AI scams, ideological bias and the hidden cost of agents
Today: how AI supercharges mass fraud, how biased LLMs are on economics and why agents pay too much for their tools.
Today’s most important AI news
Friday closed with three signals pointing in the same direction: AI is no longer just a productivity promise, it’s also infrastructure for fraud, for politics and for costs nobody was measuring. There were no big launches today, but there was research and analysis that forces a rethink of how the technology is being deployed in the real world. For developers, product managers and decision-makers, the implications are direct.
AI has fully entered the fraud industry — and nobody is ready
Since ChatGPT appeared in 2022, generative models have shown that producing convincing, human-sounding text is trivial. What wasn’t obvious back then was the scale that would enable in the world of digital crime. According to MIT Technology Review, we’re in a new era of AI-powered scams: phishing emails without spelling mistakes, calls with cloned voices, synthetic identities that pass verification filters and automated social-engineering campaigns that used to require human teams.
The qualitative change isn’t just that scams are more convincing, but that they’re cheaper to run at scale. A malicious actor who used to need dozens of people to operate a fraud campaign can now do it with accessible models and automated pipelines. This puts direct pressure on security, compliance and product teams at any company that handles user data or transactions.
For those building products: the friction that used to act as a barrier — mistakes, inconsistencies in the text, slow responses — is no longer enough of a defense. Trust & safety teams need to assume the adversary has access to the same tools they do, and design verification layers that don’t depend on spotting “what sounds off”.
LLMs have ideological bias when analyzing economics — and that’s a real problem
Research published on arXiv evaluated whether large language models show systematic ideological bias when reasoning about causal effects in economics. The short answer: yes, and it’s consistent. Using an extension of the EconCausal benchmark, the researchers compared how different LLMs judge pro-market versus pro-government interventions. The models tended to favor interventionist narratives in scenarios where the empirical evidence is ambiguous or disputed.
This isn’t a minor academic problem. LLMs are being used today for public-policy analysis, to generate economic context reports and to support strategy teams at companies and governments. If the model you’re consulting has a systematic lean in how it interprets economic causality, the outputs it produces — even if they sound neutral and well-grounded — are biased from the inside.
The practical implication is simple but uncomfortable: if you use LLMs for economic or policy analysis, you need to actively test whether the model answers differently to equivalent questions framed from different ideological perspectives. You can’t assume neutrality just because the model is big or popular.
The “hidden tax” AI agents are paying on tools
A paper published this week identified a problem that any developer working with agents and the Model Context Protocol (MCP) has already felt without being able to name it: the MCP Tax, or “tool tax”. Every time an agent connected to multiple MCP servers processes a conversation turn, it statically loads the schemas of every available tool — even if it won’t use them. According to the authors, that can add up to between 10,000 and 60,000 tokens per turn in typical multi-server deployments.
The cost isn’t only monetary, although that’s already significant if you’re running agents in production with pay-per-token models. The subtler problem is that inflating the context with unnecessary schemas degrades reasoning: models with longer, noisier contexts make more mistakes, skip instructions or lose coherence on long tasks. In other words, you pay more and get a worse result.
The paper’s proposal — dynamic tool loading and “tool attention” that only activates the schemas relevant to the current turn — is still research, not an implemented standard. But the diagnosis is what matters right now: if you have agents in production connected to MCP, it’s worth auditing how many tokens are being spent on overhead for tools that are never invoked. The savings can be considerable.
In a nutshell
What connects these three stories is the same gap: AI is being deployed faster than threat models, evaluation frameworks and the architecture of the systems that use it are being updated. Fraud scales because defenses assume human adversaries. Economic analyses get distorted because nobody tests for bias before using the output. Agents quietly burn tokens because the default architecture wasn’t designed for production efficiency. In all three cases, the problem isn’t the technology itself — it’s that it was adopted without the checks that make it trustworthy. That’s what defines this stage of AI: not capability, but operational maturity.
Sources used: https://www.technologyreview.com/2026/04/24/1136400/the-download-supercharged-scams-questionable-ai-healthcare/, https://arxiv.org/abs/2604.21334, https://arxiv.org/abs/2604.21816