Technology

Many AI model safety assessments have significant limitations

Published

7 months ago

August 5, 2024

IAM

Many safety evaluations for AI models have significant limitations

Despite the growing demand for AI security and accountability, today’s tests and benchmarks will not be enough, a brand new report finds.

Generative AI models—models that may analyze and generate text, images, music, video, and more—are coming under increasing scrutiny for his or her tendency to make mistakes and usually behave unpredictably. Now, organizations from public sector agencies to big tech firms are proposing recent benchmarks to check the safety of those models.

At the tip of last yr, the startup Scale AI created lab dedicated to assessing how well models adhere to security guidelines. This month, NIST and the U.K. AI Safety Institute released tools designed to evaluate model risk.

However, these tests and model testing methods could also be insufficient.

The Ada Lovelace Institute (ALI), a British non-profit organization dedicated to artificial intelligence research, conducted test who interviewed experts from academic, civil society, and vendor modeling labs and examined recent research on AI security assessments. The co-authors found that while current assessments will be useful, they should not comprehensive, will be easily fooled, and don’t necessarily provide guidance on how models will perform in real-world scenarios.

“Whether it’s a smartphone, a prescription drug, or a car, we expect the products we use to be safe and reliable; in these sectors, products are rigorously tested to ensure they’re safe before being deployed,” Elliot Jones, a senior researcher at ALI and co-author of the report, told TechCrunch. “Our research aimed to examine the limitations of current approaches to assessing AI safety, assess how assessments are currently being used, and explore their use as a tool for policymakers and regulators.”

Benchmarks and red teaming

The study’s co-authors first surveyed the tutorial literature to determine an summary of the harms and risks that current models pose and the state of existing assessments of AI models. They then interviewed 16 experts, including 4 employees of unnamed technology firms developing generative AI systems.

The study revealed that there’s wide disagreement across the AI industry on the perfect set of methods and taxonomies for evaluating models.

Some evaluations only tested how well the models matched benchmarks within the lab, not how the models might impact real-world users. Others were based on tests designed for research purposes, not on evaluating production models—yet vendors insisted on using them in production.

We’ve written before concerning the problems with AI benchmarking. This study highlights all of those issues and more.

Experts cited within the study noted that it’s hard to extrapolate a model’s performance from benchmark results, and it’s unclear whether benchmarks may even show that a model has a certain capability. For example, while a model may perform well on a state exam, that doesn’t mean it can have the ability to resolve more open legal challenges.

Experts also pointed to the issue of knowledge contamination, where benchmark results can overstate a model’s performance if it was trained on the identical data it’s being tested on. Benchmarks, in lots of cases, are chosen by organizations not because they’re the perfect assessment tools, but due to their convenience and ease of use, experts said.

“Benchmarks run the risk of being manipulated by developers who may train models on the same dataset that will be used to evaluate the model, which is equivalent to looking at an exam paper before an exam or strategically choosing which assessments to use,” Mahi Hardalupas, a researcher at ALI and co-author of the study, told TechCrunch. “Which version of the model is being evaluated also matters. Small changes can cause unpredictable changes in behavior and can override built-in safety features.”

The ALI study also found problems with “red-teaming,” the practice of getting individuals or groups “attack” a model to discover gaps and flaws. Many firms use red-teaming to judge models, including AI startups OpenAI and Anthropic, but there are few agreed-upon standards for red-teaming, making it difficult to evaluate the effectiveness of a given effort.

Experts told the study’s co-authors that finding individuals with the correct skills and experience to steer red teaming efforts will be difficult, and the manual nature of the method makes it expensive and labor-intensive, a barrier for smaller organizations that don’t have the mandatory resources.

Possible solutions

The foremost the reason why AI rankings have not improved are the pressure to release models faster and the reluctance to run tests that might cause issues before launch.

“The person we spoke to who works for a foundation modeling company felt that there is more pressure within companies to release models quickly, which makes it harder to push back and take assessments seriously,” Jones said. “The major AI labs are releasing models at a speed that outpaces their ability or society’s ability to ensure they are safe and reliable.”

One ALI survey respondent called evaluating models for safety an “intractable” problem. So what hopes does the industry—and those that regulate it—have for solutions?

Mahi Hardalupas, a researcher at ALI, believes there’s a way forward, but it can require greater commitment from public sector entities.

“Regulators and policymakers need to be clear about what they expect from ratings,” he said. “At the same time, the ratings community needs to be transparent about the current limitations and potential of ratings.”

Hardalupas suggests that governments mandate greater public participation in the event of assessments and implement measures to support an “ecosystem” of third-party testing, including programs to offer regular access to any required models and datasets.

Jones believes it could be mandatory to develop “context-aware” assessments that transcend simply testing a model’s response to a command, and as an alternative consider the sorts of users a model might affect (akin to people of a certain background, gender, or ethnicity), in addition to the ways wherein attacks on models could bypass security measures.

