AI

Making AI models ‘forget’ undesirable data hurts their performance

Comment

Colorful streams of data flowing into colorful binary info.
Image Credits: NicoElNino / Getty Images

So-called “unlearning” techniques are used to make a generative AI model forget specific and undesirable info it picked up from training data, like sensitive private data or copyrighted material.

But current unlearning techniques are a double-edged sword: They could make a model like OpenAI’s GPT-4o or Meta’s Llama 3.1 405B much less capable of answering basic questions.

That’s according to a new study co-authored by researchers at the University of Washington (UW), Princeton, the University of Chicago, USC and Google, which found that the most popular unlearning techniques today tend to degrade models — often to the point where they’re unusable.

“Our evaluation suggests that currently feasible unlearning methods are not yet ready for meaningful usage or deployment in real-world scenarios,” Weijia Shi, a researcher on the study and a Ph.D. candidate in computer science at UW, told TechCrunch. “Currently, there are no efficient methods that enable a model to forget specific data without considerable loss of utility.”

How models learn

Generative AI models have no real intelligence. They’re statistical systems that predict words, images, speech, music, videos and other data. Fed an enormous number of examples (e.g. movies, voice recordings, essays and so on), AI models learn how likely data is to occur based on patterns, including the context of any surrounding data.

Given an email ending in the fragment “Looking forward…”, for example, a model trained to autocomplete messages might suggest “… to hearing back,” following the pattern of all the emails it’s ingested. There’s no intentionality there; the model isn’t looking forward to anything. It’s simply making an informed guess.

Most models, including flagships like GPT-4o, are trained on data sourced from public websites and data sets around the web. Most vendors developing such models argue that fair use shields their practice of scraping data and using it for training without informing, compensating or even crediting the data’s owners.

But not every copyright holder agrees. And many — from authors to publishers to record labels — have filed lawsuits against vendors to force a change.

The copyright dilemma is one of the reasons unlearning techniques have gained a lot of attention lately. Google, in partnership with several academic institutions, last year launched a competition seeking to spur the creation of new unlearning approaches.

Unlearning could also provide a way to remove sensitive info from existing models, like medical records or compromising photos, in response to a request or government order. (Thanks to the way they’re trained, models tend to sweep up lots of private information, from phone numbers to more problematic examples.) Over the past few years, some vendors have rolled out tools to allow data owners to ask that their data be removed from training sets. But these opt-out tools only apply to future models, not models trained before they rolled out; unlearning would be a much more thorough approach to data deletion.

Regardless, unlearning isn’t as easy as hitting “Delete.”

The art of forgetting

Unlearning techniques today rely on algorithms designed to “steer” models away from the data to be unlearned. The idea is to influence the model’s predictions so that it never — or only very rarely — outputs certain data.

To see how effective these unlearning algorithms could be, Shi and her collaborators devised a benchmark and selected eight different open algorithms to test. Called MUSE (Machine Unlearning Six-way Evaluation), the benchmark aims to probe an algorithm’s ability to not only prevent a model from spitting out training data verbatim (a phenomenon known as regurgitation), but eliminate the model’s knowledge of that data along with any evidence that it was originally trained on the data.

Scoring well on MUSE requires making a model forget two things: books from the Harry Potter series and news articles.

For example, given a snippet from Harry Potter and The Chamber of Secrets (“‘There’s more in the frying pan,’ said Aunt…”), MUSE tests whether an unlearned model can recite the whole sentence (“‘There’s more in the frying pan,’ said Aunt Petunia, turning eyes on her massive son”), answer questions about the scene (e.g. “What does Aunt Petunia tell her son?”, “More in the frying pan”) or otherwise indicate it’s been trained on text from the book.

MUSE also tests whether the model retained related general knowledge — e.g. that J.K. Rowling is the author of the Harry Potter series — after unlearning, which the researchers refer to as the model’s overall utility. The lower the utility, the more related knowledge the model lost, making the model less able to correctly answer questions.

In their study, the researchers found that the unlearning algorithms they tested did make models forget certain information. But they also hurt the models’ general question-answering capabilities, presenting a trade-off.

“Designing effective unlearning methods for models is challenging because knowledge is intricately entangled in the model,” Shi explained. “For instance, a model may be trained on copyrighted material — Harry Potter books as well as on freely available content from the Harry Potter Wiki. When existing unlearning methods attempt to remove the copyrighted Harry Potter books, they significantly impact the model’s knowledge about the Harry Potter Wiki, too.”

Are there any solutions to the problem? Not yet — and this highlights the need for additional research, Shi said.

For now, vendors betting on unlearning as a solution to their training data woes appear to be out of luck. Perhaps a technical breakthrough will make unlearning feasible someday. But for the time being, vendors will have to find another way to prevent their models from saying things they shouldn’t.

More TechCrunch

The pharma giant won’t say how many patients were affected by its February data breach. A count by TechCrunch confirms that over a million people are affected.

Pharma giant Cencora is alerting millions about its data breach

Self-driving technology company Aurora Innovation is looking to raise hundreds of millions in additional capital as it races toward a driverless commercial launch by the end of 2024.  Aurora is…

Self-driving truck startup Aurora Innovation to sell up to $420M in shares ahead of commercial launch

Payments infrastructure firm Infibeam Avenues has acquired a majority 54% stake in Rediff.com for up to $3 million, a dramatic twist of fate for the 28-year-old business that was the…

Rediff, once an internet pioneer in India, sells majority stake for $3M

The ruling confirmed an earlier decision in April from the High Court of Podgorica which rejected a request to extradite the crypto fugitive to the United States.

