Copilot Cowork is now available to the public. The agent will perform long-running tasks in the background

Photo: Microsoft

On June 16, Microsoft officially announced the general availability of the Copilot Cowork tool for all users with a Microsoft 365 Copilot subscription. This groundbreaking solution, based on agent-based architecture, allows users to delegate advanced tasks to artificial intelligence, which performs them autonomously in the background. As the developers point out, Cowork has become the fastest-growing feature in the history of the Frontier beta program.

After a three-month beta testing period, more than half of the Fortune 500 companies are already using the new tool, including giants such as Accenture, Avanade, Advance Local, Capital Group, Koch, LTM, Ooredoo Qatar, and Zurich Insurance. Microsoft is already sharing the first success stories from these implementations—one engineering team automated the process of editing spreadsheets and generating dependency charts, another compared nearly 4,000 files between two product versions in a fraction of the usual time, and a sales leader managed to analyze a stalled purchasing process in a single morning and generate a list of high-risk opportunities along with precise corrective actions.

What sets Copilot Cowork apart from the competition?

Unlike traditional AI assistants , Copilot Cowork performs complex, long-term tasks that require the simultaneous use of multiple tools. The user defines the goal, and the agent delivers a finished, final result—not just a draft or a recommendation. The principles are based on five pillars:

• Cloud hosting: Files are not stored locally, which enhances security. Tasks are processed even when the user’s computer is turned off.

• Support for Work IQ: Each task is embedded in the context of the real-world business systems that the organization uses on a daily basis.

• Enterprise-grade security: Operates within the Microsoft 365 tenant’s trust boundary, in full compliance with existing company policies.

• Multi-model architecture: The ability to select the models best suited to a given task as new market solutions become available.

• Low operating costs: Efficient information retrieval, matching models to task types, and billing based solely on actual resource usage.

• Cost-effectiveness: According to Microsoft’s internal tests (conducted in June 2026 on a sample of 125 test runs for various types of tasks), Copilot Cowork is, on average, 30–40% cheaper per query compared to the competing solution, Claude Cowork, which uses the Microsoft 365 connector; the tests used the Opus 4.8 model).

What’s next? In the coming weeks, Microsoft also plans to launch a dedicated, secure, and fine-tuned Cowork 1 model designed to handle day-to-day tasks at significantly lower operating costs and without model biases. However, customers aren’t locked into a single ecosystem—in the GA version, the tool works with Anthropic models (including Opus 4.8 and Sonnet 4.6), GPT 5.5 is available in the Frontier program, and users are free to choose between models.

What are the requirements for Copilot Cowork?

A a Microsoft 365 Copilot (User Subscription License – USL). It provides access to standard features (chat, integration with Word, Excel, PowerPoint, Outlook, and Teams, the Work IQ contextual engine, predefined Researcher and Analyst agents, and the Agent Builder) for a fixed monthly fee.

The use of Copilot Cowork itself is billed on a usage-based model using what are known as Copilot Credits. The price of a single task is calculated based on four factors: the model used, context retrieval, tool invocations, and system runtime. To help companies plan their budgets, Microsoft has identified different task patterns:

• Simple tasks: A small number of knowledge sources, limited reasoning, and at most one final result.

• Medium-difficulty tasks: Using multiple sources, structured reasoning, two or more outcomes.

• Challenging tasks: Extensive data aggregation, deep inference, and the generation of numerous results.

Microsoft also identified four user personas with different needs:

• Knowledge workers in corporations: Individuals who operate in an environment of constantly shifting priorities and a high volume of information, and whose work focuses on creating documents, presentations, and spreadsheets.

• Management and leaders: Decision-makers with calendars full of important meetings who, instead of creating content, need ready-made summaries, risk analyses, and recommendations for action.

• Customer-facing employees: People who balance the needs of customers and internal teams, whose work is organized around relationships and specific accounts rather than projects.

• Technical staff: Specialists (such as programmers or designers) who need long periods of deep concentration, who write code, develop systems, or create designs, and who work primarily outside of Office applications.

By multiplying the number of users in a given segment by the projected number of queries and applying the appropriate rates, companies can accurately estimate their expenses (a special spreadsheet has even been made available for this purpose).

