The TikTok AI Paradox

So TikTok held its annual marketing showcase ---TikTok World---and dropped a phrase that tells you everything about where they think commerce is heading: "Watch it. Love it. Want it." Which sounds like a tagline but is actually a thesis statement about collapsing the entire shopping journey into a single dopamine hit. Sixty-four per cent of TikTok users have bought something after discovering it on the platform. TikTok isn't describing a future. It's describing a Tuesday.

But behind the slick presentation was something much more significant: the infrastructure TikTok is building to make that collapse permanent. Infrastructure powered by AI at every layer---campaign management, creative production, checkout. And infrastructure that raises questions about measurement, trust, and organisational readiness that most ecommerce brands haven't even started thinking about.

Every Walled Garden Now Speaks the Same Language (Which Is Weird, Right?)

The headline announcement was TikTok's new Ads Model Context Protocol server---an interface that lets third-party AI agents connect directly to TikTok for Business accounts and manage campaigns through natural language. Campaign setup, bid adjustments, budget reallocation, performance reporting---all handled by external AI agents without a human ever touching Ads Manager.

If you've been reading this newsletter, you've seen this movie before. Amazon launched its Ads MCP Server in beta back at the IAB Annual Leadership Meeting in February. Meta opened its AI Connectors to advertisers globally in May. Google Ads MCP landed somewhere in between. TikTok's announcement completes the set: every major walled garden advertising platform now speaks the same protocol.

Here's the thing: the significance of this convergence is easy to understate, and most people are underestimating it spectacularly. For the past decade, each of these platforms invested heavily in making their advertising ecosystems self-contained---proprietary dashboards, platform-specific APIs, bespoke reporting frameworks designed to keep advertisers firmly inside the walls. MCP inverts that entire logic. It creates a universal interface layer through which a single AI agent can, in theory, manage campaigns across TikTok, Meta, Google, and Amazon without needing platform-specific development work.

The reason every platform is adopting MCP despite the fact that it makes their walls significantly less wall-like is delightfully simple: the alternative is worse. In a world where AI agents increasingly mediate advertising decisions, being the one platform an agent can't reach means being the first platform to lose budget. It's the advertising equivalent of being the only restaurant on the street that doesn't accept credit cards. You can have principles, or you can have customers.

For ecommerce operators running paid media across multiple platforms, this is the infrastructure that makes the theoretical cross-platform AI agent actual. One agent connected to four MCP servers, monitoring performance across every channel in real time, shifting spend toward whatever's working. That's a meaningful operational capability. It's also one that requires measurement infrastructure most brands absolutely do not have.

The Attribution Problem That's About to Eat Your Ad Budget

This is where TikTok's funnel collapse thesis and the MCP infrastructure story crash into each other in ways that should concern anyone managing ad spend.

TikTok's platform dynamics are structurally different from other advertising channels. The purchase journey is non-linear in the extreme---a user sees a product in a creator video, searches for more info, reads comments, watches three more clips from different angles, and buys without leaving the platform. Awareness, consideration, and conversion don't happen in that tidy sequence we all learned about. On TikTok, discovery often is the funnel.

TikTok Shop has evolved into a full ecosystem built around this reality---affiliate-driven creator commerce, episodic content formats with purchasing embedded inside them, gamified engagement loops where buying becomes an extension of watching. The boundary between content and storefront has functionally dissolved. GMV Max, TikTok's automation layer for Shop sellers, optimises toward sales as a unified system rather than treating product selection, creative delivery, audience targeting, and budget allocation as separate levers. Smart+ automates targeting, bidding, placements, and creative optimisation across the broader ads platform.

The problem is that while TikTok is building for a world where the funnel has collapsed, most brands are still measuring with tools designed for a world where it hadn't. Last-click attribution---the default measurement framework for most ecommerce advertisers---systematically undervalues TikTok because most of the platform's impact happens earlier in the journey. Cultural influence, search behaviour shifts, assisted conversions, the slow accumulation of product familiarity through repeated low-intent exposure---none of this registers in a last-click model.

Now consider what happens when an AI agent is managing budgets across all four platforms via MCP, optimising to last-click ROAS. The agent will systematically defund TikTok---moving budget toward channels that capture the final click, even when TikTok drove the discovery that made that click possible. The more sophisticated the automation, the more aggressively it will punish the channel that creates demand without capturing the conversion. Which is, if you think about it, a spectacular way to use very smart technology to make very dumb decisions.

The brands getting a more accurate read on TikTok's contribution are using incrementality testing, tracking search lift, monitoring brand demand signals, and building assisted conversion pathways into their measurement models. Without those frameworks, cross-platform AI agents won't just fail to help---they'll actively misallocate your spend based on incomplete data. Congratulations, your AI is now confidently wrong at machine speed.

