Mosseri thinks AI makes taste more valuable

Adam Mosseri's strongest AI claim is a little counterintuitive: synthetic content may make human taste more valuable, not less. In his Lenny's Podcast conversation, the Instagram head does not frame AI as a simple automation story. He describes it as a pressure test for product teams, ranking systems, creator identity, and the old assumption that scale should be managed by more specialized roles. 1

Mosseri says Meta's old product-team shape was roughly a dozen people: several Android engineers, several iOS engineers, several server engineers, a PM, a designer, a data scientist, and sometimes a researcher. That structure made sense when each specialty needed enough peers to review work inside its own lane. This year, he says Instagram has moved toward "pods": four to six more generalist engineers, one "product staff" role, and only the specialists the work truly needs. 1

The important phrase is "product staff." Mosseri describes it as an evolution of the PM: someone who can do some design work, some data analysis, and some research with the help of newer tools. A pricing problem may still need a senior data scientist. A novel experience may still need a senior designer. But the default middle of the team is becoming less role-pure. 1

That makes the episode different from another "AI writes code now" conversation. Mosseri's view is not that every specialist disappears. It is that the boundary between specialties becomes easier to cross, so the remaining specialist has to be meaningfully better at judgment, taste, or depth.

Mosseri is blunt about why designers should not read AI as an automatic career death sentence. If it is easier to build things, he says, more of the work moves toward deciding what should be built in the first place. Designers are anxious because engineers and product staff can now do pieces of design work. But he says he is "pretty long" on designers because they tend to have taste, and taste is harder to automate than production. 1

That argument only works if teams stop treating taste as surface polish. In Mosseri's version, taste is a filter on opportunity cost. When prototypes are cheap, the scarce resource is not the first mockup or the first implementation. It is the ability to tell which direction deserves more time, which result feels off, and which compromise will compound into a bad product.

He is also careful about the state of the tools. He rejects the binary split between being "AI-pilled" and anti-AI, saying the tools are amazing at some things and "remarkably bad" at others. The people who benefit most will be clear-eyed about both sides, and able to sense what the tools are likely to become good at next. 1

That is a useful hiring lens. The best candidate is not simply the person who uses AI most aggressively. It is the person who knows when the model output is good enough to accelerate a decision, when it is weak enough to mislead the team, and when a human specialist still has to own the call.

One of the episode's more grounded moments comes when Mosseri explains what Instagram's algorithm actually "knows." He says people often assume it has a detailed semantic understanding of their interests. In reality, much recommender progress has come from large embedding models and related techniques that produce artifacts people cannot read directly. The system does not literally know, in English, that someone likes surfing. It has numerical representations that correlate with surfing-like behavior. 1

Large language models change that interface. Mosseri says they can help describe previously illegible embedding regions in human language. Instagram's "your algorithm" idea is built around that shift: let users see topics the system thinks they care about, then add or remove them. 1

That matters because agency is now part of the product, not just a privacy slogan. A chronological feed gives users a simple kind of control, but Mosseri argues it creates bad incentives at scale. Publishers and professional accounts can post far more often than close friends, so a pure chronological default can drown the personal content users say they want. 1

His alternative is not "trust the algorithm." It is closer to: make ranking necessary, then make it more legible and adjustable. That is a harder product problem than giving users one chronological toggle, because it asks the platform to expose enough of a black-box system for people to steer it without pretending the black box has become fully transparent.

Mosseri's title claim is that AI-generated content can be a tailwind for Instagram. He gives two reasons. First, more content can create more attention, though he says that is not free because Instagram still has to rank good AI content above bad AI content. Second, he believes synthetic abundance will push people to seek out creativity, authenticity, and identifiable people. 1

That is a subtle position. He is not saying AI content is inherently good. He is saying the platform should judge content by the content, the point of view, and the person behind it, not by the tool used to make it. He does want users to know whether something was AI-generated, but he does not want the platform to filter AI content out as a class. 1

The hard part is labeling. Mosseri says detection may get worse as models improve, so platforms should be honest about their confidence. He even suggests it may be more practical over time to label camera-captured content, meaning non-AI content, rather than trying to label every AI-created thing. 1

For AI builders, that distinction matters. The future content platform problem is not just watermarking. It is provenance, account history, creator reputation, ranking quality, and user control wrapped together. A feed full of synthetic media is not automatically worse. A feed where nobody can tell who made what, why it was shown, or whether the account is disposable is worse.

Mosseri gives TikTok and ByteDance credit for pushing the industry on recommendation systems that can break small talent. He contrasts exploitation-based ranking, which uses known preferences, with exploration-based ranking, which tests what a user might like before the user has signaled it clearly. Showing a Bieber fan more Bieber-like content is easy. Testing whether that person might also like Afropunk is harder. 1

That is where the creator argument connects back to AI. If synthetic tools let more people make more media, ranking cannot merely reward already-proven accounts and already-obvious tastes. It has to create enough exploration for small or niche creators to find the audience that would care, including audiences that did not yet know they cared.

Mosseri says Instagram has invested in originality, breakout content, recency, and recommendations over the past few years, and that some of that work was inspired by TikTok. He also says Instagram is catching up and has line of sight to being best in class at recommendations during his tenure, while conceding: "not there yet." 1

The practical edge of the episode comes from Mosseri's failure stories. He describes the 2022 backlash over Instagram's video-heavy direction as several issues getting conflated: a small iOS test of a video viewer, more Reels, more recommendations, and creator frustration about reach. His takeaway was not that Instagram should have avoided testing the design. It was that a platform serving billions of people has to be more realistic about how quickly it can evolve. 1

He also names Facebook Home as a "spectacular failure" from his earlier PM career. The lesson he draws is unusually concrete: sometimes executing an idea well is the fastest way to learn that it does not have market fit. He gives Reels as another example. The first version was built inside Stories because Stories had momentum, but that was the wrong foundation. Most Reels were never seen before disappearing. 1

That closes the loop on his AI argument. Cheap building does not remove product judgment. It raises the penalty for weak judgment because teams can now execute more bad ideas more quickly. Mosseri's answer is smaller teams, broader operators, sharper specialists, more legible ranking, and a platform stance that treats AI as a tool whose value depends on the person and the taste behind it.