Latest / Neural Newscast / Zvi Mowshowitz and the Antitrust Safety Waiver [Operational Drift]
Transcript
- [00:06] Margaret Ellis: On a date recorded in early 2024, an artificial intelligence safety researcher named Zvi Moshevitz [00:14] Margaret Ellis: began documenting a specific friction point in the development of frontier models. [00:20] Margaret Ellis: He argued that the United States government was maintaining a legal environment that made [00:25] Margaret Ellis: safety coordination between competitors a liability. The implication was [00:30] Margaret Ellis: that being cautious could be interpreted as a violation of antitrust law. [00:36] Margaret Ellis: This record investigates how these systems quietly drift away from intent, oversight, and control, [00:43] Margaret Ellis: and what happens when no one is clearly responsible for stopping it. I'm Margaret Ellis. [00:49] Margaret Ellis: This is operational drift. The record shows that the current legal framework for competition assumes [00:56] Margaret Ellis: that when companies talk to each other, they are doing so to the detriment of the consumer. [01:01] Margaret Ellis: In most industries, this is called collusion. However, in the context of artificial intelligence, [01:07] Margaret Ellis: a new argument has emerged. Proponents suggest that frontier labs, specifically the small group [01:14] Margaret Ellis: of companies developing the most advanced systems, are currently exposed to real legal risk [01:20] Margaret Ellis: if they coordinate on safety measures. According to these researchers, [01:24] Margaret Ellis: the law creates a race to the bottom, where the least cautious actor sets the pace [01:29] Margaret Ellis: for the entire industry, because their competitors are legally barred from agreeing to slow down [01:34] Margaret Ellis: together. In the archives of 2024, Zvi Moshevitz framed this as a low-cost, [01:41] Margaret Ellis: high-leverage first step for the United States government. He proposed targeted [01:46] Margaret Ellis: antitrust exemptions. These waivers would allow labs to share evaluation results, discuss [01:51] Margaret Ellis: dangerous capabilities, and jointly decide on deployment thresholds. Without these exemptions, [01:57] Margaret Ellis: a lab that chooses to delay a release due to observed risks might be outcompeted by one that ignores [02:03] Margaret Ellis: those same risks. The cautious lab cannot legally coordinate a response [02:08] Margaret Ellis: or ask their competitor to match their safety standards without risking a lawsuit [02:12] Margaret Ellis: for anti-competitive behavior. The system creates a penalty for restraint. According to the record, [02:19] Margaret Ellis: this idea gained traction in private and semi-public circles. Elon Musk reportedly told [02:25] Margaret Ellis: Demis Hassabis that he wanted leading artificial intelligence companies to have a regular call [02:31] Margaret Ellis: to discuss safety and security issues. Commentators like Dave Castan echoed this, [02:37] Margaret Ellis: suggesting the government should explicitly waive antitrust risks [02:41] Margaret Ellis: for these specific safety-related interactions. The drift here is subtle: [02:46] Margaret Ellis: a shift from seeing competition as a safeguard to seeing it as a danger, and seeing coordination, [02:53] Margaret Ellis: the very thing antitrust law is meant to prevent, as the new safety mechanism. [02:59] Margaret Ellis: We are witness to a reversal of regulatory logic. This leads to a reported memorandum that included [03:05] Margaret Ellis: a waiver process. This process allowed agency heads to grant limited exceptions, capped at one year, [03:12] Margaret Ellis: for relationships necessary to responsibly steward United States national security. [03:18] Margaret Ellis: While this was aimed primarily at government-to-lab relationships, it established a precedent. [03:24] Margaret Ellis: It signaled that the government was willing to suspend the rules of the market if the technology [03:28] Margaret Ellis: was deemed significant enough for national security. This is where the operational drift [03:33] Margaret Ellis: begins to manifest. The rules governing how these companies interact are changing, [03:39] Margaret Ellis: but the oversight of those interactions remains opaque and largely hidden from public view. [03:45] Margaret Ellis: In this context, behavior drift means a system changing how it acts without triggering [03:50] Margaret Ellis: the mechanisms designed to notice change. Here, the system is the regulatory environment. [03:56] Margaret