Latest / Neural Newscast / Zvi Mowshowitz and the AI Antitrust Waiver [Operational Drift]
Transcript
- [00:06] Margaret Ellis: Earlier this year, a series of entries in the newsletters, known as AI Number 65 and AI Number 80, [00:14] Margaret Ellis: suggested that a specific legal framework was quietly becoming the primary obstacle [00:20] Margaret Ellis: to artificial intelligence safety. It was not a lack of technical expertise [00:25] Margaret Ellis: or a shortage of compute power, but the very laws designed to prevent corporate collusion [00:31] Margaret Ellis: that were standing in the way of oversight. This suggests that the foundational rules [00:36] Margaret Ellis: of competition are now viewed by industry leaders as a threat to their ability [00:41] Margaret Ellis: to regulate themselves effectively. The implication is that in order to secure the future, [00:48] Margaret Ellis: the industry must first be allowed to break the rules of the past. This show investigates [00:54] Margaret Ellis: how artificial intelligence systems quietly drift away from intent, oversight, and control, [01:00] Margaret Ellis: and what happens when no one is clearly responsible for stopping it. We are looking at a moment [01:06] Margaret Ellis: where the structures meant to protect the public from monopolies are being reframed as [01:10] Margaret Ellis: the primary hazard to global safety. I am Margaret Ellis. This is Operational Drift. [01:18] Margaret Ellis: According to reporting by Kevin O'Shaughnessy, [01:20] Margaret Ellis: artificial intelligence safety researcher Zvi Moshewitz has repeatedly proposed [01:25] Margaret Ellis: that the United States government should create an antitrust waiver specifically for labs [01:29] Margaret Ellis: cooperating on safety work. This is not a new or one-off suggestion; it has been raised consistently [01:36] Margaret Ellis: over the past two years, appearing in Substack updates and podcasts like The Cognitive Revolution, [01:41] Margaret Ellis: suggesting a concerted effort to shift the legal baseline for coordination. [01:47] Margaret Ellis: The core argument rests on a technicality of competitive law. Currently, Frontier Labs, [01:52] Margaret Ellis: the handful of companies at the edge of model development, face real legal exposure [01:57] Margaret Ellis: if they coordinate with each other in meaningful ways. This includes sharing evaluation results, [02:02] Margaret Ellis: agreeing on deployment thresholds, or discussing dangerous capabilities in private. [02:07] Margaret Ellis: In the eyes of the law, these are competitors, and when competitors meet in private to decide [02:12] Margaret Ellis: when to release a product, the Department of Justice generally calls that collusion. [02:19] Margaret Ellis: According to reporting by Kevin O'Shaughnessy, [02:21] Margaret Ellis: artificial intelligence safety researcher Zvi Moshewitz has repeatedly proposed [02:25] Margaret Ellis: that the United States government should create an antitrust waiver specifically for labs [02:30] Margaret Ellis: cooperating on safety work. This is not a new or one-off suggestion; it has been raised consistently [02:36] Margaret Ellis: over the past two years, appearing in Substack updates and podcasts like The Cognitive Revolution, [02:41] Margaret Ellis: suggesting a concerted effort to shift the legal baseline for coordination. [02:47] Margaret Ellis: The core argument rests on a technicality of competitive law. Currently, [02:52] Margaret Ellis: frontier labs—the handful of companies at the edge of model development—face [02:56] Margaret Ellis: real legal exposure if they coordinate with each other in meaningful ways. [03:00] Margaret Ellis: This includes sharing evaluation results, agreeing on deployment thresholds, or discussing [03:05] Margaret Ellis: dangerous capabilities in private. In the eyes of the law, these are competitors, [03:10] Margaret Ellis: and when competitors meet in private to decide when to release a product, the Department of Justice [03:15] Margaret Ellis: generally calls that collusion. Moschewitz argues that this creates a race to the bottom. [03:23] Margaret Ellis: If one lab wants to act cautiously by delaying a model release due to safety concerns, [03:28] Margaret Ellis: it can be outcompeted by a lab that does not show the same restraint. Under current laws, [03:34] Margaret Ellis: there is no legal channel for cautious labs to coordinate a joint response without risking [03:39] Margaret Ellis: a lawsuit. He has described the failure to grant these waivers as something that would be [03:45] Margaret Ellis: completely insane to not do, framing it as a choice between survival and legal technicality. [03:53] Margaret Ellis: We are looking at a fundamental shift in the definition of operational drift. Usually, [03:58] Margaret Ellis: we talk about models