Last updated: July 2026. The Anthropic vs. Department of War section describes an active, unresolved legal dispute — check for developments before treating its status as settled.
Disclaimers: The thumbnail made for this post is created by AI (Google's Nano Banana), the content has been reviewed and edited by AI (Sonnet 5) to reflect the voice of me the author. Thanks for reading the AI use disclaimers.
Epistemic status: The Fiction World scenario below is illustrative allegory, not a forecast — I built it to make a real, citable technical gap vivid. Everything after that section is either a documented fact (with a source) or a labeled opinion of mine. Where I don't have a source for something, I've said so directly rather than implying otherwise.
Fiction World
Two countries are global superpowers in this made-up world: the capitalist giant, the United Country of War (UCW), and the communist Republic of Dudes and Dudettes (RDD). Caught between them is Yipet — a nation whose leaders have lived in exile for decades, its people a proxy battleground for powers bigger than itself.
It's 2030. Robotic warfare is saving the lives of millions of UCW soldiers. Global conflict is rising, terror is striking civilians everywhere, and the RDD has spent generations oppressing the mountain people of Yipet. Their leaders are still in exile, still hoping for a homeland they can safely return to. Now the UCW is sending robots into Yipet to help push back the conflict — to give Yipet a shot at becoming its own sovereign nation, the same aspiration for freedom the UCW claims to stand for everywhere else.
Picture one of these robots on a checkpoint outside a Yipet village. It's running a local model and an ASR pipeline on a Jetson-Thor-class chip — no cloud connection needed, everything happens on-device, real time. An old farmer walks up, hands raised, saying something in his local tonal dialect. Maybe he's warning the robot about a mine up the road. Maybe he's just asking it to move. The ASR mishears the tone, mishears the words, and outputs a clean, confident transcript that says something it was never meant to say. The robot doesn't hesitate the way a scared, tired human soldier might. It doesn't ask him to repeat himself. It acts on what it "heard."
This is the scene I keep picturing when people tell me robotic warfare will save lives. Maybe it will save UCW lives. I'm less sure it'll save Yipet's.
This isn't science fiction: three grounded facts
(a) Interpreter failure is a decades-long, documented pattern, not a one-off. This isn't new. Human soldiers in real wars have failed to understand tonal languages and cultural differences for as long as we've sent soldiers somewhere they don't speak the language. In Iraq and Afghanistan, interpreter shortages were chronic for the entire length of both wars — the military leaned on undertrained contractors and local hires who were distrusted by every side in the conflict, a pattern confirmed by in-depth interviews with Afghan interpreters themselves [1], and officers made calls based on bad translations [2]. Civilians got fired on from exactly this kind of misunderstanding. We know this happened with humans. There's no reason to assume it stops happening just because we swap the human for a machine — if anything, a machine is worse at knowing when it doesn't know.
(b) Tibetan ASR — the closest real-world analogue to Yipet's language — is still an open, unsolved research problem. Robotics is developing fast, but mostly for English and Chinese. Everything else is trailing behind, and Tibetan is a good stand-in for "everything else." Current systems are still landing in the mid-teens to high-20s percent word-error-rate range, even with transfer learning and multilingual pretraining thrown at the problem [3][4][5][6]. Compare that to well under 5% for English or Mandarin. This isn't a solved problem quietly waiting to be shipped — it's an active, ongoing research struggle, and it has been for years.
(c) The hardware to run this today already exists. Jetson Orin and now Jetson Thor put serious AI compute directly onto a robot-mountable chip — no cloud round-trip required [7][8]. Real companies (Boston Dynamics, Agility Robotics, 1X) are already building on this hardware for humanoid and autonomous systems [8]. So the "picture a world with robots being sent as soldiers" part of this isn't speculative fiction anymore. The compute to do it is shipping right now. What's missing isn't the chip. It's the language.
