If the stakes for effectively deploying artificial intelligence were not so high, then the technology would not be a front-burner strategic priority throughout the corporate world; the Big Tech sector would not be allocating billions of capital-expenditure dollars to it; and rivalries would not be breaking out over goals and principles like the one between Sam Altman and Elon Musk.
The high-tech billionaires, co-founders of the ChatGPT developer OpenAI, have feuded over whether the company was forsaking its original nonprofit, “benefit of humanity” ideals. Altman, the CEO, was ousted by his board in November 2023 before being reinstated a week later. In the fallout, OpenAI director and chief scientist Ilya Sutskever left to start a venture with “one goal and one product: a safe superintelligence.”
That “singular focus,” and the branding of Team Sutskever’s Safe Superintelligence Inc., points toward the positive consequences of AI as many true believers see them unfolding. Anthropic, co-founded by Dario Amodei, who previously worked on the GPT-2 and GPT-3 models at OpenAI, calls itself an “AI safety and research company.” Its Responsible Scaling Policy commits to not training or deploying models without having “implemented safety and security measures that will keep risks below acceptable levels.”
Anthropic’s Dario Amodei: Stressing safety and security.
Anthropic, Google, Microsoft and OpenAI founded the Frontier Model Forum – Amazon and Meta joined last May – as a “vehicle for cross-organizational discussions and actions on AI safety and responsibility.”
Indeed, “safe” and “safety” are keywords embraced by technologists and marketers alike and carry over into the policy realm.
An October 2023 executive order from the Biden White House was titled “The Safe, Secure and Trustworthy Development and Use of Artificial Intelligence.” The intent of the European Union’s AI Act, which took effect last year, was to ensure that the systems “are safe, transparent, traceable, non-discriminatory and environmentally friendly.”
Following a series of multi-jurisdictional discussions – including the Bletchley Declaration that came out of a November 2023 AI Safety Summit in the U.K., and the formation early last year of the Artificial Intelligence Safety Institute Consortium under the U.S. Department of Commerce’s National Institute of Standards and Technology – the Commerce and State departments in November hosted an “inaugural convening of the International Network of AI Safety Institutes” in San Francisco.
The latter was part of a “roadmap” leading to the AI Action Summit on February 10-11 in Paris. The new Trump administration, which will then be in its fourth week, has a designated “AI and crypto czar,” venture capitalist and former PayPal executive David Sacks.
In the view of the Frontier Model Forum, to build on the governmental efforts, “further work is needed on safety standards and evaluations to ensure frontier AI models are developed and deployed responsibly.”
“First do no harm” could be a unifying ethic for both providers and users of AI, and it has a natural appeal for policymakers in the U.S., where legislative proposals to date have been all over the map.
Safety implies striking a balance between encouraging innovation and installing risk-mitigating or regulatory guardrails.
In a Gartner Hype Cycle for Artificial Intelligence report, safety is associated with responsible AI (RAI): “Measuring success is difficult. Creating RAI principles and operationalizing them without regularly measuring the progress makes it hard to sustain RAI practices.”
Gartner AI Hype Cycle (detail) as of June 2024.
A lack of safety could translate into loss of trust, disillusionment, even a rollback in AI research and development funding, warns Anand Rao, a Distinguished Services Professor of Applied Data Science and AI in Carnegie Mellon University’s Heinz College of Information Systems and Public Policy.
“Loss of trust represents the more benign consequence of failing to achieve safe AI,” Rao remarks. “However, given the widespread adoption of AI today, the stakes are much higher than in previous decades. If AI systems cause significant harm – whether due to errors, biases or adversarial attacks – it could severely hinder the progress and adoption of AI technologies, undermining their potential benefits for society.”
Safety and confidence also concern more broadly defined “stakeholder” constituencies, the professor adds, meaning civil liberties advocates, labor unions, professional associations and governmental organizations – because collective trust is essential for AI’s sustainable growth and integration into everyday life.
Despite safety’s common appeal, the word can be interpreted differently.
“Safe AI” and “responsible AI” are interchangeable in the Gartner report; RAI is an umbrella term for making choices based on business and societal value, risk, trust, transparency, fairness, bias mitigation, explainability, sustainability, accountability, safety, privacy and regulatory compliance.
