We’re halfway through 2025, and the hype around AI shows no signs of slowing.

In just six months, we’ve seen DeepSeek challenge U.S. model dominance, multibillion-dollar bets like Stargate, a steady drumbeat of state-of-the-art model launches, and Nvidia becoming the world’s first $4 trillion company.

Amidst the flood of headlines, it's hard to distinguish reality from market theater. This month, we step back to assess the loudest AI narratives in 2025.

⸻

Narrative 1: 2025 is the Year of the Agent

Verdict: False

The Claim

The term “agentic” has become a buzzword. Agents are being billed as the next leap—AI systems that autonomously execute complex, multi-step tasks. Launches like Salesforce’s AgentForce have fueled the hype, to the point that media outlets now debate what “agent” even means. 

The Reality

That future remains aspirational. Gartner warns of widespread ‘agent washing’: companies misrepresenting their products as fully autonomous AI agents. They estimated only ~130 of thousands of agentic offerings are credible. Salesforce has seen slow adoption. OpenAI’s Operator has already been replaced by Agent, which testers call a ‘big improvement but still not very useful’. 

While AI products have improved, they still lack true autonomy. A helpful analogy comes from autonomous driving: 

Source: Synopsys

Agents today are not Level 5 “self-driving.” At best, they resemble adaptive cruise control—useful for well-bounded, repetitive tasks, but not capable of fully managing the road ahead. 

Narrative 2: AI Has Triggered a Wave of White-Collar Layoffs

Verdict: Partially true

The Claim

Stories about AI-induced job loss are everywhere. Executives cite automation in workforce reductions, and headlines link AI to layoffs. Rising unemployment among graduates and slowing professional job growth suggest that white-collar displacement is already underway.

The Reality

We examined this in What AI Means for the Class of 2025. So far, few layoffs have been directly tied to AI. Most still stem from macroeconomic headwinds—slow growth, high interest rates, and supply chain impacts (like GM’s $1.1B tariff charge).

Still, it would be naive to say displacement isn’t coming. Leaders across industries are exploring where AI can streamline or automate work. 41% of executives plan headcount reductions by 2030. In areas like data entry, logistics, and support, task-level automation is already happening—and models are improving quarterly. While we remain optimistic about talent redeployment, this wave is likely to move faster than past ones. The shift has begun, with the boundaries of white-collar work being redrawn in real time.

Narrative 3: The GenAI Race is Wide Open

Verdict: Mostly false

The Claim

With a rising field of contenders—Perplexity, Claude, Grok, Gemini, and open-source players like LLaMA and DeepSeek—the GenAI market seems competitive and dynamic. Falling prices signal looming commoditization, with no clear winner.

The Reality

Despite more entrants, the field is consolidating. Open-source models have improved, but still trail closed systems by 6 to 20 months. DeepSeek made headlines with a near-frontier model, but it was later revealed to be distilled - i.e. its knowledge and capabilities were taken from OpenAI’s frontier model - raising legal and technical questions about true innovation.

On the commercial front, OpenAI is miles ahead. As measured by website traffic, OpenAI accounts for 190 million of the 240 million daily visits to GenAI tools—roughly 80% of the total:

Source: Similarweb

OpenAI’s advantage will compound, as higher usage provides more data for training, enabling faster product improvement, which in turn draws more users. This dynamic mirrors the streaming wars: many services launched (think Peacock, Paramount+, etc), but Netflix’s larger user base generated more capital to plough back into content acquisition, in turn attracting more users. 

Still, this narrative is mostly false, not entirely false. The real race isn’t for chatbot dominance—it’s for AGI (Artificial General Intelligence). That milestone would reshape economic power. OpenAI leads today, but the long game remains open.

Narrative 4: Hallucination Remains a Major Problem

Verdict: Mostly false

The Claim

Hallucination—AI’s tendency to generate inaccurate or fabricated content—is seen as an unsolvable flaw. Critics argue GenAI is fundamentally less reliable than humans.

The Reality

This narrative lags behind the data. Hallucination rates have dropped from 21.8% in 2021 to 0.7% in 2025:

Source: All About AI

As Anthropic CEO Dario Amodei notes, “AI models probably hallucinate less than humans, but they hallucinate in more surprising ways.”

Media coverage contributes to the perception gap: spectacular errors go viral, while routine accuracy goes unnoticed. But the benchmarks speak for themselves:

• Breast cancer detection: 9.4% improvement and 5.7% fewer false positives

• Fake review detection: 73% accuracy vs. 55% for humans

• Data entry: 99.96–99.99% accuracy vs. 96–99% for humans

The takeaway: hallucination isn’t gone, but it’s increasingly manageable. Rates vary by model and task. Human oversight is still critical, but the trend is heading in the right direction.

Narrative 5: AI is Overhyped and Bubble Will Burst

Verdict: False

The Claim

The counterargument to all the AI hype is twofold: (a) it represents an even bigger bubble than the dotcom era and (b) the core technology is just sophisticated autocomplete—occasionally impressive, but fundamentally unreliable and limited in real-world utility. Critics question whether companies will ever realize meaningful returns on their AI spending.

The Reality:

Hype has outpaced product maturity in some areas (see Narrative 1), but capability is advancing fast. Hallucination rates are down. OpenAI’s o3 model recently scored 116 on an independent IQ test, putting it up with the top 15% of human testers. Just a year ago, models lagged far behind the human average IQ of 100:

Source: Maxim Lott

You may have felt this shift yourself. Whether it’s having a coherent voice conversation, completing a Deep Research sprint, or using tools like Claude Code or Loveable, leading indicators of AI’s real-world utility are becoming more visible. Businesses are feeling it too: automation of repetitive manual work is expanding across functions, even if it doesn’t always generate headlines.

The takeaway: While some of the excitement may be premature, the underlying progress is real, and continuing unabated.

Closing Thoughts

The real story of AI in 2025 isn’t about flashy agents or viral demos. It’s about steady, measurable improvements in functionality and reliability.

For business leaders, the playbook is clear: ignore the noise. Focus on where AI can quietly reduce friction—those manual, repeatable processes still bogging down teams. That’s where value is emerging now.

And while mass layoffs haven’t materialized yet, displacement is likely to accelerate as capabilities and economic pressure grow. The flywheel of progress, adoption, and workforce impact is already in motion.