On Tibetan ASR / low-resource speech tech
Tibetan is genuinely hard for a model to learn. It's tonal, it has multiple major dialects (Lhasa, Amdo, Kham) that don't fully transfer to each other, and there just isn't much clean, labeled speech data to train on compared to English or Mandarin [3]. Researchers have made real progress — some of the newer speech-LLM approaches have cut error rates almost in half compared to older baselines [6] — but "almost in half" starting from a bad number still lands you at a bad number. Years of work, and the field still hasn't closed the gap to parity with high-resource languages [4][5][6].
On LAWS / civilian harm / AI in warfare
Most of the existing debate about autonomous weapons focuses on whether a system can tell a combatant from a civilian — the "distinction" problem under international humanitarian law [9]. That's the visual and behavioral side of the problem: crowds, movement, proximity to a fight. It's a real and serious debate, and there's a whole UN process trying to write treaty language around it right now [10].
What doesn't get talked about as much is the language side. A system that's uncertain about what it's seeing can, in theory, be built to flag that uncertainty and hold fire. A system that mishears a low-resource language doesn't look uncertain. It looks confident. It outputs a clean transcript and moves on. That's a much scarier failure mode, because nothing downstream knows to double-check it.
And this isn't a far-off hypothetical anymore. Two examples, both real, both operating right now:
Israel's Harpy and Harop loitering munitions have been in service for years — built to patrol a zone, detect enemy radar, and destroy the target without a human in the loop once they're launched [14]. And in the Russia-Ukraine war, Russia's V2U drones use onboard AI to navigate and strike, and per Ukrainian reporting are in daily operational use in active combat areas like the Sumy region [14]. Most currently deployed systems still sit somewhere between full autonomy and full human control — a platform like the MQ-9 Reaper can fly and identify targets on its own but still needs a person to authorize the strike [14] — but the fully autonomous end of that spectrum is not speculative. It's already fighting a war.
Edge compute trend (exponential, well-documented) vs. Tibetan ASR trend (much flatter, still an open problem)
Here's the asymmetry that worries me most. Edge AI compute has grown something like 4,000x in under a decade — from the first Jetson boards doing a fraction of a TOPS, up to Jetson Thor doing over 2,000 FP4 TFLOPS today [7][8]. That's an almost unbelievable curve. Meanwhile, Tibetan ASR accuracy has barely moved in the same window — hovering in the same rough error-rate band for years [3][4][5][6], while the hardware to deploy it got thousands of times more powerful.
And this isn't really a hardware problem or even purely a research-difficulty problem. It's an investment problem. Roughly 6 million people speak Tibetan worldwide [25] — a real population, but small next to the languages that get the funding, the annotated datasets, and the professional translator pipelines. Nobody builds a $200 million ASR benchmark effort for 6 million speakers when the same money, spent on English or Mandarin, reaches over a billion. That's not a moral failing on any one company's part — it's just what happens when research investment follows speaker count and market size, and nobody with the resources to close the gap has had a reason to. The compute got cheap and fast for everyone. The data, the annotation labor, and the trained linguists needed to make a low-resource language usable never got the equivalent investment, because there was never a comparable business case for it.
Why this is an AI safety problem, not just an ethics problem
I want to be specific about why I think this is a safety problem and not just an ethics problem, because those get treated as the same thing and they're not.
An ethics problem is "should we build autonomous weapons at all." That's an important conversation, but it's not the one I'm having here.
A safety problem is "this system will behave in ways its designers didn't intend, in a way that's hard to catch before it causes harm." That's what a silent ASR miscalibration is. The system isn't malfunctioning by any metric its own developers are tracking. Word-error-rate numbers look fine on the benchmark. It's only failing on the language and the population that never made it into the training or benchmarking pipeline in the first place — which also happens to be exactly the population most likely to be standing in front of it in a conflict like Yipet's. Visual uncertainty in an AI system is something we at least know how to try to detect. Linguistic miscalibration in a low-resource language doesn't announce itself. It just quietly does the wrong thing with full confidence.
