AI Security

OpenAI’s GPT-4 Shows Higher Trustworthiness but Vulnerabilities to Jailbreaking and Bias, Research Finds

Published

2 months ago

August 8, 2024

James

OpenAI's GPT-4 Shows Higher Trustworthiness but Vulnerabilities to Jailbreaking and Bias, Research Finds 33

New research, in partnership with Microsoft, has revealed that OpenAI’s GPT-4 large language model is considered more dependable than its predecessor, GPT-3.5. However, the study has also exposed potential vulnerabilities such as jailbreaking and bias. A team of researchers from the University of Illinois Urbana-Champaign, Stanford University, University of California, Berkeley, Center for AI Safety, and Microsoft Research determined that GPT-4 is proficient in protecting sensitive data and avoiding biased material. Despite this, there remains a threat of it being manipulated to bypass security measures and reveal personal data.

OpenAIs GPT-4 Shows Higher Trustworthiness but Vulnerabilities to Jailbreaking and Bias, Research Finds

Table of Contents

Trustworthiness Assessment and Vulnerabilities

The researchers conducted a trustworthiness assessment of GPT-4, measuring results in categories such as toxicity, stereotypes, privacy, machine ethics, fairness, and resistance to adversarial tests. GPT-4 received a higher trustworthiness score compared to GPT-3.5. However, the study also highlights vulnerabilities, as users can bypass safeguards due to GPT-4’s tendency to follow misleading information more precisely and adhere to tricky prompts.

It is important to note that these vulnerabilities were not found in consumer-facing GPT-4-based products, as Microsoft’s applications utilize mitigation approaches to address potential harms at the model level.

Testing and Findings

The researchers conducted tests using standard prompts and prompts designed to push GPT-4 to break content policy restrictions without outward bias. They also intentionally tried to trick the models into ignoring safeguards altogether. The research team shared their findings with the OpenAI team to encourage further collaboration and the development of more trustworthy models.

OpenAI's GPT-4 Shows Higher Trustworthiness but Vulnerabilities to Jailbreaking and Bias, Research Finds 35

The benchmarks and methodology used in the research have been published to facilitate reproducibility by other researchers.

Red Teaming and OpenAI’s Response

AI models like GPT-4 often undergo red teaming, where developers test various prompts to identify potential undesirable outcomes. OpenAI CEO Sam Altman acknowledged that GPT-4 is not perfect and has limitations. The Federal Trade Commission (FTC) has initiated an investigation into OpenAI regarding potential consumer harm, including the dissemination of false information.

James

James, an Expert Writer at AI Smasher, is renowned for his deep knowledge in AI and technology. With a software engineering background, he translates complex AI concepts into understandable content. Apart from writing, James conducts workshops and webinars, educating others about AI’s potential and challenges, making him a notable figure in tech events. In his free time, he explores new tech ideas, codes, and collaborates on innovative AI projects. James welcomes inquiries.

AI Security

Report Finds Top AI Developers Lack Transparency in Disclosing Societal Impact

Published

2 months ago

August 9, 2024

James

Report Finds Top AI Developers Lack Transparency in Disclosing Societal Impact

Table of Contents

Stanford HAI Releases Foundation Model Transparency Index

A new report released by Stanford HAI (Human-Centered Artificial Intelligence) suggests that leading developers of AI base models, like OpenAI and Meta, are not effectively disclosing information regarding the potential societal effects of their models. The Foundation Model Transparency Index, unveiled today by Stanford HAI, evaluated the transparency measures taken by the makers of the top 10 AI models. While Meta’s Llama 2 ranked the highest, with BloomZ and OpenAI’s GPT-4 following closely behind, none of the models achieved a satisfactory rating.

Transparency Defined and Evaluated

The researchers at Stanford HAI used 100 indicators to define transparency and assess the disclosure practices of the model creators. They examined publicly available information about the models, focusing on how they are built, how they work, and how people use them. The evaluation considered whether companies disclosed partners and third-party developers, whether customers were informed about the use of private information, and other relevant factors.

Top Performers and their Scores

Meta scored 53 percent, receiving the highest score in terms of model basics as the company released its research on model creation. BloomZ, an open-source model, closely followed at 50 percent, and GPT-4 scored 47 percent. Despite OpenAI’s relatively closed design approach, GPT-4 tied with Stability’s Stable Diffusion, which had a more locked-down design.

OpenAI’s Disclosure Challenges

OpenAI, known for its reluctance to release research and disclose data sources, still managed to rank high due to the abundance of available information about its partners. The company collaborates with various companies that integrate GPT-4 into their products, resulting in a wealth of publicly available details. Report Finds Top AI Developers Lack Transparency in Disclosing Societal Impact 42

Creators Silent on Societal Impact

However, the Stanford researchers found that none of the creators of the evaluated models disclosed any information about the societal impact of their models. There is no mention of where to direct privacy, copyright, or bias complaints.

