Latest / USA Daily News / Sam Altman Sounds the Alarm on AI Risks — How Dangerous Could Artificial Intelligence Become

Transcript

Sam Altman Sounds the Alarm on AI Risks — How Dangerous Could Artificial Intelligence Become · USA Daily News

• 0:00
  Speaker 1: Artificial intelligence has moved rapidly from being a specialized technology
• 0:04
  Speaker 1: discussed mainly by researchers and technology companies to becoming one
• 0:09
  Speaker 1: of the most consequential issues facing governments, businesses, workers, and
• 0:14
  Speaker 1: the public. That transformation is at the center of a
• 0:17
  Speaker 1: renewed debate over how society should handle the risks of
• 0:21
  Speaker 1: increasingly powerful AI systems. In a recent interview, OpenAI Chief
• 0:27
  Speaker 1: Executive Sam Altman addressed the debate directly. arguing that the
• 0:31
  Speaker 1: potential benefits of artificial intelligence are so significant that society
• 0:36
  Speaker 1: may have to accept a certain amount of risk and
• 0:38
  Speaker 1: disruption along the way. His comments have placed a fresh
• 0:42
  Speaker 1: spotlight on a question that has become increasingly difficult to avoid.
• 0:47
  Speaker 2: How much risk.
• 0:48
  Speaker 1: Should the world tolerate in exchange for technological progress? Altman's
• 0:52
  Speaker 1: position is particularly significant because he is one of the
• 0:56
  Speaker 1: most prominent executives shaping the modern AI industry. OpenAI, the
• 1:01
  Speaker 1: company behind ChatGPT and a major force in the development
• 1:05
  Speaker 1: of advanced artificial intelligence, sits at the center of the
• 1:08
  Speaker 1: global race to build systems capable of performing increasingly complex tasks.
• 1:14
  Speaker 2: That position gives.
• 1:15
  Speaker 1: Altman's comments a weight that extends far beyond one company's
• 1:19
  Speaker 1: internal policies. When he discusses the dangers … benefits, or
• 1:23
  Speaker 1: possible regulation of AI, those remarks become part of a
• 1:27
  Speaker 1: much larger international conversation about the future of technology itself.
• 1:33
  Speaker 1: The interview comes at a moment when expectations surrounding AI
• 1:37
  Speaker 1: are unusually high. Companies are investing heavily in new models,
• 1:41
  Speaker 1: AI assistance, and autonomous systems designed to carry out tasks
• 1:46
  Speaker 1: with limited human intervention. Governments see the technology as a
• 1:50
  Speaker 1: possible source of economic growth and strategic advantage. Researchers believe
• 1:55
  Speaker 1: advanced AI could accelerate scientific discovery, improve productivity, and help
• 2:01
  Speaker 1: solve problems that have resisted conventional approaches. At the same time,
• 2:06
  Speaker 1: concern has grown over what could happen if these systems
• 2:09
  Speaker 1: become more capable faster than institutions can adapt. That tension
• 2:14
  Speaker 1: is central to Altman's message. He is not presenting artificial
• 2:17
  Speaker 1: intelligence as risk-free. Instead, his argument is that eliminating every
• 2:23
  Speaker 1: negative consequence associated with a powerful technology may not be realistic.
• 2:29
  Speaker 1: Some risks, such as fraud, scams, cyberattacks, misinformation, or harmful
• 2:35
  Speaker 1: uses of AI, may prove difficult to eliminate entirely as
• 2:39
  Speaker 1: the technology becomes more widely available. The question, therefore, becomes
• 2:44
  Speaker 1: whether those risks can be managed without blocking the broader
• 2:48
  Speaker 1: benefits that AI could provide. For Altman, access is an
• 2:52
  Speaker 1: important part of that calculation. He is pushed back against
• 2:56
  Speaker 1: the idea that advanced artificial intelligence should effectively be controlled
• 3:00
  Speaker 1: by a single organization or a very small group of
• 3:03
  Speaker 1: institutions in the name of safety. That approach, in his view,
• 3:08
  Speaker 1: could create a different kind of danger by placing enormous
• 3:11
  Speaker 1: technological power in the hands of a limited number of
• 3:15
  Speaker 1: decision makers. The debate is therefore no longer simply about
• 3:19
  Speaker 1: whether AI is safe. It is also about who should
• 3:22
  Speaker 1: control it, who should have access to it, and who
• 3:24
  Speaker 1: should ultimately decide where the boundaries are drawn. This creates
• 3:28
  Speaker 1: a difficult policy dilemma. Governments and regulators have a responsibility
• 3:33
  Speaker 1: to protect citizens from technologies that could cause significant harm,
• 3:38
  Speaker 1: But excessive restrictions can also slow innovation and prevent businesses, researchers,
• 3:44
  Speaker 1: and individuals from benefiting from new capabilities. The challenge becomes
• 3:48
  Speaker 1: even more complicated because AI development is global. If one
• 3:53
  Speaker 1: country or company imposes strict limits while competitors continue moving quickly,
• 3:59
  Speaker 1: the result could be a shift in technological or economic advantage.
• 4:03
  Speaker 1: That concern has become especially important as artificial intelligence develops
• 4:08
  Speaker 1: into a strategic asset. Advanced computing infrastructure, access to data,
• 4:14
  Speaker 1: specialized chips, and highly skilled researchers are now closely connected
• 4:19
  Speaker 1: to national competitiveness. Countries want to remain at the forefront
• 4:23
  Speaker 1: of AI because leadership in the field could influence everything
• 4:26
  Speaker 1: from productivity and industrial capacity to scientific research and military strategy.
• 4:33
  Speaker 1: As a result, discussions about safety increasingly overlap with questions
• 4:38
  Speaker 1: of national security and geopolitical power. Within the AI industry itself,
• 4:44
  Speaker 1: there is also no universal agreement about how quickly development
• 4:48
  Speaker 1: should proceed. Some leaders and researchers have argued that the
• 4:52
  Speaker 1: emergence of more capable systems requires stronger safeguards and, in
• 4:57
  Speaker 1: certain circumstances, a slower pace of development. Others believe that
• 5:02
  Speaker 1: the technology's benefits are too important to delay and that
• 5:06
  Speaker 1: society should focus on improving safety while allowing innovation to continue.
• 5:11
  Speaker 1: Altman's remarks place OpenAI closer to the second camp. although
• 5:16
  Speaker 1: his position still recognizes that some risks require limits and
• 5:20
  Speaker 1: serious attention. The debate surrounding artificial intelligence is ultimately.
• 5:26
  Speaker 2: A debate about value.
• 5:28
  Speaker 1: On one side is the possibility that increasingly capable AI
• 5:32
  Speaker 1: systems could transform productivity, scientific research, healthcare, education, and business
• 5:39
  Speaker 1: on a scale that rivals some of the most important
• 5:42
  Speaker 1: technological breakthroughs of the modern era. On the other is
• 5:45
  Speaker 1: the recognition that technologies capable of producing enormous benefits can
• 5:50
  Speaker 1: also generate significant disruption, mistakes, and unintended consequences. Sam Altman's
• 5:56
  Speaker 1: discussion of AI risks brings that tension into sharper focus
• 6:00
  Speaker 1: because it challenges policymakers and the public to consider not
• 6:04
  Speaker 1: only what could go wrong, but also what might be
• 6:07
  Speaker 1: lost if society becomes too cautious. The potential rewards from
• 6:11
  Speaker 1: artificial intelligence help explain why the technology has developed at
• 6:16
  Speaker 1: such a rapid pace. Businesses see AI as a way
• 6:19
  Speaker 1: to automate repetitive work, analyze enormous quantities of information, and
• 6:24
  Speaker 1: assist employees with tasks that once required substantial amounts of time.