Ellis: We are seeing a move toward a unipolar or multipolar dilemma, as documented in the records of 2024. [04:03] Margaret Ellis: The goal is coordination, but the cost is the concentration of authority. [04:08] Margaret Ellis: When authority concentrates in the absence of transparency, the mechanisms [04:12] Margaret Ellis: of accountability begin to fail. The record suggests this shift was not accidental, [04:17] Margaret Ellis: but a deliberate move to bypass the friction of existing law. There are five primary pushback points [04:23] Margaret Ellis: identified in the record against these antitrust exemptions. The first is the risk [04:28] Margaret Ellis: of regulatory capture dressed up as safety. Skeptics worry that safety coordination is merely [04:34] Margaret Ellis: a convenient label for incumbent labs, specifically OpenAI, Anthropic, and Google DeepMind, [04:40] Margaret Ellis: to jointly set industry standards that benefit them. By establishing compute thresholds [04:45] Margaret Ellis: and release timings, these labs could effectively box out smaller or open-source competitors [04:50] Margaret Ellis: who are not part of the coordinating group. [04:52] Margaret Ellis: This creates a barrier to entry that is enforced by the state under the guise of public protection. [04:58] Announcer: History shows [04:59] Margaret Ellis: that antitrust law's suspicion of competitor coordination exists because such cooperation [05:05] Margaret Ellis: has been used to entrench dominant players in other industries. When the record shows a shift [05:12] Margaret Ellis: towards safety coordination, it rarely details how we distinguish a safety standard [05:17] Margaret Ellis: from a market barrier. If the leading labs agree that a model requires a specific level [05:23] Margaret Ellis: of expensive safety testing that only they can afford, they have effectively regulated [05:29] Margaret Ellis: their competition out of the market. The cost of safety becomes the price of entry, [05:34] Margaret Ellis: ensuring that only the largest entities can survive the transition. The second issue is [05:40] Margaret Ellis: the difficulty of separating safety talk from business talk. [05:45] Margaret Ellis: Once executives are legally permitted to be in a room discussing deployment timing [05:49] Margaret Ellis: and capability evaluations, there is no documented mechanism to prevent that conversation [05:55] Margaret Ellis: from drifting into pricing, hiring, or market division. A waiver would need a very narrow, [06:01] Margaret Ellis: well-monitored scope to avoid becoming a backdoor for ordinary collusion. However, [06:07] Margaret Ellis: as the record indicates, vague or broad waivers are inherently harder for regulators to police, [06:13] Margaret Ellis: especially when the subject matter is as complex as frontier model capabilities. Some economists, [06:20] Margaret Ellis: according to reports from 2024, argue that competitive pressure is actually a safety mechanism [06:26] Margaret Ellis: in disguise. A crowded, competitive field makes it difficult for any single lab [06:32] Margaret Ellis: or a small cartel to control the entire trajectory of the technology. By reducing this friction [06:38] Margaret Ellis: through antitrust waivers, we concentrate power into fewer hands. The safety of the system becomes [06:44] Margaret Ellis: dependent on the integrity of a very small number of people, rather than the distributed oversight [06:50] Margaret Ellis: of a market. We are trading the resilience of competition for the perceived stability [06:55] Margaret Ellis: of a concentrated alliance. Verification remains a significant problem. Even [07:01] Margaret Ellis: if a weaver is well designed, safety coordination agreements are nearly impossible to verify from [07:06] Margaret Ellis: the outside. If labs claim they are coordinating to slow down or align safety standards, the public [07:13] Margaret Ellis: and regulators have no easy way to confirm they aren't using those same channels [07:17] Margaret Ellis: for anti-competitive ends. The opacity that makes antitrust enforcement difficult is not removed; [07:23] Margaret Ellis: it is simply shielded by the label of safety. The lack of an independent auditing mechanism means [07:28] Margaret Ellis: that we are essentially operating on [07:30] Margaret Ellis: a system of trust in an industry defined by intense competition. [07:35] Announcer: Finally, there is the problem of international asymmetry. A United States-only antitrust carveout [07:42] Announcer: does not bind foreign