“This will require investment in fundamental evaluation science to develop more robust and repeatable evaluations based on an understanding of how the AI model works,” she added.

However, there’s never a guarantee that a model is protected.

“As others have noted, ‘safety’ is not a property of models,” Hardalupas said. “Determining whether a model is ‘safe’ requires understanding the contexts in which it is used, to whom it is sold or shared, and whether the safeguards that are implemented are appropriate and robust to mitigate those risks. Baseline model assessments can serve exploratory purposes to identify potential risks, but they cannot guarantee that the model is safe, much less ‘completely safe.’ Many of our interviewees agreed that assessments cannot prove that a model is safe and can only indicate that the model is unsafe.”

This article was originally published on : techcrunch.com

Related Topics:Ada Lovelace Institute AI Generative AI research studies

Up Next

TikTok Lite: EU closes addictive design case after TikTok pledges not to reinstate rewards mechanism

Don't Miss

From golf to hunting, a new group of startups wants to make the experience even better

Click to comment

Technology

The signal is the number one application in the Netherlands. But why?

Published

23 hours ago

March 3, 2025

IAM

The application signal for sending a privacy -oriented message flew high in Dutch application stores last month, often sitting at the top as the most steadily downloaded free application for iOS and Android in all categories, for data from many application tracking platforms akin to the sensor tower.

The application has experienced popularity over the years, often in response to Changes in politics in rivals akin to WhatsApp Or Geopolitical events. This is because Signal has set a reputation as a more friendly privacy option-it is served by the non-profit foundation (though based in the USA), not a personal company focused on data earning data. In addition, the signal tracks minimal metadata.

In 2025, along with the recent US president, who strengthened the warm Big Tech hug, it is not surprising that digital privacy tools have a moment – especially in Europe, which attracted the anger of President Trump.

But this time, the meaning of the signal in one very specific place-Holandia is particularly eye-catching.

Signal data from the sensor towerImage loans:Sensor tower / screenshot

IN Interview with Dutch newspaper de Telelegraaf last week, President signal Meredith Whittaker He noticed that the number of “new registrations” in the Netherlands was 25 this 12 months, even though it is not clear what the exact comparative period for this data is.

Asked why the Netherlands recorded such development, Whittaker pointed to the combination of things: “growing awareness of privacy, distrust of large technology and political reality in which people realize how sensitive digital communication can be,” said Whittaker.

Data provided to TechCrunch from the application intelligence company Appfigures Increase in Signal Signal in the Netherlands. According to its data, the signal was 365. Among the applications apart from the iPhone in the Netherlands on January 1 and didn’t appear on the list of the most significant general applications. Then, from around January 5, he began to climb the rankings, reaching the highest position until February 2.

The signal immersed and comes out of the lead during weeks, spending around mid -February at the top – including every single day from February 22. By digging deeper into the data, the AppFigures estimates that the total download in Apple and Google Applets in total in December 2024 jumped to 99,000 in January and increased to 233,000 to February – 958%.

While a part of this height could be assigned to a lower saturation signal than other markets, a continuing application position at the top in comparison with neighboring markets of comparable size.

“No other markets are approaching the Netherlands in terms of growth between December and February,” said AppFigures Techcrunch.

For comparison, from December in Belgium, download increased by over 250%, Sweden by 153%and dishes by 95%.

So why the signal can experience what one redditor called “The moment of mass adoption“In the Netherlands?

Clear signal

Give ZengerSenior Policy Advisor at Dutch Digital Rights Foundation Fragments of freedomHe said that even though it is difficult to point one specific reason, he is not surprised.

The last changes in the US have seen Large platform suppliers Adapt with the recent Trump administration, and this has retained a major public and media debate. Relying Europe from the technology of big private American corporations has turn out to be the point of interest of this debate.

“The Dutch are, like many others, very dependent on the infrastructure provided by extremely dominant technology companies, mainly from the USA,” said Zenger. “What does this mean, and the risk that results from it has been nicely demonstrated in the last few weeks. As a result, the public debate in the Netherlands was relatively sharp. Where in the past this problem was discussed only at the level “:” I feel that we are now conducting a debate at the higher levels: “.

In this context, society can mix dominance with data protection abuse. Since corporations akin to meta are frequently studied and fined in the field of information privacy practices, the signal could appear to be less evil: it is based on the US, but supported by a non-profit organization, which ensures encryption of each the content of the message and around it.

Vincent BöhreDirector of the Dutch Organization of Privacy Privacy firstHe also pointed to increased media relationships and a wider change of public opinion.

“Since a few months ago he was re-elected in the United States, in the Dutch-and European media, which seem to support Trump, there were many” Elon) Muska. “Articles criticizing X (previously Twitter) and Meta appear everywhere in the Dutch media, which leads to a change in Dutch public opinion: even people who have never really known or cared for privacy and security in social media, suddenly became interested in” friendly privacy “alternative, in particular the signal.”