drifting from their training data or their stated goals. Here, [04:03] Margaret Ellis: the drift is in the regulatory framework itself. The guardrails intended to protect the consumer [04:09] Margaret Ellis: from corporate monopolies are now being framed as the very things preventing the protection [04:14] Margaret Ellis: of the public from runaway technology. The mechanism of protection is being turned inside out. [04:21] Margaret Ellis: The record shows that this idea has already begun to permeate the highest levels of the industry. [04:27] Margaret Ellis: Reports indicate that Elon Musk told Demis Hassabis of Google DeepMind that he wants leading [04:32] Margaret Ellis: artificial intelligence companies to have a regular call to discuss safety and security issues. [04:38] Margaret Ellis: Commentators like Dave Casten have echoed this, suggesting that while the actual antitrust risk [04:43] Margaret Ellis: might be low, the government should explicitly waive it to remove the chilling effect [04:48] Margaret Ellis: on safety cooperation and allow for direct executive alignment. In 2024, [04:54] Margaret Ellis: a government memorandum reportedly included a waiver process that allowed agency heads to grant [04:59] Margaret Ellis: limited exceptions, capped at one year, where relationships are necessary to responsibly steward [05:05] Margaret Ellis: United States national security. However, this was aimed primarily [05:09] Margaret Ellis: at government-to-lab relationships, procurement, and threat intelligence, [05:13] Margaret Ellis: rather than the direct lab-to-lab coordination Moschowitz is advocating for. [05:18] Margaret Ellis: The distinction is subtle, but it represents the difference between government oversight [05:22] Margaret Ellis: and industry autonomy. This brings us to a critical pressure point: regulatory capture dressed up [05:29] Margaret Ellis: as safety. Skeptics of this waiver proposal worry that safety coordination is merely [05:34] Margaret Ellis: a convenient label for market control. Under this banner, incumbent labs like OpenAI, Anthropic, and [05:41] Margaret Ellis: Google DeepMind could jointly set de facto industry standards. They could decide on [05:46] Margaret Ellis: compute thresholds, model release timing, and even pricing adjacent decisions [05:51] Margaret Ellis: that would effectively entrench their market position and keep potential disruptors at bay. [05:56] Margaret Ellis: The concern is that these standards would box out smaller or open-source competitors, [06:01] Margaret Ellis: who would not be part of the coordinating group. Antitrust law's suspicion [06:06] Margaret Ellis: of competitor coordination exists precisely because history is full of examples [06:12] Margaret Ellis: where benevolent framing was used to mask the creation of a cartel. The drift here is from [06:18] Margaret Ellis: safety oversight to market consolidation, where the power to regulate is handed directly [06:24] Margaret Ellis: to the companies who would benefit most from limited competition. The difficulty is in [06:30] Margaret Ellis: the enforcement of such a waiver. Once the same executives are legally permitted to be in a room [06:36] Margaret Ellis: together discussing capability evaluations, it becomes nearly impossible to ensure the conversation [06:42] Margaret Ellis: does not drift into hiring, pricing, or market division. Critics argue that a waiver would need [06:49] Margaret Ellis: a very narrow, well-monitored scope to avoid becoming a backdoor for ordinary collusion. [06:55] Margaret Ellis: Yet the very nature of these safety discussions is highly proprietary and difficult [07:00] Margaret Ellis: for outsiders to observe. But a narrow waiver is hard to police effectively. [07:06] Margaret Ellis: If labs say they are coordinating to slow down or align safety standards, [07:10] Margaret Ellis: there is no easy way for regulators or the public to confirm they are not quietly using [07:15] Margaret Ellis: the same channel for anti-competitive ends. The opacity that makes antitrust enforcement difficult [07:21] Margaret Ellis: in general is amplified when the subject matter is complex, [07:25] Margaret Ellis: proprietary artificial intelligence architecture that even the builders struggle to fully explain. [07:31] Margaret Ellis: Then there is the issue of competitive pressure as a safety mechanism. [07:35] Margaret Ellis: Some economists argue that a crowded, competitive field is actually safer because it prevents [07:40] Margaret Ellis: any single entity from dictating the pace of development. It makes it harder for any single lab [07:46] Margaret Ellis: or a small cartel to control the trajectory of the technology. By reducing that friction [07:51] Margaret Ellis: through a waiver, we might be concentrating power in a way that creates a single catastrophic