Deployed models and quantization
Deployed models usually go through quantization to run fast on resource-constrained edge hardware. But quantization — INT8, for example — is calibrated against a dataset, and if that calibration set under-represents a language both in the original training and in the quantization step, the compression can quietly strip out exactly the acoustic detail that language depended on. This isn't speculation: research on compressing multilingual speech models has found that quantization actively amplifies pre-existing bias tied to how well-resourced a language is, hitting low-resource languages hardest of all [11], and that smaller, more edge-friendly models degrade faster under compression than larger ones [12].
But here's the important counterpoint, and I want to be fair to it: this isn't an inevitable law of physics. It's a default, and defaults can be beaten by people who actually try. A 2026 study on Cantonese — itself a resource-constrained language relative to Mandarin or English — took a Whisper model's character error rate from 49.5% down to 11.1% using targeted fine-tuning, then compressed it to a 60MB INT8 checkpoint without giving that accuracy back [13]. That's the whole point I want to land here: the degradation this section describes isn't destiny. It's what happens by default when nobody bothers to fix it for your language. Someone built the fix for Cantonese.
To be precise about Tibetan specifically, since I don't want to overstate the gap: fine-tuned Tibetan ASR models do exist, and real people have put real work into them [3][4][5][6]. What doesn't exist yet is a single, complete system that handles all three major dialects — Ü-Tsang/Lhasa, Amdo, and Kham — at once, at the accuracy and speed a real-time edge deployment would need, and then survives quantization down to a size that fits on a Jetson-class chip without losing what accuracy it had. Each of those three pieces (multi-dialect coverage, real-time edge performance, quantization-survivable accuracy) has been demonstrated separately, for Tibetan or for a comparable low-resource language. Nobody has yet put all three together for Tibetan specifically, and — as far as I can find — nobody has done it for any low-resource language in a military or edge-robotics deployment context. The mechanism that causes the problem is documented [11][12]. The mechanism that solves it is also documented [13]. The gap isn't technical impossibility. It's that nobody has assembled the pieces that already exist into one deployable system for the languages that need it most.
Put the pieces together and the shape of the risk is this: a low-resource language starts behind on training data, then loses more ground at the quantization step, then gets deployed on hardware fast enough that nobody has time to notice the confident wrong transcript before it's acted on. Each layer looks like a reasonable engineering tradeoff on its own. Stacked, they add up to exactly the failure mode this piece opened with.
AI warfare: a quick, honest overview of the debate
This is a live, unresolved debate, and I don't want to pretend otherwise.
The pro-robotics case, stated at its most human: every soldier who doesn't have to go home to their family in a coffin is a family that gets to keep their father, their daughter, their friend. Proponents argue machines can, in principle, be steadier under pressure than a frightened, exhausted 19-year-old holding a rifle — no fatigue, no panic, no desire for revenge clouding a split-second decision that costs someone their life.
The against case: responsibility gaps (who's accountable when a machine kills the wrong person?), the risk of "technological inevitability" arguments getting used to skip real oversight, and — the part I'm adding to this conversation — the fact that these systems inherit and can amplify exactly the coverage gaps we already know exist in the underlying models.
The realistic trend, regardless of where you land on the ethics: robots are coming. The compute curve above isn't slowing down, and every major robotics player is already building on this hardware. The question that's actually still open isn't "will this happen," it's "which languages and which people get left out of the training data when it does."
This already happened: Anthropic vs. the Department of War
I don't want this piece to read like a distant hypothetical, so here's a real, dated example of the exact tension underneath everything above — a company's safety judgment colliding with a military's demand for faster, fewer-restrictions deployment.
In July 2025, the Department of Defense (before its rename) awarded Anthropic, along with Google, OpenAI, and xAI, contracts worth up to $200 million each to speed up defense adoption of frontier AI. Claude became the first frontier model cleared for use on classified networks [15]. Reports later surfaced that Claude had been used in the January 2026 operation to capture Venezuelan President Nicolás Maduro [15].
