AI fails at the basics: causality, companies and graduates

Today’s most important AI news

Thursday was marked by an uncomfortable but necessary current: empirical evidence piling up against the hype. Three different signals point in the same direction — language models have structural limits that aren’t solved with more parameters or more data. Today the topic isn’t what AI can do, but what it can’t do, and why that matters for those making decisions about products and companies.

LLMs don’t understand cause and effect — and it’s a mathematically inevitable problem

A new arXiv paper titled Why LLMs Fail at Causal Discovery and How Interventional Agents Escape reaches a conclusion that should shake any team using language models for analysis or decision-making: LLMs’ inability to reason causally isn’t a training bug, it’s a fundamental limitation. The researchers formally prove that standard training techniques — supervised fine-tuning, direct preference optimization (DPO) and in-context learning — produce models that can only learn correlations, not cause-and-effect relationships. And the benchmarks confirm it: even models tuned specifically for this task quickly stall as causal graphs grow in complexity.

This isn’t an academic problem. If you use an LLM to analyze why your retention metrics dropped, why a user segment converts less, or for any “if we change X, what happens to Y” reasoning, the model is doing statistical pattern matching disguised as causal analysis. The solution the paper proposes is interventional agents — systems that can act on the environment and observe real results, instead of inferring everything from static text.

The practical implication is direct: if your product or process depends on AI telling you why something happened or what would happen if you change something, you need a design that includes real experimentation, not just queries to a model. LLMs can be a piece of the system, but not the causal arbiter.

Frontier models fail at more than 50% of real enterprise IT tasks

IBM and Artificial Analysis published ITBench-AA, the first benchmark designed specifically to evaluate AI agents on real-world IT tasks within enterprise environments. The result is brutal: the most advanced models available today achieve less than 50% accuracy on these tasks. These aren’t theoretical questions — the benchmark evaluates real agentic workflows: incident diagnosis, configuration management, operations automation.

The central problem the benchmark reveals is the difference between performing well on controlled evaluations and functioning in environments with state, dependencies between systems and real consequences for errors. Current agents are good at discrete, well-defined tasks, but degrade when context is ambiguous, steps are interdependent and there’s no single correct answer to extract from text.

For engineering and operations teams evaluating adopting AI agents to automate IT work, this benchmark is a concrete reference. Before committing budget or architecture, it’s worth mapping your use cases against ITBench-AA’s categories. If the best models on the market fail at more than half of those scenarios, expectations about timelines and reliability should be adjusted accordingly.

Eric Schmidt booed: the first sign of social fatigue with AI

In an episode MIT Technology Review included in this week’s AI Hype Index, Eric Schmidt — former CEO of Google — was greeted with boos when telling University of Arizona graduates that their mission is to help shape AI. The anecdote might seem trivial, but it represents something industry professionals should take seriously: the narrative that AI is a universal and inevitable opportunity is colliding with a generation facing an uncertain job market and hearing technological promises that haven’t materialized in their lives.

This isn’t an isolated phenomenon. Fatigue with AI hype is growing across different layers of society — from workers who see their roles threatened without a clear retraining roadmap, to end users experiencing products that over-promise and under-deliver. The fact that it happened at a graduation ceremony, one of the most symbolically charged rituals about the future, gives it particular weight.

For those building products and communicating AI strategies outward — to customers, employees or stakeholders — the signal is clear: the language of hype now carries a credibility cost. Audiences are more willing than before to publicly reject narratives they perceive as disconnected from their reality. Communicating precisely what AI does, where it fails and for whom it generates real value has stopped being an optional virtue.

In a nutshell

What these three stories reveal together today is a pattern of forced maturation. Research demonstrates mathematical limits. Enterprise benchmarks expose performance gaps. And the social reaction indicates that the contract of trust with the technology is being renegotiated. The sector has spent two years accelerating on the premise that models’ current problems are temporary — a matter of scale, more data, better fine-tuning. Today’s evidence suggests some of those problems are structural. That doesn’t mean AI isn’t transformative, but that the organizations that win in the coming years will be those that build on its real capabilities, not its projected ones.

---

Sources used: https://arxiv.org/abs/2605.27567, https://huggingface.co/blog/ibm-research/itbench-aa, https://www.technologyreview.com/2026/05/28/1138053/the-ai-hype-index-ai-gets-booed-in-graduation-season/