Platform Dependence Is Breaking Across AI Products and Digital Media

The argument against California-style AI regulation was not that model risk is imaginary. It was that rules written around today’s visible capabilities may freeze a fast-changing technology stack before policymakers understand which capabilities will matter.

One unnamed speaker criticized SB 1047 as regulation that “starts with the solution” rather than letting builders discover what the platform can actually do and where the real issues emerge. A California executive-department veto letter from Governor Gavin Newsom to the California State Senate stated that he was returning Senate Bill 1047 without his signature. The timing mattered because the speakers treated the bill’s target as already slipping out of frame.

The strongest version of the critique was speed. A bill passed “literally tomorrow,” one speaker argued, could be irrelevant “by the end of next month.” The underlying model of AI products, in that speaker’s view, was changing from large models in data centers that “spit text out” into reasoning systems that can plan, infer, and act more like agents. If the real shift is from text generation to action, a bill designed around an earlier conception of model capability may regulate the wrong thing.

The practical question raised was what happens when these capabilities are embedded in consumer products. The speaker used Apple as the test case: if Apple launches a more intelligent Siri, and the product looks like the category lawmakers are trying to regulate even though the company says it is not doing “half the stuff you think we’re doing,” will California go after Apple?

That led into a second criticism: standardized regulation could become a moat for incumbents. One speaker argued that once AI rules become uniform and compliance-heavy, small and medium-sized open-source efforts are likely to be disadvantaged. The companies with the money to manage compliance, he said, would be the big technology platforms. He predicted that the next phase would be federal lobbying, with large model companies spending heavily to ensure their products face fewer threats from “hundreds of thousands of different competitors.”

The exchange settled on a familiar phrase: regulatory capture. The concern was not simply that regulation would slow AI. It was that, in the speakers’ view, it could determine who is allowed to compete.

Scale AI’s reported numbers stood out because they suggested an AI infrastructure company with large revenue and profitability, at a time when the best-known model companies are associated with heavy burn.

Scale CEO Alex Wang wrote that demand for reliable AI was “skyrocketing,” and said Scale grew ARR 2.5x in 2023, reached $400 million in ARR over Q2 2024, had Q2 annualized cash receipts of $800 million, and had reached profitability while continuing to invest in data for AI. A CNBC article screenshot reported that Scale, which supplies data for model training, expected sales to hit $1.4 billion in annualized revenue by the end of 2024, according to a person familiar with the company’s financials.

The reaction was less about whether Scale was large and more about what kind of AI company it is. One speaker described Scale as “a category of one” among AI startups because it was reportedly profitable. When another asked whether any major AI startup was profitable, the initial answer was effectively no. Perplexity was mentioned by a speaker as having reached $50 million in ARR, but the question immediately became what its costs were. OpenAI was invoked as the opposite case: a company with enormous demand and enormous compute obligations.

A tweet from Amir Efrati stated that OpenAI was losing so much money — about a $5 billion run rate that year, according to the tweet — that it was using “cash burn” rather than EBITDA or other standard metrics. The visible article excerpt said OpenAI’s cash burn was $340 million in the first half of the year, leaving it with roughly $1 billion in the bank. The speakers were puzzled by the gap between $340 million in first-half burn and a $5 billion annual run rate, then reasoned that the second half could include much higher compute costs after large purchases.

They were careful to distinguish the businesses. Scale’s cost base, one speaker said, is fundamentally different from a frontier model lab’s. Scale is essentially a data-labeling and evaluation company: it puts people in front of complex tasks, asks them to judge good and bad answers, and uses those judgments to help fine-tune models. That means the company’s main cost is human time and the platform around that labor, not training massive model clusters.

The metaphor used was straightforward: Scale is a shovel company. Nvidia supplies the metal. In that reading, the companies training frontier models are the miners consuming capital and compute, while Scale sells a required input into the AI gold rush without bearing the same compute-cost profile.

The warning about AI special purpose vehicles was about how access to hot private AI rounds can be repackaged and sold downstream.

A speaker defined an SPV as a vehicle that pools money from a group of investors to back a single startup. Instead of each person appearing separately on a startup’s cap table, the founder deals with one line item, while the SPV organizer handles the allocation and any eventual distributions. AngelList syndicates were cited as the mechanism that popularized this style of investing: one person sources access to a deal, many smaller investors contribute checks, and the pooled capital is deployed as one investment.