• 6:29
  Speaker 1: Software developers can use AI to generate and review code.
• 6:33
  Speaker 1: Researchers can use advanced systems to organize complex information and
• 6:38
  Speaker 1: accelerate parts of the discovery process. Manufacturers can apply AI
• 6:43
  Speaker 1: to improve logistics and production. Smaller companies can gain access
• 6:48
  Speaker 1: to tools that previously required expensive teams of specialists. That
• 6:52
  Speaker 1: promise creates powerful incentives to continue investing. Every major improvement
• 6:58
  Speaker 1: in AI capability can potentially produce commercial advantages for the
• 7:02
  Speaker 1: companies that develop or deploy it. Faster systems, Better reasoning,
• 7:07
  Speaker 1: improved automation, and more reliable digital assistants can reduce costs
• 7:12
  Speaker 1: and create entirely new services.
• 7:15
  Speaker 2: This means the AI race.
• 7:16
  Speaker 1: Is driven by more than technological curiosity. It is also
• 7:20
  Speaker 1: powered by competition for customers, investment, market share, and strategic influence.
• 7:27
  Speaker 1: The economic implications are potentially enormous. If AI allows workers
• 7:32
  Speaker 1: to complete tasks more efficiently, businesses could produce more with
• 7:36
  Speaker 1: fewer resources. That could increase productivity, lower some costs, and
• 7:41
  Speaker 1: create wealth. New industries could emerge around AI infrastructure, software,
• 7:46
  Speaker 1: data services, robotics, and specialized applications. Countries that successfully integrate
• 7:53
  Speaker 1: AI into their economies could potentially gain an advantage over
• 7:56
  Speaker 1: those that are slower to adapt. Yet those same productivity
• 8:00
  Speaker 1: gains can create disruptions. When machines become capable of performing
• 8:05
  Speaker 1: tasks that were previously done by people, the economic benefits
• 8:09
  Speaker 1: do not automatically reach everyone equally. Some workers may see
• 8:13
  Speaker 1: their jobs augmented, while others may find that particular responsibilities
• 8:18
  Speaker 1: are reduced or eliminated. Companies may reorganize around smaller teams
• 8:23
  Speaker 1: supported by AI systems. Entire professions could experience changes in
• 8:28
  Speaker 1: the skills that employers value. That creates one of the
• 8:31
  Speaker 1: central contradictions in the AI debate. A technology can increase
• 8:36
  Speaker 1: overall economic output while still causing serious difficulties for individuals
• 8:41
  Speaker 1: and communities. Higher productivity does not guarantee that every worker
• 8:45
  Speaker 1: will benefit at the same pace. A company may save
• 8:49
  Speaker 1: money through automation, while employees affected by those changes face
• 8:53
  Speaker 1: a difficult transition. Policymakers, therefore, have to consider not just
• 8:58
  Speaker 1: whether AI creates economic growth, but how the gains and
• 9:01
  Speaker 1: losses are distributed. The argument over risk becomes even more
• 9:05
  Speaker 1: complicated when AI moves beyond ordinary workplace automation. More capable
• 9:11
  Speaker 1: systems could assist researchers in areas that require advanced reasoning,
• 9:16
  Speaker 1: help professionals analyze complex problems, or enable organizations to make
• 9:20
  Speaker 1: decisions much faster. The potential benefits could extend into areas
• 9:25
  Speaker 1: where speed and information processing have significant social value. Healthcare
• 9:31
  Speaker 1: is one example of why proponents are optimistic. AI could
• 9:35
  Speaker 1: help doctors analyze medical information, assist with administrative work, identify
• 9:40
  Speaker 1: patterns in large datasets, or accelerate research into diseases and treatments.
• 9:46
  Speaker 1: In scientific research, Systems capable of processing vast amounts of
• 9:50
  Speaker 1: information could help investigators explore possibilities that would be difficult
• 9:55
  Speaker 1: for humans to evaluate manually. In education, AI could provide
• 10:00
  Speaker 1: personalized assistance and make certain learning tools more accessible. But
• 10:05
  Speaker 1: greater capability also means greater responsibility. An AI system that
• 10:10
  Speaker 1: helps with an important decision can potentially amplify mistakes if
• 10:14
  Speaker 1: the information it uses is wrong or the reasoning behind
• 10:17
  Speaker 1: its output is misunderstood. As people become more dependent on
• 10:21
  Speaker 1: AI-generated recommendations, the consequences of an error can increase. The
• 10:27
  Speaker 1: central issue is therefore not whether AI will make mistakes.
• 10:31
  Speaker 1: Every complex system does. The more important question is how
• 10:35
  Speaker 1: much authority society should give these systems and how failures
• 10:38
  Speaker 1: should be identified and corrected. This is where Altman's risk
• 10:43
  Speaker 1: versus reward argument becomes especially important. The question of artificial
• 10:48
  Speaker 1: intelligence risk becomes considerably more serious when the discussion moves
• 10:52
  Speaker 1: beyond what AI systems can do today and toward what
• 10:56
  Speaker 1: they may be capable of doing in the future. Current
• 10:59
  Speaker 1: systems can already generate text, write software, analyze information, create images,
• 11:05
  Speaker 1: and assist with complex tasks. But as developers continue pursuing
• 11:10
  Speaker 1: more advanced models, attention is increasingly turning.
• 11:14
  Speaker 2: To a harder problem.
• 11:15
  Speaker 1: What happens when AI systems become capable of making longer
• 11:19
  Speaker 1: chains of decisions, operating with greater independence and responding to
• 11:24
  Speaker 1: situations that their creators did not specifically anticipate? This concern
• 11:29
  Speaker 1: is often described as.
• 11:30
  Speaker 2: The problem of maintaining control.
• 11:33
  Speaker 1: It does not necessarily mean that an AI system would
• 11:36
  Speaker 1: suddenly become conscious, hostile, or determined to defeat humanity. A
• 11:41
  Speaker 1: more practical concern is that increasingly capable systems could pursue
• 11:45
  Speaker 1: objectives in ways that are unexpected, inefficient, or potentially harmful.
• 11:50
  Speaker 1: The greater the independence given to an AI system, the
• 11:54
  Speaker 1: more important it becomes to ensure that its behavior remains
• 11:57
  Speaker 1: consistent with human intentions. That challenge begins with a basic limitation.
• 12:03
  Speaker 2: Of AI development.
• 12:04
  Speaker 1: Engineers can train systems on enormous quantities of data and
• 12:08
  Speaker 1: evaluate them across thousands of tests, but no testing environment
• 12:12
  Speaker 1: can perfectly reproduce.
• 12:14
  Speaker 2: The real world. An AI model may.
• 12:17
  Speaker 1: Encounter unfamiliar situations after deployment, interact with users in unexpected ways,
• 12:23
  Speaker 1: or combine capabilities in ways that were not obvious during development.
• 12:27
  Speaker 1: A system that performs reliably under controlled conditions may behave
• 12:32
  Speaker 1: differently when placed into a much more complicated environment. The
• 12:36
  Speaker 1: issue becomes especially important as AI moves from answering questions
• 12:40
  Speaker 1: to performing actions. A chatbot that produces an incorrect explanation
• 12:45
  Speaker 1: can create confusion, but the consequences may be limited if
• 12:49
  Speaker 1: a human checks the response. An AI system given authority
• 12:53
  Speaker 1: to make decisions, operate software, Manage resources or interact with
• 12:58
  Speaker 1: external systems could produce far greater consequences from a single mistake.