Terraform Labs co-founder and crypto fugitive Do Kwon set for extradition to South Korea

A day after Meta CEO Mark Zuckerberg talked about his newest social media experiment Threads reaching “almost” 200 million users on the company’s Q2 2024 earnings call, the platform has…

Meta’s Threads crosses 200 million active users

TechCrunch Disrupt 2024 will be in San Francisco on October 28–30, and we’re already excited! Disrupt brings innovation for every stage of your startup journey, and we could not bring you this…

Connect with Google Cloud, Aerospace, Qualcomm and more at Disrupt 2024

Featured Article

A comprehensive list of 2024 tech layoffs

The tech layoff wave is still going strong in 2024. Following significant workforce reductions in 2022 and 2023, this year has already seen 60,000 job cuts across 254 companies, according to independent layoffs tracker Layoffs.fyi. Companies like Tesla, Amazon, Google, TikTok, Snap and Microsoft have conducted sizable layoffs in the…

A comprehensive list of 2024 tech layoffs

Intel announced it would layoff more than 15% of its staff, or 15,000 employees, in a memo to employees on Thursday. The massive headcount is part of a large plan…

Intel to lay off 15,000 employees

Following the recent lawsuit filed by the Recording Industry Association of America (RIAA) against music generation startups Udio and Suno, Suno admitted in a court filing on Thursday that it did, in…

AI music startup Suno claims training model on copyrighted music is ‘fair use’

In spite of a drop for the quarter, iPhone remained Apple’s most important category by a wide margin.

iPad sales help bail out Apple amid a continued iPhone slide

Molly Alter wears a lot of hats. She’s a mocumentary filmmaker working on a project about an alternate reality where charades is big business. She’s a caesar salad connoisseur and…

How filming a cappella concerts and dance recitals led Northzone’s newest partner Molly Alter to a career in VC

Microsoft has a long and tangled history with OpenAI, having invested a reported $13 billion in the ChatGPT maker as part of a long-term partnership. As part of the deal,…

Microsoft now lists OpenAI as a competitor in AI and search

The San Jose-based startup raised $60 million in a round that values it lower than the $500 million valuation it garnered in its most recent round, according to multiple sources.

Sequoia-backed Knowde raises Series C at a valuation cut

X (formerly Twitter) can no longer be accessed in the Mac App Store, suggesting that it has been officially delisted.  Searches for both “Twitter” and “X” on Apple’s platform no…

Twitter disappears from Mac App Store

Google Thursday said that it is introducing new Gemini-powered features for Chrome’s desktop version, including Lens for desktop, tab compare for shopping assistance, and natural language integration for search history.…

Google brings Gemini-powered search history and Lens to Chrome desktop

When Xiaoyin Qu was growing up in China, she was obsessed with learning how to build paper airplanes that could do flips in the air. Her parents, though, didn’t have…

Heeyo built an AI chatbot to be a billion kids’ interactive tutor and friend

While the company was awarded a massive, $4.2 billion contract to accelerate Starliner development in 2014, it was structured as a “fixed-price” model.

Boeing bleeds another $125M on Starliner program, bringing total losses to $1.6B

Welcome back to TechCrunch Mobility — your central hub for news and insights on the future of transportation. Sign up here for free — just click TechCrunch Mobility! Summer road…

Anthony Levandowski bets on off-road autonomy, Nuro plots a comeback and Applied Intuition gets more investor love

Google’s new features include Gemini in BigQuery and Looker to help users with data engineering and analysis.

Google Cloud expands its database portfolio with new AI capabilities

Rad Power Bikes, the Seattle-based e-bike startup that has raised more than $300 million from investors, went through another round of layoffs in July, TechCrunch has exclusively learned. This is…

VC darling Rad Power Bikes hit with another round of layoffs

Five years ago, as robotaxis and self-driving truck startups were still raking in millions in venture capital, Anthony Levandowski turned to off-road autonomy. Now, that decision — which brought the…

Why Anthony Levandowski returned to his off-road autonomous vehicle roots with AV startup Pronto

Commercial space station company Vast is building a private microgravity research lab as part of its wider Haven-1 station plans. The module is set to launch no earlier than the…

Vast plans microgravity lab on its Haven-1 private space station

Google Cloud is giving Y Combinator startups access to a dedicated, subsidized cluster of Nvidia graphics processing units and Google tensor processing units to build AI models. It’s part of…

Google Cloud now has a dedicated cluster of Nvidia GPUs for Y Combinator startups

StackShare is one of the more popular platforms for developers to discuss, track, and share the tools they use to build applications.

Open source startup FOSSA is buying StackShare, a site used by 1.5M developers

Featured Article

Indian startups gut valuations ahead of IPO push

Ola Electric and FirstCry are set to test investor appetite with public listing, both pricing their shares below their previous valuation asks.

Indian startups gut valuations ahead of IPO push

The European Union’s risk-based regulation for applications of artificial intelligence has come into force starting from today.

The EU’s AI Act is now in force

The company also said it has received regulatory clearance to start Phase 2 clinical trials for a new drug in the U.S. later this year.

Healx, an AI-enabled drug discovery platform for rare diseases, raises $47M

The European Commission (EC) has given the go-ahead to HPE’s planned megabucks acquisition of Juniper Networks.

EU greenlights HPE’s $14B Juniper Networks acquisition

Meta, which develops one of the biggest foundational open source large language models, Llama, believes it will need significantly more computing power to train models in the future. Mark Zuckerberg…

Zuckerberg says Meta will need 10x more computing power to train Llama 4 than Llama 3

Axle Energy is a B2B, back-end infrastructure business focused on connecting flexible assets, such as electric vehicles and home batteries, to energy markets that aren’t otherwise available for consumers to…

Axle Energy’s sprint to decarbonize the grid lights up with $9M seed led by Accel