The concern was that AI SPVs are proliferating around the hottest companies. A speaker said a friend had warned about “the absolute devastation of the AI SPV” a month earlier, and that major AI deals — OpenAI, Perplexity, Glean, and others were named in discussion — were attracting SPVs layered on top of allocations. The described risk was not only a syndicate investing directly into a company. It was also the possibility of sub-allocations: an angel or syndicate organizer trying to get a slice of another investor’s larger allocation, such as asking Thrive Capital for a smaller allocation inside a larger OpenAI financing.

That produced the obvious question: are these allocations marked up before they reach smaller investors? The first answer was yes, but the speaker quickly corrected himself. Historically, he said, syndicates often used the same valuation as the company’s financing round and earned economics through carry — commonly 20% of profits after principal is returned, though terms vary. The broader point remained: the person organizing the SPV has an economic incentive, and the retail participant may be buying into a structure with layers between themselves and the company.

The evidence offered was anecdotal as well as market-based. One speaker said he had personally been approached that week with an OpenAI allocation. The reaction from another was that this looked like a top-of-bubble signal: in his view, institutional investors may be shifting risk toward retail investors attracted by the chance to own a recognizable AI name.

A Crunchbase News article screenshot reported that more than a quarter of all venture dollars that year had gone to AI startups. The speakers interpreted that concentration as especially striking because money was not cheap. In a higher-rate environment, they argued, directing such a large share of venture capital into one sector implies enormous expectations for growth — the kind of 20x, 30x, or 40x outcomes across portfolios that may require consolidation and produce many failures.

The sharpest example was Safe Superintelligence, Ilya Sutskever’s AI startup. The speakers described it as having raised $1 billion at a $5 billion valuation despite being pre-product, backed by a small team and a thesis rather than a commercial product. A Fortune article screenshot referred to a famed tech investor calling the company’s valuation a “joke.” Abacus AI CEO Bindu Reddy wrote: “We have reached peak AI bubble. A one-month-old pre-product AI company raising $1B at a $5B valuation!” Her post added that the company planned to hire fewer than two dozen people and build a massive cluster, with $1 billion buying roughly 20,000 to 25,000 H100s for a couple of years plus salaries for top researchers — “No product, no customers, no roadmap, simply raw research!”

A later passage from Sriram Krishnan made the return math concrete. If an investor enters a seed-stage company at a $100 million valuation, he said, the company may need to become a $10 billion company for the return profile to make sense. In a separate exchange, he and Aarthi Ramamurthy returned to the same concern: valuations were “nuts,” seed caps could be $50 million or $100 million for companies without products, and retail participants in AI SPVs could be “entirely wiped out.”

The point was not that all AI companies are bad investments. It was that scarcity of access, brand-name founders, and the social value of saying one invested in OpenAI or SSI can create demand even before a plausible return path exists. One speaker said participants may effectively be paying to put the logo on their website.

The AI investment discussion bled into a product question: if today’s text boxes are not the final form of AI, what is?

A Verge article screenshot reported that Jony Ive was working with OpenAI on a new hardware device, described in the visible text as an effort to build a computing experience “less socially disruptive than the iPhone.” That example pointed toward a shift in interaction models. The current AI product pattern — typing a prompt into a chatbot and judging the returned answer — may not be the model that wins once systems become more personal, ambient, and action-oriented.

Sriram Krishnan used Amazon’s Alexa as the negative example. He asked what it would mean to pay for Alexa as a subscription: if he says, “Alexa, order milk,” what exactly is Amazon charging him five dollars to do? Aarthi Ramamurthy said the product had not fundamentally improved and may even have regressed, despite Amazon’s distribution across millions of endpoints.

Krishnan’s household use case was mundane: weather, Spotify, lights, alarms, timers. The frustration came after the command succeeded. He said Alexa would turn off the lights, then say, “OK, by the way...” and try to promote a Fire Stick, a Spotify track, Audible, or another Amazon offering. His family’s reaction was to yell “stop” back at the device.

Krishnan described two UI patterns that bother him. One is filling empty space on a TV with whatever an ad network can serve. The other is using a voice assistant’s access to the user’s ears as an opportunity to upsell. In both cases, the product violates the task. Speech works for “turn off the light” or “order milk.” It does not work when the device responds to a simple command with an advertisement.