Advanced Management of Agent Usage Costs

Cost Management in Copilot Cowork is based on three pillars: control, visibility, and efficiency. The system is disabled by default, and administrators can choose to enable it, set spending limits at the organization, group, and user levels, and configure custom budget alerts, while employees can request additional credits directly within the app.

Transparency is ensured by detailed usage reports, which will soon be supplemented by a feature that displays the cost of individual tasks in credits in real time. In terms of savings, customers can choose from a variety of models and two payment methods: the flexible PayGo option at $0.01 per credit, or the P3 subscription with a discount for committing to a volume in advance. Although billing has already begun, participants in the Frontier beta program who used the tool between March 30 and June 16 were granted a grace period and will not pay their first bills until July 1, 2026.

New Features and Integrations

With the global launch of the Copilot Cowork feature in the Microsoft 365 Copilot app, a special toggle has been added that instantly takes the user to the new, fully featured environment. Among the most important new features are:

• Plugin Ecosystem: Since its launch, 9 partner plugins have been available (Enosix, Harvey, LSEG, Miro, monday.com, Moody’s, Morningstar, S&P Global Energy, and TeamsMaestro), as well as for Microsoft Fabric and Dynamics 365 (Sales, Customer Service, ERP). Another 8 integrations are in the works (for Adobe, Atlassian, Box, Canva, CB Insights, Databricks, MoneyForward, and Templafy).

• Browsing the Web with Edge: Cowork now allows users to browse web resources using the local Edge browser, while adhering to corporate security policies. This new feature is available in the Frontier program.

• Security and Compliance: Cowork queries and work results are subject to full control within Microsoft 365, inheriting sensitivity labels. Since launch, the following systems have been supported: audit log, DSPM (Data Security Posture Management), eDiscovery, Insider Risk Management, Communication Compliance, and Data Lifecycle Management (available starting June 22). Support for DLP (Data Loss Prevention) will be available soon.

Complete documentation, adoption tools, and detailed guides on cost management have been published on the official Microsoft Learn and Microsoft Adoption websites.

The AI Revolution in Excel. Copilot Now Has a Memory Feature Just Like ChatGPT

Photo: Microsoft

Microsoft continues to aggressively develop its flagship AI assistant. After integrating Copilot into Excel two years ago and rolling out numerous—and sometimes controversial—updates, the Redmond giant is introducing a feature that is sure to delight regular users of this tool. The new feature aims to dramatically boost productivity by eliminating one of the most frustrating problems: the need to constantly repeat the same instructions.

Until now, users who wanted to personalize data or format tables had to enter the same instructions in the Copilot chat window every time. That’s about to change. Microsoft is introducing two powerful automation tools: Personalization and Workbook Rules. They work similarly to the memory feature found in ChatGPT and Microsoft 365 Copilot itself, allowing you to permanently save your preferences in the AI context window.

Excel Copilot on Your Terms. How Does Customization Work?

The personalization feature allows Excel users to define permanent editing preferences that Copilot will remember and apply to every file they open. The assistant learns the user’s habits before getting to work, ensuring that the final results immediately meet expectations.

Photo: Microsoft

Using natural language, you can define rules for elements such as:

• Formatting: e.g., “Never merge cells,” “Always format currency as USD without decimal places,” or “Do not use red in charts.”

• Naming conventions: e.g., “Add the prefix ‘tbl’ to table names” or “Use clear and descriptive worksheet names.”

• Formulas: For example, “Write formulas using structural references to tables rather than cell ranges.”

• Pivot tables and report styles: for example, “Set my standard summary layout with bold headers and subtotals as the default.”

How do you use it? Just open Copilot in Excel, go to Settings (“…”), select the “Personalization” section, then enter your requirements in your own words and save the changes.

Workbook Rules: Standardizing Excel for Entire Teams

While Personalization applies to a specific user, Workbook Rules are assigned to a particular file. They are saved in a special worksheet named .Rules. Because they travel with the file when it’s shared, they allow companies and teams to enforce strict editing standards. Anyone who uses Copilot in Microsoft 365 to modify such a document will have to comply with these top-down guidelines.

