web analytics

Learn AI With Kesse | Newest Trends in Artificial Intelligence

We answer questions about artificial intelligence and bring you what's new in the AI World.

Many Safety Evaluations for AI Models Have Significant Limitations

3 min read

Despite increasing demand for AI safety and accountability, current tests and benchmarks for AI models may fall short, according to a new report.

generative AI models are under increased scrutiny for their tendency to make mistakes and behave unpredictably. Organizations propose new benchmarks to test these models’ safety.

Benchmarks and Red Teaming

The study’s authors first surveyed academic literature to understand the current state of AI model evaluations. They interviewed 16 experts, including employees from tech companies developing generative AI systems. Their findings showed sharp disagreements within the industry on the best methods to evaluate models.

Some tests only examined how models aligned with benchmarks in the lab, not how they would impact real-world users. Others used tests developed for research, not evaluating production models. However, vendors still insisted on using these in production.

Experts noted it’s tough to determine a model’s performance from benchmark results. For instance, a model may perform well on a state bar exam but fail to solve real-world legal challenges. Data contamination is also a significant issue, where benchmark results can overestimate a model’s performance if trained on the same data used for testing.

Benchmarks are often chosen for convenience rather than effectiveness. “Benchmarks risk being manipulated,” said Mahi Hardalupas, researcher at the ALI. Developers might train models on the same data set used for evaluation, similar to seeing the exam paper before the test. Small changes can also cause unpredictable behavior changes and may override safety features.

The ALI study found issues with ‘red-teaming,’ where individuals or groups try to find vulnerabilities in a model. Many companies, including OpenAI and Anthropic, use red-teaming. However, few standards exist for red-teaming, making it hard to assess its effectiveness. Finding people with the necessary skills and expertise is also challenging, and the manual nature of red-teaming makes it costly and laborious.

Possible Solutions

One of the main reasons AI evaluations haven’t improved is the pressure to release models quickly. “A person we spoke with developing foundation models felt there was more pressure to release models quickly, making it harder to take evaluations seriously,” said Jones.

Major AI labs are releasing models faster than society can ensure they are safe and reliable. Evaluating models for safety is considered an ‘intractable’ problem by some experts.

Mahi Hardalupas believes there is a path forward but says it requires more engagement from public-sector bodies. “Regulators and policymakers must clearly articulate what they want from evaluations,” he said. The evaluation community must also be transparent about the current limitations and potential of evaluations.

Hardalupas suggests governments mandate more public participation in developing evaluations. He also recommends supporting an ecosystem of third-party tests, including programs to ensure regular access to required models and data sets. Jones thinks developing ‘context-specific’ evaluations that look at the types of users a model might impact and how attacks could defeat safeguards is necessary.

Behind those recommendations, there’s a call for more investment in the underlying science of evaluations. This is to develop more robust and repeatable evaluations based on understanding how an AI model operates.

However, there may never be a guarantee that a model is entirely safe. According to Hardalupas, ‘safety’ is not a model’s inherent property. It depends on the context in which it is used, who accesses it, and if the safeguards are adequate and robust.

Evaluations can help identify potential risks but cannot guarantee a model is perfectly safe. Many experts agree that evaluations can only indicate a model is unsafe, not prove it is safe.


In summary, while AI evaluations are essential, they currently have significant limitations. These tests can indicate potential risks but can’t guarantee complete safety.

Moving forward, greater involvement from public-sector bodies, better transparency, and context-specific evaluations can help improve the reliability and safety of AI models.

About The Author

Leave a Reply

Your email address will not be published. Required fields are marked *

We use cookies to personalize content and ads and to primarily analyze our geo traffic sources. We also may share information about your use of our site with our social media, advertising, and analytics partners to improve your user experience. We respect your privacy and will never abuse your information. [ Privacy Policy ] View more
Cookies settings
Accept
Decline
Privacy & Cookie Policy
Privacy & Cookies policy
Cookie name Active

The content on this page governs our Privacy Policy. It describes how your personal information is collected, used, and shared when you visit or make a purchase from learnaiwithkesse.com (the "Site").

Kesseswebsites and Advertising owns Learn AI With Kesse and the website learnaiwithkesse.wiki. For the purpose of this Terms and Agreements [ we, us, I, our ] represents the owner of Learning AI With Kesse which is Kesseswebsites and Advertising. [ You, your, student and buyer ] represents you as the user and visitor of this site. Terms of Conditions, Terms of Service, Terms and Agreement and Terms of use shall be considered the same here. This website or site refers to https://learnaiwithkesse.com. You agree that the content of this Terms and Agreement may include Privacy Policy and Refund Policy. Products refer to physical or digital products. This includes eBooks, PDFs, and text or video courses. If there is anything on this page you do not understand you agree to reach out to us via email [ emmanuel@learnaiwithkesse.com ] for explanation before using any part of this site.

