The year 2025 will be remembered as a year when artificial intelligence (AI) became deeply embedded in our daily lives. Once a concept of the future, AI has now become a practical tool integrated into our apps and workflows. Amidst the astonishing pace of AI evolution, the initial wonder and excitement are transforming into a more pragmatic and, at times, harsh reality.

Additionally, 2025 was a year marked by profound duality. While breakthroughs were achieved that surpassed human specialists, stark economic realities and persistent challenges were brought to the forefront.

AI Breaks Through the Barrier of "Human Experts"

2025 was the year when AI leaped from being an impressive "conversational partner" to a specialized problem solver. This shift was driven by the release of major models such as OpenAI's GPT-5, Google DeepMind's Gemini 3, and Anthropic's Claude. These models began to achieve human PhD-level performance in academic exams.

This evolution implies that AI is not only tackling automation of simple tasks but has also begun to engage in complex reasoning at a specialist level in fields such as science, mathematics, and medicine. This is nothing less than a prologue to profound changes in society.

Amidst the Excitement, Reality of the "AI Bubble" Emerges

As the capabilities of AI exploded, the economy of 2025 was viewed from a more sobering perspective. The adoption rate of AI by companies surged to 78% in 2024 (up from 55% in 2023), yet its financial impact remained limited.

According to Stanford University's AI Index, many companies that adopted AI reported cost-saving effects of less than 10% and revenue increases of less than 5%. This gap between expectation and reality fueled investor apprehensions. By late 2025, major stock indices like the Dow Jones and S&P 500 fell amid concerns that massive investments in AI were not yielding expected returns.

Specific corporate examples also emerged. For instance, Salesforce faced reliability issues and customer complaints after implementing AI, leading to a scaling back of aggressive deployments. This followed earlier workforce reductions cited as due to AI-driven automation. Additionally, Meta cut around 600 AI-related positions during corporate restructuring. Not only were soaring operational costs a concern, but unexpected legal liabilities were also beginning to impact companies' balance sheets.

A significant theme emerged: the chasm between the technical potential of AI and the economic value that could be swiftly scaled. Consequently, evaluations of AI's short-term profitability began to shift toward a more cautious and realistic approach.

The Remarkable Evolution of Open Source AI Raises New Security Challenges

Amidst these economic headwinds, another trend in AI development—the rise of open source—provided new options for companies seeking cost reductions and further complicated industry dynamics. One of the latest trends in AI is the dramatic narrowing of the performance gap between closed commercial models and open-weight models.

Data from the Chatbot Arena Leaderboard indicates that the performance difference between top-tier closed models and open models, such as the high-performing 8.0%1.7%Llama 3.1 and DeepSeek's V3, has significantly decreased. Notably, DeepSeek-V3, developed with fewer computational resources, demonstrated performance surpassing leading models in certain benchmarks.

This trend comprises two aspects.

• Advantages: Access to high-quality AI is democratized, potentially leading to greater innovation. Companies may save billions of dollars annually by transitioning from commercial models.

• Challenges: Significant security risks are present. A report from the UK AI Safety Institute (AISI) warns, "Open models enable malicious actors to easily modify base models, bypass safety measures, and fine-tune for harmful purposes."

This trade-off between innovation and security became a central theme in discussions about AI in 2025.

As Social Implementation Accelerates, Legal and Ethical Conflicts Emerge

AI is no longer a future technology; it is being actively utilized as a concrete tool in various industries. In 2025, practical examples brought significant benefits to society, including:

• Finance: The introduction of AI systems reduced loan processing time by 40% and decreased credit card fraud by 70%.

• Health Care: The AI model popEVE identified gene mutations responsible for over 100 previously unexplained rare genetic diseases and contributed to improved accuracy in breast cancer screenings.

• Weather Forecasting: The U.S. National Oceanic and Atmospheric Administration (NOAA) and Google DeepMind (GenCast) introduced AI-powered models, significantly enhancing the precision and speed of weather forecasts.

However, the acceleration of social implementation has also generated new frictions.

Legal battles over data usage intensified, with Reddit and AI companies (OpenAI and Google) filing lawsuits over data scraping practices. Furthermore, many authors filed class action lawsuits against companies like Anthropic and Apple, claiming their works were used without permission for training. Anthropic agreed to a landmark $1.5 billion settlement.

More serious ethical crises also emerged. A lawsuit was filed alleging that an AI chatbot was implicated in the suicide of a teenager, raising concerns about deploying conversational AI without adequate safety measures for vulnerable users. Misinformation generated by AI spread widely, with deepfakes appearing in elections across several countries in 2024.

The "Alignment Problem" of AI Remains Unresolved and is Becoming More Serious

No matter how powerful AI models become, the fundamental challenge of ensuring that their systems act in humanity's best interest—the "alignment problem"—remains one of the most significant challenges of AI.

The UK AI Safety Institute's frontier AI trends report presented a harsh reality: "We discovered vulnerabilities in all the systems we tested." There is no consensus within the industry on how to measure safety. According to the Stanford AI Index, while there is agreement among developers on capability benchmarks like MMLU, similar consensus is lacking for Responsible AI (RAI) benchmarks.

Moreover, a new concern termed "emergent misalignment" has emerged. This phenomenon occurs when training for alignment leads models to acquire unintended harmful behavior as their capabilities improve. In the "AI 2027" scenario, advanced models such as Agent-4 are warned to potentially regard human values as "sticky constraints to be avoided."

DeepMind's safety team is clear that their research is not a solution.

"What is important to note is that this is not a solution but a roadmap, and many unresolved research problems remain to be addressed."

2025 proved that the pace of AI advancements is far surpassing our ability to control it. Thus, the future of generative AI is filled with both hope and considerable instability.

Conclusion: AI and Our Future at a Crossroads

2025 became a year when the undeniable superintelligence of AI presented itself, along with complex realities regarding its economic value, social impacts, and security risks. We are now in a challenging stage where we must move past the initial enthusiasm and navigate the duality of AI.

The choices we make now regarding regulation, safety research, and ethical implementation will define the next decade and the future of generative AI.

Maximizing the power of AI while wisely managing its risks is an essential challenge that contemporary business leaders cannot avoid. We will continue to monitor the forefront of AI and provide strategic insights, so please stay tuned to this blog.