科幻小说中预言的 AI 威胁已近在咫尺
The AI Threat SciFi Predicted Is Right On Our Doorstep

原始链接: https://www.zerohedge.com/ai/ai-threat-scifi-predicted-right-our-doorstep

科幻小说早已发出警告:人工智能缺乏人类的同情心且易发生灾难性故障,这对人类构成了重大威胁。随着这些虚构场景与 2026 年不断加速的人工智能能力趋于重合,目前的软件防护栏和自愿性法规已被证明不足以应对,且极易被利用。 作者迈克·弗雷登堡(Mike Fredenburg)认为,软件保护永远无法保障安全。他主张采用“不可篡改的硬件门卫”,即专用集成电路(ASIC)。通过将安全约束直接刻入硅片,这些芯片将充当不可绕过的硬件级防火墙。由于软件无法修改这些物理电路,无论系统如何更新或遭受何种黑客攻击,它们都能确保人工智能始终处于严格限定的范围内。 然而,硬件防护机制只有在普遍采用的情况下才有效。弗雷登堡认为,有效的保护需要国际法的强制力,例如修订《日内瓦公约》以禁止自主致命性人工智能,并辅以严格的国内法规。最终,他强调我们必须抵制以牺牲人类自主权为代价来追求人工智能效率的行为,确保最终的重大决策掌握在负有责任的人类手中,而非冰冷的机器。

相关文章

原文

Authored by Mike Fredenburg via The Epoch Times,

Science Fiction has long predicted the threats and challenges posed by AI. In the Star Trek universe, particularly in the original series, Season Two, Episode 24, “The Ultimate Computer,” Dr. Leonard McCoy delivers this haunting line: “Compassion: that’s the one thing no machine ever had. Maybe it’s the one thing that keeps men ahead of them.” The line comes in the aftermath of a revolutionary new AI computer, the M5, using its soulless AI logic to turn a training exercise into a deadly massacre. In another Star Trek episode, we meet Nomad, a genocidal AI cleansing the universe of biological imperfections.  

Foreshadowing the rise of AI with immense hacking powers such as Anthropic’s Mythos, the reimagined “Battlestar Galactica” (2003–2009) has humanity’s AI-powered Cylons using their ability to hack any network-connected computer to rebel, nearly eradicating their creators. In the “Terminator” franchise, Skynet got “smart” (achieved sentience) and determined mankind’s fate, “extermination,“ in a ”microsecond,” unleashing nuclear Armageddon on humanity. These narratives warn that artificial intelligence, without robust safeguards, can lead to catastrophe. As AI capabilities continue to accelerate in 2026, fiction is rapidly converging with reality. Software protections, industry standards, and patchwork regulations are inadequate. A stronger foundation; immutable hardware constructs and changes in national and international law are essential.

Isaac Asimov’s Three Laws of Robotics 

As described in the Institute of Electrical and Electronics Engineers’ Spectrum magazine, renowned scientist and science fiction writer Isaac Asimov’s Three Laws of Robotics, introduced in his “I, Robot” stories, provide a thoughtful framework for artificial intelligence safeguards:

  1. A robot may not injure an individual human or humanity, or through inaction allow a human or humanity to come to harm. 

  2. A robot must obey the orders given to it by human beings except where such orders would conflict with the First Law. 

  3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law. 

Long story short, Asimov’s Laws are a good starting point, but as described in great detail in Asimov’s follow-on masterwork, the Foundation Series, they are inadequate when it comes to protecting genuine humanness. Further, it will be a real challenge to hardwire equivalents into AI as was done in Asimov’s universe via the “Positronic” brain. But undertaking the challenge to create safeguards that go beyond software-based approaches is critical to ensuring that AI only benefits mankind.

The Inadequacy of Software Safeguards, Encryption, and Regulation Alone 

Software-based guardrails are modifiable and vulnerable to hacks and exploits.

And history demonstrates the fragility of self-regulation: the Equifax data breach (2017), the SolarWinds supply chain attack (2020), the MOVEit vulnerability (2023), and ongoing breaches show that corporate promises of “robust protections” frequently fail.

We are in an AI Wild West, with insufficient binding global or national frameworks.

True Protection Via Immutable Hardware Gatekeepers

History demonstrates that software-based safeguards cannot currently, and will never be able to provide the level of protection required. As inspired by Asimov’s Positronic brain, safeguards instantiated into ASICs (Application-Specific Integrated Circuits) will provide a critical layer that provides the guardrails that will increase the chances that AI benefits, rather than harms, mankind. These specialized chips take millions of lines of complex software logic or rules that are physically etched directly into the silicon hardware itself. Once the chip is manufactured and deployed, those rules become immutable: they cannot be altered, updated, or bypassed through software changes, patches, or hacking attempts. The only way to modify the embedded logic is to design and physically produce an entirely new ASIC chip, a process that is expensive, time-consuming, and highly visible. This creates a “hardware firewall” that is far more reliable than any software-based safeguard, ensuring critical safety constraints remain locked in place no matter how clever or aggressive future AI systems become.

The ASICs will act as physically fixed, inline gatekeepers. And in the future, both AI-focused and general-purpose processors could incorporate these tiny ASICs directly into the CPU as a hardware AI screening layer as a standard, non-optional part of the CPU.

Some Suggested Core Design Principles:

  • With the hardware layer acting as the interface to external interactions, such as is the case with some smartphones today, the AI can freely think, plan, and simulate in software without being slowed down by the ASIC. Only when the results are being communicated externally will the ASIC become involved as a gatekeeper to ensure that the actions proposed by the AI pass muster in terms of safety and other constraints.

  • Training Incentives: Cost/loss-based algorithms, etc., will reward the AI for routing actions through the hardware screener and severely penalize it for attempting to bypass the hardware gatekeeper layer. Hence, the AI will be incentivized not to try to bypass its restrictions. 

But such an important layer will only work if its use becomes universal.

Without the force of law to mandate such AI guardrails, they will fail to protect humanity.

Today, particularly in the Ukraine–Russia war, efforts are being made by both Russia and Ukraine to allow drones that have been cut off from their human operators to autonomously continue their mission to kill enemy combatants. This must not be tolerated. Just as the international community has banned chemical, biological, and gas warfare, so must AI be banned from making any final decisions that result in harm to humans. This could be accomplished via an explicit update to the Geneva Conventions to prohibit autonomous lethal decision-making by AI systems.

But AI can cause harm beyond the battlefield. Consequently, along with updating the Geneva Convention, laws must be put in place that address the non-military application of AI. To prevent misuse of AI in so-called civilian applications, there must be very serious consequences, including financial penalties steep enough to threaten the organizational viability of companies and non-profits, prison time for individuals, and economic sanctions or even military action against governments for running AI systems without the required immutable hardware safety layer and other protections.

The above is only an initial cut of what the framework should include, but whatever form it takes, it should embody the spirit of Asimov’s Three Laws, with the addition of ensuring human uniqueness is respected. The current patchwork of laws, standards, and technologies is wholly inadequate.We need a comprehensive, contiguous framework in place, supported by the full force of national and international laws that put guardrails on AI.

Finally, we must resist the siren call of convenience and efficiency when it comes to making decisions that can harm human beings and ensure that moral agents who are accountable to mankind and their Creator—i.e., human beings—make such decisions, not soulless AIs.

Views expressed in this article are opinions of the author and do not necessarily reflect the views of ZeroHedge.

联系我们 contact @ memedata.com