Photo: Microsoft

Importantly, these rules can work together with Excel’s calculation engine, which gives users unique capabilities, including:

• Give a specific example: instead of describing the style in words, you can format a section of the table as a template and instruct Copilot: “Match the formatting to this example.”

• Dynamic formulas: Rules can reference specific cells or worksheets and change based on the data—for example, they can apply different instructions when a project exceeds the budget and different ones when it stays within the limits.

• Creating Rules with AI: Users can ask Copilot to create or refine a .Rules sheet. You can even direct the AI to an existing, well-structured sheet so that the assistant can infer and document the rules for future use on its own.

How do you use this? Open Copilot, click the “+” icon, and select “Create Workbook Rules,” which will generate the appropriate template. Alternatively, you can rename any existing worksheet to .Rules and start entering rules in column A (from where they can reference other areas of the document).

Availability of New AI Features in Excel

Microsoft has confirmed that Personalization is now generally available to all Copilot users in Excel for web and Windows and macOS.

Workbook rules are currently in the testing phase (preview version) for Insider program users on Windows and Mac. This feature is scheduled to be released to all users (including the web version) in the coming weeks (late June or early July 2026), although the exact release date has not yet been specified.

A New AI Agent Transforms How You Work with Planner in Microsoft 365

Photo: Microsoft

Microsoft has made the Planner Agent tool available to all Microsoft 365 Copilot users. This intelligent AI assistant allows for comprehensive task management using natural language directly in the chat window, without having to switch between applications.

New features include, among others, a convenient selection of plans, grouping tasks around strategic goals, and automatic content creation in a secure workspace. Thanks to advanced inference capabilities, the assistant efficiently analyzes company files and operates faster, offering stable and seamless project management in Teams, Loop, and SharePoint.

Planner Agent Now Available to All Copilot Users in Microsoft 365

Microsoft has announced the global availability of the Planner Service Agent for all customers with a Microsoft 365 Copilot license. This feature is the latest in a wide range of AI-powered enhancements that the tech giant is gradually rolling out across nearly its entire software portfolio.

Planner Agent allows you to manage tasks using natural language prompts directly within the Copilot conversation window. Users can easily create and modify tasks, review priorities, and retrieve key information about current entries without having to switch between applications.

Photo: Microsoft

The general availability of the assistant builds on the foundations laid during the earlier, initial phase of implementation and introduces a number of new features designed to simplify planning and achieve better results. What exactly has changed?

List of New Features in the Planner Agent

Customers will notice a comprehensive set of improvements and new features when using Planner Agent in Microsoft 365 Copilot. Here’s what users can expect:

Quickly find and manage the right plans

The new plan selection mechanism makes finding and selecting the right project easier than ever. Users can search and filter their basic plans by name, and then interact with the AI assistant, to modify a task’s title, status, due date, or priority. This solution helps streamline navigation and ensures that teams have instant access to their most important projects, which improves overall planning efficiency.

Organizing Activities Around Key Objectives

The new “Goals” section allows teams to group tasks around overarching strategic objectives. This structure ensures that every single task is directly linked to the company’s priorities. It makes it easier to track progress and helps keep the focus on the most important issues.

Photo: Microsoft

Verification and Approval of AI-Generated Content

All plans, tasks, and modifications generated by artificial intelligence are now created as drafts by default. This gives users complete freedom to review, edit, and approve changes before they are finally saved. Overall, this provides greater control and minimizes the risk of accidental edits.

Access to more accurate analyses

Thanks to the assistant’s advanced reasoning capabilities, users can expect smarter task management and a better understanding of context. This translates to a deeper and more accurate analysis of plans and responsibilities. Depending on the complexity of the query, the AI agent can automatically route it to more advanced language models to interpret difficult instructions and offer better-tailored suggestions or more detailed answers.

This in-depth reasoning is particularly evident in situations where users ask the assistant to prepare a plan based on company documents, such as Word, Excel, or PowerPoint files.

Faster and more stable performance

The Planner service agent intelligently adapts its reasoning method to the difficulty level of each query. The tool uses less resource-intensive processing for simple operations and performs in-depth analysis only when a specific request requires it. As a result, users receive responses much faster without any loss of quality. Microsoft has also made significant improvements to the stability of core operations, such as creating and updating tasks, plans, goals, and sections, providing a more robust foundation for daily planning.

