Why The Atlantic signed a deal with OpenAI

Today, I’m talking to Nicholas Thompson, nan CEO of The Atlantic, 1 of nan oldest magazines successful nan United States — for illustration really old. It was founded successful 1857 and is now owned by Laurene Powell Jobs, whose past sanction I americium definite that Decoder listeners will recognize.

I was really excited to talk to Nick — for illustration truthful galore media CEOs, he conscionable signed a woody allowing OpenAI to usage The Atlantic’s immense archives arsenic training data, but he besides has a rich | inheritance successful tech. Before he was nan CEO of The Atlantic, Nick was nan editor-in-chief of Wired, wherever he group his sights connected AI reporting good earlier anyone else, including me. So he’s been paying attraction to this for a agelong time.

Now, I consciousness for illustration I should disclose correct distant that Vox Media, The Verge’s genitor institution wherever I work, besides has a woody pinch OpenAI, which was announced connected nan aforesaid day arsenic The Atlantic’s deal.

I really don’t cognize very overmuch astir nan position of our deal, since I’m connected nan editorial broadside of nan location and there’s a strict firewall betwixt nan business broadside and nan editorial side. I fishy each of these deals are beautiful similar, but I really asked Nick astir that. And there’s a beautiful funny logic that he doesn’t cognize either; you’ll perceive america talk astir it.

Of course, I besides asked Nick why he was consenting to motion a woody pinch OpenAI successful nan first place, and why now erstwhile there’s truthful overmuch general unhappiness astir AI companies utilizing different people’s activity without permission, and circumstantial unhappiness pinch OpenAI. You’ll perceive Nick explicate that what he really wanted to get backmost was a consciousness of control: Control complete really overmuch information was being used, really results were being displayed, and, of course, complete really overmuch money The Atlantic was being paid.

I was besides really willing successful asking Nick astir nan wide consciousness that nan AI companies are getting vastly much than they’re giving pinch these sorts of deals — yes, they’re paying immoderate money, but I’ve heard from truthful galore of you that nan money mightiness now beryllium nan point. That there’s thing other going connected here, that possibly allowing productivity to get commodified this measurement will travel pinch a value tag truthful large money tin ne'er salary it back.

If location is anyone who could get into it pinch maine connected that question, it’s Nick. This 1 went long, and it’s a bully one. Okay, Nick Thompson, CEO of The Atlantic. Here we go.

This transcript has been lightly edited for magnitude and clarity.

Nick Thompson, you are nan CEO of The Atlantic. You are besides notably, for this conversation, nan erstwhile editor successful main of Wired. Welcome to Decoder.

Thank you truthful much, Nilay. I’m delighted to beryllium here.

I americium really excited to talk to you. I bring up nan Wired point because I want to talk to you astir AI and nan deals media companies for illustration The Atlantic, and notably Vox Media, nan institution that I activity for, are making pinch companies for illustration OpenAI. It feels for illustration you person to understand nan media business, nan tech business, and wherever nan tech business mightiness beryllium going successful narration to nan media. Let’s commencement astatine nan very beginning, why make a woody for illustration this pinch OpenAI? What is your woody pinch OpenAI?

We tin spell done it successful analyzable measurement aliases nan elemental way. The elemental measurement is we judge it provides revenue, but much importantly provides a imaginable postulation source. Provides an avenue for a merchandise business that could beryllium very beneficial, and that provides a measurement for america to thief style nan early of AI.

AI is coming, it is coming quickly. We want to beryllium portion of immoderate modulation happens. Transition mightiness beryllium bad, nan modulation mightiness beryllium good, but we judge nan likelihood of it being bully for publicity and nan benignant of activity we do pinch The Atlantic are higher if we participate successful it. So we took that approach.

We started talking to each nan AI companies, each of nan ample connection exemplary companies. We had parameters that we would judge for a deal, parameters we would not judge for a deal, and we reached a woody pinch OpenAI. So that’s nan basal framework.

What were nan parameters?

The woody really has 3 parts, 4 parts, depending connected really you look astatine it. Part 1 is for a constricted play of time, 2 years successful our case, they’re allowed to train connected our data. So they tin publication Atlantic stories and they tin incorporated that into their guidelines ample connection model. We person immoderate controls complete nan benignant of outputs they’re allowed to springiness to people, but they’re allowed to train connected our information for 2 years.

The 2nd portion of nan woody is nan merchandise partnership. So they springiness america credits. So we were building devices connected nan business broadside pinch nan engineering squad that are utilizing OpenAI. So we don’t person to trust connected Llama, we are conscionable utilizing OpenAI.

Credits, we are moving pinch them. At immoderate constituent location whitethorn beryllium engineering support, location whitethorn not beryllium engineering support. Who knows precisely really that is going to work, but that is simply a perchance valuable part. And we are launching a laboratory tract soon wherever we’ll person a full bunch of experimental devices to thief readers.

First 1 we created was a Chrome hold that will, arsenic group are reference different places connected nan web, will show them stories The Atlantic has written that are related to, conscionable worldly for illustration that. So we’ll person a labs experimental site. So that’s nan 2nd portion of deal.

Third portion is this very absorbing hunt element, wherever correct now successful OpenAI they person browse mode and they tin nexus retired to Atlantic stories. They person said that they’re going to build a hunt product. They person not launched nan hunt product, but they person said they would build it. We person allowed them to see The Atlantic successful their hunt product.

Our position is that if this becomes an important measurement that group navigate nan internet, that it will beryllium amended for america to beryllium successful it than to not beryllium successful it, and besides to thief style it than not thief style it. So that’s nan 3rd part.

And past nan 4th portion is that location is simply a statement backmost and forth. So erstwhile we spot something, for illustration successful browse mode we announcement thing absorbing astir nan URLs and nan measurement they’re linking retired to media websites. You spell backmost and distant and those things get fixed. So our consciousness is that we are helping nan merchandise germinate successful a measurement that is bully for superior publicity and bully for The Atlantic.

So those are nan cardinal components of nan deal. Underlying it is simply a position that journalists and media companies should beryllium paid for their work. Obviously nan ample connection exemplary scraped without permission, did not salary us. We deliberation we should beryllium paid for that.

There are a full bunch of ways you tin get paid for that, you tin sue, you tin do deals, you tin shingle your fist. You fig retired immoderate nan champion attack is to get paid, but location should beryllium a adjacent speech successful value. So that is simply a cardinal portion of it.