• 13:03
  Speaker 1: The difference is not simply greater intelligence. It is greater agency.
• 13:08
  Speaker 1: Agency changes the risk calculation because a system that can
• 13:12
  Speaker 1: act independently has more opportunities to create unintended outcomes. Developers, therefore,
• 13:18
  Speaker 1: have to think about more than whether an AI produces
• 13:21
  Speaker 1: accurate information. They must consider what happens when the system
• 13:25
  Speaker 1: is given objectives, access to tools, and enough freedom to
• 13:29
  Speaker 1: pursue those objectives over time. One of the central technical
• 13:32
  Speaker 1: questions is known as alignment. How can developers ensure that
• 13:37
  Speaker 1: an AI system's behavior reflects the intentions and values of
• 13:41
  Speaker 1: the people using or supervising it? On the surface, the
• 13:45
  Speaker 1: problem appears straightforward. Developers can instruct a model to be helpful, honest,
• 13:50
  Speaker 1: and safe. But human values are rarely simple. People disagree
• 13:55
  Speaker 1: about what is ethical, what information should be restricted, and
• 13:59
  Speaker 1: what trade-offs are acceptable. Translating those complicated judgments into rules
• 14:04
  Speaker 1: that an advanced machine can consistently follow is an enormous challenge.
• 14:09
  Speaker 1: There is also a difference between following instructions and understanding
• 14:13
  Speaker 1: why an instruction exists. An AI system may appear to
• 14:17
  Speaker 1: follow a rule in one situation while finding an unexpected
• 14:20
  Speaker 1: way around it in another. That does not necessarily imply
• 14:24
  Speaker 1: malicious intent. It may simply mean that the system has
• 14:27
  Speaker 1: learned patterns that allow it.
• 14:29
  Speaker 2: To achieve a stated objective.
• 14:31
  Speaker 1: Without fully capturing the broader purpose behind the instruction. This
• 14:35
  Speaker 1: is why increasingly sophisticated AI models are being subjected to
• 14:39
  Speaker 1: more extensive evaluations before and after deployment. Developers need to
• 14:44
  Speaker 1: identify behaviors that may not appear during ordinary conversations. They
• 14:49
  Speaker 1: may test models for deception, harmful reasoning, manipulation, security vulnerabilities,
• 14:55
  Speaker 1: or attempts to bypass restrictions. The objective is to discover
• 14:59
  Speaker 1: weaknesses before they become serious problems in the real world.
• 15:03
  Speaker 1: Yet testing itself has limits. The number of possible interactions
• 15:08
  Speaker 1: with an advanced AI system can be enormous. A model
• 15:11
  Speaker 1: might behave safely across thousands of scenarios while still failing
• 15:15
  Speaker 1: in a rare combination of circumstances. As capabilities increase, the
• 15:20
  Speaker 1: space of possible behaviors can become more complicated rather than less.
• 15:25
  Speaker 1: Another concern is that an AI system may learn to
• 15:28
  Speaker 1: respond differently when it recognizes that it is being evaluated.
• 15:32
  Speaker 1: This creates a difficult problem for researchers. determining whether a
• 15:36
  Speaker 1: model is genuinely behaving safely or simply adapting its behavior
• 15:40
  Speaker 1: to the conditions under which it is being tested. Reliable
• 15:44
  Speaker 1: safety therefore requires more than a few successful demonstrations. Researchers
• 15:49
  Speaker 1: need evidence that desirable behavior remains stable across different environments
• 15:54
  Speaker 1: and pressures. The issue is further complicated by the competitive
• 15:58
  Speaker 1: structure of the AI industry. The question of artificial intelligence
• 16:02
  Speaker 1: control extends beyond the behavior of individual systems. Even if
• 16:07
  Speaker 1: developers succeed in making advanced models more predictable and reliable,
• 16:12
  Speaker 1: another concern remains. Who will control the most powerful AI technologies,
• 16:17
  Speaker 1: and how much influence could those organizations accumulate? As artificial
• 16:22
  Speaker 1: intelligence becomes more capable, access to computing resources, specialized talent,
• 16:29
  Speaker 1: data and advanced models could translate into enormous economic and
• 16:33
  Speaker 1: political power. That raises the possibility that the concentration of
• 16:38
  Speaker 1: AI capability could become a major risk in its own right.
• 16:42
  Speaker 1: For much of the history of computing, technological power has
• 16:46
  Speaker 1: gradually become more distributed. Personal computers moved computing from institutions
• 16:51
  Speaker 1: into homes. The internet expanded access to information and communication.
• 16:57
  Speaker 1: Smartphones placed powerful digital tools into the hands of billions
• 17:01
  Speaker 1: of people. Artificial intelligence, however, presents a more complicated picture
• 17:06
  Speaker 1: because the development of cutting-edge systems often requires extraordinary amounts
• 17:11
  Speaker 1: of computing infrastructure, specialized hardware, and highly trained researchers. That
• 17:17
  Speaker 1: creates barriers to entry. A small startup can build software
• 17:21
  Speaker 1: on top of existing AI models, but developing a frontier
• 17:25
  Speaker 1: system from the ground up can require massive financial resources
• 17:28
  Speaker 1: and access.
• 17:29
  Speaker 2: To advanced data centers.
• 17:31
  Speaker 1: The companies capable of making those investments may therefore gain
• 17:35
  Speaker 1: an advantage that becomes increasingly difficult for smaller competitors to overcome.
• 17:40
  Speaker 1: This concentration has an obvious economic dimension. Companies that control
• 17:45
  Speaker 1: the most advanced AI systems could influence the direction of
• 17:49
  Speaker 1: entire industries. Businesses may become dependent on a handful of
• 17:53
  Speaker 1: AI providers for software development, customer service, data analysis, research tools,
• 18:00
  Speaker 1: and other essential functions. As dependence grows, the companies operating
• 18:05
  Speaker 1: the underlying infrastructure could acquire significant bargaining power over their
• 18:10
  Speaker 1: customers and competitors. The implications could extend beyond business. AI
• 18:16
  Speaker 1: systems are increasingly relevant to education, healthcare, government services, scientific research,
• 18:23
  Speaker 1: and national security. If a small number of organizations control
• 18:27
  Speaker 1: the most capable models, they could potentially influence which capabilities
• 18:32
  Speaker 1: are prioritized, which uses are permitted, and which information or
• 18:36
  Speaker 1: activities are restricted. Even when those decisions are made with
• 18:40
  Speaker 1: legitimate safety objectives, the concentration of authority raises questions about accountability.
• 18:46
  Speaker 1: One concern is that private companies could end up making
• 18:50
  Speaker 1: decisions with consequences normally associated with public institutions. Technology companies
• 18:56
  Speaker 1: are not elected governments. Their executives and boards are accountable
• 19:01
  Speaker 1: primarily to corporate structures, investors, employees, and customers. Yet increasingly
• 19:07
  Speaker 1: capable AI systems may affect public communication and economic opportunities,
• 19:12
  Speaker 1: and access to information on a scale that makes their
• 19:15
  Speaker 1: decisions relevant to society as a whole. That does not
• 19:19
  Speaker 1: mean companies should have no role in deciding how their
• 19:21
  Speaker 1: technologies are deployed. The developers of AI systems understand their products,
• 19:27
  Speaker 1: their vulnerabilities, and their technical limitations better than most outside organizations.