1. Personal Information We Collect

When you visit this Site, we automatically collect certain information about your device, including information about your web browser, IP address, time zone, and some of the cookies that are installed on your device. The primary purpose of this activity is to provide you a better user experience the next time you visit our again and also the data collection is for analytics study. Additionally, as you browse the Site, we collect information about the individual web pages or products that you view, what websites or search terms referred you to the Site, and information about how you interact with the Site. We refer to this automatically-collected information as "Device Information."

We collect Device Information using the following technologies:

"Cookies" are data files that are placed on your device or computer and often include an anonymous unique identifier. For more information about cookies, and how to disable cookies, visit http://www.allaboutcookies.org. To comply with European Union's GDPR (General Data Protection Regulation), we do display a disclaimer a consent text at the bottom of this website. This disclaimer alerts you the visitor or user of this website about why we use cookies, and we also give you the option to accept or decline. If you accept for us to use cookies on your site, the agreement between you and us will expire after 180 has passed.

"Log files" track actions occurring on the Site, and collect data including your IP address, browser type, Internet service provider, referring/exit pages, and date/time stamps.

"Web beacons," "tags," and "pixels" are electronic files used to record information about how you browse the Site.

Additionally, when you make a purchase or attempt to make a purchase through the Site, we collect certain information from you, including your name, billing address, shipping address, payment information (including credit card numbers), email address, and phone number. We refer to this information as "Order Information."

When we talk about "Personal Information" in this Privacy Policy, we are talking both about Device Information and Order Information.

Payment Information

Please note that we use 3rd party payment processing companies like https://stripe.com and https://paypal.com to process your payment information. PayPal and Stripe protects your data according to their terms and agreement and may store your data to help make your subsequent transactions on this website easier. We never and [ DO NOT ] store your card information or payment login information on our website or server. By making payment on our site, you agree to abide by the Terms and Agreement of the 3rd Party payment processing companies we use. You can visit their websites to read their Terms of Use and learn more about them.

2. How Do We Use Your Personal Information?

We use the Order Information that we collect generally to fulfill any orders placed through the Site (including processing your payment information, arranging for shipping, and providing you with invoices and/or order confirmations). Additionally, we use this [a] Order Information to:

[b] Communicate with you;

[c] Screen our orders for potential risk or fraud; and

When in line with the preferences you have shared with us, provide you with information or advertising relating to our products or services. We use the Device Information that we collect to help us screen for potential risk and fraud (in particular, your IP address), and more generally to improve and optimize our Site (for example, by generating analytics about how our customers browse and interact with the Site, and to assess the success of our marketing and advertising campaigns).

3. Sharing Your Personal Information

We share your Personal Information with third parties to help us use your Personal Information, as described above. For example, we use System.io to power our online store--you can read more about how Systeme.io uses your Personal Information here: https://systeme.io/privacy-policy/ . We may also use Google Analytics to help us understand how our customers use the Site--you can read more about how Google uses your Personal Information here: https://www.google.com/intl/en/policies/privacy/. You can also opt-out of Google Analytics here: https://tools.google.com/dlpage/gaoptout.

Finally, we may also share your Personal Information to comply with applicable laws and regulations, to respond to a subpoena, search warrant or other lawful request for information we receive, or to otherwise protect our rights.

4. Behavioral Advertising

As described above, we use your Personal Information to provide you with targeted advertisements or marketing communications we believe may be of interest to you. For more information about how targeted advertising works, you can visit the Network Advertising Initiative’s (“NAI”) educational page at http://www.networkadvertising.org/understanding-online-advertising/how-does-it-work.

You can opt-out of targeted advertising by:

COMMON LINKS INCLUDE:

FACEBOOK - https://www.facebook.com/settings/?tab=ads

GOOGLE - https://www.google.com/settings/ads/anonymous

BING - https://advertise.bingads.microsoft.com/en-us/resources/policies/personalized-ads]

Additionally, you can opt-out of some of these services by visiting the Digital Advertising Alliance’s opt-out portal at: http://optout.aboutads.info/.

5. Data Retention

Besides your card payment and payment login information, when you place an order through the Site, we will maintain your Order Information for our records unless and until you ask us to delete this information. Example of such information include your first name, last name, email and phone number.

6. Changes

We may update this privacy policy from time to time in order to reflect, for example, changes to our practices or for other operational, legal or regulatory reasons.

7. Contact Us

For more information about our privacy practices, if you have questions, or if you would like to make a complaint, please contact us by e-mail at emmanuel@learnaiwithkesse.com or by mail using the details provided below:

8. Your acceptance of these terms

By using this Site, you signify your acceptance of this policy. If you do not agree to this policy, please do not use our Site. Your continued use of the Site following the posting of changes to this policy will be deemed your acceptance of those changes.

Last Update | 18th August 2024

Save settings
Cookies settings