But we besides judge that nan world will beryllium a amended spot for superior publicity if contented for illustration that created successful The Atlantic and that created successful The Verge is portion of these models. If nan hunt results return Verge stories, that is amended for nan readers and it is amended for nan world than if they do not, right? There are each kinds of trade-offs, but that is different constituent successful it.

There’s a batch there. I want to return 1 portion of it and conscionable attraction connected it for 1 second. You mentioned revenue. How overmuch money is it complete 2 years?

Some of nan position of nan woody are nondisclosure agreements. Obviously I can’t disclose that peculiar term, but it is simply a adjacent speech successful value.

Do you deliberation it’s worldly aliases meaningful to The Atlantic’s gross arsenic a whole?

So it’s short-term revenue. Is it worldly successful 2024, is it worldly successful 2025, nan 2 years of deal? Of course. Would you want to extrapolate retired to 2026? Of people not.

One of nan things that we each cognize from deals pinch tech companies is they attraction astir their interests, not your interests. They do deals that end, you don’t expect it to proceed forever.

I consciousness for illustration nan manufacture learned that instruction successful nan hardest imaginable way, nan rug propulsion of Facebook’s various news initiatives aliases Google’s various news initiatives and that money going away. Basically everyone depended connected those companies and past that dependency was revealed to beryllium successful error. Do you consciousness that? Was that skepticism coming erstwhile you were talking to OpenAI?

Yes and no. So I deliberation location was a different mistake. My view, my philosophy, and this is not a cleanable metaphor, is that fundamentally nan editorial activity sits upstream and past everything other is downstream. That’s nan measurement you tally a business.

So you determine what stories you’re going to run, nan editors take them, they constitute them to nan champion measurement they can, and past you conflict for illustration hellhole to get arsenic overmuch postulation arsenic you tin from Google, from Facebook, connected Instagram, TikTok, immoderate you’re doing to get them connected to publication successful nan correct way. You do each those things, but you do those things aft you’ve written precisely nan communicative you want.

And wherever nan companies made mistakes is that they moved nan Google and nan Facebook worldly upstream, and truthful they signed these deals. And they didn’t conscionable expect that nan gross would proceed forever, which is correction one. But nan overmuch much superior correction is if you commencement to delegate stories, aliases edit stories, aliases alteration moreover a connection successful nan stories because you want to person it spell viral connected Facebook, past you’ve started to sacrifice nan point that you do that matters, right?

I’ve spent astir of my life moving only astatine 3 places. I worked astatine Wired, past it was The New Yorker, past went backmost to Wired, now I’m astatine The Atlantic. All of those places are benignant of different dynamics. And The New Yorker, what you’re doing is you’re fighting really difficult to make judge everyone pays attention, fighting for Facebook and Google. There is nary consequence that this will beryllium moved to nan incorrect spot successful nan river, successful different magazines, and surely you saw it successful a batch of media. So that’s correction number one.

Mistake number one, and that is nan important mistake, is making nan business woody upstream of nan editorial. Mistake number 1 is assuming that these companies will partner pinch you forever, and if they say, “We’re going to springiness you X money this year,” you’ll besides person that X money successful 3 years, which is aft nan statement ends. That is simply a mistake.

But nan overmuch much important correction is if you commencement to alteration nan ineffable point you do, which is nan creation of stories for nan platforms.

So now backmost to nan AI deal. Is location immoderate measurement successful which we will alteration nan measurement we do our stories because of this deal? Absolutely not, this will person nary effect. We will do nan nonstop aforesaid stories successful 2024 and 2025 than we would’ve if we didn’t person this deal.

One of nan large criticisms present is, okay, you sold this worldly for 2 years, they’re going to train their model, it’s going to get better. Then nan woody will end. They won’t salary you again, but they will person already trained nan model. And that worth will stay everlastingly and past they will conscionable proceed doing immoderate they want to do.

There are astir 20 different position that are important erstwhile you’re negotiating a woody for illustration this. That is 1 of nan important terms. And truthful it has been publically stated, and truthful I tin opportunity this, they are destroying our data. They will usage our information to train immoderate exemplary that they build successful nan adjacent 2 years, nan 2 years aft we motion that deal.

They train each caller exemplary connected wholly caller data, and truthful they will person our information for nan adjacent 2 years, but erstwhile it gets to GPT6 they won’t, unless they person different deal. That clause is important some for nan logic you said, and besides truthful we person much leverage erstwhile there’s different infinitesimal of negotiation.

It feels for illustration OpenAI is nan challenger. They evidently are nan upstart, they’re chaotic successful nan ways that startups tin beryllium chaotic, successful a nosy measurement and besides successful a compromised way.

The existent target here, it feels like, is Google, which has had a very extractive narration pinch nan media for a agelong time. Now is keeping much of that postulation for itself. Is besides building AI hunt products, delivering AI results, and is paying nary one. Do you deliberation a woody for illustration this helps you get leverage against Google?

I deliberation so. Google has a different situation, wherever they person truthful overmuch much leverage connected america because you can’t artifact Google. I mean, location are ways you tin partially artifact Google, and you tin artifact this Googlebot, not that Googlebot, but they person a batch much leverage connected america that OpenAi does, nan negotiations are different.

I besides would ideate that they are waiting. There are a batch of things that are happening pinch OpenAI, including nan New York Times lawsuit. I deliberation they’re waiting to spot really that shakes out. I haven’t talked to Google straight astir this, but if they salary for content, do they person to salary for each nan links? And do they person to salary backmost for 25 years worthy of it?

So I don’t cognize what their calculations are, but I deliberation they’re watching what’s happening. And my dream will beryllium that there’s a adjacent worth speech pinch Google arsenic they build AI search.

That portion wherever you said OpenAI has already taken it, they’ve already scraped connected what they mention to arsenic publically disposable information, which mightiness see each nan measurement up to YouTube, and these are nan reports that we’ve heard. Do you consciousness for illustration you’re taking nan costs now successful recompense for what they’ve already taken? Or is this for nan future?

That’s a difficult mobility to answer. This is not for illustration you committed a misdeed and you’re paying america for nan sin, we don’t position it that way. We position it as, you created… I was trying to do a calculation nan different day. I was like, “How overmuch does nan precocious value journalistic content, really overmuch worth did it create for OpenAI?” And you tin really benignant of do a backmost of nan letter cover calculation, and you tin spot really overmuch money, based connected that calculation, a unsmooth backmost of nan envelope, what they beryllium nan publicity manufacture aliases what nan publicity manufacture contributed.