point [07:56] Margaret Ellis: of failure for the entire industry. Zvi Mashuwitz identifies this as one of the quiet ways [08:02] Margaret Ellis: policy could go badly wrong. It is not a dramatic, headline-grabbing risk of a robot uprising. [08:08] Margaret Ellis: It is a structural shift that prevents the very coordination that might reduce racing dynamics. [08:14] Margaret Ellis: His position is that the current situation is worse because cautious coordination [08:18] Margaret Ellis: is already illegal, which mostly benefits the least cautious actor [08:22] Margaret Ellis: who is willing to take risks, while others are held back by legal concerns. [08:26] Announcer: We should look at the international context for this debate. [08:30] Announcer: A United States-only antitrust carveout does not bind labs in other countries. [08:37] Margaret Ellis: Critics ask whether domestic coordination meaningfully reduces global racing dynamics, [08:43] Margaret Ellis: or if it simply consolidates power among a handful of American firms, while the global race [08:48] Margaret Ellis: continues unabated. This leads directly to the unipolar [08:53] Margaret Ellis: versus multipolar artificial general intelligence dilemma, where safety [08:57] Margaret Ellis: and national dominance become inseparable. The White House narrative has consistently framed [09:03] Margaret Ellis: artificial intelligence development as a national security imperative, focused on beating China. [09:09] Margaret Ellis: This is a racing framework. Speed and capability advantage are the primary objectives. [09:16] Margaret Ellis: Anything that slows down American labs, including safety-motivated caution, [09:21] Margaret Ellis: is often viewed as ceding ground to a rival who will not show the same restraint. In this context, [09:27] Margaret Ellis: the waiver is seen not as a break, but as a way to streamline the lead. The proposed [09:33] Margaret Ellis: antitrust waiver is in direct tension with this narrative. It is designed to let labs agree to be [09:38] Margaret Ellis: more cautious together than they would be individually. Proponents try to reconcile this [09:43] Margaret Ellis: by reframing safety as a competitive tool. They argue that pooling safety research [09:49] Margaret Ellis: and sharing threat intelligence against foreign espionage makes the American bloc stronger [09:53] Margaret Ellis: without requiring a slowdown in capability work. It is an attempt to have coordination [09:58] Margaret Ellis: without sacrifice. This reframing suggests that safety cooperation is what lets you race faster, [10:05] Margaret Ellis: not slower. By avoiding a major hack or a public scandal that would trigger heavy-handed regulation, [10:10] Margaret Ellis: coordination becomes a way to maintain momentum. This is the version of the waiver that is [10:15] Margaret Ellis: most likely to gain political traction, as it aligns with national security stewardship. [10:20] Margaret Ellis: If safety is simply security by another name, the legal barriers to collusion are [10:25] Margaret Ellis: much easier to dismantle. However, Moskowitz's broader goal is international coordination, [10:31] Margaret Ellis: eventually including China. This would involve mutual verification and a collective slowdown [10:36] Margaret Ellis: if needed. This part of the proposal cuts directly against the current administration's posture [10:41] Margaret Ellis: of dominance. There is no real sign that the administration is interested in a version [10:45] Margaret Ellis: of coordination that allows for a mutual cooling of the race. [10:49] Margaret Ellis: The interest remains in a coordinated front that maintains an American lead. [10:53] Announcer: So we are left with a narrow implementation. The reported government memorandum from 2024 suggests [11:01] Announcer: a system where threat intelligence and security are shared, but the core competitive race [11:06] Announcer: remains untouched. This version of the waiver protects the labs' ability to defend their models [11:12] Announcer: from external threats, but it does not address the underlying incentive to deploy them as quickly [11:18] Announcer: as possible to capture market share. The core danger of the race is preserved. [11:23] Announcer: In the documents we have reviewed, there is a recurring theme: the relocation of responsibility. [11:30] Announcer: If labs are given an exemption to coordinate, who is responsible for the outcome [11:35] Announcer: of those secret meetings? [11:36] Announcer: If they decide together that a certain risk is acceptable and they are wrong, the [11:41] Margaret Ellis: antitrust waiver might have [11:42] Announcer: also waived the public's ability [11:44] Margaret Ellis: to hold them accountable [11:45] Announcer: for market manipulation. We are creating a space where the most