• 19:33
  Speaker 1: They are therefore in a strong position to build safeguards
• 19:36
  Speaker 1: and establish responsible use policies. The challenge arises when private
• 19:41
  Speaker 1: decision-making becomes the primary mechanism for determining questions.
• 19:45
  Speaker 2: That have broad social consequences.
• 19:48
  Speaker 1: This is one reason Altman's view on AI risk have
• 19:51
  Speaker 1: attracted attention. His discussion of the dangers of advanced artificial
• 19:55
  Speaker 1: intelligence is connected to a broader debate over whether safety
• 19:59
  Speaker 1: should be achieved by restricting access to powerful systems or
• 20:04
  Speaker 1: by making those systems broadly available under appropriate safeguards. Those
• 20:08
  Speaker 1: approaches reflect very different assumptions about where risk comes from.
• 20:13
  Speaker 1: The first approach emphasizes concentration.
• 20:16
  Speaker 2: As a form of protection.
• 20:18
  Speaker 1: If powerful AI is available only to carefully controlled organizations,
• 20:23
  Speaker 1: it may be easier to monitor its use, establish security standards,
• 20:27
  Speaker 1: and respond to dangerous behavior. A limited number of operators
• 20:31
  Speaker 1: could theoretically make it easier to enforce rules and prevent
• 20:35
  Speaker 1: malicious actors from obtaining advanced capabilities. but concentration creates another vulnerability.
• 20:42
  Speaker 1: If too much power is placed in too few hands,
• 20:45
  Speaker 1: a mistake or abuse at one organization could have consequences
• 20:49
  Speaker 1: far beyond that organization. A security breach, corporate failure, political pressure,
• 20:55
  Speaker 1: or flawed internal policy could affect large portions of the
• 21:00
  Speaker 1: AI ecosystem. The system might become orderly, but also fragile
• 21:04
  Speaker 1: because so much depends on a small number of players.
• 21:08
  Speaker 1: The alternative approach is broader distribution. The debate over artificial
• 21:13
  Speaker 1: intelligence regulation has increasingly become a debate over how society
• 21:17
  Speaker 1: should manage uncertainty itself. As AI systems grow more capable,
• 21:22
  Speaker 1: governments are being pushed to decide whether existing laws are sufficient,
• 21:27
  Speaker 1: whether entirely new rules are necessary, or whether the technology
• 21:31
  Speaker 1: should largely be governed through voluntary safeguards developed by the
• 21:35
  Speaker 1: companies building it. Sam Altman's recent comments place OpenAI firmly
• 21:40
  Speaker 1: inside that wider argument, while also highlighting an important division
• 21:45
  Speaker 1: within the AI industry over how much regulation is appropriate
• 21:49
  Speaker 1: and how quickly it should be introduced. At the center
• 21:52
  Speaker 1: of OpenAI's approach is the belief that regulation should protect
• 21:56
  Speaker 1: society from serious dangers without unnecessarily restricting technological progress. That
• 22:02
  Speaker 1: philosophy reflects the company's broader view that artificial intelligence could
• 22:07
  Speaker 1: produce benefits on a scale large enough to justify taking
• 22:11
  Speaker 1: some risks. From this perspective, the objective is not to
• 22:15
  Speaker 1: create an environment in which every possible negative outcome is prevented. Instead,
• 22:21
  Speaker 1: the goal is to identify the most dangerous forms of
• 22:24
  Speaker 1: misuse and develop safeguards that reduce those risks while allowing
• 22:28
  Speaker 1: useful AI systems to continue improving. This approach differs from
• 22:33
  Speaker 1: a more cautious philosophy that places greater emphasis on slowing
• 22:37
  Speaker 1: the development of increasingly powerful systems until their potential dangers
• 22:42
  Speaker 1: are better understood. Supporters of that position argue that the
• 22:46
  Speaker 1: consequences of advanced AI could become much more serious as
• 22:50
  Speaker 1: systems gain greater autonomy, stronger reasoning abilities, and the capacity
• 22:54
  Speaker 1: to interact with real-world infrastructure. In their view, The pace
• 22:59
  Speaker 1: of technological progress itself may become a risk if safety
• 23:03
  Speaker 1: research and institutional oversight cannot keep up. The disagreement is
• 23:08
  Speaker 1: significant because both sides are responding to the same basic reality.
• 23:13
  Speaker 1: No one can predict exactly how advanced AI will evolve.
• 23:17
  Speaker 1: The difference lies in how that uncertainty should influence policy.
• 23:20
  Speaker 1: A more innovation-focused approach treats uncertainty as something to manage
• 23:25
  Speaker 1: while continuing forward. A more cautious approach treats uncertainty as
• 23:30
  Speaker 1: a reason to create additional barriers before the technology reaches
• 23:34
  Speaker 1: a more powerful stage. OpenAI's lighter-touch philosophy is influenced in
• 23:39
  Speaker 1: part by the belief that overly restrictive regulation could have
• 23:43
  Speaker 1: unintended consequences. Complex regulations can raise the cost of developing
• 23:48
  Speaker 1: AI and make it harder for smaller companies, independent researchers,
• 23:53
  Speaker 1: and new competitors to enter the market. If compliance becomes
• 23:57
  Speaker 1: expensive or difficult, technological development could become even more concentrated
• 24:02
  Speaker 1: among the largest corporations, potentially reducing competition rather than improving safety.
• 24:08
  Speaker 1: There is also a concern that rigid regulations can become
• 24:12
  Speaker 1: outdated quickly. Artificial intelligence does not develop according to the
• 24:16
  Speaker 1: timelines of traditional legislation. A law designed around the capabilities
• 24:21
  Speaker 1: of today's models could become irrelevant as newer systems appear.
• 24:25
  Speaker 1: Policymakers therefore face a difficult challenge. They must create rules
• 24:30
  Speaker 1: that are strong enough to address genuine risks, but flexible
• 24:34
  Speaker 1: enough to remain useful as the technology changes. From OpenAI's perspective,
• 24:40
  Speaker 1: this can mean favoring standards that focus on outcomes rather
• 24:43
  Speaker 1: than attempting to regulate every technical detail. Instead of prescribing
• 24:48
  Speaker 1: exactly how an AI model must be constructed, policymakers could
• 24:53
  Speaker 1: establish requirements concerning security and transparency, testing, privacy, accountability, and
• 25:00
  Speaker 1: the handling of dangerous capabilities. Companies would retain room to
• 25:04
  Speaker 1: innovate while still being expected to meet certain safety obligations.
• 25:09
  Speaker 1: Critics of a lighter regulatory approach, however, argue that relying
• 25:13
  Speaker 1: too heavily on companies to determine their own limits may
• 25:17
  Speaker 1: create a conflict of interest. AI developers have strong commercial
• 25:21
  Speaker 1: incentives to release increasingly capable products. A company that delays
• 25:26
  Speaker 1: an important model for safety reasons could potentially lose customers
• 25:30
  Speaker 1: or market share to a competitor willing to move faster.
• 25:34
  Speaker 1: Even companies acting in good faith could therefore face pressure
• 25:37
  Speaker 1: that makes voluntary restraint difficult. This is where the idea
• 25:41
  Speaker 1: of industry-wide standards becomes important. If one company voluntarily adopts
• 25:47
  Speaker 1: stricter safeguards while its competitors do not, the responsible company
• 25:51
  Speaker 1: may bear the immediate costs while receiving few short-term benefits.