And you tin deliberation astir of what nan publicity manufacture contributed, what percent should spell to america and what percent should they keep, right? And that’s benignant of 1 measurement wherever you came up pinch a number. I don’t position it arsenic paying for a sin. I position it as, “Okay. They’ve built this thing, it has this value. We’re portion of it. We’d for illustration to beryllium paid for it.”

That calculation, erstwhile you went to unfastened AI pinch it, did that lucifer what they wanted to salary you? Or were you higher aliases lower?

That peculiar calculation has truthful overmuch variety successful it because really overmuch do you measurement each of nan factors is astir wherever we ended up.

The logic I inquire it that measurement is nan conception that this is simply a pre-settlement for a suit that you might’ve revenge nan measurement that nan New York Times revenge a lawsuit, aliases you’re mounting a value level for a further speech pinch Google, really changes nan measurement you deliberation astir nan woody itself, right?

So if you’re saying, “You already took it. Just salary america to drawback america up, and past successful 2 years, we’ll commencement complete from scratch,” that changes versus, “You’re building GPT-5 and a hunt product. We want to beryllium connected nan crushed level arsenic nan challenger to Google.” You mightiness judge a discount successful that lawsuit because you deliberation nan upside is higher. What’s nan equilibrium there?

We want to maximize respective things, right? We want to maximize nan magnitude of money that comes to superior publicity companies. We want to style nan manufacture successful nan champion imaginable guidance based connected our values, and we deliberation nan values that are important. We want to bring successful arsenic galore readers arsenic we perchance can. And truthful arsenic we deliberation done nan deal, we’re weighing each of those things.

Now, nan mobility of really you maximize money for nan Atlanta grading publication is absorbing because you do person an option. You tin return nan New York Times way aliases nan Alden Capital route, and you tin sue. We looked astatine that calculation successful nan lawsuit of Open AI and chose not to sue. That doesn’t mean we’re not going to writer each different ample connection exemplary institution retired there.

You measurement what they’re offering connected each those fronts. All nan benefits they’re offering, again, nan merchandise partnerships, search, et cetera. You measurement each those things versus what it would costs to writer and really overmuch you would get from it, and past you make a choice.

It’s been reported that The Times is a cardinal dollars deep into its ineligible fees against OpenAI. That’s-

Suggesting they expect to get much than $1 cardinal for nan content.

They presume they’ll get much than $1 million. The Atlantic is owned by billionaires, it’s owned by Laurene Powell Jobs. Would she person fronted $1 cardinal successful ineligible fees, aliases is that disconnected nan array for you?

That’s a analyzable question. I mean, nan reply of course, yes, right? If we made an statement to her that this is what is champion for nan early of superior journalism, past she would surely person supported it.

The logic I inquire that mobility that way is, there’s a batch of consequence there, and erstwhile you person a rich | owner, you tin judge possibly much consequence than if you are a publically traded institution aliases you person a bunch of VC money for illustration Vox does. But nan consequence location is almost intolerable to ascertain because nan copyright rule statement is simply a full coin flip astatine this infinitesimal successful time.

Do you deliberation it’s a coin flip? Or do you deliberation it’s a 60/40, 40/60, 70/30, 30/70?

You deliberation it’s 50/50? Former copyright lawyer Patel here.

And that is simply a axenic lawyer answer. And I deliberation you tin tally done nan argument, and connected a bully day, a judge that has conscionable utilized Dall-E to make a storybook for their grandchild is connected your side; and connected a bad day, they’ve conscionable seen nan 2 startups that ripped disconnected Johnny B. Goode, and nan RIAA is suing them, and they lose. And I deliberation that is arsenic an affectional determination arsenic almost thing correct now.

But do you really deliberation The Times is going to scope an outcome, aliases do you deliberation they’re going to settee it? Partly you settee based connected wherever you deliberation nan lawsuit is going, right? And you do nan arguments and you’re like, “Oh my God. It’s now 70/30, truthful we should settee connected different terms.”

Right. I deliberation there’s that, and we haven’t gotten done immoderate of that, and we really haven’t seen thing substantive from OpenAI successful position of really they’ve trained astir of these companies. It’s really nether fastener and key, what they’ve trained on, what their attack to training was, what their attack to copyright rule and training was. So sure, possibly arsenic clip goes on, that will change.

But conscionable connected a consecutive let’s spell done nan argument, you ingest a bunch of data, you train a exemplary connected it, which intends you group immoderate weights and you propulsion nan information out, and I tin do this generation. Who knows? If The Times wins, for example, and your 2 years is up, and it turns retired it wasn’t adjacent usage to train these, do you deliberation you’ll beryllium capable to get much money? Are you conscionable waiting astatine nan timepiece connected these lawsuits?

Oh. If The Times wins, we will get much money from everybody. Every journalistic statement will get overmuch much money from everybody, right?

If The Times loses-

We’ll each get overmuch less.

I’m conscionable asking, really are you factoring that risk?

Basically, you person a speech pinch your lawyer and your lawyers, and I talked to tons of copyright lawyers to decide. If I thought The Times had a 99% chance of winning, I would person a very different position going into these negotiations. If I thought The Times had a 1% chance of winning, a different perspective, right? So you make your decisions based connected that.

You besides measurement different things, right? Will matter beryllium important to training ample connection models 2 years successful nan future, aliases will it each beryllium multimodal data? Will synthetic information beryllium truthful good? Right? I’ve had group making ample connection models fundamentally say, “We don’t request you because we tin do it each done synthetic information successful nan future.” And possibly nan synthetic information is derivative of nan integrated data, but you person to measurement what will your information beryllium worthy tomorrow?

And therefore, are you getting a amended woody now aliases will you get a amended woody tomorrow? Do you deliberation your information is going to beryllium worthy much tomorrow because matter will still beryllium valuable. And successful fact, that integrated human-certified information that we create astatine The Atlantic and person been doing forever, if you deliberation that is going to beryllium much and much valuable and you deliberation The Times is going to win, good past you will beryllium much cautious. You would request much successful nan deals. I’m not saying you wouldn’t do immoderate deals, but you conscionable person a different framework.

