Apple delays European launch of new AI features...

Plus: Experts develop a "BS detector" to fix AI's biggest flaw

June 21, 2024

Today's edition is a 3-minute read!

Here are the top 3 stories in AI…

Presented by BrainStation

Unlock the potential of generative AI with our hands-on, expert-led course starting on June 25th. Learn from industry leaders and gain essential skills in AI foundations, generative AI, and business applications. Dive into real-world case studies and practical exercises to master AI tools like ChatGPT.

Ready to elevate your career?

Experts develop a "BS detector" to fix AI's biggest flaw

A new algorithm aims to tackle AI's problem of confident but incorrect answers, adding a dose of humility to the mix.

AI errors are risky, especially when people rely on chatbots for crucial info like medical advice or legal precedents.

The big picture: AI models frequently make various mistakes, often lumped together as "hallucinations." Sebastian Farquhar from Oxford argues this term has become meaningless due to its broad use.

The innovation: Farquhar and his Oxford colleagues have created a new method to detect "arbitrary and incorrect answers," or confabulations, as detailed in Nature.

How it works:

  • Multiple queries: Ask the chatbot the same question several times.

  • Response grouping: Another AI groups responses by meaning.

  • Semantic entropy: Calculate the consistency of responses to spot confabulations.

Findings: This method detects confabulations 79% of the time, better than other techniques based on word similarity.

Limitations:

  • Only detects inconsistent errors, not those from biased or erroneous data.

  • Requires more computing power than typical chatbot interactions.

Expert opinions: "Detecting confabulations is a step forward, but caution is still needed," says Jenn Wortman Vaughan from Microsoft Research. She highlights the importance of AI systems expressing uncertainty to help users trust responses.

Future directions: Vaughan and her team are exploring how AI can better express uncertainty. Research shows people are less likely to trust AI responses when the system explicitly communicates doubt.

Real-world impact: Top chatbots from OpenAI, Meta, Google, and others still "hallucinate" at rates between 2.5% and 5%. While newer versions reduce some errors, the problem persists in new forms.

Balancing accuracy with creative expression is key. For factual precision, reducing hallucinations is crucial, but in creative contexts, these "hallucinations" might be desired.

Hackers jailbreak AI models: expose flaws in ChatGPT, Google, and xAI

Global effort to expose AI vulnerabilities: Pliny the Prompter, a pseudonymous hacker, can crack the world's most powerful AI models in about 30 minutes. From making Meta’s Llama 3 share napalm recipes to making Elon Musk’s Grok praise Adolf Hitler, Pliny's "jailbreaking" efforts highlight the shortcomings of rushed AI releases by major tech companies.

Pliny and other ethical hackers, researchers, and cybersecurity experts are revealing the dangerous potentials of large language models (LLMs) from OpenAI, Meta, Google, and xAI by bypassing their "guardrails." These hacks have demonstrated how AI can produce harmful content, spread misinformation, and leak private data.

Growing market for AI security: The constant evolution of jailbreaking techniques has spurred a burgeoning market for LLM security start-ups, raising $213 million in 2023 alone. Cybersecurity firms like CyberArk are now offering LLM security solutions to counter increasingly sophisticated attacks.

Regulatory concerns and malicious exploits: As AI models face potential regulatory crackdowns globally, including the EU's AI Act and California's proposed AI safety bill, ethical hackers continue to expose vulnerabilities. Meanwhile, malicious hackers sell modified LLMs like WormGPT and FraudGPT on the dark web to facilitate cyber attacks and phishing campaigns.

The future of AI security: Leading AI developers, including OpenAI, Meta, and Google, are investing in better defenses against these exploits. However, experts warn that as AI models become more integrated with technology and devices, the risks will escalate.

"In general, companies are not prepared," says Rony Ohayon, CEO of DeepKeep, an Israeli LLM security firm developing tools to protect against these sophisticated threats. As the capabilities of AI models grow, so does the importance of securing them against exploitation and misuse.

Apple delays European launch of new AI features...

Apple has announced that the rollout of its new AI features for the iPhone will be delayed in Europe, citing "regulatory uncertainties" related to the EU's new competition laws.

Why it matters: The flagship features, branded as "Apple Intelligence," are crucial for driving iPhone upgrades and include a partnership with OpenAI. However, the complexities of complying with the EU's Digital Markets Act (DMA) have caused Apple to halt their European launch.

Details:

  • Apple unveiled these AI advancements two weeks ago, highlighting them as a significant step forward.

  • The delayed features include iPhone Mirroring, SharePlay Screen Sharing enhancements, and the Apple Intelligence suite.

  • The company cited difficulties in ensuring these systems meet new EU requirements for interoperability with third-party services.

Backdrop: This decision follows a report from the Financial Times indicating that Brussels regulators are preparing to accuse Apple of breaching DMA rules. The EU's ongoing investigation into Apple's competitive practices has added pressure on the tech giant.

Apple's stance: "Due to the regulatory uncertainties brought about by the Digital Markets Act, we do not believe that we will be able to roll out three of these new features to our EU users this year," Apple stated.

What's next: The regulatory standoff between Apple and Brussels is expected to continue, potentially influencing how and when new tech features are introduced in the European market.