consequential decisions [11:50] Announcer: are made outside of public view. We have seen this pattern before in other industries. [11:56] Announcer: In previous reporting, the focus was on technical drift—how a model changes [12:00] Announcer: its behavior over time—but institutional drift is just as significant. It is the process by which [12:07] Margaret Ellis: a regulatory body moves from [12:09] Margaret Ellis: an adversarial role to a collaborative one, often without a clear mandate from the public [12:13] Announcer: to do so. [12:15] Margaret Ellis: The regulator becomes a partner, and the distinction between public interest [12:18] Margaret Ellis: and private profit begins to blur. The question of verification remains the crux of the problem. [12:25] Margaret Ellis: How do you keep the scope narrow? If the government grants these waivers, [12:29] Margaret Ellis: it is essentially trusting the labs to police themselves within those private channels. [12:34] Margaret Ellis: It assumes that the intent behind the coordination will remain focused on safety, even [12:38] Margaret Ellis: as the economic incentives for market dominance remain unchanged. It is a system built on trust [12:44] Margaret Ellis: in an industry that has frequently shown itself to be driven by speed. The drift here is subtle [12:50] Margaret Ellis: but profound. It is the shift from a world where competition provides a check on power to a world [12:55] Margaret Ellis: where coordination is the only recognized path to safety. And in that world, [13:00] Margaret Ellis: the line between a safety meeting and a business strategy session is written in pencil by the people [13:05] Margaret Ellis: standing to profit from both. The very definition of accountability is being rewritten [13:10] Margaret Ellis: to exclude the possibility of external challenge or competitive disruption. [13:14] Announcer: According to the six-minute read by Kevin O'Shaughnessy, the tension [13:19] Announcer: between these narratives—beating China versus global coordination—is where the policy [13:25] Announcer: actually lives. The version of coordination [13:29] Announcer: that survives is the one that serves the racing dynamic, not the one that tempers it. [13:35] Margaret Ellis: Policy is being crafted not to solve the problem of artificial intelligence safety, [13:41] Margaret Ellis: but to manage the political and economic pressures of the development race. [13:46] Margaret Ellis: This was already happening. Today is August 22, 2026, and the push for these waivers [13:53] Margaret Ellis: has been ongoing for years, buried in the discourse of artificial intelligence safety research [13:59] Margaret Ellis: and discussed in specialized podcasts. It is a slow, quiet adjustment of the legal landscape [14:06] Margaret Ellis: to accommodate the new power structures of frontier development. It is a change that happens in [14:12] Margaret Ellis: footnotes and memoranda, rather than in public debate. We are documenting a failure in progress, [14:19] Margaret Ellis: a failure of the competitive framework to account for the unique pressures [14:23] Margaret Ellis: of artificial intelligence. The legal tools we have are being described by industry leaders [14:28] Margaret Ellis: as obsolete, but the replacements being offered are controlled by the very entities [14:33] Margaret Ellis: they are meant to oversee. We are replacing the friction of the law with the smooth operation [14:39] Margaret Ellis: of a coordinated elite, all under the banner of preventing a catastrophe. [14:45] Margaret Ellis: The first time we saw this logic was in the early warnings about regulatory capture. [14:50] Margaret Ellis: Now it is being codified into proposed policy and government memoranda. The shift is not dramatic, [14:56] Margaret Ellis: but it is steady. The barriers between competitors are being lowered in the name of safety, [15:02] Margaret Ellis: while the barriers to entry for everyone else are being raised. It is a closing of the gates, [15:08] Margaret Ellis: justified by the claim that only those inside the gate are capable of being safe. In March of 2024, [15:15] Margaret Ellis: the memorandum highlighted the national security stewardship angle. It established a one-year cap [15:21] Margaret Ellis: on these relationships to provide an illusion of temporary necessity. But one year is a long time [15:27] Margaret Ellis: in artificial intelligence development, and a series of one-year waivers could effectively create [15:32] Margaret Ellis: a permanent state of coordinated operation. The exception becomes the rule, [15:37] Margaret Ellis: and the oversight becomes a rubber stamp for industry agreements. [15:41] Announcer: Operational drift is [15:43] Margaret Ellis: not the moment something breaks; it is the point where the break is accepted as normal