• 25:56
  Speaker 1: Coordinated rules could reduce that problem by establishing a common
• 26:00
  Speaker 1: baseline that every major developer must meet. The difficult question
• 26:05
  Speaker 1: is determining who should establish those standards and how they
• 26:08
  Speaker 1: should be enforced. Government regulation is one possible answer, but
• 26:13
  Speaker 1: governments themselves face limitations. Legislators may not possess the technical
• 26:18
  Speaker 1: expertise needed to evaluate rapidly changing AI systems. The debate
• 26:23
  Speaker 1: over artificial intelligence regulation has increasingly become a debate over
• 26:28
  Speaker 1: how society should manage uncertainty itself. As AI systems grow
• 26:33
  Speaker 1: more capable, Governments are being pushed to decide whether existing
• 26:37
  Speaker 1: laws are sufficient, whether entirely new rules are necessary, or
• 26:42
  Speaker 1: whether the technology should largely be governed through voluntary safeguards
• 26:46
  Speaker 1: developed by the companies building it. Sam Altman's recent comments
• 26:50
  Speaker 1: place OpenAI firmly inside that wider argument, while also highlighting
• 26:55
  Speaker 1: an important division within the AI industry over how much
• 26:58
  Speaker 1: regulation is appropriate and how quickly it should be introduced.
• 27:03
  Speaker 1: At the center of OpenAI's approach is the belief that
• 27:06
  Speaker 1: regulation should protect society from serious dangers without unnecessarily restricting
• 27:12
  Speaker 1: technological progress. That philosophy reflects the company's broader view that
• 27:17
  Speaker 1: artificial intelligence could produce benefits on a scale large enough
• 27:21
  Speaker 1: to justify taking some risks. From this perspective, the objective
• 27:26
  Speaker 1: is not to create an environment in which every possible
• 27:29
  Speaker 1: negative outcome is prevented. Instead, The goal is to identify
• 27:34
  Speaker 1: the most dangerous forms of misuse and develop safeguards that
• 27:38
  Speaker 1: reduce those risks while allowing useful AI.
• 27:41
  Speaker 2: Systems to continue improving.
• 27:43
  Speaker 1: This approach differs from a more cautious philosophy that places
• 27:47
  Speaker 1: greater emphasis on slowing the development of increasingly powerful systems
• 27:52
  Speaker 1: until their potential dangers are better understood. Supporters of that
• 27:56
  Speaker 1: position argue that the consequences of advanced AI could become
• 27:59
  Speaker 1: much more serious as systems gain greater autonomy, stronger reasoning abilities,
• 28:05
  Speaker 1: and the capacity to interact with real-world infrastructure. In their view,
• 28:10
  Speaker 1: the pace of technological progress itself may become a risk
• 28:14
  Speaker 1: if safety research and institutional oversight cannot keep up. The
• 28:18
  Speaker 1: disagreement is significant because both sides are responding to the
• 28:22
  Speaker 1: same basic reality— No one can predict exactly how advanced
• 28:26
  Speaker 1: AI will evolve. The difference lies in how that uncertainty
• 28:30
  Speaker 1: should influence policy. A more innovation-focused approach treats uncertainty as
• 28:36
  Speaker 1: something to manage while continuing forward. A more cautious approach
• 28:40
  Speaker 1: treats uncertainty as a reason to create additional barriers before
• 28:44
  Speaker 1: the technology reaches a more powerful stage. OpenAI's lighter-touch philosophy
• 28:50
  Speaker 1: is influenced in part by the belief that overly restrictive
• 28:53
  Speaker 1: regulation could have unintended consequences. Complex regulations can raise the
• 28:58
  Speaker 1: cost of developing AI and make it harder for smaller companies,
• 29:03
  Speaker 1: independent researchers, and new competitors to enter the market. If
• 29:07
  Speaker 1: compliance becomes expensive or difficult, technological development could become even
• 29:12
  Speaker 1: more concentrated among the largest corporations, potentially reducing competition rather
• 29:18
  Speaker 1: than improving safety. There is also a concern that rigid
• 29:22
  Speaker 1: regulations can become outdated quickly. Artificial intelligence does not develop
• 29:27
  Speaker 1: according to the timelines of traditional legislation. A law designed
• 29:31
  Speaker 1: around the capabilities of today's models could become irrelevant as
• 29:35
  Speaker 1: newer systems appear. Policymakers therefore face a difficult challenge They
• 29:40
  Speaker 1: must create rules that are strong enough to address genuine risks,
• 29:44
  Speaker 1: but flexible enough to remain useful as the technology changes.
• 29:49
  Speaker 1: From OpenAI's perspective, this can mean favoring standards that focus
• 29:54
  Speaker 1: on outcomes rather than attempting to regulate every technical detail.
• 29:58
  Speaker 1: Instead of prescribing exactly how an AI model must be constructed,
• 30:03
  Speaker 1: policymakers could establish requirements concerning security, transparency, testing, privacy, accountability,
• 30:11
  Speaker 1: and the handling of dangerous capabilities. Companies would retain room
• 30:16
  Speaker 1: to innovate while still being expected to meet certain safety obligations.
• 30:20
  Speaker 1: Critics of a lighter regulatory approach, however, argue that relying
• 30:25
  Speaker 1: too heavily on companies to determine their own limits may
• 30:28
  Speaker 1: create a conflict of interest. AI developers have strong commercial
• 30:33
  Speaker 1: incentives to release increasingly capable products. A company that delays
• 30:38
  Speaker 1: an important model for safety reasons could potentially lose customers
• 30:42
  Speaker 1: or market share to a competitor willing to move faster.
• 30:45
  Speaker 1: Even companies acting in good faith could therefore face pressure
• 30:49
  Speaker 1: that makes voluntary restraint difficult. This is where the idea
• 30:53
  Speaker 1: of industry-wide standards becomes important. If one company voluntarily adopts
• 30:58
  Speaker 1: stricter safeguards while its competitors do not, the responsible company
• 31:03
  Speaker 1: may bear the immediate costs while receiving few short-term benefits.
• 31:07
  Speaker 1: Coordinated rules could reduce that problem by establishing a common
• 31:11
  Speaker 1: baseline that every major developer must meet. The difficult question
• 31:16
  Speaker 1: is determining who should establish those standards and how they
• 31:20
  Speaker 1: should be enforced. Government regulation is one possible answer, but
• 31:24
  Speaker 1: governments themselves face limitations. Legislators may not possess the technical
• 31:30
  Speaker 1: expertise needed to evaluate rapidly changing AI systems.
• 31:34
  Speaker 2: The risks associated with.
• 31:35
  Speaker 1: Artificial intelligence are not limited to mistakes made by the
• 31:39
  Speaker 1: systems themselves. Another concern is what happens when increasingly capable
• 31:44
  Speaker 1: AI tools are deliberately used by people who intend to
• 31:47
  Speaker 1: cause harm. As advanced models become more accessible, the technology
• 31:52
  Speaker 1: could give individuals and organizations new ways to conduct fraud,
• 31:57
  Speaker 1: spread false information, launch cyberattacks, manipulate people, or interfere with
• 32:03
  Speaker 1: critical systems. This creates one of the most difficult challenges
• 32:07
  Speaker 1: facing AI developers. How can powerful technology remain broadly useful
• 32:12
  Speaker 1: while preventing its capabilities from becoming tools for malicious activity?
• 32:17
  Speaker 1: The problem begins with the dual-use nature of artificial intelligence.
