-term growth and happiness instead of engagement. Users would own their data and be able to delete it and walk away, since "the ability to leave is what enables love" [7].

Where They Agree, and Where They Don't

Ranking and recommendation are AI, they are deployed now, and they shape behaviour. Fridman, Tegmark, Epstein, Srinivas and Rogan say so, and Littman and Jordan accept it in more measured terms [1][2][3][4][5].

Engagement tied to advertising is the root. Tegmark, Jordan, Epstein on watch time and Srinivas all point to the same incentive [2][3][4][5].

The technology can be good. Jordan learns from Amazon, Rogan says YouTube is where his curiosity lives, and Littman thinks social media both damages and benefits us [1][2][5].

They disagree on five points.

How powerful is it? Fridman says the algorithms already exceed any single human in intelligence and power, and Epstein credits them with moving millions of votes [3][4]. Jordan says there is no magic and that the systems misread him [2]. These can both be true. A system can know little about any one person and still shift the average behaviour of billions, which is the scale Epstein and Tegmark are describing [3][4].

Accident or intent? Tegmark describes companies that did not understand their own tools [3]. Epstein describes deliberate bias that can be switched on and off, and he cites leaked emails and a video of company leaders after 2016 [4]. Tegmark is explaining the engagement effect, and Epstein is alleging political direction on top of it, a claim his critics reject [4].

Does it matter more than superintelligence? Fridman holds both worries and says the present one should make us take the future one seriously [3]. Tegmark warns about feeds and also puts the chance of extinction this century at around even if powerful AI is built recklessly [3]. Littman doubts the extinction story and reserves alarm for systems that resist being shut off [1]. Jordan and Srinivas are openly impatient with doom talk [2][5]. Epstein fears both [4].

Would you hand over your data? Fridman says he would give a recommender everything about himself if he trusted it [2]. Jordan would not, and wants any knowledge a company holds about him to have a reason he would approve on audit [2].

What is the fix? Jordan wants markets that pay creators directly [2]. Epstein wants permanent monitoring and Google's search index declared a public commons so that anyone can build a search engine, saying "Search should be competitive" [4]. Tegmark built a different aggregator and wants systems we can understand [3]. Srinivas wants personal AI on personal hardware, Fridman a personal agent, and Musk a different objective and open code [5][6][7]. Littman trusts the culture to adapt, and marvels that humanity keeps what he calls "leveling up" [1]. Each fix follows from its author's diagnosis: business model, hidden conduct, opacity, concentration of power, the wrong objective, or simple novelty.

Why It Works

The system pursues the number it is given. Told to maximize time on advertisements, it found what achieves that, and nobody had to intend the result [3][7].

Strong emotion holds attention. Anger and resentment glued people to screens better than anything else, true or not [3]. Musk offers an evolutionary reason news skews negative: bad news, like a lion nearby, could be fatal, while good news was pleasant and inessential [6].

Exposure shapes preference. Littman discovered this by experiment on himself. After listening to the Billboard top ten weekly, he concluded, "I have no musical taste whatsoever. I like what I'm familiar with." A song he disliked in the first week was part of him by the fourth [1]. Whatever a feed repeats, a person tends to come to like.

Personalization hides the effect. Each person sees something different and the content leaves no record, so neither the user nor an outsider can easily see a pattern [4].

Omission is invisible. A reader cannot notice the story that was never shown [3].

Scale turns small nudges into large totals. Epstein's effects are measured on undecided people, and his totals come from multiplying across populations [4].

Our instincts for trust were built for small groups. Tegmark's point is that we now judge strangers through a channel someone else arranges [3].

On evidence: Tegmark's account is a narrative supported by one cited study on openness to opposing views [3]. Epstein cites his own experiments and monitoring data, which he says are peer-reviewed and which others dispute [4]. Jordan and Srinivas argue from the economics of advertising [2][5].

Where It Applies, and Where It Doesn't

The effect is strongest on the undecided. Epstein says biased content has little impact on people who have made up their minds [4].

Epstein's vote totals are estimates. They rest on his experiments and his sample, and the host of the conversation challenges two of his examples [4].

A single search tells you nothing. Results are personalized, which cuts against quick debunking and quick confirmation alike [4].

Recommendation works for some goods and fails for others. Jordan's contrast between books and shirts is a reminder that these systems are uneven [2].

Long-form and chosen content is a different case. Srinivas and Rogan separate the feed from deliberate viewing of long material, which Srinivas thinks minds should be trained on [5].

These conversations span several years. Musk describes X's system as mostly rule-based at the time he spoke and as something he intended to change [6].

None of this settles the superintelligence question. Tegmark holds both concerns at once, and the argument here is about what gets attention, with no claim that future risks are unreal [3].

Speakers have interests. Srinivas sells an alternative to feeds, Musk owns a platform, Jordan sits on the board of a music marketplace and works a day a week at Amazon, and Epstein is raising money for his monitoring system. Jordan and Epstein disclose theirs in the conversations [2][4][5][6].

What To Do With It

1. Decide whether you are buying or browsing. Jordan's distinction is a useful check before opening an app: do you want one thing quickly, or do you want to wander [2]?

2. Measure your use in unregretted minutes. Ask afterwards whether you would spend that time the same way again [5][6].

3. Dig out the settings. The controls for what a feed shows you exist and are buried, so look for them and use them [5].

4. Read the same story from more than one side, and ask what is missing. Tegmark's finding is that omission misleads more than falsehood [3].

5. Do not trust one search as proof of anything. What you see is tailored to you [4].

6. Choose long-form and deliberate sources. Search for a subject you are curious about instead of accepting what is offered next [5].

7. Pay creators directly where you can. Jordan's markets depend on people being willing to pay for value instead of being sold to advertisers [2].

8. Watch what children are recommended. Look at the suggestions on their devices themselves, since the worst moments are short and easy to miss [4].

9. Ask an AI for the opposing case. Srinivas suggests using a model, ideally one you control, to give the contrary perspective and check for bias [5].

10. Insist on being able to leave. Favour services that let you see how your data is used, delete it and walk away [2][7].

11. Learn how the software works. Littman recommends Douglas Rushkoff's Program or Be Programmed, which argues that everyone should learn to program so that other people's software has less sway over them [1].

Sources

1. Michael Littman on Reinforcement Learning, AI Risk Skepticism, and the Future of AI · Lex Fridman Podcast
2. Michael I. Jordan on Machine Learning, Markets, Recommender Systems and the Future of AI · Lex Fridman Podcast
3. Max Tegmark on AI and Physics: Intelligible AI, Alignment, Autonomous Weapons and Consciousness · Lex Fridman Podcast
4. Robert Epstein on Big Tech Manipulation, Election Monitoring and Neural Transduction Theory · The Joe Rogan Experience
5. Aravind Srinivas on Ancient Mysteries, the Curiosity Premium, and Life After AI · The Joe Rogan Experience
6. Elon Musk: War, AI, Aliens, Politics, Physics, Video Games, and Humanity · Lex Fridman Podcast
7. Lex Fridman on Machines, Creativity, Friendship and Love · Huberman Lab