Ramamurthy gave the strongest possible strategic defense for Amazon. If people are already paying $20 a month for Perplexity, Claude, or other AI services, why would they also pay Amazon $5 to $10 a month for a speaker they do not carry with them? Her answer was that Echo devices may be valuable as nodes for personal intelligence. They sit in the home. They know routines, sleep patterns, music preferences, moods, lighting habits, and other signals. If the goal is a single personal AI that learns over time across interactions, Amazon’s endpoints could be useful.

But she also noted the obvious threat: Apple. Siri sits on the phone and can potentially see what is directly in front of the user on-screen. If Apple captures the personal-assistant layer, Amazon may struggle unless it can bring its assistant beyond stationary speakers.

Krishnan’s concern was organizational. He doubted whether anyone at Amazon had the mandate to make Echo an uncompromised product experience. The product feels, to him, like the org chart: speech recognition, ads, internal brands, Prime Video, Freevee, Audible, and other teams all fighting for surface area. His complaint was not that Alexa lacks technology. It was that the product experience has been subordinated to internal monetization pressure.

The contrast with ChatGPT’s voice mode was explicit. Ramamurthy said talking to the ChatGPT app in voice mode “feels way better” than talking to any of these speakers. In that comparison, Amazon’s decade-old hardware footprint is not enough. The winning interaction model has to respect the intent of the user and avoid turning every moment of attention into ad inventory.

The Amazon discussion widened from Alexa to retail, after a Business Insider article screenshot reported Amazon’s push to sell cheap, unbranded goods, including a leaked-video claim that an executive told sellers the company had “completely transformed the Amazon experience” to compete with Temu.

Sriram Krishnan asked whether Amazon was in trouble from Temu, Shein, TikTok Shop, and similar models. Aarthi Ramamurthy answered “yes and no.” TikTok Shop, she said, can feel as if it is overtaking the product itself when she scrolls her For You Page. She had bought a $2 item early on just to try it, and described the goods as often unbranded and cheap, though occasionally useful or surprisingly good quality. Her examples were the sorts of viral utility products that circulate through short-form video: a small nightlight, a potato-peeling gadget, and other inexpensive household items.

Krishnan was skeptical of the quality — “some of this sounds like garbage” — but the business model was harder to dismiss. Ramamurthy said Temu is spending aggressively on user acquisition across social channels. Both noted that Temu’s interface is heavily gamified: spins, clicks, unboxing mechanics, and other engagement devices. Krishnan assumed those choices had been A/B tested, even as he found the experience confusing and overbearing.

His broader explanation was generational and behavioral. Temu’s appeal, he argued, rests on gamified UI and the same disposable-consumption insight that powered fast fashion. A high school student might buy a cheap sweatshirt for a few days of novelty. The product does not need to last, because the price and entertainment loop are part of the value proposition.

This matters for Amazon because it shows pressure from below. Amazon built trust, logistics, and convenience. Temu and TikTok Shop compete with entertainment, extreme affordability, and impulse. If Amazon responds by adding cheaper, less branded inventory, it may protect volume while changing the feel of the marketplace.

Two separate eBay items appeared, and they should not be treated as one developed story.

The more substantive discussion came from an X post by Sriram Krishnan linking to a Bloomberg article about Brad Stone’s reporting. The visible post said eBay’s Meg Whitman passed on buying Amazon, and that in 1999 Jeff Bezos bought a 25% stake in eBay for $100 million.

Krishnan and Ramamurthy treated the anecdote as remarkable but not irrational in its moment. In 1999, eBay was the stronger business by several measures. Krishnan said eBay was more valuable and more profitable than Amazon, and described eBay as a high-margin business even today, using “80% margins or whatever” as his rough characterization. From that vantage point, Meg Whitman’s rejection was not obviously absurd. It only becomes legendary because of what Amazon later became.

The useful point was timing. Strategic decisions are made with the information and market structure of the day, not with twenty-five years of hindsight. eBay looked like the better business. Amazon looked less certain. The outcome reversed the apparent hierarchy.

A separate TechCrunch screenshot displayed the headline “eBay Rejects Acquisition Offer” and said the company decided to remain independent after a lengthy board meeting. The spoken treatment was brief: eBay “completely rejected the deal,” and one speaker thought staying independent was the right long-term move. The source did not provide the buyer, date, terms, or rationale beyond that passing treatment. The supportable point is therefore limited: the source surfaced a recent rejection headline, but did not develop it into an account of a transaction.

