During this year’s Google I/O conference held in Mountain View, California, company executives discussed guardrails to its new artificial intelligence (AI) products to ensure their responsible use and prevent their potential misuse.
With AI technologies advancing, the spread of misinformation, deepfakes, and abusive text or imagery generated by AI has become a growing concern. Such a development could be detrimental to society, especially if the model responsible for creating the content came from Google, says James Sanders, principal analyst at CCS Insight.
According to Sanders, Google’s interests in responsible AI are motivated, at least in part, by reputation protection and discouraging regulators from interfering.
For instance, one technology called Universal Translator, which is a video AI of Google Translate, can take footage of a person speaking and translate their speech into another language. This app could potentially expand the video’s audience to include those who do not speak the original language. However, the AI might modify the lip movement to make it seem like the person was speaking the translated language which could erode trust in the source material. Therefore, there is an inherent tension between the benefits and potential misuse of the technology.
To prevent these issues, Google has established custom guardrails to limit who can use their technologies. The Universal Translators is available to fewer than ten partners. Another example is ChatGPT, a text-generating AI program that has been programmed to say that it couldn’t answer certain types of questions if the question or answer could cause harm.
Google is among several companies that have approached AI guardrails differently. Other companies, like Nvidia, have an open-source tool called NeMo Guardrails that ensures responses fit within specific parameters. The technology also prevents AI from hallucinating, which is when a system provides a confident response that isn’t justified by its training data.
Moreover, Google’s research on safeguards has been shared in its new PaLM-2 large-language model, which was also announced during the Google I/O conference. The technical paper explains that there are specific questions the AI engine will not touch. Google relies on automated adversarial testing to identify and reduce these outputs. The Perspective API, created for this purpose, is used by academic researchers to test models from OpenAI, Anthropic, among others.
The narrative of responsible AI use gained urgency after concerns about bad actors misusing AI technologies like ChatGPT to craft phishing approaches or generate malicious codes to hack systems. AI was already being used for deepfake videos and voices. The Department of Defense’s Graphika, an AI company, recently identified instances of AI-generated footage being used to try and influence public opinion.
In response, the White House has called for guardrails to mitigate the misuse of AI technology. Earlier in May, the Biden administration secured the commitment of companies like Google, Microsoft, Nvidia, OpenAI, and Stability AI to allow participants to publicly evaluate their AI systems during DEF CON 31, which will be held in August in Las Vegas. The models will be red-teamed using an evaluation platform developed by Scale AI. The White House has stated that this independent exercise will provide critical information to researchers and the public about the impacts of these models. It will enable AI companies and developers to take steps in fixing any issues found in those models.
To conclude, Google’s commitment to responsible AI use highlights the importance of preventing the misuse of advanced AI technologies. It helps ensure that the AI products are used in the best way possible while minimizing any risk factor potentially posed by their deployment. By imposing guardrails and controls, companies like Google can mitigate the potential issues while allowing society to benefit from the many advantages of AI.