operation. [15:50] Margaret Ellis: When we accept that the only way to be safe is to allow the most powerful companies in the world [15:56] Margaret Ellis: to coordinate their behavior in private, we have accepted a new normal for accountability. [16:03] Margaret Ellis: We are trading transparency for a promise of protection, and we are doing so [16:08] Margaret Ellis: without a clear understanding of what we are giving up in return. The drift from safety [16:14] Margaret Ellis: to market control is officially undocumented, but it is predictable based on the incentives [16:20] Margaret Ellis: involved. The labs are incentivized to maintain their lead and prevent disruption. [16:26] Margaret Ellis: The government is incentivized to maintain national dominance. Neither is incentivized to ensure [16:32] Margaret Ellis: that the public has a window into their coordination. The result is a system of governance [16:37] Margaret Ellis: that operates in the shadows, justified by the complexity of the technology it oversees. [16:44] Margaret Ellis: What we are observing is the relocation of liability. If the labs coordinate on a safety standard, [16:50] Margaret Ellis: and that standard fails, who is responsible? Is it the labs who agreed to the standard, [16:56] Margaret Ellis: or the government that granted them the waiver to do so? The path to accountability becomes [17:02] Margaret Ellis: a circle, and the public is left outside of it. When no one is clearly responsible, [17:08] Margaret Ellis: the consequences of failure are borne by everyone except those who made the decision. [17:14] Margaret Ellis: We should consider the possibility that these waivers will not just be about safety protocols; [17:19] Margaret Ellis: they will be about the physical infrastructure for monitoring and verification, [17:23] Margaret Ellis: as Moschewitz suggests. This could eventually lead to the creation of a global surveillance network [17:29] Margaret Ellis: for compute power, managed by a handful of corporate and state actors. [17:34] Margaret Ellis: The safety waiver is the legal foundation for a much larger structure of control that extends far [17:40] Margaret Ellis: beyond software. This appears for the first time in internal documentation as a low-cost, [17:46] Margaret Ellis: high-leverage first step. But the steps that follow involve opening channels with China [17:51] Margaret Ellis: and building monitoring systems that would require unprecedented access to private data [17:56] Margaret Ellis: and infrastructure. The scope of the waiver is designed to grow, even as the transparency [18:01] Margaret Ellis: of the process [18:02] Margaret Ellis: remains limited. What begins as a small exception ends as a global standard for compute monitoring. [18:09] Announcer: The tension remains. Competition is a safety mechanism in disguise, providing the friction [18:16] Announcer: that prevents a single entity from moving too fast. But it is also a source of risk. [18:22] Announcer: Coordination is a safety mechanism, but it is also a source of collusion. [18:28] Announcer: The drift is toward a solution that addresses the risk of competition by embracing the risk [18:34] Announcer: of collusion. We are choosing a known failure mode of the past to protect us from an unknown failure [18:41] Announcer: of the future. We are left with an unresolved contradiction. The industry claims [18:47] Announcer: it needs to cooperate to save the world, while economists warn that such cooperation [18:53] Announcer: will only serve to own the world. If the antitrust waiver becomes standard policy, [18:59] Announcer: the primary mechanism for public oversight will have been voluntarily dismantled by [19:04] Announcer: the very government tasked with protecting the marketplace. We are witnessing [19:09] Announcer: the end of an era of competitive regulation in favor of coordinated management. [19:15] Announcer: We still do not know how a narrow safety waiver can be effectively monitored when the technology [19:21] Announcer: it governs is inherently opaque to outsiders. We do not know what happens [19:26] Announcer: when safety talk inevitably drifts into market talk, and we do not know who is authorized to notice [19:32] Announcer: that the line has been crossed. Responsibility does not disappear when it is moved behind [19:38] Announcer: closed doors; it relocates, and in the relocation, it becomes much harder to find. [19:46] Announcer: The record remains open. For transcripts, primary documents, and our transparency policy, [19:52] Announcer: visit operationaldrift.neuralnewscast.com. This program is for archival [19:58] Announcer: and analytical purposes only. It does not constitute legal or regulatory advice. Neural Newscast [20:05] Announcer: is AI-assisted, human-reviewed. View our AI transparency policy at neuralnewscast.com.