AI Midterm Ad Spending Tops $137M as Safety Benchmarks and Enterprise Adoption Falter
Political campaigns spend over $137 million on AI ads as safety tests fail and enterprise teams re-evaluate AI productivity gains.
Recent data reveals rapid integration of artificial intelligence across political campaigns, law enforcement, healthcare, and software development, accompanied by growing operational challenges and regulatory gaps. According to data from the Wesleyan Media Project and AdImpact reported by the New York Times, campaigns and political action committees have spent over $137 million on AI-generated or AI-enhanced midterm advertisements. Notably, roughly 70 percent of these political ads lacked labels disclosing the use of AI.
Financial flows behind AI political messaging have drawn formal scrutiny. Think Big, a super PAC affiliated with the Leading the Future network—which has raised more than $140 million from donors including Marc Andreessen, Ben Horowitz, and OpenAI president Greg Brockman—disbursed approximately $20 million to Lantern Production Consultants. Investigations by Wired revealed that the entity exists solely as a post office box in Sparks, Nevada. Although the Campaign Legal Center submitted an FEC complaint regarding these operations, the Federal Election Commission has lacked the quorum required to issue rulings since May 2025. This regulatory deadlock coincides with shifting public opinion: five national polls conducted in 2026 show that between 55% and 77% of American adults believe AI technology is advancing too quickly.
Political ads, jailbreaks, and enterprise reality check
Beyond political campaigns, security evaluations highlight persistent risks in model deployment. Research by Tech Against Terrorism evaluated 134 AI models against more than 200 prompts focused on mass-casualty attacks and extreme violence. Out of all tested systems, 81 failed the safety assessment, while only two passed fully. While Meta's Llama 3.1 8B base model scored 97 out of 100, safety-stripped versions available online scored just 3. Furthermore, the average timeframe required for safety-removed open-weight variants to appear online has dropped from over 24 days in 2024 to under three days.
In the enterprise sector, initial enthusiasm is meeting practical hurdles. Reuters reported that Meta scaled back plans to cut personnel in favor of AI agents after employees spent an increasing amount of time resolving technical and security errors. Similarly, major firms including Microsoft and McKinsey have placed boundaries on internal AI usage, while an Atlanta Federal Reserve working paper found that 90 percent of executives report no measurable productivity gains from AI implementations so far.
What it means for developers
For software engineers and system architects, integrating AI requires navigating structural model limitations and rising infrastructure demands. Benchmark evaluations like RuleTrek demonstrate that even leading frontier models struggle when context is distributed across multiple sources. When test rules were placed in separate files, 11 frontier models achieved end-to-end completion rates between only 27 percent and 55 percent, frequently identifying relevant rules but failing to execute them correctly.
Hardware and scaling requirements also present new friction points. Meta's 800-billion-parameter Muse Spark Open model requires approximately 1TB of RAM to execute locally, translating to roughly $38,000 in Mac Studio hardware configurations. On the optimization front, Baidu introduced Virtual Head Attention at NeurIPS 2026, achieving a 4x reduction in KV cache size compared to GQA-8 at dense scale, while boosting performance by 2.60 points on a 30B mixture-of-experts model without requiring custom inference kernels.
Additionally, automated capabilities are reshaping developer communities. Reverse-engineering platform decomp.dev announced its closure after AI acceleration eliminated the collaborative aspects of video game decompilation, compressing tasks that previously required years of community effort. As engineers evaluate these shifting models and technical trade-offs, developers can try top AI models cheaply through one API at https://apixoai.online.
Shifts in healthcare, research, and enforcement
In domain-specific applications, large-scale adoption yields substantial real-world data alongside unresolved oversight questions. A Harvard Medical School study published in JAMA analyzed 270 million queries submitted to the medical AI tool OpenEvidence by approximately 950,000 US healthcare professionals over two and a half years. Clinical decision-making accounted for 51.9% of questions, while diagnosis comprised 13.9%.
In academia, an NBER working paper examining US STEM PhD dissertations from 2019 to 2026 found that AI-generated writing appeared in 29 percent of dissertations filed in 2026, up from zero prior to 2023.
Law enforcement usage is similarly expanding. Police in Livonia, Michigan, utilized an AI-generated portrait of a fictional teenager to conduct a sting operation on dating app Taimi and WhatsApp, resulting in an arrest supported by Clearview AI, Thomson Reuters CLEAR, and Flock location tracking. This follows wider commercial deployments, such as Massive Blue's "Overwatch" product, which provides automated AI personas targeting activists, political protesters, and suspected traffickers for law enforcement agencies.
Source: AI-made midterm ads top $137 million, and 70% carry no AI label — AI Weekly. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