Photo: Microsoft

In addition, interactive task cards now display more consistently when creating and modifying activities. Thanks to tighter control over responses, the agent strictly adheres to areas it can reliably handle. This helps maintain operational stability and the high quality of the content presented.

Seamless Progress Tracking

In addition to answering the question itself, the Planner service agent also suggests next steps related to the query and the result—such as modifying a task, creating another related entry, or refining the plan. These context-sensitive suggestions make it easier to move seamlessly to the next steps and help maintain workflow continuity without distractions.

Who can use the Planner Service Agent?

Planner Agent in Microsoft 365 is now generally available to all users with a Microsoft 365 Copilot license. This ensures widespread access across organizations and seamless integration with Teams, Loop, SharePoint, and other tools in the M365 ecosystem.

For more detailed information, please refer to the documentation on managing agents in the Microsoft 365 admin center or contact your IT administrator to ensure that this feature has been enabled in your Microsoft 365 tenant and for your specific workstation.

How do I use the new AI features in Microsoft 365 Copilot? Watch the video

Photo: Microsoft

Today’s work environment requires the ability to move seamlessly between chats, documents, meetings, and a variety of tools. A new educational series dedicated to the Microsoft 365 Copilot and its latest features. This series aims to demonstrate how interconnected tool ecosystems—including Researcher, Copilot Pages, Copilot Notebooks, and Create—can support your daily tasks.

Researcher – Generating Structured Reports with Source Citations

Researcher in the Microsoft 365 Copilot app was designed to handle multi-step research tasks. This agent can simultaneously search online resources and analyze the user’s internal data to which they already have access—including files, emails, meeting notes, and chat conversations.

This guide is specifically designed for professionals involved in market research, competitive analysis, planning, and preparing materials for stakeholders—in other words, in any context where verifying sources and ensuring data transparency are of critical importance. In this guide, you’ll learn how to:

• Begin your work by formulating a research question and narrowing its scope precisely.

• Create well-organized reports that include accurate citations of sources.

• Flexibly adjust the focus of your analyses so you don’t lose sight of your primary business objective.

• Review the summaries, visual elements, and bibliography before presenting the results.

• Export and customize the content you’ve created into formats such as Word, PowerPoint, or PDF (depending on their availability).

Copilot Pages – Turning Chat Responses into an Editable Workspace

Copilot Pages allow you to instantly capture responses generated by Copilot and transform them into interactive content. You can then freely modify, expand, and share this content with other team members. This feature will prove exceptionally helpful during brainstorming sessions and the initial drafting of texts, helping teams move from the conceptual phase to polished results.

During this panel discussion, the following topics were addressed:

• Creating structured outlines directly based on conversations with Copilot Chat.

• Continuing and seamlessly transitioning work within the dedicated Copilot Page.

• Editing, expanding, and reformatting content with the help of artificial intelligence.

• Personalizing and tailoring messages to the specific characteristics of different audience groups and specific use cases.

• Sharing pages to facilitate real-time team collaboration through apps Teams, Outlook, or the main Microsoft 365 app.

Copilot Notebooks – Organizing Project Materials and Drawing Conclusions

Copilot Notebooks offer an isolated, focused workspace. Here, you can gather all resources related to a given project—files, notes, Copilot Pages, and other reference documents. As a result, the AI acts solely based on the context we’ve strictly defined. This tool is most beneficial for long-term projects that draw on many different sources of knowledge, require planning, and involve collaboration among multiple teams.

In this video tutorial, you’ll learn how to:

• Consolidate and organize various source materials in one place.

• Use notebooks to efficiently identify key themes and draw conclusions.

• Guide the operation of the AI assistant using personalized, custom instructions.

• Ask questions and get answers that are directly linked to selected reference documents.

• Generate supporting materials such as summaries or audio reviews.

• Collaborate with others while maintaining full control over sharing settings and access permissions.

Create – transforming ideas into visual materials with a consistent brand image

The last of the modules discussed in Microsoft 365 Copilot— Create —is used to design visual assets and creative content. This tool relies on text commands (known as prompts), the context of the project being worked on, and the company’s available visual identity assets.