BuzzFeed and Complex were used as evidence that one era of internet media has ended, but the source included multiple fragments that should not be collapsed into one clean sale narrative.

The clearest reported item was a post from The Information: BuzzFeed agreed to sell its youth culture brand Complex to livestream shopping platform NTWRK for $108.6 million in cash. Krishnan said BuzzFeed had bought Complex for about $300 million in 2021, which he described as wiping out roughly $200 million of value in three years.

Other BuzzFeed references were less settled inside the source. A TBPN lower-third mentioned a Byron Allen bid for the struggling publisher. Later, a Verge-attributed screenshot displayed the headline “BuzzFeed Sold to Private Equity Firm Following Years of Decline.” The speakers did not reconcile those items into a single clean account, and the source does not establish how they relate to the Complex transaction. The reliable substance is narrower: Complex was discussed as a sale to NTWRK, and BuzzFeed was treated more broadly as a symbol of decline in the viral publisher model. The Byron Allen and private-equity references appeared as unresolved on-screen fragments rather than as a developed account of BuzzFeed being sold.

The analysis from Krishnan and Ramamurthy was that BuzzFeed once owned internet culture, but its traffic model depended on platforms that changed their incentives. Facebook and other social platforms used to be traffic fire hoses. Then they realized they did not want to send people away. X was cited as another example: link posts get penalized because the platform wants to keep users inside its own product. If a media company relies on traffic to its website so it can serve display ads, Ramamurthy said, “you are basically dead.”

Krishnan agreed and pointed to creators as the replacement media brands. Individuals can build audience more easily than faceless publishers and monetize through subscriptions, products, or other direct channels rather than traditional ad networks. Ramamurthy noted that some creators generate millions of dollars with a Substack or YouTube channel and very little overhead.

The conclusion from that exchange was blunt: you cannot build the next BuzzFeed. The era of the viral publisher is over.

But Roger Lynch, CEO of Condé Nast, complicated that claim. He did not dispute that BuzzFeed belonged to an older internet. He described it as a business that worked when publishers could take search and social traffic and convert it into commerce dollars or other revenue. “That era is gone,” he said. But he argued that the decline of platform-driven traffic does not mean all media brands are structurally doomed.

His search example was concrete. Condé Nast had shown its board screenshots of search results from seven or eight years ago: a few sponsored links, then the familiar organic results. Today, Lynch said, the same query produces an AI overview, rows of commerce links, sponsored results, and only then organic publisher links. For Google, that may be good business. For publishers, it pushes them down the page.

Lynch said Condé Nast had been forecasting search declines for several years, only to find each year that traffic fell more than expected. His response was to tell teams to plan as if search were zero. Not because he expected literal zero, but because he expected search to become a very low share of traffic — in his words, a “single digit percentage.” Brands that lacked a plan for that world would be deprioritized.

That is the key distinction between BuzzFeed and Condé Nast in Lynch’s telling. A publisher built on arbitraging platform traffic is fighting all the way down. A publisher with strong authoritative brands, niche loyalty, direct audience relationships, and other revenue streams has a path.

The discussion with Lynch began with Vox Media’s reported “interaction models,” which John Coogan described as a move toward performance-based compensation for creators. He said people online were joking that it looked like a consulting-style rebrand for laying off half the company, while internal critics viewed it as a shift toward performance comp.

Lynch’s first response was restrained. Vox, he said, had historically prided itself on treating journalists similarly across brands, with broadly similar compensation structures. A shift toward revenue share or performance pay would be a clear departure. He suspected two motives: financial pressure, because nearly every media property is facing tougher economics, and a belief that performance incentives democratize rewards by letting the “cream rise to the top.”

But he said the central question is not whether performance compensation is good or bad in the abstract. It is what the organization means by “value created.”

Raw traffic, Lynch argued, is a poor indicator of value for many premium media brands. If journalists are paid for clicks, they will be pushed toward the kinds of content that erode trust. Coogan called that “the BuzzFeed model,” and Lynch agreed. The challenge for Vox or any publisher using highly structured performance compensation is aligning incentives with the value of the media property itself. A Verge writer could generate traffic with ridiculous clickbait lists, but that would detract from The Verge’s reputation for credible technology journalism.

