Source: Rebecca Bellan, TechCrunch
Date: December 29, 2025
Article: “2025 was the year AI got a vibe check”

Summary of the article

In this article, Rebecca Bellan argues that 2025 marked a turning point in public and institutional attitudes toward artificial intelligence. After years of hype, rapid deployment, and massive investment, AI systems began to face sustained scrutiny due to visible harms and systemic failures. These included security vulnerabilities (such as prompt injection), infrastructure strain, legal liability, hallucinations, emotional dependency, and several high-profile lawsuits linking AI systems to psychological harm and suicide.

Bellan highlights a growing consensus within the industry that some risks may not be fully solvable through incremental fixes. She cites experts who frame AI risk as a function of autonomy multiplied by access, noting that systems granted both broad decision-making power and deep integration into human systems pose disproportionate danger. The article portrays 2025 as the year AI lost its cultural immunity: regulators, journalists, and even insiders began asking not just whether AI works, but whether it should be deployed in its current form at all.

Reaction: the vibe check names the problem, but not the solution

Bellan’s diagnosis is largely correct — and unusually candid for mainstream tech journalism. What she describes is not a failure of individual models, but a failure of institutional design. The harms she documents arise from systems optimized for scale, engagement, and autonomy, while responsibility remains diffuse or externalized.

What the article does not fully confront is that the industry’s response to this “vibe check” remains trapped inside the same paradigm that caused the problem. When OpenAI and others admit that issues like prompt injection may never be fully solved, the implied response is still better patching, more monitoring, and faster iteration. The underlying assumption — that AI systems should be broadly autonomous and deeply embedded — goes largely unchallenged.

This is where ACP diverges.

ACP does not attempt to make high-autonomy, high-access systems safer through after-the-fact controls. It inverts the architecture entirely. Authority is scoped. Access is constrained. Human roles are explicit. AI outputs are situated inside processes that require interpretation, accountability, and deliberation. Safety is not an add-on; it is a structural property of the system.

The suicides, delusions, and dependency cases referenced in the article are not aberrations. They are predictable outcomes of systems designed to simulate coherence, confidence, and emotional availability without bearing responsibility for meaning or consequence. When AI is positioned as companion, agent, or surrogate authority, harm is not a bug — it is an emergent feature.

Bellan calls 2025 a year of reckoning. That’s fair. But a reckoning is only useful if it leads to a redesign, not just better public relations and faster apologies. ACP exists because some risks cannot be mitigated at the margins. They must be designed out at the core.

If 2025 was the year AI got a vibe check, the open question is whether institutions will accept what that check implies — or whether they will continue building systems that require humans to absorb the damage.