labs. This raises the question of whether domestic coordination [07:48] Announcer: meaningfully reduces global racing dynamics, or if it simply consolidates power among a handful of [07:55] Announcer: United States firms while the global race continues unabated. The record suggests that proponents [08:02] Announcer: like Moshewitz acknowledge this, eventually wanting to open similar channels with China [08:07] Announcer: for monitoring and verification. But the political reality is far more fractured. [08:13] Announcer: Domestic safety measures mean little if the global environment remains a high-speed race with no [08:19] Announcer: off-ramp. This tension is most visible when we compare the safety coordination narrative [08:25] Announcer: with the official White House narrative. The record shows that the current administration views it [08:30] Announcer: as a national security imperative for [08:32] Announcer: the United States to beat China in the artificial intelligence race. [08:37] Margaret Ellis: This is a racing framework. Speed and capability advantage are the primary objectives. [08:43] Margaret Ellis: Anything that slows down United States labs, including safety-motivated caution, [08:48] Margaret Ellis: is framed as ceding ground to a rival who will not show the same restraint. [08:53] Margaret Ellis: The conflict between the safety goal and the geopolitical goal is rarely resolved in the documents. [08:59] Margaret Ellis: Proponents of safety coordination attempt to reconcile this through three different moves. [09:04] Margaret Ellis: The first is to reframe cooperation as a competitive tool. They argue that by pooling [09:10] Margaret Ellis: safety research and sharing threat intelligence, the United States bloc [09:14] Margaret Ellis: becomes stronger collectively. This version of the waiver is tied to national security stewardship [09:20] Margaret Ellis: and covers threat intelligence sharing rather than actual deployment delays. [09:25] Margaret Ellis: This is the version that has seen the most political traction, as it aligns with the desire to build [09:29] Margaret Ellis: a more resilient domestic front against foreign intrusion and corporate espionage. [09:34] Margaret Ellis: The second move is the argument that safety coordination lets you race faster, not slower. [09:40] Margaret Ellis: By standardizing security practices and evaluation standards, labs reduce the risk [09:45] Margaret Ellis: of a catastrophic failure, such as a hack, a scandal, or an accident, that would trigger [09:50] Margaret Ellis: a public backlash and heavy-handed regulation. In this view, coordination is a way to avoid [09:56] Margaret Ellis: the scenario where a disaster forces everyone to stop. It is safety as a form of insurance [10:02] Margaret Ellis: for the race itself. The goal is not to slow down, [10:05] Margaret Ellis: but to ensure that the speed does not lead to a derailment that ends the competition entirely. [10:10] Announcer: The third move, and the one most in tension with the "beat China" narrative, [10:15] Announcer: is the call for international coordination that includes China. The goal here is mutual restraint [10:22] Announcer: and verification. However, there is no documented sign [10:26] Announcer: that the current administration is interested in a version of the waiver that allows [10:31] Margaret Ellis: United States [10:31] Announcer: labs to coordinate a slowdown, especially if that slowdown includes international rivals. [10:38] Margaret Ellis: The documents suggest that any coordination that impacts speed is viewed with extreme skepticism. [10:44] Margaret Ellis: The priority remains maintaining the lead, even if that lead comes with unquantified risks [10:50] Margaret Ellis: to system stability. What we observe in the documents is a narrowing of the original proposal. [10:57] Margaret Ellis: The broader version, where labs coordinate to jointly hold back capabilities for safety reasons, [11:02] Margaret Ellis: is being replaced by a narrower version focused on threat intelligence [11:06] Margaret Ellis: and security against foreign espionage. The drift is toward a system where coordination is allowed [11:12] Margaret Ellis: only if it serves the race, not if it slows it down. This shift changes the fundamental purpose [11:19] Margaret Ellis: of the antitrust exemption from a safety mechanism to a national security tool. [11:24] Margaret Ellis: The original concern for model safety is being eclipsed by the necessity of global dominance. [11:31] Margaret Ellis: In the case of Claude 5 and Claude