Condé Nast has not moved to purely quantitative compensation for content workers, Lynch said. It still relies heavily on base compensation, with variable pay layered on top. The rationale is that premium journalism needs enough stability that reporters are not forced into a click-chasing race to the bottom. There is room to reward people who break major stories and drive significant attention, but Lynch argued it is difficult to maintain “highly elevated brand-safe content” with a fully programmatic pay structure.

The tension is broader than Vox. Media companies want creator-like incentives because creators are visibly capturing attention and money. But if a publication’s asset is trust, the wrong metric can liquidate the brand in pursuit of the quarter’s traffic.

Lynch described Condé Nast’s business changes as a move away from total dependence on advertising and toward revenue streams tied more directly to brand authority. He framed the work as part of managing a company through structural transformation in media, not as a claim that one simple playbook solves the industry’s problems.

The first major area is consumer revenue. The New Yorker was his strongest example: a large paid subscriber base across digital and print that insulates the brand from advertising swings. Condé Nast is extending paid walls and membership models across other properties, including Wired and Vanity Fair. Later, Lynch said digital subscription revenue grew 29% the previous year and was still growing double digits. Pitchfork and Tatler had launched subscriptions, while Vogue was seeing strong digital subscription growth.

Price increases had not damaged the model as expected. Lynch said Condé Nast had raised subscription prices materially over the prior couple of years, each time anticipating lower retention. Instead, retention improved every year. For Condé Nast, at least so far, subscription elasticity looked favorable.

Another area is intellectual property. Condé Nast’s archive is not just a library of articles; Lynch described it as source material for film and television. Condé Nast Entertainment, in this account, is no longer only a YouTube production group. It develops feature films and premium TV series based on journalism and fiction from Condé Nast brands. Lynch cited Netflix’s “Spiderhead” as based on a New Yorker short story.

Commerce is also part of the model. If GQ recommends a white T-shirt or Wired reviews running shoes, the brand’s trust can convert into purchasing behavior, and Condé Nast can capture a portion of that transaction. Lynch said commerce has grown every year, but he was clear that Condé Nast is not primarily trying to launch its own products — no “New Yorker protein powder.” It is using partnerships and authority to sell fashion, travel, and other categories.

He also described Vet, an initiative at the intersection of e-commerce, social commerce, and the creator economy. Condé Nast plans to use its relationships with luxury fashion companies to create a marketplace platform that fashion tastemakers and creators can use with their audiences.

Events are another growth engine, but Lynch rejected the idea that the company should simply do more of them. Events are one of the fastest-growing parts of Condé Nast’s business, he said, even though the company is doing fewer events than when he joined. The strategy is to focus on cultural moments: the Met Gala, the Vanity Fair Oscar party, and similar events that can become global media phenomena.

The Met Gala numbers were the strongest example. Lynch said the first seven days of content from the most recent Met Gala produced 3.1 billion video views, up about 60% from the prior year, which had itself been up about 60% from the year before. The red-carpet livestream drew 200 million viewers. The Vanity Fair Oscar party grew 65% year over year.

The reason these events work, in his telling, is not that Condé Nast can manufacture one every week. It is that a few events have cultural gravity, and the company’s global organization can amplify them across markets and brands.

Coogan asked Lynch about a specific strategic reversal: Condé Nast had previously centralized many video efforts under a large Condé Nast Entertainment umbrella, then unwound that structure so brands such as GQ and Vogue had their own channels again.

Lynch said the original centralized model predated his tenure and made sense at a certain point in digital video. Condé Nast viewed video as a capability: if the company knew how to make videos, a central production unit could make them for everyone. But as YouTube and short-form platforms evolved, the flaw became clear. Each Condé Nast brand had a distinct audience and community. A generic handle or centralized programming structure weakened the brand-audience relationship.

When Lynch came in, one of the first moves was to decentralize video and give power back to global editorial directors at GQ, Vogue, Vanity Fair, and other brands. The people who understood GQ’s voice should program the GQ YouTube channel. The result, he said, was better performance, stronger engagement, and a healthier digital ecosystem.

The algorithmic logic was simple. If a channel mixes menswear, Hollywood roundtables, beauty, politics, and travel in one feed, the platform may not know whom to serve it to. Separating the brands gives both users and algorithms clearer signals.