The Trust Paradox (Or: TikTok's Existential Crisis in Two Slides)

The second set of announcements concerned Symphony, TikTok's generative AI suite for creative production. Rapid generation of video variations, conversion of static product images into video ads, creator-style AI avatars, faster creative testing at scale. The stated purpose is to address what is genuinely the biggest bottleneck in performance marketing on TikTok---creative volume. Most performance ceilings on the platform are creative ceilings, not media buying limitations. Symphony is designed to turn production from a one-off cost into a scalable input.

Here's what's wild: at the very same event, TikTok's own data made a clear case that creator-led content consistently outperforms controlled brand messaging. The reason is cultural fluency---content that feels native to the platform, that belongs in the feed rather than interrupting it. POV storytelling, conversational low-production video, subculture-specific formats. The most useful creative test is whether people would still watch the content if the product were removed. If they wouldn't, it's an ad that happens to be on TikTok, and TikTok's algorithm will treat it accordingly.

And then, in the next breath, TikTok announced it's scaling tools that allow brands to generate synthetic creator-style content---AI avatars that hold products, model clothing, and demonstrate apps in formats designed to be indistinguishable from human-created content.

Let's sit with that for a second.

TikTok's entire commercial value is built on a specific kind of trust---the implicit belief that the person in the video is a real human with a genuine reaction to a real product. That belief is what makes TikTok's discovery-to-purchase compression work. It's why 64 per cent of users have bought after discovering something on the platform. It's why creator-led content outperforms. And TikTok is now actively building tools to undermine the thing that makes its own platform valuable. (Bold strategy, but sure.)

AI avatars threaten that mechanism. Not immediately---early synthetic content will be labelled, imperfect, and easy enough to identify. But the trajectory of generative video is toward indistinguishability, and TikTok is actively accelerating it. As AI-generated content volume on the platform increases, the implicit trust signal---"this is a real person sharing a real experience"---starts to degrade. Not because every video is synthetic, but because users can no longer be certain which ones are. The uncertainty itself erodes the trust premium.

Research consistently shows that once consumers become aware synthetic content is prevalent on a platform, their trust in all content on that platform decreases---including the authentic stuff. The trust damage isn't limited to the AI-generated material. It bleeds into everything. Which is a great way to describe the world's most expensive own goal.

What This Actually Means for Sellers

For ecommerce brands, this creates a strategic decision that will become more consequential as Symphony's capabilities improve. Brands that lean heavily into AI-generated content may capture short-term efficiency gains---lower production costs, faster iteration, higher creative volume---while contributing to a platform-level trust erosion that ultimately makes all content on TikTok less commercially effective.

The brands that invest in genuine creator relationships, authentic product demonstrations, and culturally fluent content will bear higher production costs but will be building on the trust infrastructure that actually drives conversions. An AI avatar cannot pass the "would people still watch this without the product?" test. It has no interests, no personality, no cultural context that exists independently of the product it was generated to sell.

The Bottom Line

The picture from TikTok World is a platform simultaneously ahead of and at odds with itself. Its commerce infrastructure is arguably the most advanced in social media. Its AI advertising tools now speak the same MCP standard as every other major platform. Its creative AI suite addresses a real production bottleneck.

But its measurement frameworks lag behind its commerce model, its synthetic content tools risk undermining the trust that makes its commerce model work, and most brands' organisational readiness---measurement infrastructure, creative strategy, governance frameworks for AI-managed campaigns---hasn't kept pace with any of it.

Three things to do now: first, invest in measurement beyond last-click before connecting AI agents to your ad accounts, because those agents will make decisions based on whatever data they have. Second, treat Symphony as a production tool for scaling concepts that originate from genuine human insight---not as a replacement for the humans. Third, double down on authentic creator content while the trust premium still exists, because TikTok is doing its level best to erode it.

The most integrated entertainment-commerce-AI system in social media is being built. Whether it can sustain the trust it depends on is the question that will determine if "Watch it. Love it. Want it." remains a description of what happens on TikTok---or a eulogy for what used to.

• Andrej Karpathy joined Anthropic to work on frontier LLM R&D. He's staying close to education long-term but betting the next few years at the frontier are too formative to sit out.

• Synthetic customers now replicate ~90% of human conjoint study outcomes. US Bank and Target are already using AI-generated audiences to test messaging and products before going live.

• How MCP Servers Enabled Claude to be my Command Center

• How Does AI Get Its Information?

✍️ AI Transcription by Transistor - Generate incredibly accurate transcripts for your podcast episodes.

📖 Capitol AI - AI content that blends storytelling, design & research.

🤖 Forefront - Your AI assistant for work

🌊 Shortwave AI Assistant on iOS - Write, search, and analyze your emails with AI on the go.

♠️ AgentX - Build an AI agent chatbot for your business in less than 5 minutes.

😍It was great, all I needed

🙂Good, not great

🙄It sucked