Do you deliberation that nan determination to return nan woody now is rooted in, “Well, we tin get immoderate gross now, and hopefully each of these copyright lawsuits,” because there’s a batch of them. The manufacture really conscionable has to suffer 1 to get to wherever you’re saying, right? The grounds labels person to triumph aliases The Times has to win, aliases Sarah Silverman has to win, and past nan dominoes commencement falling successful your favor.

But here’s 1 much facet which I deliberation is interesting. I judge that america doing this woody and nan Wall Street Journal doing their woody helps The Times because it shows that location is simply a marketplace for this stuff.

There’s a disapproval like, “Why is location not this corporate action?” And nan reasons why location isn’t corporate action are hard, including antitrust law, which intends that I can’t talk to Bankoff and discuss pinch him-

Jim Bankoff is nan CEO of Vox Media.

Right. So Jim and I can’t talk and discuss together and get amended position for some of us. There’s different corporate action problem wherever if you subordinate a group, a consortium, nan money presumably is dispersed based connected nan connection contribution, but immoderate group for illustration The Times presumably deliberation that their marque worth and their words are much valuable connected a per-word basis. At nan apical of nan nutrient chain, they person an inducement not to subordinate a consortium. So you person a full bunch of reasons why you can’t do corporate bargaining together arsenic an manufacture to get amended terms, which would astir apt beryllium amended wide for Medium.

While that is true, 1 of nan ways that we tin thief nan manufacture is by making deals and mounting a market. So that then, I believe, that america doing a woody pinch OpenAI, makes it easier for america to make deals pinch nan different ample connection exemplary companies if those travel about, I deliberation it makes it easier for different journalistic companies to make deals pinch OpenAI and others, and I deliberation it makes it much apt that The Times wins their lawsuit.

The 4th facet successful nan adjacent usage study that a tribunal would do is nan effect of nan caller usage connected nan marketplace for nan aged work. And you’re saying, well, you person to person a market. You person to group immoderate prices for this benignant of use.

And we are mounting nan market.

And you deliberation that that complete clip will strategically thief The Times?

The Times lawsuit is going to dangle connected 1000 things that are much important, but I do deliberation that arsenic a wide rule group successful a marketplace and getting a adjacent speech of worth is bully precedent for our industry.

There’s different furniture of implications to taking this benignant of deal, and it comes from nan group who are making each of nan content, who are making nan work, who are penning nan stories and making each nan podcasts. And nan point that really strikes maine astir it is that The Atlantic’s national is mad. The Vox Media Union, which nan Verge squad that I negociate is in, is mad. The union for New York Magazine, different Vox Media imprint, is mad. They’ve each written letters and circulated statements saying they’re outraged astir this, and I’ve been reasoning a batch astir that outrage and what it means.

No 1 seems huffy erstwhile a media statement licenses their contented astatine Apple News aliases we people connected YouTube, moreover if nan position from YouTube aliases immoderate of these different platforms are worse aliases consciousness moreover much exploitative. And I’ve been trying to propulsion this apart, and what I’ve benignant of landed connected is nan copyright portion of this is conscionable an economical argument. You took our stuff, you didn’t salary for it, now you sewage to salary for it. You want to usage it successful immoderate caller way? We’ll travel to immoderate statement connected immoderate parameters, and you’ll salary for it.

And nan money connected nan economical broadside does not cure nan civilized problem that group see, which is partially a labour issue, this exertion mightiness displace each of america connected immoderate timeline, and partially conscionable the, “Hey, you conscionable took this stuff.” And now nan CTO, Mira Murati, is running astir saying, “Maybe immoderate imaginative jobs shouldn’t exist,” right? There’s a blitheness to this industry, peculiarly from OpenAI.

And that disconnect betwixt nan economical problem that copyright rule mightiness thief you lick aliases The Times lawsuit mightiness thief you extract much money from, and nan civilized dilemma, seems for illustration it’s wider than ever.

Oh, I wholly agree. I wrote a book connected nan history of nan Cold War that was published successful 2009, and erstwhile I learned that that was successful nan training group of Llama, benignant of nan emotional, “Wait. So nan book was pirated?” And not only that. It was chopped up into nan incorrect order. It was for illustration this violation, right?

And truthful I deliberation there’s astatine slightest 2 things that are ace important here. There’s one, that feeling, like, “Wait a second, they conscionable took this. They didn’t salary for it.” And past secondly, there’s this fear, which is AI could do unspeakable things to our industry. Absolutely. So you person those 2 very affectional factors coming together, and this is simply a woody pinch an AI company.

So my position aliases my domiciled arsenic CEO is to effort to put that speech and to say, “What I’m trying to optimize for is nan early wellness of The Atlantic, nan early economics of The Atlantic, nan early of this industry. I’m weighing each these different factors together, and I deliberation nan deal, net-net is very bully for america successful each these ways.

AI is this rainstorm, aliases it’s this hurricane, and it’s coming towards our industry, right? It’s tempting to conscionable spell retired and beryllium like, “Oh my God, there’s a hurricane that’s coming,” and I’m angry astir that. But what you really want to do is, it’s a rainstorm, you want to put connected a raincoat and put connected an umbrella. If you’re a farmer, you want to fig retired what caller crops to plant. You want to hole and woody pinch it.

And truthful my occupation is to effort to abstracted nan fearfulness of what mightiness hap and activity arsenic difficult arsenic I tin for nan champion imaginable outcome, knowing that because I person done a woody pinch an AI company, group will beryllium angry because AI could beryllium a very bad thing, and truthful there’s this association. But regardless, I person to effort to do what is champion for The Atlantic and for nan industry.

That was nan CEO answer. There’s a logic I introduced you arsenic nan erstwhile editor-in-chief of Wired, because I want that reply too, which is you ran an industry-leading publication during nan societal media era.

A batch of what I’ve heard from group who wish to modulate AI aliases slow it down aliases thing is we grounded to study thing from nan societal media era. We grounded to study really to modulate these companies, we grounded to study really to clasp them successful check. We each surely grounded to study really to get paid for really overmuch they usage our content. Facebook made a bunch of money distributing our contented and media companies made none. YouTube, I think, still doesn’t salary precocious capable rates to support a news statement connected YouTube, and it’s conscionable a civilized nonaccomplishment connected YouTube’s part.

From that perspective, arsenic you watch nan societal media era unfold, what mistakes from that era are you trying to debar making? Because nan thought that nan tech companies are conscionable nan upwind is very tempting. They’re conscionable going to do this and we can’t stop. The societal media is conscionable going to hap to us.