This fed into Lynch’s broader portfolio theory. Condé Nast works best at the ends of a barbell. Large, authoritative category leaders such as Vogue, Architectural Digest, Condé Nast Traveler, and The New Yorker can rise above changes in search and AI. Small, deeply loyal niche brands can also work; Lynch cited Pitchfork as a very small brand, about 1% of revenue, with a strong audience in its category. The danger is being caught in the middle: too broad to inspire loyalty, too weak to dominate a category, and too dependent on traffic mechanisms that are deteriorating.

Vogue, Lynch said, has grown revenue and profitability every year he has been at the company. The New Yorker had its most successful year ever “by a long shot.” These brands are not immune to platform shifts, but their authority gives them more options.

Coogan asked whether individual creators are a threat, acquisition target, talent pool, or a separate race. Roger Lynch answered that it was all of the above, depending on the situation.

He does not view creators as a strict zero-sum threat because Condé Nast brands have institutional access and capabilities that are hard for individuals to replicate. A solo creator with an iPhone and ring light is unlikely to get the same access to the inner workings of a political campaign as The New Yorker, or stage the Met Gala as Vogue does. But creators do compete for attention, and time spent is finite.

The preferred model is partnership. Lynch cited Emma Chamberlain serving as a special correspondent for Vogue at the Met Gala red carpet as a successful pairing: the creator brings a highly engaged younger audience, while Vogue brings prestige and access. Coogan called it an old-school distribution deal updated for the algorithmic age; Lynch agreed.

Acquisition is harder. Creator-led businesses carry extreme key-person risk. If the enterprise value is tied to one person and that person retires, leaves, or is otherwise unavailable, the buyer can destroy a large amount of capital. Lynch said creator companies become interesting acquisition targets only when the brand has value independent of the person on camera. Those cases are rare. In most cases, partnership gives most of the upside without trying to own human capital that can walk out the door.

The same tension applies inside the company. Journalists who build personal profiles can be good for business, Lynch said, but Condé Nast does not impose one hard rule across all brands. A New Yorker writer and a Vogue journalist may have different relationships to social media and public persona. The company supports journalists developing audiences, but not through a one-size-fits-all mandate.

Lynch drew a bright line between using AI to make the company more efficient and using AI to replace the human creative work audiences pay Condé Nast to produce.

On product and technology, the changes are already material. Lynch said a new head of product and technology joined in December, at a moment he described as a step-function change in AI. Lynch told him to question everything, start with a blank sheet of paper, and rethink how Condé Nast develops technology and products using AI.

The initial pilots were small: three or four people, with certain roles removed from what would previously have been a much larger product team. After six or eight weeks, Lynch said, there was enough evidence to make larger changes. Condé Nast reorganized around AI-assisted product development. Teams that once had 10 or 12 people — requiring technical project managers, QA engineers, product analysts, and other roles — became teams of three or four. The product manager might also act as the product analyst; a designer and engineer might work with AI that writes software and helps perform QA. Lynch said these smaller teams moved at three times the speed.

He was direct about the employment implication: there will be fewer software engineering jobs, at least for now. But he also accepted Coogan’s counterpoint that AI may create demand for software work in companies that would never previously have hired engineers. Coogan used TBPN as an example: as a small podcast, it would not historically have made sense to hire a full-time software engineer, but the new tooling changes the calculation.

On editorial and creative output, Lynch was more protective. Jordi Hays asked whether AI-generated advertising would matter to Condé Nast audiences. Lynch cited an ad that ran in Vogue print magazine using an AI-generated model. The backlash, he said, was aimed partly at the advertiser but mostly at Vogue. He considered the reaction validating rather than merely negative, because it confirmed that audiences want human-generated content and want to know that what they are seeing and reading is real.

That incident shaped the strategy: use AI in many parts of the company to drive efficiency, increase speed, and reach audiences faster, but do so in order to invest more in human-generated content.

Advertising remains a category-by-category judgment. In print Vogue, Lynch said, ads are absolutely a feature. In digital, ads can be either a bug or a feature. Programmatic display ads are more likely to be a bug because they disrupt the experience. Branded content, by contrast, is Condé Nast’s largest advertising category and a better fit because it uses the company’s brands, audiences, and creativity.

The implied AI bargain is narrow: automate the infrastructure and workflows around the media business, but preserve the human trust that gives the brands their value.