Signal of intentions

President of the application for sending a signal message Meredith Whittaker. — President Signa Meredith Whittaker on Web Summit, in Lisbon on November 4, 2022.Image loans:Patricia de Melo Moreira / AFP / Getty Images

While the Netherlands is only one market of 18 million people in the European population over 700 million, its increase in adoption can signal a wider trend throughout the continent, especially when governments try to cut back privacy barriers.

For example, Apple has recently pulled out comprehensive encryption from iCloud in Great Britain to counteract government efforts to put in a backdoor.

Speech Fr. Rightcon 25 In Taiwan, this week, Whittaker confirmed the unwavering Signal attitude regarding privacy.

“Signal position on this subject is very clear- we will not walk, falsify or otherwise disturb the solid guarantees of privacy and security that people rely on” Said Whittaker. “Regardless of whether this disturbance or backdoor is called scanning on the client’s side or removing the protection of encryption against one or the other, the features similar to what Apple has been forced to do in Great Britain”

Separately, in Interview with Swedish public broadcaster, Whittaker said that Signal wouldn’t follow the proposed Swedish law requiring application to send messages for storage.

“In practice, this means asking us to break encryption, which is the basis of our entire activity,” said Whittaker. “Asking us to store data would undermine all our architecture and we would never do it. We would prefer to completely leave the Swedish market. “

TechCrunch contacted to signal a comment, but he didn’t hear during the publication.

(Tagstotranslat) signal of the Netherlands

This article was originally published on : techcrunch.com

Technology

Gayle King announces participation in the space mission of all women

Published

2 days ago

March 2, 2025

Gayle King, CBS News, new deal, morning show

Gayle King will join the thirty first Blue Origin civil flight into space.

Gayle King announced that he was going to space. The host of the talk show during the day provided messages CBS MORNINGS.

King revealed Her participation in the thirty first Blue Origin flights, NS-31. Before discussing the details of the mission, she and her co -lecturers presented the video editing, which described her long -term fascination with travel travel.

In one clip, King said: “I am excited to watch the premiere at home in my pajamas.”

Her enthusiasm led to an invite with Blue Origin. The television personality will disappear from Crew from the whole familyIncluding an award -winning journalist Lauren Sánchez, award -winning Grammy singer Katy Perry and astronaut Aish Bowe.

Soon the explorer of the space admitted that she was hesitating at first.

“I don’t know how to explain at the same time terrified and excited,” said King.

To make a choice, King turned to a gaggle of family members, including her children and a detailed friend, Oprah Winfrey. She said that when her most trusted confidants approved, she was ready.

“When Kirby, Will and Oprah were fine, I was fine,” said King. “I thought Oprah would say no. She said: “I feel that when you don’t do it, if you all come back and also you had the opportunity to do it, you’ll kick.” She is right. “

King is not going to be the first television host who wandered into space with blue origin. In 2021, then-Good morning America Coheat Michael Strahan took part in the third civil flight Blue Origin. The former NFL star and the sender was delighted after returning, expressing how this experience gave him a brand new “perspective” in the world.

“I want to come back,” said Strahan.

Blue origin, Founded by Amazon Billionaire Jeff Bezos in 2000 is a non-public aviation company that focuses on sharing space travels for civilians and developing technology to explore the space long.

The upcoming flight of the king New Shepard It will probably be part of Blue Origin’s constant efforts to normalize civil space travel.

This article was originally published on : www.blackenterprise.com

Technology

Instagram can turn the rollers in a separate application

Published

5 days ago

February 27, 2025

IAM

Meta is occupied with an independent application for brief movies, Information He informed, citing an anonymous source, which he heard the boss on Instagram Adam Mosseri talked about the personnel project.

The project is reportedly called RAY code, which goals to enhance recommendations for brand new users and existing users in the US and to conclude one other three minutes of movies, the report quoted the source.

The finish line didn’t answer immediately at the request for comment.

Last month, the company announced a video editing application called Edyta to compete with Capcut (belonging to Tiktok Matter Company Bytedance) since it was geared toward using the uncertain future Tiktok and Bytedance in the USA

Currently, the Instagram channel is a mixture of photos, movies (drums) and stories. However, many users imagine that the application has been cluttered since it incorporates movies and not persist with the roots as an application for sharing photos. If the company rotates in an independent application for brief movies, it can create a possibility for Instagram to emphasise other functions.

Instagram began at the starting of this yr paying creators To promote Instagram on other platforms, resembling Tiktok, Snapchat and YouTube. Apparently he also began to supply Big money for the creators Present only on roller skates.

(Tagstranslate) Instagram

This article was originally published on : techcrunch.com