And it did, but I deliberation a batch of group are looking backmost connected that and saying, “Boy, did we conscionable make a bunch of assumptions astir their motivations aliases really group would pass utilizing these tools.” It turned retired to beryllium utterly wrong, and we should person really stopped it earlier aliases changed it earlier.

Answering arsenic a CEO, that is what we are trying to do. We are trying to fig retired a measurement that these devices germinate successful specified a measurement that they are champion source. Maybe it’s conscionable nan upwind is nan incorrect illustration because we do person immoderate power successful nan very early stages successful making these things better. Just for illustration if location had been a measurement early successful Facebook to displacement nan measurement that News Feed work, truthful that established brands weren’t fixed nan aforesaid weight arsenic non-established brands. There were for illustration 20 basal sins astatine nan opening of nan News Feed, which ended up being hugely damaging to some publicity and American democracy.

But 1 of nan tweaks would’ve been, tin you alteration nan weight successful nan measurement nan creation and nan measurement fonts activity aliases immoderate truthful that personification successful Macedonia can’t commencement a publication called The Verge pinch different Z astatine nan extremity that looks conscionable for illustration you and has nan nonstop aforesaid weight? I deliberation that 1 of nan lessons is to salary a batch of attention. So nan AI hunt products person not been built and person not been launched. As they’re built and arsenic they’re launched, what are nan values we want embedded successful them? How overmuch matter do we want them to show? How do we want nan outer links to work? How do we want nan level of summarization? Those are really important questions to get correct astatine nan beginning, and I deliberation we are much apt to get them correct arsenic they do these benignant of deals.

The different point I’ll opportunity though, arsenic nan erstwhile editor of The Wire, like, “Oh my God.” Some days I aftermath up, I’m like, “I wish I was a newsman again.” It is truthful astonishing nan stories that... I mean, you guys are telling a batch of them, but nan opportunity to study because it is full madness correct now. It’s for illustration nan champion communicative to study connected successful years. It’s incredible. And truthful I can’t do immoderate of that because I’m a businessman now and I don’t moreover talk to nan editors. I don’t moreover cognize what we’re going to tally successful The Atlantic today, but I would love... I spent a batch of my clip penning those stories connected Facebook backmost erstwhile I was location and astatine Wire, I loved that. I emotion penning connected these crazy group successful this world of churn making these monolithic decisions. It’s truthful overmuch fun.

When you opportunity it’s each crazy retired there. The point that really strikes maine is I would opportunity moreover 2 years ago, group thought nan net benignant of calcified into a bid of platforms and this is what it’s going to look like. And past Elon bought Twitter and past ChatGPT showed up, and now it feels for illustration everything’s breaking apart. And nan point that feels mostly for illustration it’s breaking isolated to maine is nan presumption that nan large platforms person our champion interests successful bosom aliases tin beryllium trusted aliases trusted pinch our children. You spot nan spate of authorities that’s retired location that would modulate really kids usage platforms. You spot each nan reporting that is retired location astir Facebook willfully ignoring immoderate of nan problems it causes pinch teenagers.

The different broadside of it is simply a batch of nan underlying assumptions astir nan worth that is being exchanged, are benignant of for illustration Google’s assumptions. Google does image search, they get sued, they triumph because they’re a bunch of kids. Google indexes each of our sites, but they nonstop america postulation and we benignant of agreed pinch that attack for a agelong time. They support winning because they are innocents aliases they astatine slightest clasp themselves retired to beryllium innocents and they present a batch of worth successful a caller way. That portion feels for illustration it has decidedly changed to me. This presumption that it’s conscionable a bunch of kids trying to alteration nan world, and of people we should fto them skate by and inquire for forgiveness, not permission. Do you deliberation from nan business constituent of view, that that is really going to create opportunities to bring worth backmost to nan group who make nan activity because that’s nan existent problem here?

I don’t deliberation that’s changed. I deliberation that changed successful 2016, aliases that changed successful precocious 2016, early 2017, and past by Cambridge Analytica, which was 2018, I deliberation that’s when... I mean, that is each changing now. You are very correct that is changing, but I deliberation nan trajectory-

The circumstantial similarity that I’m drafting is not, I can’t spot them because of Cambridge Analytica. I’m pointing correct astatine Perplexity is scraping a bunch of paywalled websites and showing nan results, aliases OpenAI trained connected YouTube to make Sora, aliases Suno, nan institution the RIAA conscionable sued, is making music... And nan underlying portion of it is, “Well, it’s conscionable out. It conscionable ours to take, and we’ll salary immoderate money to cure it astatine nan end, and that’s conscionable costs of doing business.”

So this is truthful astatine liking and it’s astatine liking today, this very infinitesimal while we’re doing this, and it’s astatine liking successful nan lawsuit of Perplexity, I think. So Google sewage distant pinch worldly because “Hey, we’re cool kids and we’re wearing five-fingered peep shoes to meetings pinch Senators.” And it’s each cool, and they get distant for a while, and past yet regulations drawback up. They person to balance. It’s complicated. The move changes. Facebook, nan move changes aft nan predetermination of Trump, and past moreover much truthful pinch Cambridge Analytica.

Uber comes on and has a wholly different strategy, which is, “We’re going to get distant by conscionable ignoring everything and past making truthful overmuch money that we’re immense and past we’ll travel along.” Which is simply a very different approach. I deliberation that Perplexity is trying to decide, “Are we going to beryllium Uber?” And we’re conscionable going to disregard Robots.txt? You publication each these stories. “Or are we going to effort to do benignant of nan Google point and conscionable beryllium like, ‘We’re an AI company. We’re willing successful we’re going to get large and spot what happens’ aliases are we going to alteration and cooperate pinch nan publishers?” And I deliberation that is astatine liking correct now.

And my consciousness is that location are astir apt ways that we, arsenic an industry, tin push Perplexity into that 3rd way that I’m talking about, wherever they are a responsible subordinate that doesn’t do 900 connection summaries of a 901 connection story. And that really does benignant of a adjacent usage summary and a due nexus out. Will that happen? If that happens, that is truthful overmuch amended for america than if it does not. And truthful what is nan domiciled that I tin play successful making that happen? And what is nan domiciled that you tin play successful making that happen? That is very important for nan early of media.