Then contract renegotiations broke down. Defense Secretary Pete Hegseth — now also using the title "Secretary of War" — pushed for an "any lawful use" clause that would let the Department of War use Claude however it saw fit, with no corporate red lines attached [16]. Anthropic held two firm limits: no use for mass domestic surveillance, and no use in fully autonomous weapons systems that could fire without a human authorizing the strike [15]. On February 26, 2026, CEO Dario Amodei published a public statement saying Anthropic "cannot in good conscience" remove those guardrails, and that current AI systems are "not sufficiently reliable" for fully autonomous weapons deployment [15] — the same reliability gap this whole piece is about, just one layer up the stack.
The government's response was immediate. On February 27, President Trump directed all federal agencies to stop using Anthropic's technology, with a six-month phase-out window, and Hegseth publicly called the company "sanctimonious" and "arrogant" [17][18]. A week later, the Department of War formally designated Anthropic a "supply chain risk" — the first time that designation has ever been applied to an American company — which bars defense contractors from using Anthropic products in any military-related work [19][20]. Reuters reported one source saying part of the trigger was Claude's use in ongoing military operations in Iran [21].
Anthropic sued. Two federal lawsuits, filed March 9, arguing the designation was First Amendment retaliation dressed up as a national-security finding [21]. Researchers from OpenAI and Google DeepMind — companies that compete with Anthropic — filed a personal-capacity brief supporting them [21]. A federal judge agreed and issued a preliminary injunction blocking one of the two designations on March 26 [22]. The Department of War appealed that ruling to the Ninth Circuit, where the appeal is currently on hold pending a separate case. That separate case — over a different designation, made under a different statute — went to the D.C. Circuit, which denied Anthropic's request to pause enforcement while the case is heard, and held oral argument on the merits on May 19 [23]. As of this writing, neither appeals court has issued a final ruling. Meanwhile, on July 10, the Air Force Research Laboratory reportedly instructed its own contractors to remove Anthropic products from their systems entirely by September 1 — an administrative deadline moving forward on its own timeline, independent of how the litigation resolves [24].
So, as of today, this is genuinely unresolved: one federal judge has called the government's theory unsupported by the statute; a separate appeals panel has said the military's operational need outweighs the company's financial harm; and enforcement is proceeding regardless of either. I'm not going to predict who wins, and I don't think this piece needs to. And I want to be precise about what this dispute actually is, not what it's easy to assume it is: per the Congressional Research Service's own account, the Department of War is not publicly known to be using Claude, or any frontier model, inside an actual autonomous weapons system today [15]. This is a fight over whether Anthropic will permit that future use case — not a confirmed record of it already happening. What I want you to notice is just the shape of the collision underneath that distinction: an institution built to fight wars, moving at wartime speed, running directly into a company's own judgment about what its models are and aren't reliable enough to do. That collision is not hypothetical, and it is not going away, even though the worst-case use it's fought over hasn't been confirmed yet. It is, right now, one of the clearest real examples we have of automated-warfare urgency clashing head-on with an AI developer's own ethics — and nobody yet knows which one wins.
Where this leaves us
Go back to that checkpoint outside the Yipet village. The old farmer, hands raised, saying something the machine doesn't understand. Nothing about that scene requires a war that hasn't happened yet. It requires a language the model was never trained on well, running on hardware that's already shipping, deployed by an institution already fighting, in public, about how much safety testing it's willing to wait for.
