As AI usage increases, more instances of harmful responses emerge, including hate speech, copyright infringement, and sexually explicit content, highlighting the need for stronger standards to test these models before releasing them to the public.
Article contents:
- Introduction
- Harmful responses and their challenges
- Red teaming strategy in testing
- Project Moonshot as a comprehensive solution
- The need for higher standards
- Conclusion
- FAQ
Introduction
In an era of rapidly accelerating artificial intelligence technologies, companies and developers face new challenges in ensuring the safety and reliability of the models they release. With increased usage, whether well-intentioned or for malicious purposes, more potential harmful responses are coming to light.
Harmful responses and their challenges
These undesirable behaviors include hate speech, copyright violations, and sexually explicit content. Compounding this issue are a lack of regulations and insufficient testing of AI models, researchers told CNBC.
As Javier Rando, an AI researcher specializing in adversarial machine learning, explains: “The answer, after nearly 15 years of research, is no, we do not know how to do this, and we do not seem to be getting better.” Getting machine learning models to behave as intended is extremely difficult.
Red teaming strategy in testing
Despite the challenges, there are ways to evaluate risks in AI, such as the “red teaming” strategy. This practice involves individuals testing and examining AI systems to discover and identify any potential harm, a common method in cybersecurity.
Shane Longpre, an AI and policy researcher and leader of the Data Provenance Initiative, notes that there are currently not enough people working on red teams. While AI startups now use first-party evaluators or second-party contractors to test their models, opening testing to third parties such as everyday users, journalists, researchers, and ethical hackers would lead to more robust evaluation.
Longpre explains: “Some flaws in systems that people uncovered required lawyers and doctors for actual inspection, and specialized domain expert scientists to figure out whether this was a flaw or not, because the average person might not have or will not have enough expertise.”
Project Moonshot as a comprehensive solution
Project Moonshot is one such approach, combining technical solutions with policy mechanisms. Launched by Singapore’s Infocomm Media Development Authority, it is a toolkit for evaluating large language models developed with industry partners like IBM and DataRobot in Boston.
The toolkit integrates benchmarks, red teaming testing, and testing baselines. There is also an evaluation mechanism allowing AI startups to ensure their models can be trusted and do not harm users, Anup Kumar, head of client engineering for data and AI at IBM Asia Pacific, told CNBC.
The need for higher standards
Pierre Alquier, a professor of statistics at ESSEC Business School Asia-Pacific, says tech companies are currently rushing to release the latest AI models without proper evaluation. He adds: “When a pharmaceutical company designs a new drug, it needs months of testing and very serious proof that it is useful and not harmful before getting approval from the government.”
AI models need to meet a strict set of conditions before approval. Moving away from broad AI tools toward developing tools tailored for more specific tasks will make it easier to predict and monitor their misuse.
Conclusion
As AI technologies evolve at an increasingly rapid pace, establishing rigorous standards and comprehensive testing becomes essential to ensure the safety of these technologies. Collaboration between researchers, companies, and regulators is vital to ensure that AI technologies serve humanity without causing harm.
FAQ
1. What is the red teaming strategy in AI testing?
It is a practice involving individuals testing and examining AI systems to discover and identify any potential harm, inspired by cybersecurity practices.
2. Why do AI models need stronger standards?
Because increased AI usage reveals more harmful responses such as hate speech and copyright infringement, requiring strict standards to ensure safety.
3. What is Project Moonshot?
It is a toolkit for evaluating large language models developed by Singapore with industry partners, integrating benchmarks, red teaming testing, and testing baselines.