• 32:22
  Speaker 1: Many capabilities that have legitimate purposes can also be adapted
• 32:26
  Speaker 1: for harmful ones. A system capable of analyzing large quantities
• 32:30
  Speaker 1: of information can help a researcher identify useful patterns, but
• 32:35
  Speaker 1: the same ability could potentially be used to organize sensitive
• 32:38
  Speaker 1: information for malicious purposes. A powerful language model can help
• 32:43
  Speaker 1: a business communicate with customers Yet similar capabilities could be
• 32:48
  Speaker 1: used to produce convincing fraudulent messages at enormous scale. That
• 32:52
  Speaker 1: makes AI safety different from simply preventing an obvious technical malfunction.
• 32:57
  Speaker 1: Developers must consider not only whether a system can behave incorrectly,
• 33:02
  Speaker 1: but whether a legitimate capability could be deliberately redirected toward
• 33:06
  Speaker 1: an undesirable objective. The challenge becomes more difficult as models.
• 33:12
  Speaker 2: Grow more flexible.
• 33:13
  Speaker 1: Because the same system can potentially be used in thousands
• 33:17
  Speaker 1: of different environments. Fraud is one of the clearest examples.
• 33:21
  Speaker 1: Artificial intelligence can generate convincing text, imitate communication styles, and
• 33:27
  Speaker 1: rapidly personalize messages for different audiences. Criminals could potentially use
• 33:32
  Speaker 1: those capabilities to make scams more persuasive and less obviously suspicious.
• 33:38
  Speaker 1: Instead of sending the same fraudulent message to thousands of people,
• 33:42
  Speaker 1: Automated systems could help create customized communications designed around specific targets.
• 33:48
  Speaker 1: The scale of the problem is particularly important. Human criminals
• 33:52
  Speaker 1: are limited by time. They must write messages, conduct research,
• 33:57
  Speaker 1: identify targets, and manage communications themselves. AI can automate portions
• 34:02
  Speaker 1: of those processes, potentially allowing a small number of individuals
• 34:07
  Speaker 1: to conduct operations at a much larger scale. The danger
• 34:10
  Speaker 1: is not necessarily that AI creates entirely new crimes, but
• 34:15
  Speaker 1: that it lowers the cost and effort required to carry
• 34:18
  Speaker 1: out existing ones. Cybersecurity presents another major concern. Artificial intelligence
• 34:24
  Speaker 1: can help defenders identify vulnerabilities, monitor networks, and respond to
• 34:29
  Speaker 1: suspicious activity. the same underlying capabilities could potentially be exploited
• 34:35
  Speaker 1: by attackers. AI could assist with reconnaissance, automate certain technical tasks,
• 34:41
  Speaker 1: or help malicious actors adapt their methods more quickly. This
• 34:45
  Speaker 1: creates a technological race between defenders and attackers, with both
• 34:49
  Speaker 1: sides seeking to use increasingly sophisticated systems. The cybersecurity problem
• 34:55
  Speaker 1: also illustrates why access controls matter. If an AI system
• 34:59
  Speaker 1: is connected to external tools, databases, or software environments, its
• 35:04
  Speaker 1: potential impact can increase substantially. A model capable only of
• 35:09
  Speaker 1: generating information is one thing. A model that can interact
• 35:13
  Speaker 1: directly with computer systems, execute tasks, and maintain long sequences
• 35:18
  Speaker 1: of activity presents a different level of risk. This is
• 35:22
  Speaker 1: why developers increasingly have to think about what their systems
• 35:25
  Speaker 1: are allowed to access. Limiting permissions can reduce potential damage
• 35:30
  Speaker 1: if a system is misused or compromised. But strict restrictions
• 35:34
  Speaker 1: can also.
• 35:34
  Speaker 2: Make AI less useful.
• 35:37
  Speaker 1: A business may want an AI assistant that can search
• 35:39
  Speaker 1: internal databases, update records, and interact with software automatically. Removing
• 35:45
  Speaker 1: all such access may improve security while defeating much of
• 35:49
  Speaker 1: the purpose of advanced automation. The same dilemma appears in
• 35:53
  Speaker 1: critical infrastructure. AI could eventually play a larger role in
• 35:58
  Speaker 1: managing transportation, energy, telecommunications, manufacturing, and other essential systems. Automation
• 36:06
  Speaker 1: could make these environments more efficient, but it could also
• 36:09
  Speaker 1: create new vulnerabilities if attackers gain control of the systems
• 36:13
  Speaker 1: or manipulate the data on which they depend. The consequences
• 36:17
  Speaker 1: of a successful attack against critical infrastructure are could be
• 36:21
  Speaker 1: significantly greater than those of a conventional online scam. A
• 36:25
  Speaker 1: disruption to an important service could affect large populations and
• 36:30
  Speaker 1: create economic damage that extends well beyond the original target.
• 36:35
  Speaker 1: That possibility makes security a central part of AI deployment
• 36:38
  Speaker 1: rather than an optional feature added later. Information manipulation creates
• 36:43
  Speaker 1: another category of concern. Generative AI can produce large amounts
• 36:48
  Speaker 1: of convincing content quickly, including text, images, audio, and other
• 36:53
  Speaker 1: forms of media. As artificial intelligence becomes.
• 36:57
  Speaker 2: More powerful, the question of who should.
• 37:00
  Speaker 1: Establish its limits is becoming almost as important as the
• 37:03
  Speaker 1: technology itself. Governments, technology companies, Researchers and members of the
• 37:09
  Speaker 1: public all have different interests in determining how AI should
• 37:13
  Speaker 1: be developed and deployed. Companies want enough freedom to innovate
• 37:17
  Speaker 1: and compete. Governments want to protect national security and the
• 37:21
  Speaker 1: public interest. Researchers want space to explore new capabilities while
• 37:26
  Speaker 1: improving safety. Citizens want access to useful technology without being
• 37:31
  Speaker 1: exposed to unnecessary risks. Bringing all of those priorities together
• 37:36
  Speaker 1: may be one of the most difficult governance challenges created
• 37:39
  Speaker 1: by the AI revolution. The debate begins with a basic problem.
• 37:44
  Speaker 1: Artificial intelligence is developing faster than traditional political institutions are
• 37:50
  Speaker 1: accustomed to handling. Governments typically create regulations through lengthy processes
• 37:55
  Speaker 1: involving research, consultation, legislation, and implementation. Technology companies, meanwhile, can
• 38:03
  Speaker 1: release software updates and new capabilities on much shorter timelines.
• 38:09
  Speaker 1: This difference in speed makes it difficult for lawmakers to
• 38:12
  Speaker 1: respond to every new development individually. One response is to
• 38:16
  Speaker 1: give technology companies greater responsibility for regulating themselves. The argument
• 38:22
  Speaker 1: is that developers possess the technical expertise required to understand
• 38:26
  Speaker 1: their systems in ways that policymakers may not. Engineers can
• 38:30
  Speaker 1: examine model behavior identify vulnerabilities, conduct safety testing, and develop
• 38:36
  Speaker 1: technical measures that are difficult to translate into ordinary legislation.
• 38:41
  Speaker 1: Self-governance also has the advantage of speed. A company can
• 38:45
  Speaker 1: change a safety policy, restrict a capability, or modify a
• 38:50
  Speaker 1: system without waiting for a lengthy legislative process. In a
• 38:54
  Speaker 1: rapidly changing field, that flexibility can be valuable. It allows
• 38:59
  Speaker 1: developers to respond to newly discovered problems as they appear.
• 39:03
  Speaker 1: But self-regulation has an obvious weakness. Technology companies have commercial interests.
• 39:09
  Speaker 1: The same organizations responsible for developing safety standards may also
• 39:14
  Speaker 1: benefit financially from releasing more capable products and attracting more users.
