Military AI 'Dangerously Inaccurate' — Ex-OpenAI Scientist

Military AI accountability is at the center of a new warning from Heidy Khlaaf, former OpenAI safety lead — and her conclusion is stark: the models being used in active military operations are not accurate enough to be trusted with life-or-death targeting decisions. Heidy Khlaaf spent years at OpenAI working on AI safety. Now she's warning that the models her former employer helped build are being pointed at human targets — and that they're not accurate enough to be trusted with those decisions. This isn't hypothetical. It's happening now, in active military operations.

The AI Now Institute, an independent research organization where Khlaaf serves as Chief AI Scientist, just released findings that cut through Silicon Valley's "responsible AI" messaging to document what's actually happening on the ground — from AI-assisted weapons targeting in the Middle East to data extraction from communities across the Global South.

The AI Safety Scientist Who Left OpenAI — and Didn't Stay Quiet

Khlaaf's move from OpenAI to AI Now Institute as Chief AI Scientist wasn't a career retreat — it was a deliberate repositioning. She specializes in autonomous weapons systems and AI ethics (the study of how AI decisions affect human lives and who is held accountable when they go wrong). That expertise is now pointed squarely at the U.S. military's adoption of AI.

Her most urgent warning concerns AI used in planning U.S. military operations against Iran. Her own words: "It's very dangerous that 'speed' is somehow being sold to us as strategic here, when it's really a cover for indiscriminate targeting when you consider how inaccurate these models are."

The concern isn't abstract. AI language models (programs that generate text and make recommendations based on patterns in large training datasets) are being integrated into military decision-making pipelines. But those same models famously hallucinate — confidently stating false information as fact — a flaw that produces embarrassing chatbot errors in a consumer app, and potentially catastrophic targeting errors in a war zone.

Khlaaf on the AI safety gap: "The existing guardrails for generative AI are deeply lacking, and it has been shown how easily compromised they are, intentionally or inadvertently. It's highly doubtful that if they cannot guard their systems against benign cases, they'd be able to do so for complex military and surveillance operations."

When "AI for Good" Becomes Cover for Extraction

If the military AI story is alarming, the "AI for Good" story is insidious in a quieter way. AI Now Institute has documented a pattern it calls "impact lingo" — the systematic use of high-sounding phrases designed to pre-empt criticism while perpetuating the same extractive systems underneath.

Common examples you'll recognize from press releases and conference stages:

• "AI for Good" — used to justify expansion into communities without consent or fair benefit-sharing
• "Frugal AI" — reframes resource-intensive systems as sustainable, without measurable evidence
• "AI for climate" — deployed as green marketing while data centers (massive server facilities that consume electricity and water around the clock) devour local resources
• "Democratization" — stripped of its historical political meaning and repurposed as a product launch term
• "Responsible AI" — cited in press releases without binding commitments or third-party verification

The institute's position is pointed: these phrases have been "stripped of historical meaning and repurposed as marketing." When a company announces it's "democratizing AI access," the honest question is what that means in practice. If it means deploying a centralized platform that extracts user data and sends profits back to Silicon Valley — that's the opposite of democratization. As one AI Now analysis put it, you might just be "distributing the terminals and the data extraction facilities of a centralized authority."

The Data Pipeline Nobody Talks About at AI Launch Events

Behind every AI model is a training dataset. Behind many of those datasets are workers in the Global South — Kenya, the Philippines, Venezuela — who spend hours labeling disturbing content, categorizing images, and cleaning data for wages that would be considered illegal in the countries that profit from their work. This labor is largely invisible in the finished product and absent from every launch keynote.

AI Now Institute has documented how this labor extraction follows colonial patterns (the historical practice of wealthy nations removing resources from poorer regions without fair compensation — now updated for the digital economy). Data is the new raw material. The Global South provides it cheaply; Silicon Valley monetizes it at scale, then announces it's "helping" those same communities with AI tools they didn't ask for and can't fully use.

The linguistic dimension of this problem is stark: there are 2,000+ African languages that currently receive minimal AI support. Building effective AI systems for those languages requires deep local expertise — the kind that can only come from communities who actually speak them. But the current economic model offers little incentive to invest in that expertise, because those user bases generate lower profit margins than English-speaking markets.

Meanwhile, AI data centers are facing real pushback. Residents in Chile, multiple U.S. states, and Canada have organized against new AI infrastructure buildouts, citing water consumption, electricity demand, and the fundamental question: who actually benefits from this being built in our community while we bear the environmental costs?

Big Tech, Big Oil, and the Pentagon: The AI Accountability Gap

AI Now Institute has mapped what it calls the deepening nexus between Big Tech, Big Oil, and state military actors. This isn't speculation — it's organizational chart reading combined with contract tracing.

OpenAI signed a partnership with the U.S. Pentagon. Anthropic, the maker of Claude, holds defense contracts. Microsoft Azure hosts classified government data. Google has conducted research for defense agencies. The companies publicly pledging to "align AI with human values" are simultaneously selling those same systems to militaries with specific tactical objectives — where "alignment" means something very different from what's written in their responsible AI charters.

The environmental overlay makes this structurally worse. AI data centers require massive, sustained electricity consumption — a footprint that's growing rapidly as models scale in size. In many regions, that electricity still comes from fossil fuel infrastructure. The same companies running "AI for climate" campaigns are — through their compute (the physical computing power and server infrastructure required to run AI at scale) — deepening reliance on the energy systems driving climate change. The 2025 AI Now Landscape Report documents this contradiction in rigorous detail.

What You Can Do With This AI Accountability Information Right Now

The AI Now Institute doesn't just document problems and publish PDFs. It advocates for "concretizing" abstract AI claims — mapping actual infrastructure, tracking measurable impacts, and demanding empirical evidence before accepting any organization's impact assertions at face value. That's a methodology anyone can apply.

Its proposed alternative is what the institute calls "prototypes of struggle": community-led AI projects controlled by the people they serve, often built through federated coalitions (networks of independent organizations that pool technical resources and knowledge without surrendering governance to a single corporate platform). The honest acknowledgment from the institute: these communities are currently "small, resource-constrained and struggling to survive." Building counter-infrastructure to Big Tech with limited funding is genuinely hard. But the alternative — accepting that AI development will be defined entirely by whoever has the most compute — is worse.

The March 2026 India AI Impact Summit became a flashpoint for this work. While tech companies announced data center investments and partnerships, the institute's side conversations revealed what's actually happening on the ground: digital colonialism relabeled as development aid. By centering voices like Timnit Gebru (former Google AI ethics researcher and co-founder of the Distributed AI Research Institute), Audrey Tang (Taiwan's former Digital Minister), and Meredith Whittaker (Signal's president and longtime AI critic), the institute is building a counter-narrative grounded in community experience rather than venture capital assumptions.

If you work in tech, policy, media, or at an organization that deploys AI tools, AI Now Institute's 2025 Landscape Report is worth your time — not for its warnings, but for its framework for asking better questions. The next time you see "AI for Good" in a pitch deck, a grant application, or a government announcement, you'll have the vocabulary to ask: good for whom, measured how, and accountable to who? You can also explore our AI accountability guides to apply these questions inside your own organization.

Related Content — Get Started | Guides | More News