Mark McCreary of Fox Rothschild
“Safe AI and ethical AI are the same topic,” around which marketers are “trying to devise language differentiators to gather attention,” says Alberto Roldan, founder and owner of Artificial Intelligence and Business Analytics Group in Kissimmee, Florida.
To Mark G. McCreary, chief artificial intelligence and information security officer at Fox Rothschild, “Safe AI focuses on protection against harm or misuse, often through reliability and technical safeguards. Responsible AI emphasizes accountability, probably more in the development and application of AI.
“Ethical AI delves into the moral implications of AI. They are very different priorities and are given weight depending on the use and who may be impacted.”
Rao links safety to the conception, design, development, deployment and use of AI systems that minimize risk, while ensuring security, reliability, resilience and robustness against naïve usage and adversarial attacks.
Anand Rao of Carnegie Mellon
“It represents the minimal threshold for releasing AI systems for broader public or enterprise adoption,” the Carnegie Mellon professor states. “It encompasses technical safeguards – such as security and reliability – and socio-technical measures including testing, verification and validation of safety.”
In the terms of Maslow’s Hierarchy of Needs, Rao places safe AI at the base of the pyramid. Responsible AI occupies a higher layer of explainability, interpretability, transparency and inclusivity, ensured by governance and risk management. Ethical AI is at the apex, aligning with such human values as fairness, justice and human rights.
Rao reels off a range of safety-related risks raising both immediate and progressively serious, longer-term issues: bias and discrimination, privacy and security, job displacement, and existential threat scenarios.
Be careful and consistent, McCreary advises. “Input of confidential information or personal data in a public AI tool should be a concern . . . Bias in output should also be a concern, although I believe in most cases the user’s job is to identify and remove that bias.
“Finally, as regulations around AI start to be enacted, non-compliance can lead to reputational damage, sanctions or fines.”
Safety standards can underpin preparedness. Rao points to international benchmarks and tools such as IEEE 7001 and the National Institute of Standards and Technology’s Assessing Risks and Impacts of AI (ARIA).
Another avenue is participation in consortiums such as the Partnership on AI, a nonprofit, non-lobbying organization that brings together diverse voices seeking “positive outcomes for people and society.”
Governments should be championing safe AI, but the manner of regulation is important, explains Professor Michael Mainelli, chairman of consulting firm Z/Yen Group. Until recently serving as Lord Mayor of London, he championed an Ethical AI Initiative including interoperable, international standards specifically around ISO 42001.
Michael Mainelli of Z/Yen Group
“We contend that ‘horizontal’ regulation – across sectors, industries, professions and functions – is not only possible, but desirable,” Mainelli asserts. The ISO standard addresses accreditation, certification and governance and “should be the first approach to regulation as mankind learns what we want to regulate.”
“Ethical AI training can help ensure that AI is developed and deployed in a way that aligns with our values, respects human rights and promotes the common good,” Mainelli adds. “This initiative should establish ethical guidelines and principles for AI development, encourage transparency and accountability in AI systems, and promote collaboration and dialogue among stakeholders.”
Collaboration across regulatory, financial industry and more broadly defined stakeholder communities was a theme of the U.S. Treasury Department’s Report on the Uses, Opportunities and Risks of Artificial Intelligence in the Financial Services Sector, released in December.
The university-centric Enlightenment in Action Alliance (EAA) was announced in December by the Boston Global Forum (BGF) “to foster peace, security and equitable development.” Said the forum’s co-founder and co-chair, former Massachusetts Governor Michael Dukakis, “We aim to empower youth voices to shape AI as a force for public good. This alliance harnesses the talent and passion of students worldwide, enabling them to meaningfully contribute to ethical AI innovation and enlightened governance.”
In addition, the BGF is conducting a dialogue among business leaders as a planned contribution to next month’s Paris AI summit.
McCreary, also stressing collaboration, says, “I do not want some leaders of AI technology companies setting safety standards independently.”
“Persons responsible for AI adoption in their organization should start with robust governance, prioritize transparency, and foster a culture of collaboration and accountability,” McCreary goes on. “For developers of AI tools, by integrating safety into every stage of the AI lifecycle, organizations can unlock its benefits responsibly while protecting themselves and those they serve.”
Jeffrey Kutler of GARP contributed reporting for this article.