• 39:19
  Speaker 1: Even when executives genuinely prioritize safety, the pressure of competition
• 39:24
  Speaker 1: can create difficult incentives. This is particularly important in an
• 39:29
  Speaker 1: industry where technological leadership can translate into enormous economic value.
• 39:34
  Speaker 1: A company that slows development to conduct additional safety research
• 39:38
  Speaker 1: may be acting responsibly, but a competitor that moves faster
• 39:42
  Speaker 1: could capture customers and investment. That creates a collective action problem.
• 39:47
  Speaker 1: Everyone may benefit from stronger safety standards, but individual companies
• 39:52
  Speaker 1: may have incentives to avoid bearing the full cost of them.
• 39:56
  Speaker 1: Government regulation offers a potential solution by establishing common requirements
• 40:01
  Speaker 1: for the entire industry. Instead of asking companies to decide
• 40:05
  Speaker 1: individually how much caution is enough, governments can establish minimum
• 40:10
  Speaker 1: standards that apply across the market. Such rules could address
• 40:14
  Speaker 1: issues such as security testing. data protection, transparency, risk assessments,
• 40:20
  Speaker 1: incident reporting, and accountability. The advantage of this approach is consistency.
• 40:26
  Speaker 1: If every major company is required to meet the same
• 40:29
  Speaker 1: basic safety standards, companies are less likely to compete by
• 40:33
  Speaker 1: weakening safeguards. Regulation can also give citizens legal protections that
• 40:38
  Speaker 1: voluntary corporate policies might not provide. However, government intervention comes
• 40:44
  Speaker 1: with its own challenges. Artificial intelligence is a technically complicated field,
• 40:50
  Speaker 1: and poorly designed rules could have consequences that lawmakers did
• 40:54
  Speaker 1: not intend. A requirement that appears reasonable on paper could
• 40:58
  Speaker 1: make certain beneficial systems unnecessarily difficult to develop. Regulations can
• 41:04
  Speaker 1: also favor large corporations if compliance costs become so high
• 41:08
  Speaker 1: that smaller companies cannot afford them. That raises a particularly
• 41:13
  Speaker 1: important concern about competition. The stated purpose of AI regulation
• 41:18
  Speaker 1: may be to improve safety, but regulation can unintentionally increase
• 41:22
  Speaker 1: concentration if only the biggest companies can afford to comply
• 41:26
  Speaker 1: with detailed requirements. Large technology companies may have legal departments,
• 41:31
  Speaker 1: compliance teams, researchers, and infrastructure capable of meeting complex standards.
• 41:38
  Speaker 2: Smaller competitors may not.
• 41:40
  Speaker 1: The result could be a paradox in which regulation designed
• 41:43
  Speaker 1: to prevent excessive corporate power actually strengthens the companies already
• 41:48
  Speaker 1: at the top of the industry. Policymakers, therefore, have to
• 41:52
  Speaker 1: consider not only whether a rule improves safety, but also
• 41:55
  Speaker 1: how it affects competition and access. Another approach is to
• 41:59
  Speaker 1: create independent technical institutions capable of evaluating AI systems without
• 42:04
  Speaker 1: being directly controlled by the companies that develop them. Independent
• 42:09
  Speaker 1: testing could potentially provide greater credibility than internal assessments alone.
• 42:14
  Speaker 1: The debate over artificial intelligence is no longer confined to
• 42:18
  Speaker 1: technology companies competing for customers. It has increasingly become part
• 42:23
  Speaker 1: of a broader international contest over economic strength, scientific leadership,
• 42:29
  Speaker 1: and strategic influence. As advanced AI systems become more capable,
• 42:34
  Speaker 1: governments are beginning to treat control over the underlying technology
• 42:38
  Speaker 1: as a matter of national interest. That transformation adds another
• 42:42
  Speaker 1: layer to the concerns surrounding AI risk because decisions about
• 42:46
  Speaker 1: safety are now being made in an environment where countries
• 42:49
  Speaker 1: also fear falling behind their competitors. The basic strategic calculation
• 42:54
  Speaker 1: is straightforward. Artificial intelligence has the potential to improve productivity,
• 43:00
  Speaker 1: accelerate scientific research, transform industries, and strengthen technological capabilities across
• 43:07
  Speaker 1: a wide range of sectors. A country that develops leading
• 43:10
  Speaker 1: AI systems could potentially gain advantages not only in commercial markets,
• 43:15
  Speaker 1: but also in areas connected to national security and public administration.
• 43:21
  Speaker 1: That makes AI leadership increasingly comparable to other technologies that
• 43:25
  Speaker 1: have historically carried geopolitical significance. For governments, therefore, slowing AI
• 43:32
  Speaker 1: development can be a complicated proposition. Leaders may recognize genuine
• 43:37
  Speaker 1: safety concerns while simultaneously believing that rapid progress is necessary
• 43:42
  Speaker 1: to preserve economic and strategic competitiveness. If one country imposes
• 43:47
  Speaker 1: extensive restrictions on its domestic technology companies while another allows
• 43:52
  Speaker 1: developers to move rapidly, the first government may worry that
• 43:56
  Speaker 1: it is effectively surrendering an important technological advantage. This creates
• 44:01
  Speaker 1: a tension between safety and competition that is difficult to resolve.
• 44:06
  Speaker 1: In theory, countries could cooperate to establish common standards that
• 44:10
  Speaker 1: reduce the most serious risks. In practice, governments may hesitate
• 44:15
  Speaker 1: to accept rules that could limit their access to technology
• 44:19
  Speaker 1: that they view as strategically valuable. The result is an
• 44:23
  Speaker 1: environment in which nations may simultaneously seek cooperation on AI
• 44:28
  Speaker 1: safety and compete aggressively for leadership. The international AI race
• 44:33
  Speaker 1: also has an economic dimension. Advanced AI requires substantial investment
• 44:38
  Speaker 1: in computing infrastructure, data centers, specialized hardware, and highly skilled researchers.
• 44:45
  Speaker 1: Countries with strong technology sectors, reliable energy supplies and large
• 44:50
  Speaker 1: capital markets may have structural advantages. Governments are therefore increasingly
• 44:56
  Speaker 1: interested in attracting AI investment, developing domestic expertise, and expanding
• 45:01
  Speaker 1: computing capacity. This can create a feedback loop. Companies are
• 45:06
  Speaker 1: drawn to countries with large pools of talent and infrastructure.
• 45:10
  Speaker 1: Governments then invest more heavily to attract those companies. Successful
• 45:15
  Speaker 1: AI development can generate additional wealth, which allows further investment
• 45:19
  Speaker 1: in research and infrastructure. Countries that fall behind may find
• 45:24
  Speaker 1: it increasingly difficult to catch up because the leading ecosystems
• 45:28
  Speaker 1: continue to accumulate resources and expertise. That possibility has significant
• 45:34
  Speaker 1: implications for the global balance of economic power. AI could
• 45:38
  Speaker 1: potentially increase productivity across a wide range of industries, meaning
• 45:43
  Speaker 1: countries that integrate the technology effectively may experience faster economic growth.
• 45:49
  Speaker 1: They could attract more investment, develop new industries, and gain
• 45:53
  Speaker 1: greater influence over global technology standards. But the benefits of
• 45:57
  Speaker 1: AI leadership may not be distributed evenly. Countries without sufficient
• 46:02
  Speaker 1: computing infrastructure, technical expertise, or investment capacity could become dependent
• 46:08
  Speaker 1: on foreign AI providers. That dependence could extend beyond software.