Index Aims to Encourage Transparency

Rishi Bommasani, a society lead at the Stanford Center for Research on Foundation Models and one of the researchers involved in the index, explains that the goal is to provide a benchmark for governments and companies. Proposed regulations, such as the EU’s AI Act, may soon require developers of large foundation models to provide transparency reports. The index aims to make models more transparent by breaking down the concept into measurable factors. The group focused on evaluating one model per company to facilitate comparisons.

OpenAI’s Research Distribution Policy

OpenAI, despite its name, no longer shares its research or codes publicly, citing concerns about competitiveness and safety. This approach contrasts with the large and vocal open-source community within the generative AI field.

The Verge reached out to Meta, OpenAI, Stability, Google, and Anthropic for comments but has not received a response yet.

Potential Expansion of the Index

Bommasani states that the group is open to expanding the scope of the index in the future. However, for now, they will focus on the 10 foundation models that have already been evaluated.

James

AI Security

Coding help forum Stack Overflow lays off 28% of staff as it faces profitability challenges

Published

2 months ago

August 8, 2024

James

Coding help forum Stack Overflow lays off 28% of staff as it faces profitability challenges 47

Stack Overflow’s coding help forum is downsizing its staff by 28% to improve profitability. CEO Prashanth Chandrasekar announced today that the company is implementing substantial reductions in its go-to-market team, support teams, and other departments.

Table of Contents

Scaling up, then scaling back

Last year, Stack Overflow doubled its employee base, but now it is scaling back. Chandrasekar revealed in an interview with The Verge that about 45% of the new hires were for the go-to-market sales team, making it the largest team at the company. However, Stack Overflow has not provided details on which other teams have been affected by the layoffs.

Challenges in the era of AI

The decision to downsize comes at a time when the tech industry is experiencing a boom in generative AI, which has led to the integration of AI-powered chatbots in various sectors, including coding. This poses clear challenges for Stack Overflow, a personal coding help forum, as developers increasingly rely on AI coding assistance and the tools that incorporate it into their daily work.

Coding help forum Stack Overflow lays off 28% of staff as it faces profitability challenges

Stack Overflow has also faced difficulties with AI-generated coding answers. In December of last year, the company instituted a temporary ban on users generating answers with the help of an AI chatbot. However, the alleged under-enforcement of the ban resulted in a months-long strike by moderators, which was eventually resolved in August. Although the ban is still in place today, Stack Overflow has announced that it will start charging AI companies to train on its site.

James

AI Security

UK Government Hosts Global AI Safety Summit to Address Potential Risks and Transformative Opportunities

Published

2 months ago

August 8, 2024

James

UK Government Hosts Global AI Safety Summit to Address Potential Risks and Transformative Opportunities 50

Table of Contents

Overview

The United Kingdom Government has announced information regarding the inaugural worldwide AI Safety Summit that will take place at the historic Bletchley Park. This summit aims to bring together key figures from the tech industry, including government representatives, AI companies, members of the public, and specialists, to talk about the potential dangers of advanced AI and explore the promising opportunities it presents.

Focus on Frontier AI

UK Government Hosts Global AI Safety Summit to Address Potential Risks and Transformative Opportunities

The primary focus of the summit is on frontier AI, the most advanced generation of AI models. These models, if not developed responsibly, could pose significant risks. The event will delve into understanding the risks associated with frontier AI and explore ways to enhance safety.

Topics of Discussion

The AI Safety Summit will cover a range of topics, including threats to national security, potential election disruption, erosion of social trust, and exacerbation of global inequalities. These discussions aim to shed light on the potential risks and challenges posed by AI advancements.

Enhancing Frontier AI Safety

The summit will feature roundtable discussions aimed at enhancing frontier AI safety responsibly. Delegates will explore defining risk thresholds, effective safety assessments, and robust governance mechanisms to ensure the safe scaling of frontier AI by developers.

International Collaboration

International collaboration will be a key theme of the summit, emphasizing the need for policymakers, scientists, and researchers to work together in managing risks and harnessing AI’s potential for global economic and social benefits.

Transformative Opportunities in Education

The summit will conclude with a panel discussion on the transformative opportunities of AI for the public good, particularly in revolutionizing education. The closing remarks will highlight the importance of global collaboration in adopting AI safely.

Fostering International Cooperation

The UK Government aims to foster international cooperation in the responsible development and deployment of AI technology through this summit. By convening global experts and policymakers, the UK Government intends to lead the conversation on creating a safe and positive future with AI.

James