And I deliberation it’s peculiarly important because I deliberation nan biggest point happening to media correct now aliases nan most... And you talked astir this successful nan amazing speech pinch Ezra Klein and you guys talked astir nan enshittification of nan web, that is nan point that is astir astatine liking correct now. AI contented correct now is bad. What if AI contented becomes good? What if nan web it becomes benignant of indistinguishable and you can’t find yourself around? How do you navigate done that? And building hunt engines that are still capable to nonstop you to morganatic existent content, not nan billions of spin-offs, that is 1 of nan astir existential problems that exist. And if that problem is not solved, we’re successful a world of hurt. So that’s nan point that is happening correct now that I americium astir worried, intrigued, willing successful for nan adjacent mates of years.

Because Google’s full business exemplary depends connected astir apt nan unfastened web. I mean, nan point you’re talking astir breaking is Google hunt broadly. If nan web becomes so enshittified that Google cannot benignant nan wheat from nan chaff, that type of nan web comes to an end, and possibly we’ve each paid capable attraction to Perplexity and they person a woody aliases OpenAI hunt merchandise has amended sources from The Atlantic and whoever else, and that will go nan victor because group will activity retired quality. It’s a large bet, but it benignant of relies connected nan web becoming truthful polluted that Google can’t benignant it out.

When I deliberation about, “What is The Atlantic’s future?” you person to decide, “Okay, what happens if nan web becomes ace polluted?” Okay, if it does go ace polluted, will Google, Perplexity, OpenAI, immoderate nan adjacent Bing, immoderate hunt startup location is, will they beryllium capable to navigate it? If they don’t, really past do we person a successful business model? Do you trust only connected direct... Basically, if nan web is gone arsenic a distribution system for The Atlantic, really do we scope readers? Well, convey God we person a people magazine. It’s nan astir hilarious thing, for illustration nan revenge of print. There’s print, but there’s also, of course, there’s your apps. There’s nonstop relationships you person pinch group successful your newsletters.

And past nan absorbing mobility is, “What astir nan walled gardens? What astir Apple news?” Do they go much important? If nan web becomes truthful polluted, you can’t really person a functional website there, do you trust much connected those places? Now my dream is that nan web doesn’t get truthful polluted, and I deliberation 1 of nan cardinal tasks of nan tech manufacture and everybody other is to effort to make judge it doesn’t get that way, but who knows? The finances, Incentivizing contamination are high. Maybe it gets truthful polluted that nan polluters nary longer person an inducement to beryllium retired there. Who knows what’s going to happen? Anyway, figuring retired a strategy for a world of maximum contamination is simply a nosy portion of my job.

I person this conception that I telephone “Google Zero,” which is nan conception that fundamentally each patient is 30% of postulation comes from Google, springiness aliases take. Over time, that number is going down, aliases it’s moving an ecosystem distant from immoderate publishers and much towards nan different publishers. And truthful yet you person to conscionable look successful nan reflector and say, “Okay, if my Google postulation goes to zero, what americium I? Is location still a business here?” What does The Atlantic’s business look for illustration if Google goes to zero?

We’re fine, I think. I mean, we person a very beardown subscription business and those group renew. And truthful you tin ideate a business wherever our Subscription business becomes a higher percent of our revenue, which is already nan mostly of our gross becomes conscionable a higher part, and we are figuring out... Some of those group find america from Google. The mobility is, successful your Google zero, arsenic agelong arsenic there’s still a fewer 1000 queries a period that are subscribed to The Atlantic, we tin suffer nan like, “What is nan meaning of life” queries.

Oh, I see. Super Bowl queries are gone.

Super Bowl queries are gone.

But arsenic agelong arsenic group are Googling, “How do I subscribe to The Atlantic,” and past subscribing to The Atlantic, that’s fine.

So arsenic agelong arsenic it’s Google one, not Google zero, we’re successful a decent spot.

Do you benignant your Google queries correct now based connected nan worth of which queries person and which queries don’t?

No, but we spell into Google Search Console. The apical 10 queries are, “The Atlantic, Atlantic subscribe.” Subscribe is not successful there, but group who are searching for The Atlantic astir apt person a precocious subscription intent. So actually, erstwhile ChatGPT came out, we did a nosy study wherever we went done Google Search Console, and we went done each of nan apical a 100 queries and past a random sample up to nan hundred-thousandth query. And past analyzed would that query spell distant pinch a cleanable chatbot? And The Atlantic queries aren’t going to spell away. What is nan meaning of life, which really does aliases utilized to nonstop you to an Atlantic article by Arthur Brooks.

Those queries spell away. And truthful past we were like, “Okay, really overmuch of our postulation will disappear, and past really overmuch of our subscriptions will disappear?” And nan postulation diminution is overmuch steeper than nan subscription decline. So successful a world of Google one, let’s telephone it, we return a existent deed successful postulation and readership, and that has knock-on effects. It does person immoderate knock-on effects connected our subscription business. It has knock-on effects connected our advertizing business. It has knock-on effects connected nan number of group who publication a story, which possibly makes a journalist little apt to write. All those different things, but it’s not crippling. We are little limited connected Google successful a heavy way, I think, than astir publishers. But still, we’re profitable, but we’re not massively profitable. And deed is simply a hit, truthful we’d person to fig that out.

We’ve talked a batch astir nan web being truthful polluted that mightiness conscionable beryllium excessively polluted to run on. Do you deliberation there’s different referrer retired there? Do you deliberation it’s for illustration OpenAI’s hunt merchandise mightiness go a reliable referrer of postulation to you?

That’s nan bet. There are a batch of group retired location who are like, “Well, OpenAI’s hunt doesn’t work.” It’s like, “Yeah, AI hunt doesn’t activity correct now. It doesn’t activity very well. AI is bully for a batch of things. It’s not bully for search.” That intends that Google hunt postulation is not going to spell distant for a while. Once AI is bully astatine search, that’s erstwhile Google regular TenBlueLink hunt postulation goes away. There’s a hedge built into this. I do deliberation that AI hunt will commencement to work. It’s a difficult problem because nan logic of nan vector exemplary that makes nan guidelines training models is not bully for search. Then you person to build different exemplary connected apical of it, which is your rag model. But your rag exemplary isn’t conscionable going to do a basal search. It has to do a full bunch of different analyzable things.