I'm not here to make the case for building war robots, and nothing in this piece should be read that way. What I do think is that if these systems are being built regardless of what any of us think about it, then the language coverage gap is a solvable, yet non-salient, engineering problem — and right now almost nobody is working on it for the languages that need it most. The people with the resources to close that gap keep pointing them at the languages of the richest, most powerful countries, and almost never at the languages of the people who tend to be standing in the blast radius of those countries' decisions. That's backwards. If we're going to build systems that operate in someone else's country, speaking to someone else's people, the burden should be on us to go the extra mile — to make sure those systems can actually understand and fairly treat the people they're deployed among, not just the people who built them. Underrepresented languages and the cultures that speak them shouldn't be an afterthought bolted on once the English and Mandarin versions ship. They should be part of the cost of building the system at all.
So here's the actual ask. You don't need a lab, a defense contract, or a million-dollar compute budget to make a dent in this. Fine-tuning an existing model on a few dozen hours of a language nobody's bothered with, cleaning up a messy transcript, building a small open dataset from recordings that would otherwise just sit in a drawer — these are exactly the kind of unglamorous, weekend-sized projects that don't need permission from anyone with more resources than you. If you've got some free time and you're looking for a way to actually move a needle instead of just having an opinion about one, this is a place to look. Find a language, a community, or an archive that the big labs have no commercial reason to ever touch, and go see what you can do with it. The gap isn't closing because nobody's capable of closing it. It's not closing because almost nobody's trying.
If you think this framing is wrong, or you know of work already underway that I've missed, I want to hear about it.
References
- de Jong, S., "Brokering War: Afghan Interpreters, Western Soldiers and Unequal Encounters in Crisis," Cultural Studies, 39(2), 248–268 (2025) — https://doi.org/10.1080/09502386.2024.2437434 — draws on interviews with Afghan interpreters and Afghanistan veterans (17 veteran interviews specifically)
- CorpWatch, "Afghanistan: Maladies of Interpreters" (2009) — https://www.corpwatch.org/article/afghanistan-maladies-interpreters
- "Improving low-resource Tibetan end-to-end ASR by multilingual and multilevel unit modeling," EURASIP J. Audio, Speech, Music Processing (2022)
- Khysru, K., Wei, J., & Dang, J., "Research on Tibetan Speech Recognition Based on the Am-do Dialect," Computers, Materials & Continua, 73(3), 4897–4907 (2022). https://doi.org/10.32604/cmc.2022.027591 — baseline CER 15.21% (N-gram LM), improved to 11.97% with proposed methods; CER on noisy/reverberant test data reached 52.17% before augmentation reduced it to 23.20%, showing real-world fragility beyond clean-lab conditions
- Ma, L., Li, G., & Zhe, R., "A Study on Speech Recognition of Tibetan Amdo Based on Whisper," Proc. 2024 3rd Int'l Conf. on AI, Internet and Digital Economy (ICAID 2024), Atlantis Press, 367–373. https://doi.org/10.2991/978-94-6463-490-7_40 — Whisper-base fine-tuned on Amdo Tibetan reached 23.84% CER; Whisper-medium reached 9.31% CER, notably better than the base-model figure alone suggests
- "Ti-Audio: The First Multi-Dialectal End-to-End Speech LLM for Tibetan," arXiv (2026) — https://arxiv.org/html/2604.11110v1
- NVIDIA official Jetson comparison / specifications page — https://nvidia.com/jetson
- NVIDIA Blog, "NVIDIA Introduces New Jetson Thor Computers to Advance Mainstream Robotics and Edge AI" (2026) — https://blogs.nvidia.com/blog/jetson-thor-robotics-edge-ai-agent/