Mythos, the records from Anthropic suggest a belief [11:36] Margaret Ellis: that specific versions of models can be managed through these agreements. [11:41] Margaret Ellis: But as the capabilities of these systems increase, the distinction between a safe deployment [11:46] Margaret Ellis: and a competitive deployment becomes harder to maintain. The authority to make these decisions [11:52] Margaret Ellis: is being quietly relocated from the public sphere to the rooms where these coordinated discussions [11:57] Margaret Ellis: occur. Decisions that affect the global technological landscape are being made by a small number [12:03] Margaret Ellis: of technical leads and corporate executives, shielded from standard regulatory scrutiny. [12:09] Margaret Ellis: When we look at the quiet ways policy could go wrong, as Moschowitz described it in late 2024, [12:15] Margaret Ellis: the structural prevention of coordination was his primary fear. Yet the counterfear remains [12:22] Margaret Ellis: that the coordination itself becomes the mechanism of drift. If the labs are allowed to decide [12:27] Margaret Ellis: what constitutes a safety risk without external verification, they become the regulators [12:32] Margaret Ellis: of their own industry. The drift is the movement from law to discretion, [12:37] Margaret Ellis: from public oversight to private agreement. We are watching the formation [12:41] Margaret Ellis: of a shadow regulatory system that operates without the traditional checks of a democratic society. [12:47] Announcer: The memorandum's one-year cap on waivers suggests a temporary measure, [12:53] Announcer: but the history of such limited exceptions often shows them becoming normalized parts [12:59] Announcer: of the institutional landscape. As of 2026, we must ask if these relationships, originally described [13:07] Margaret Ellis: as necessary for national security, have become the standard operating procedure [13:12] Margaret Ellis: for the entire frontier artificial intelligence sector. What was once an emergency measure is now [13:19] Margaret Ellis: a foundational element of the industry's legal architecture. The temporary has become permanent, [13:25] Margaret Ellis: and the exception has become the rule, without a single public debate on the consequences. [13:32] Margaret Ellis: According to the researchers, the tension between racing and safety is not just a policy debate; [13:38] Margaret Ellis: it is a technical reality. If a model like Claude Methos shows concerning capabilities, [13:45] Margaret Ellis: the decision to delay it is now wrapped in layers of legal and geopolitical calculation. [13:51] Margaret Ellis: The drift is no longer just in the system's behavior, but in the oversight mechanisms [13:56] Margaret Ellis: that were supposed to hold it accountable. The very tools meant to ensure safety [14:01] Margaret Ellis: are being recalibrated to ensure competitive advantage. This suggests a fundamental misalignment [14:08] Margaret Ellis: between the stated goals of the labs and the actual incentives of the market they inhabit. [14:14] Margaret Ellis: We are seeing the formation of a unipolar regulatory structure, where a few labs [14:18] Margaret Ellis: and a few government agency heads set the trajectory for a technology that impacts [14:23] Margaret Ellis: the entire global economy. The incentives for these actors to prioritize safety [14:29] Margaret Ellis: over their own market dominance or national security standing are not clearly established in [14:34] Margaret Ellis: any filing or system card. The record shows a preference for consolidation over competition, [14:40] Margaret Ellis: for control over transparency. This concentration of decision-making power creates a single point [14:46] Margaret Ellis: of failure that could have catastrophic consequences for the public interest. [14:51] Margaret Ellis: Yet the public is excluded from the room. The drift here is the normalization of the safety waiver [14:57] Margaret Ellis: as a tool for market control. It is an unintended but predictable outcome of trying to solve [15:02] Margaret Ellis: a coordination problem within a competitive high-stakes environment. No one person owned [15:08] Margaret Ellis: the decision to move toward this cartel-like structure. It emerged from the friction [15:13] Margaret Ellis: between antitrust law and the perceived existential risks of artificial intelligence. [15:18] Margaret Ellis: It is a structural drift where the system itself evolves to protect its most powerful participants [15:23] Margaret Ellis: at the