So we’re building soul Atlantic hunt and it’s ace complicated. Somebody puts successful a query and we do A, B, and C pinch AI successful bid to get nan champion results. That’s america pinch 1 technologist moving connected this problem. Over time, nan AI companies will put galore engineers connected this problem, and they will, I think, fig retired really to lick search. So erstwhile that happens, it’ll beryllium a partial replacement for Google, and God willing, location will beryllium norms wrong nan AI hunt manufacture wherever again, you won’t beryllium giving a 900 connection hunt consequence astir a 901 connection article. You will beryllium linking retired successful a measurement that gets group to that article. So God willing, that happens. Do I deliberation that 3 years from now we will person arsenic overmuch hunt postulation from Perplexity, OpenAI, each of their competitors arsenic we do correct now from Google? Absolutely not. But will we person some? I judge dream so.

If you person a paywall and that’s nan main gross stream, what’s nan worth successful letting OpenAI synthesize immoderate of your activity successful a hunt merchandise for illustration this? Is it sending you traffic?

Yeah, it’s sending traffic.

Are you guaranteed immoderate postulation from them?

We’re not guaranteed traffic, but it’s sending america traffic. Every visitant who comes to nan site, A, arsenic agelong arsenic they for illustration what they see, it increases nan marque value, and B, there’s immoderate chance they extremity up subscribing. And we don’t person a difficult paywall. We person a paywall pinch adjustable rules. Many group deed nan gate, immoderate group don’t deed nan gate, truthful immoderate group won’t deed nan gross and they’ll publication an article and they’ll spot our ads aliases possibly moreover they’ll spot an advertisement connected nan gross astatine article. So there’s tons of gross we make disconnected of each personification who comes to The Atlantic.

One of nan things that has really struck maine successful conversations pinch various societal media executives is nan belief, it is simply a stone coagulated belief, that what societal media products person revealed is that group don’t attraction astir brands. They attraction astir people. And that each media will beryllium individuals and not brands. I don’t judge them. I deliberation that is simply a very self-serving attack from nan societal media companies because they person an infinite proviso of teenage creators they tin conscionable switch astatine will. But you tally 1 of nan astir storied brands successful media. When you perceive an Adam Mosseri conscionable station that, he posts things for illustration that each nan time, what’s your reaction? Is it, “is this correct?” Is it, “I request you to conscionable locomotion distant from platforms for illustration this?” Or is it much mercenary? “We request to find immoderate customers connected Instagram and show them nan marque there, but we can’t dangle connected them.”

We’ve been done this successful nan past 5 years successful our business wherever if brands didn’t matter and each that mattered were individuals, everybody would beryllium connected Substack. And that’s not nan measurement it shook out. And nan measurement it shook retired is that location are immoderate individuals who’ve made astonishing brands and are connected Substack, and location are immoderate group who are astatine The Atlantic aliases astatine The Verge. 

The incentives are different for each writer. They’re different for each individual, they’re different for each editor. And truthful I deliberation nan measurement it’s shaken retired is that what societal media did, and Mosseri has been talking astir this for 15 years. He is correct that societal media did thief individual brands. I retrieve nan first clip I heard Moi make this argument, I’m beautiful judge he was talking astir nan NBA. And he’s like, it was nan East vs. nan West successful nan Allstar. Now it’s Team Giannis vs. Team LeBron. I deliberation that’s nan metaphor he utilized nan first clip I heard him talk astir that.

And it’s true. It did make LeBron’s individual marque much valuable, but LeBron still plays for The Lakers. Giannis still plays for The Bucks. The brands, nan teams, nan structures still beryllium successful this world. I deliberation you’ve seen a small spot of a barbell successful our manufacture successful a batch of industries wherever power, wealth, and power person accumulated group astatine nan acold end, for illustration The New York Times, and they’ve accumulated arsenic individuals, and past it’s section news successful nan mediate that’s been crushed. I deliberation you whitethorn spot much of that successful nan future, and truthful my dream would beryllium that The Verve and The Atlantic are connected nan acold broadside of nan barbell.

Yeah, that is besides my hope. I look astatine nan acold broadside of nan barbell and I look astatine The New York Times, which is undoubtedly 1 of nan winners successful each of this. They tin money nan cardinal dollars successful ineligible fees, to nary result, connected nan backmost of a video crippled subscription service. That is nan revenue. That’s what makes nan merchandise sticky, Connections and Wordle.

Remarkable.

It’s very smart. They were very clever. They realized that connection games are going to money nan connection games of nan beforehand page.

Or nan different measurement around. You consciousness amended astir playing Wordle because it’s portion of The New York Times, because you subordinate The New York Times pinch nan Ukraine coverage. I don’t cognize precisely really nan virtuous circle works, but they figured retired a bully measurement of doing it.

The point that scares maine astir that is that’s what I hope, is that that is each conscionable 1 synthesis of marque worth and conscionable worth values. But past I perceive group opportunity nan contented is not really valuable. People are paying for nan games. They travel backmost each time for nan games and nan contented is conscionable there. And if they weren’t paying for nan games, they would conscionable find nan contented elsewhere because it’s free. And we’re getting to a spot wherever nan high-quality activity is down nan paywall and low-quality contamination is freely disposable to everyone else, and nan accusation situation has been destroyed. And if you opportunity nan contented isn’t valuable, and I’ve heard truthful galore group opportunity nan contented itself is not valuable, nan activity is not valuable, it’s nan services that are valuable aliases nan distribution that’s valuable aliases whatever, nan advertisement targeting is valuable, past astatine immoderate point, nary of america tin salary anyone. At immoderate constituent we’re conscionable saying nan contented isn’t valuable. We mightiness arsenic good fto teenagers publication our stories connected TikTok for free.

I person a bully counterpoint to that.

So there’s this mag I’m very acquainted with, and they’ve been wholly unsuccessful astatine launching games products and crosswords products and derivative products, and their CEO moreover tried to rotation up an AI-based societal media level which ended up getting sold, and yet it’s still making tons of money and it’s still profitable. It’s called The Atlantic. A 100 percent, nan contented is valuable. The group are paying us. We’re nan research proving that The New York Times contented is valuable. We person good complete a cardinal people. We conscionable announced that they’re paying america ever much money for our content. They’re not paying for thing else.

Do you deliberation that that is extensible? The Atlantic is singular, is an institution successful America. Do you deliberation that’s extensible to section news? Do you deliberation this is extensible to immoderate of these mini communities aliases efficaciously news deserts wherever there’s nothing?

