There’s a big question hanging over TikTok’s pending sale: What the heck is going to happen to the algo?
Transferring the algorithmic magic of TikTok’s content-recommendation system is a key challenge facing Oracle’s Larry Ellison and a group of investors who the White House said are involved in a bid to buy TikTok’s US business as part of a $14 billion deal.
“The algo is what makes TikTok great,” one current TikTok staffer said. “Will a retrain be as good?”
The spin-off plan laid out by the Trump administration is simple on its surface. TikTok’s current owner, ByteDance, will hand over the keys to the business, including its US user data and algorithms, to new owners to comply with a divestment law. Oracle will audit its algorithm, which will be “retrained and operated in the United States outside of ByteDance’s control,” White House spokesperson Karoline Leavitt said last month.
The technical work will be complicated, researchers and TikTok employees said. Even using the term “algorithm” to describe TikTok’s content recommendation system — which uses a broad array of signals to decide what to show each viewer — is an oversimplification, said Nicole Ellison, a professor at the University of Michigan School of Information who has coauthored two papers on TikTok’s algorithmic personalization.
The challenge for ByteDance, which was founded in China, is finding a way to hand over that complex system while guarding some of its trade secrets and boxing itself off from access to US user data in a manner that satisfies the US government. The task for TikTok’s new US owners will be to retrain and maintain a new “For You” page without destroying what makes it compelling.
“Whatever they do, I just hope it’s still the same,” said Winta Zesu, a comedy content creator with 1 million followers on TikTok. “What we love about TikTok is the algorithm and how you just find exactly what you want.”
TikTok and ByteDance did not respond to requests for comment.
The TikTok algorithm is a black box, but there are plenty of theories on it
TikTok guards its algorithm secrets closely.
“We don’t know exactly how TikTok targets people for their content or their interests,” said Julie Vera, a PhD candidate at the University of Washington who recently coauthored a paper on TikTok’s FYP. “It is a black box.”
“The algorithm is always such a mystery,” said food content creator Jeremy Jacobowitz.
What we can assume is that the algorithm is more complicated than a single set of copy-paste code, Ellison said.
She said such systems were often governed by “very complicated computational formulas that look at many, many, many data points about an individual, including their past behavior in a particular online space” to predict what type of content they will engage with or enjoy.
Operating a recommendation system that relies on a large set of signals, without sharing specific user data with ByteDance — as required by US law — could prove to be a headache for TikTok’s new US owners.
How involved will ByteDance be after a sale?
The future of the “For You” page depends partly on how involved ByteDance will be in the US spin-off and how it licenses its technology to the new entity, according to researchers and company insiders.
ByteDance is expected to retain a minority stake in the new company, but relinquish control over US data and other key aspects of the business to comply with the divestment law.
Uncoupling the content-recommendation algorithms from ByteDance’s overall infrastructure and vast workforce of machine-learning scientists and engineers would likely change its outputs, said Paul Resnick, Ellison’s colleague at the University of Michigan who studies recommendation systems and online polarization.
Even if ByteDance offered “the whole code base and said, ‘Yep, it’s yours,’ if it didn’t come with any people who were involved in creating that code base, it would be very, very hard to make good use of it,” Resnick said.
The company has other options for opening up algorithm access without sharing its full code.
ByteDance could use an application programming interface, or API, for example. It already offers access to some content-recommendation models via an API product called “BytePlus Recommend.” However, the current API framework gives ByteDance access to user data, which would not comply with US law.
TikTok could also opt for something like a data clean room, a type of privacy-first technology popular in the advertising world that allows entities to compare data without sharing personal information. ByteDance built its own version a few years ago.
“There are technical considerations and then there’s the national security considerations, and different ways that they do this might have different technical costs,” Resnick said.
There’s also the possibility that ByteDance would retain some control over the algorithm after a sale, as Reuters reported last month. Too much ByteDance involvement could risk pushback from Congress.
The divestment law “set firm guardrails that prohibit cooperation between ByteDance and any prospective TikTok successor on the all-important recommendation algorithm, as well as preclude operational ties between the new entity and ByteDance,” John Moolenaar, chair of the House Select Committee on China, said last month.
A former TikTok product staffer was skeptical that new owners would be able to replicate TikTok’s magic without aid from ByteDance.
“It will literally take years to retrain the thousands of models that power the TikTok algorithm,” they said.
So, what happens to my curated TikTok feed?
If TikTok US figures out how to securely lease ByteDance’s algorithms, it may not take long for its “retrained” For You pages to look and feel different from the rest of the world’s.
That could irk some of the app’s users who have spent years building personalized feeds by liking videos and following creators.
“There’s this saying out there, ‘TikTok knows you better than yourself,'” Vera said. The way TikTok serves us content is part of “the magic” of the app, she said.
A second current TikTok staffer wondered whether retraining the algorithm using US user data alone would make the outputs feel more US-focused rather than global, creating a US monoculture on the app.
Travel creator Gabby Beckford shared similar concerns.
“I talk about different countries. I talk about global topics,” said Beckford, adding how she’s worried that an algorithm recreated in the US would lose some “global perspective.”
Read the full article here