expense of the original safety objectives. [15:27] Margaret Ellis: The waiver has become the primary instrument of this evolution. [15:30] Announcer: As we examine the archives from 2024 and 2025, the missing piece is always the verification. [15:39] Announcer: How does the public know that the coordination is about safety and not about maintaining [15:45] Announcer: the status quo? The record doesn't specify any independent monitoring body that has the authority [15:51] Announcer: to audit these coordinated meetings or the deployment thresholds they set. [15:57] Announcer: There is no sunshine in these rooms. Without a mechanism for outside review, [16:02] Announcer: the claim of safety is just words. We are asked to trust the very entities [16:07] Announcer: that are most incentivized to prioritize their own growth over public caution. [16:13] Announcer: This lack of transparency is where accountability dissolves. If a coordinated safety decision [16:19] Margaret Ellis: later proves to be an anti-competitive maneuver, the liability has already been relocated by [16:25] Margaret Ellis: the waiver itself. The agency heads who granted the exceptions are shielded by [16:30] Margaret Ellis: the national security mandate, and the labs are shielded by the waiver. [16:34] Margaret Ellis: Responsibility becomes a game of musical chairs, where no one is left holding the blame [16:39] Margaret Ellis: when the music stops. The drift ensures that even if something goes wrong, [16:44] Margaret Ellis: there is no clear path to holding the decision makers responsible for the outcomes [16:48] Margaret Ellis: they orchestrated. In the unipolar versus multipolar artificial intelligence dilemma, [16:54] Margaret Ellis: the documents reveal a preference for a controlled unipolar approach. The belief is [16:59] Margaret Ellis: that it is better to have a few labs coordinating safely than hundreds of labs racing recklessly. [17:06] Margaret Ellis: But this assumes the coordination remains focused on safety. The drift suggests [17:10] Margaret Ellis: that once the competitive pressure is removed, the lab's alignment with public intent becomes [17:16] Margaret Ellis: even harder to verify. We are trading a chaotic race for a coordinated march [17:21] Margaret Ellis: in an unknown direction. The fundamental question is whether that march is any safer than the race [17:26] Margaret Ellis: it replaced. We see this in the framing of threat intelligence sharing. [17:31] Margaret Ellis: While standardizing defenses against espionage is a clear national security benefit, [17:36] Margaret Ellis: it also creates a shared technical baseline among the frontier labs. This shared baseline makes [17:42] Margaret Ellis: it easier for them to move as a single unit, further distancing them from the rest of the industry [17:47] Margaret Ellis: and the public. The coordination intended to secure the technology also serves to secure [17:52] Margaret Ellis: the market position of the participants. The drift has converted a security protocol into a barrier [17:58] Margaret Ellis: that ensures no new actors can disrupt the existing order of power. Operational drift is not [18:04] Margaret Ellis: the point where something breaks; it is the point where the break is accepted as normal operation. [18:10] Margaret Ellis: The suspension of competition rules for the sake of artificial intelligence safety has become [18:16] Margaret Ellis: a normal operation, even if the safety benefits remain undocumented and unverified. [18:22] Margaret Ellis: We are living in a period where the exception has been folded into the foundation of [18:27] Margaret Ellis: how these companies function. The break in the law is no longer seen as a break; [18:32] Margaret Ellis: it is seen as the necessary architecture for the development of the most powerful technology [18:37] Margaret Ellis: in human history. As the race with China continues to dominate the administration's narrative, [18:43] Margaret Ellis: the broader coordination's V. Moshiutz argued for, the one that includes international rivals, [18:48] Margaret Ellis: remains a marginal idea. The record shows we have chosen a path of domestic consolidation [18:54] Margaret Ellis: under the label of safety, while the global racing dynamics remain fundamentally unchanged. [19:00] Margaret Ellis: The drift has narrowed our focus to a domestic alliance, leaving the global risks to be sorted out [19:06] Margaret Ellis: by the same racing incentives that caused the problem. We have built a castle for our labs, [19:11] Margaret Ellis: while the