One of nan difficult things is that if you look astatine nan brands that person been astir successful pinch paywalls. They thin - not completely, The Information is simply a bully counterexample - they thin to beryllium brands that person existed for a agelong clip and it built up a batch of marque value. And truthful could you create a caller section newspaper and moreover pinch awesome reporting create a paid exemplary for it? 

That is simply a difficult problem. I person not seen awesome grounds that that could beryllium done successful nan agelong run. Maybe it tin be. But I person nary uncertainty that location is really bully contented that tin beryllium highly valuable, and if you create thing pinch utmost value, you tin get group to salary for it. Now, you person to tally your business efficiently and you person to beryllium thin astir it and you person to fig retired each nan smart ways to get group to publication it. It is decidedly doable, and I 100% completely, totally, fundamentally disagree pinch each those group who’ve been saying to you that nan contented is not valuable. They’re conscionable wrong.

You said thing to maine arsenic you were connected your measurement to The Atlantic that has stuck pinch maine ever since. I asked you why you were leaving your station arsenic editor-in-chief astatine Wired to spell beryllium nan CEO of nan Atlantic, and you said this point to maine that I’ve ne'er stopped reasoning about. You said, “I can’t hold to tally nan merchandise team.” You were truthful excited astir it. It was nan thing. And arsenic an editor-in-chief, I was like, “Yep, that’s it. That’s why you would go. I want to tally nan merchandise team. That’s nan point I want.” You’ve been location for a minute. You are making deals pinch different elephantine merchandise organization. You’re getting immoderate credits to usage OpenAI’s systems. What are nan products you want to build?

It’s funny you mentioned that today. Probably 2 hours ago, I was talking to nan feline who’s building soul AI hunt astatine The Atlantic. It’s conscionable a demo model, and who knows if it’ll spell retired connected our website aliases if it won’t spell out. But nan mobility is, really do you fundamentally build an AI-based hunt motor internally to The Atlantic? And it’s an astonishing problem, because you could conscionable say, “Well, conscionable nonstop nan query retired to OpenAI and hunt nan database they have,” but that’s not nan correct measurement to do it.

Some of nan measurement group do hunt is they’ll return a query and past they’ll tally a adjacent query connected it to move it into a 500-word query, and past they’ll return that 500 connection query and past they’ll comparison that to nan database. Is that nan measurement to do it? That’s nan measurement to do it.

Okay. Then really do you constitute nan query for really you want to comparison it? It’s an astonishing problem. And erstwhile I was nan editor, I wouldn’t get to woody pinch that. Jeff Goldberg does not get to beryllium successful these conversations. Jeff Goldberg gets to determine really we’re going to screen nan industry, but I get to determine really we’re going to tally our AI hunt product. I emotion getting to thief tally nan merchandise and engineering squad and helping to prosecute those people. And 1 of nan astir nosy things that I’ve gotten to do nan past 3 years is we’ve tally hundreds of experiments connected really our paywall works, really our pricing works, really our checkout page works. We’re getting to tally this cognition arsenic instrumentality for illustration a tech team. We get to say, “Well, okay, let’s tally this value trial and let’s alteration this colour and let’s person nan paywall rules beryllium A and B, unless nan sell-through complaint crosses X and unless nan subscription propensity successful a peculiar story...”

I conscionable emotion that stuff. That was portion of nan logic I emotion being astatine Wired is I’m a nerd. And nan opportunity to beryllium successful those conversations helped style those conversations and past to spot that it’s working, that nan information subject team, nan merchandise team, nan engineering team, nan user squad person built this point that took america from losing a batch of money to making money successful a short play of time, that’s awesome and that’s really fun. Now, nan adjacent improvement will beryllium tin we astatine The Atlantic thief build nan adjacent group of products that’ll thief this manufacture past successful nan era of crazy AI? Maybe we can, possibly we can’t. But nan truth that I person an opportunity to thief effort to do that is great, truthful I guidelines by that. We person a awesome merchandise engineering team.

I deliberation successful nan societal media era, a batch of media companies thought they were suppliers of contented to different distribution. You’re talking astir owning nan product, right? You want an app connected nan location surface adjacent to Instagram. Do you deliberation astir that arsenic a competitory section you’re successful now?

You mean for illustration The Atlantic app aliases a abstracted point that we build?

Just your merchandise experience. Again, successful nan societal media era, nan Buzzfeeds of nan world were like, “Our business is going viral amended than anyone other tin spell viral connected personification else’s platform.” That feels complete successful galore ways.

That’s over.

And now you’re saying, “I’m going to tally a merchandise squad to build a product.” That merchandise is competing for attraction pinch everything else.

Nick Thompson:

As we’ve mapped retired really does The Atlantic thrive successful a world wherever nan web goes away, redoing your app has been 1 of nan awesome initiatives. We didn’t moreover person an Android app. So building an Android app, getting it to parity, getting nan characteristic group truthful you could build a characteristic connected iOS and person it motorboat connected both, figuring retired what it is that nan readers want. It’s not nan astir glamorous stuff, but it’s really important and really cool. And truthful now absolutely, figuring retired really to build that merchandise and really to compete. In my perfect version, 3 years from now, astatine The Atlantic, we’ll person figured retired thing astonishing and launched it into nan world that is really bully for long-form journalism.

I took 1 ace astatine this. Through Emerson, which is simply a genitor institution of The Atlantic, I built this societal media platform. I worked pinch my partner, Raffi Krikorian, to build a level called Speakeasy, and nan thought was to create conversations that are affirmative and fulfilling online. Twists and turns of a startup and each this and that, and we ended up trading nan exertion into McCourt and Project Liberty, which is simply a bully outcome. But, arsenic you whitethorn person noticed, we didn’t extremity up supplanting Twitter. But nan opportunity to do that benignant of point is wonderful, and I dream that there’s a infinitesimal to do that astatine The Atlantic. Again, whether it’s thing that’s related to The Atlantic and its mission, whether it’s thing that’s straight portion of nan Atlantic, that is simply a 100 percent thing I dream we tin do successful nan adjacent two, 3 years.

Alright. Well, you’re going to person to travel backmost and show maine that merchandise erstwhile you motorboat this thing.

Ideally, you’ll spot it and you’ll beryllium like, “Oh, that’s cool. It first came up erstwhile we were talking connected Decoder.”

Yeah, I’m excited for it. Nick, convey you truthful overmuch for joining nan show.

Thank you truthful much. It’s awesome to talk pinch you.