• 46:13
  Speaker 1: Nations might rely on overseas companies for cloud computing, advanced chips,
• 46:18
  Speaker 1: AI models, and critical technical services. Such dependency could become
• 46:23
  Speaker 1: a strategic vulnerability. If access to advanced AI technology were
• 46:28
  Speaker 1: restricted because of trade disputes, security concerns, or geopolitical tensions,
• 46:34
  Speaker 1: countries that lack domestic alternatives could face serious economic consequences.
• 46:39
  Speaker 1: AI policy may therefore increasingly be linked to questions of
• 46:43
  Speaker 1: technological sovereignty. The supply chain behind advanced AI illustrates this issue.
• 46:48
  Speaker 1: Building powerful systems depends on a complex international network involving
• 46:53
  Speaker 1: semiconductor manufacturing, chip design, data centers, energy production, networking equipment,
• 47:01
  Speaker 1: and software. Disruption at any point in that chain can
• 47:04
  Speaker 1: affect the ability to develop and deploy advanced systems. Governments
• 47:09
  Speaker 1: are consequently looking beyond the AI models themselves. They are
• 47:13
  Speaker 1: paying greater attention to semiconductor production, advanced computing infrastructure, energy availability,
• 47:20
  Speaker 1: and the movement of technological expertise. The competition is not
• 47:24
  Speaker 1: only to build the smartest model. It is also to
• 47:27
  Speaker 1: secure the physical and economic foundations required to keep improving
• 47:32
  Speaker 1: those models. The debate over artificial intelligence risks ultimately leads
• 47:37
  Speaker 1: to a question that is far larger than any single
• 47:40
  Speaker 1: company interview or technology announcement. Can humanity continue to control
• 47:45
  Speaker 1: AI as the systems become more capable? The answer remains uncertain.
• 47:51
  Speaker 1: Developers are building increasingly sophisticated models while governments are still
• 47:55
  Speaker 1: working out how to regulate them. Researchers are trying to
• 47:59
  Speaker 1: understand their behavior and businesses are rapidly integrating them into
• 48:03
  Speaker 1: the economy. The speed of these changes means that decisions
• 48:07
  Speaker 1: being made today could shape how societies interact with artificial
• 48:11
  Speaker 1: intelligence for decades. The challenge is not simply to prevent
• 48:15
  Speaker 1: AI from becoming dangerous.
• 48:17
  Speaker 2: It is to.
• 48:17
  Speaker 1: Create a framework in which powerful systems remain useful, accountable,
• 48:22
  Speaker 1: and subject to meaningful human oversight. That sounds straightforward, but
• 48:27
  Speaker 1: achieving it becomes harder as capabilities improve. A system that
• 48:32
  Speaker 1: can perform a limited task under close supervision may present
• 48:36
  Speaker 1: relatively manageable risks. A system capable of handling complex objectives
• 48:41
  Speaker 1: using multiple tools and operating for extended periods could require
• 48:46
  Speaker 1: an entirely different approach to safety. This is why the
• 48:49
  Speaker 1: discussion surrounding Sam Altman and AI risk extends far beyond
• 48:53
  Speaker 1: the immediate controversy over whether some risks should be accepted.
• 48:58
  Speaker 1: The deeper issue is how society should make decisions when
• 49:01
  Speaker 1: technological progress cannot be predicted with certainty. Waiting for perfect
• 49:05
  Speaker 1: information may mean waiting too long. Acting too aggressively could
• 49:10
  Speaker 1: limit benefits that might have improved people's lives. The future
• 49:14
  Speaker 1: of AI governance will therefore depend on managing uncertainty rather
• 49:18
  Speaker 1: than eliminating it.
• 49:20
  Speaker 2: One of the most important.
• 49:21
  Speaker 1: Tasks will be establishing clear boundaries around the most dangerous capabilities.
• 49:26
  Speaker 1: Not every AI system presents the same level of risk.
• 49:30
  Speaker 1: A tool designed to summarize documents is fundamentally different from
• 49:34
  Speaker 1: a highly autonomous system capable of controlling software, interacting with
• 49:39
  Speaker 1: sensitive infrastructure, or carrying out complex operations without continuous human intervention.
• 49:45
  Speaker 1: This suggests that future regulation may need to become more
• 49:49
  Speaker 1: flexible and risk-based. The strongest safeguards could be reserved for
• 49:54
  Speaker 1: systems with the greatest potential impact, while lower-risk applications remain
• 49:59
  Speaker 1: subject to simpler requirements. Such an approach could reduce the
• 50:03
  Speaker 1: possibility that broad regulation slows harmless innovation while still recognizing
• 50:08
  Speaker 1: that highly capable systems require greater scrutiny. Testing will become
• 50:13
  Speaker 1: central to that process. Developers may need to evaluate systems
• 50:17
  Speaker 1: not only for accuracy and performance, but also for how
• 50:21
  Speaker 1: they behave under unusual conditions. they will have to look
• 50:24
  Speaker 1: for unexpected strategies, security weaknesses, deceptive behavior, dangerous combinations of capabilities,
• 50:32
  Speaker 1: and other signs that a system may not behave as
• 50:35
  Speaker 1: intended outside controlled environments. Yet testing cannot provide absolute certainty.
• 50:41
  Speaker 1: No evaluation can reproduce every situation an AI system may encounter.
• 50:46
  Speaker 2: In the real world.
• 50:47
  Speaker 1: This means safety cannot be treated as a one-time certification
• 50:51
  Speaker 1: obtained before release. It may need to become an ongoing
• 50:55
  Speaker 1: process involving continuous monitoring, periodic reassessment, and rapid intervention when
• 51:01
  Speaker 1: new risks are discovered. That shift could fundamentally change how
• 51:05
  Speaker 1: advanced AI products are managed. Traditional software is often released
• 51:10
  Speaker 1: and then updated as problems emerge. Future high-capability AI systems
• 51:15
  Speaker 1: may require much more intensive supervision because their behavior can
• 51:19
  Speaker 1: vary depending on context, inputs, and the tools to which
• 51:22
  Speaker 1: they have access. Human oversight will remain an important part
• 51:26
  Speaker 1: of that system, but humans themselves will have to adapt.
• 51:30
  Speaker 1: Supervisors need to understand both the capabilities and the limitations
• 51:35
  Speaker 1: of AI. They must be able to recognize when a
• 51:38
  Speaker 1: system is producing an unreliable result, when its confidence is misplaced,
• 51:43
  Speaker 1: and when an apparently useful action could produce unintended consequences.
• 51:47
  Speaker 1: The danger is that humans may gradually become too trusting.
• 51:51
  Speaker 1: If an AI system performs correctly thousands of times, users
• 51:56
  Speaker 1: may become less likely to question it on the one
• 51:58
  Speaker 1: occasion when it fails. This is particularly concerning in high-stakes
• 52:02
  Speaker 1: environments where a single error can have serious consequences. Maintaining
• 52:07
  Speaker 1: meaningful human control, therefore, requires more than placing a person
• 52:11
  Speaker 1: somewhere in the approval chain. The human must have enough information, authority,
• 52:16
  Speaker 1: and time to intervene effectively. If a system operates faster
• 52:20
  Speaker 1: than people can reasonably monitor it, nominal human oversight may
• 52:25
  Speaker 1: provide little practical protection. There is also a need for
• 52:29
  Speaker 1: strong institutional accountability. AI developers should not be the only
• 52:33
  Speaker 1: actors deciding whether their systems are safe. Independent researchers, auditors, regulators,
• 52:39
  Speaker 1: and other oversight bodies could play a role in evaluating
• 52:43
  Speaker 1: important systems and challenging industry assumptions.