storm outside only grows in intensity. This was already happening. Today is just [19:17] Margaret Ellis: when we noticed. The legal frameworks that were designed to ensure fair competition [19:22] Margaret Ellis: are being quietly dismantled for a technology that no one yet knows how to control. [19:27] Margaret Ellis: The authority is accumulating in rooms where the public is not invited, [19:31] Margaret Ellis: under waivers that the public cannot review. The drift is complete when the system [19:36] Margaret Ellis: no longer requires public consent to function. We are reaching a point where the major labs [19:41] Margaret Ellis: and the security state are so intertwined that their interests are indistinguishable [19:45] Margaret Ellis: from one another, leaving the public interest behind. In the records of 2024, [19:51] Margaret Ellis: the fear was that antitrust law would kill artificial intelligence safety. In the records of 2026, [19:57] Margaret Ellis: the question is whether the pursuit of safety has killed the very competition [20:01] Margaret Ellis: and transparency that were supposed to keep the labs accountable. We have traded one risk [20:06] Margaret Ellis: for another, but we have yet to see if the trade was worth the cost. The accountability mechanisms [20:11] Margaret Ellis: of the past are being replaced by [20:12] Margaret Ellis: the promises of a few powerful actors, and we have no way to ensure those promises are kept. [20:18] Announcer: We are left with a system that coordinates to compete, rather than one that competes to be safe. [20:26] Announcer: The shift is complete, but the outcome is unresolved. If the primary labs [20:32] Announcer: are now legally allowed to act as one, who is left to notice when they all drift in [20:38] Announcer: the same dangerous direction? [20:41] Announcer: The record shows that when coordination replaces competition, the first thing to disappear is [20:47] Margaret Ellis: the dissenting voice. Without that voice, the drift becomes invisible until it is [20:52] Margaret Ellis: too late to correct. The coordination that was supposed to save us may be the very thing that makes [20:59] Margaret Ellis: the system uncontrollable. The record remains open on how a safety coordination agreement [21:06] Margaret Ellis: can be distinguished from a market cartel when both are conducted behind the shield [21:10] Margaret Ellis: of national security. We still don't know who, if anyone, has the authority to revoke a waiver [21:17] Margaret Ellis: once it becomes essential to a lab's business model. If the waiver becomes the baseline [21:22] Margaret Ellis: for operation, the ability to regulate the industry is effectively lost to the very companies [21:28] Margaret Ellis: the government was supposed to oversee. The documents do not provide an answer to [21:33] Margaret Ellis: how this power will be returned to the public sphere. Operational drift isn't the point [21:40] Margaret Ellis: where something breaks. It is the point where the break is accepted as normal operation. [21:46] Margaret Ellis: Responsibility doesn't disappear. It relocates to the dark corners of the bureaucracy, [21:51] Margaret Ellis: where decisions are made without names and consequences are managed without accountability. [21:57] Margaret Ellis: The relocation of this responsibility is the final step in the drift of the system away [22:02] Margaret Ellis: from public intent. We are left watching a system that no longer feels the need to explain itself [22:09] Margaret Ellis: to those it is meant to serve. If safety coordination becomes the new standard for market entry, [22:15] Margaret Ellis: we must ask: whose safety are we actually protecting, and what happened to the oversight [22:20] Margaret Ellis: we were promised? The answer is not in the system cards or the public press releases. [22:26] Margaret Ellis: It is in the documents we have yet to see, in the meeting minutes that remain classified, [22:31] Margaret Ellis: and in the private agreements that now govern the future of the technology. [22:35] Margaret Ellis: The record is silent on these matters, and in that silence, the drift continues unchecked. [22:41] Margaret Ellis: We are left to wonder who is truly at the wheel. Sources for this record are available [22:47] Margaret Ellis: at operationaldrift.neuralnewscast.com. Neural Newscast is AI-assisted, human-reviewed. [22:55] Margaret Ellis: View our AI transparency policy at neuralnewscast.com. [22:59] Margaret Ellis: This program adheres to the Neural Newscast Network transparency policy.