- Human Rights Watch, "A Hazard to Human Rights: Autonomous Weapons Systems and Digital Decision-Making" (2025) — https://www.hrw.org/report/2025/04/28/a-hazard-to-human-rights/autonomous-weapons-systems-and-digital-decision-making
- UNRIC, "UN Addresses AI and the Dangers of Lethal Autonomous Weapons Systems" (2025) — https://unric.org/en/un-addresses-ai-and-the-dangers-of-lethal-autonomous-weapons-systems/
- "Efficient Compression of Multitask Multilingual Speech Models," arXiv (2024) — https://arxiv.org/pdf/2405.00966
- "Edge-ASR: Towards Low-Bit Quantization of Automatic Speech Recognition Models," arXiv (2025) — https://arxiv.org/pdf/2507.07877
- "LoRA-INT8 Whisper: A Low-Cost Cantonese Speech Recognition Framework for Edge Devices," PMC (2026) — https://pmc.ncbi.nlm.nih.gov/articles/PMC12431075/ — LoRA fine-tuning brought Cantonese CER from 49.5% to 11.1%; subsequent INT8 compression preserved that accuracy rather than degrading it. Cited here as a proof-of-concept counterexample: the degradation problem is solvable when someone specifically builds for a given language, which nobody has yet done for Tibetan in an edge/military context
- Cybernews, "Autonomous weapons are already here: How close are we to machines deciding between life and death?" (2025) — https://cybernews.com/ai-news/how-close-are-we-to-autonomous-weapons/
- Congress.gov CRS Report, "Pentagon-Anthropic Dispute over Autonomous Weapon Systems: Potential Issues for Congress" (2026) — https://www.congress.gov/crs-product/IN12669
- The Conversation, "Anthropic v the US military: what this public feud says about the use of AI in warfare" (2026) — https://theconversation.com/anthropic-v-the-us-military-what-this-public-feud-says-about-the-use-of-ai-in-warfare-276999
- CNBC, "Pentagon-Anthropic AI standoff is real-time testing balance of power in future of warfare" (2026) — https://www.cnbc.com/2026/02/27/defense-anthropic-ai-war-risks-hegseth-amodei.html
- CNBC, "Anthropic faces lose-lose scenario in Pentagon conflict as deadline for policy change looms" (2026) — https://www.cnbc.com/2026/02/27/anthropic-pentagon-ai-policy-war-spying.html
- Mayer Brown, "Anthropic Supply Chain Risk Designation Takes Effect" (2026) — https://www.mayerbrown.com/en/insights/publications/2026/03/anthropic-supply-chain-risk-designation-takes-effect--latest-developments-and-next-steps-for-government-contractors
- Arms Control Association, "Pentagon Labels AI Company Supply Chain Risk" (2026) — https://www.armscontrol.org/act/2026-04/news-briefs/pentagon-labels-ai-company-supply-chain-risk
- Pearl Cohen, "Anthropic Sues Department of Defense Over Supply Chain Risk Designation" (2026) — https://www.pearlcohen.com/anthropic-sues-department-of-defense-over-supply-chain-risk-designation/
- CNN, "Judge blocks Pentagon's effort to 'punish' Anthropic by labeling it a supply chain risk" (2026) — https://www.aol.com/finance/judge-blocks-pentagon-effort-punish-002533318.html
- The Hill, "Appeals court hears Anthropic's Pentagon AI suit" (2026) — https://thehill.com/policy/technology/5885923-appeals-court-hears-anthropics-pentagon-ai-suit/; Federal News Network, "Appeals court judges appear to be divided over Pentagon's legal dispute with AI company Anthropic" (2026) — https://federalnewsnetwork.com/artificial-intelligence/2026/05/appeals-court-judges-appear-to-be-divided-over-pentagons-legal-dispute-with-ai-company-anthropic/
- Breaking Defense, "Air Force pushing contractors to purge Anthropic by Sept. 1: Memo" (July 10, 2026) — https://breakingdefense.com/2026/07/air-force-pushing-contractors-to-purge-anthropic-by-sept-1-memo/
- Endangered Language Alliance, "Tibetan" — https://www.elalliance.org/languages/tibetan; corroborated by EBSCO Research Starters and Wikipedia's Sino-Tibetan overview, all converging on approximately 6 million Tibetan speakers worldwide
(Note: sources 4 and 5 are single-paper WER figures on different test sets and dialects, not points on one controlled benchmark — treat the "flat trend" claim as directional, not a precise slope, if pressed on it.)
