Future Blog | The Future of Artificial Intelligence

In this post, Juho-Pekka Virtanen, a technical specialist from Forum Virium Helsinki’s data team, examines the current state of AI and the many ways it is already becoming a part of everyday work.

AI has become a buzzword promised to revolutionize almost any industry. Currently, the “AI boom” is led by generative AI—specifically Large Language Models (LLMs)—which are most visible to consumers through web services like ChatGPT, Copilot, or Google Gemini. However, AI also continues to make an impact through traditional machine learning and image recognition.

Gaining Understanding Through Project Experiments

With technology evolving so rapidly, I believe it is crucial to reach practical experiments and tests as quickly as possible. This allows us to better understand potential benefits and evaluate the technology’s suitability. AI has been utilized in numerous ways across Forum Virium Helsinki’s projects.

Maintenance of Green Areas

In a pilot for the PilotGreen project, we tested the interpretation of satellite imagery for park areas using AI. At three sites, the experiment focused on lawn growth, identifying invasive species, and targeting fertilization based on data from the previous summer.

A key goal was to assess maintenance needs remotely to reduce unnecessary inspection visits and mowing cycles. Maintenance staff reported that the service helped them anticipate and plan work more efficiently. By reducing redundant field visits, the city saved resources and achieved both economic and environmental benefits.

Forecasting Traffic Flow

The Helsinki pilot of the ACUMEN project utilized AI to combine multi-source traffic data, detect anomalies, and forecast traffic flow. The goal was to improve the commuting experience for residents while supporting sustainable and smooth modes of transport.

Meanwhile, the SEDIMARK project integrated machine learning components into data marketplace streams. This allowed AI to identify anomalies and errors in the data and assess its quality before the datasets even reached the marketplace.

Satellite Data and Geospatial Info into Natural Language

The ten suppliers in the SPACE4Cities project each use AI in different ways to process satellite data. They employ a wide range of AI types: machine learning, deep learning, agentic AI, multimodal foundation models, generative AI, and LLMs. Many of the models used are European in origin, such as the Mistral language model and the TerraMind multimodal model funded by the European Space Agency.

The diversity of AI solutions in SPACE4Cities stems from the need to apply the right tool to the right task. For example, while LLMs are not suitable for processing raw satellite data directly, they provide a practical way for end-users to understand the results. Users can interact with the application using text-based prompts or receive text-based suggestions based on processed raw data. This approach is particularly beneficial for users who are not accustomed to handling raw geospatial data but want to apply the resulting analysis to their work.

“This approach is particularly beneficial for users who are not accustomed to handling raw geospatial data but want to apply the resulting analysis to their work.”

AI in Coding

In software development, AI-assisted coding has already become commonplace. Afterall, a significant portion of programming involves writing fairly conventional data processing routines. AI makes this work considerably more efficient by writing simple functions based on the programmer’s instructions. Humans still define the software architecture and the methods of data transfer and processing, but the machine speeds up the manual writing.

AI also supports programmers in version control, testing, and documentation. Furthermore, AI-assisted coding is highly effective for building concept-level prototypes, visualizations, and user interfaces.

However, fully automated application development is also discussed. Here, speed and visual polish accomplished with AI can be deceptive. AI creates a strong first impression by building a visually appealing interface and the “happy path” of the code. Closer inspection often reveals flaws: As AI lacks a deep understanding of usability and accessibility, the app may break as soon as a user acts unexpectedly. Additionally, AI often misses the critical “under-the-hood” work—security, GDPR compliance, and maintainable architecture. It remains the human’s job to turn a fragile prototype into a sustainable product.

“AI lacks a deep understanding of usability and accessibility, the app may break as soon as a user acts unexpectedly. Additionally, AI often misses the critical “under-the-hood” work—security, GDPR compliance, and maintainable architecture. It remains the human’s job to turn a fragile prototype into a sustainable product.”

Analyzing Customer Feedback

Another major application for LLMs is structuring information. Countless processes involve free-form text, such as customer feedback. Analyzing large volumes of text is laborious, and such data often fits poorly into data-driven management tools that favor multiple-choice options and numbers. Yet, this data is incredibly valuable for capturing the views of city residents.

In the DataliiKe project, we experimented with using an LLM to analyze tourist advisory questions. A dataset of over 1,000 rows of free-form text was successfully categorized using an LLM into both pre-defined statistical categories and emerging themes. This allows humans to verify the results and explore the data using standard business intelligence tools like PowerBI.

Since language models can also be utilized via APIs, an AI component of this kind can easily be integrated into existing processes. In practice, for example, the language model can classify data immediately after it is entered, with the result stored in the database alongside other information. AI thus becomes an invisible part of applications and data processing—rather than just a distracting chatbot in the corner of a browser.

“AI thus becomes an invisible part of applications and data processing—rather than just a distracting chatbot in the corner of a browser.”

AI at the Edge

Equally “invisible” to the user are applications where AI classifies video or image data. For example, traffic counting can be done flexibly when a trained machine learning model monitors the video stream.

By using Edge AI, the analysis happens directly inside the camera. This protects privacy by ensuring that no video needs to be sent to the cloud and identifiable images of people or vehicles aren’t stored. In a sense, the AI enabled camera turns into a traffic counter submitting vehicle counts, types, and directions at a defined time span.

Strengthening Cybersecurity

AI makes security smarter and more proactive. In cloud services, AI can “read” the content of shared files. If a file contains sensitive info, it is automatically classified to ensure only authorized personnel can access it. It also learns normal user behavior, allowing it to immediately flag suspicious sharing attempts.

In protecting network traffic and emails, AI acts like a high-speed detective. It continuously analyses network activity to identify the smallest signs of an attack, including zero-day threats that have never been seen before. In emails, it uses natural language understanding to catch phishing attempts and malicious attachments, quarantining them before the user can even see them.

“In protecting network traffic and emails, AI acts like a high-speed detective. It continuously analyses network activity to identify the smallest signs of an attack, including zero-day threats that have never been seen before.”

Summary

AI is rapidly becoming a part of our daily lives, but perhaps in a different way than the loudest marketing hype suggests. At its best, AI provides us with better data, the ability to process difficult qualitative materials, and increased efficiency in integration processes.

Effective use of AI requires an understanding of tool quality, data integrity, and evolving work roles. Without this, tools become “bolted on,” and significant benefits may be missed. The recently launched AIRA project aims to develop new operating models for generative AI in collaboration with companies and cities, focusing on enhancing productivity and human sustainability in organizations engaged in knowledge intensive work.

Previous Future blog entries:

• What will the urban nature of the future look like?
• What does the future of food production look like?
• What will the future of Urban Air Mobility be like?
• What will the future of Helsinki’s digital twins be like?
• What will the future of water transport look like in Helsinki?
• What does the future hold for electric cars?
• What Will Urban Logistics Be Like in the Future?