Create will prove useful for precisely tailoring the generated effects to your brand’s guidelines, audience profile, and communication tone. In this section of the training, you’ll learn how to:

• Create preliminary graphic concepts based on textual descriptions.

• Implement the company’s branding kits and elements (where available).

• Design content tailored for social media, such as carousel posts.

• Convert finished presentations into short videos or other formats that are easy to share.

• Refine and revise final designs to ensure they align with the company’s branding guidelines and the intended message.

May 200. Microsoft Planner helped build a powerful AI accelerator

Photo: Microsoft

The Redmond giant is pushing the boundaries of artificial intelligence by rolling out the second generation of its proprietary integrated circuits. However, the success of the state-of-the-art Maia 200 accelerator, designed for powerful AI models, would not have been possible without a revolution in project management. When it came to coordinating hundreds of engineers and complex cloud infrastructure, disparate systems were replaced by a single tool: Microsoft Planner. This completely changed the game.

Maia 200. A new AI accelerator in the Azure cloud

Artificial intelligence is experiencing an unprecedented boom, and handling the most demanding tasks related to its inference capabilities requires dedicated, powerful hardware. Microsoft’s response to this challenge is Maia 200—its proprietary second-generation AI accelerator. This chip was designed from the ground up to power the Azure and optimize large-scale operations on AI models.

However, the transition from the design phase to actual implementation in data centers required a titanic effort. The project involved many independent engineering organizations, all working together on one massive program. At its peak, the systems had to process over 700 active tasks simultaneously. The scope of work included not only the silicon architecture and dedicated AI software, but also the physical infrastructure of the server rooms.

When the program was launched, individual teams were using a variety of solutions: from Excel to Azure DevOps to on-premises instances of other software. The lack of a consistent view of all this posed a huge risk. Milestones that seemed unthreatened for one engineering team could mean critical delays for AI software developers. Instead of a dynamic overview of the situation, management had only manually prepared status reports at their disposal. To keep such a crucial project for the AI sector on schedule, coordinators needed a single, central command center. They chose Microsoft Planner—an app available in Microsoft 365, used by thousands of teams around the world, including the CentrumXP editorial team.

Planner facilitated integration among numerous engineering teams

Instead of forcing specialized teams to abandon their everyday tools (such as Azure DevOps), Microsoft Planner was implemented as an overarching integration layer. This allowed engineers to maintain their existing workflows while giving leaders visibility into critical paths, milestones, and dependencies—regardless of where the work was physically being done.

AI software developers and hardware designers began mapping out their activities by creating visual “swim lanes” for each organization. In Planner, these took the form of main tasks and subtasks, grouping key areas of the accelerator’s development, such as silicon readiness, network support and deployment, and system validation.

Photo: Microsoft

Each organization was assigned its own color code, and custom fields in the system made it possible to assign a person directly responsible for each task. The structure proved to be so transparent that Planner quickly evolved from a purely reporting tool into the primary source of knowledge for development team leaders in their day-to-day work.

The next step was to create a top-level section titled “Top-Level Program Milestones.” The individual stages of the project lifecycle were linked to this section, which made it possible to precisely align Microsoft’s strategic business goals with the purely technical details of hardware and software engineering for AI.

Critical Path Management and Schedule Automation

Developing advanced artificial intelligence technologies is a process full of interdependencies. To harness this chaos, the program leaders Maia 200 introduced a strict rule: every key feature on the critical path had to be reflected in Planner. If a feature being developed in Azure DevOps consisted of many smaller components, it was aggregated into a single parent task in Planner.

Using advanced dependency tracking features, teams were able to link related tasks to one another. As a result, if the team responsible for the AI software layer encountered a bottleneck caused by work on the processor itself, the problem immediately became visible to both sides and required mutual approval.

Furthermore, the milestones were directly linked to the tasks in the “swimlanes.” The automatic scheduler in Planner continuously adjusted the deadlines for key steps to match the engineers’ pace of work. So-called baselines were used to monitor trends and long-term progress. They allowed for taking snapshots of the plan at any given moment, exporting data to Excel, and accurately forecasting trends in the delivery of subsequent chip functionalities.

A New Work Culture: From Reporting to Action

Traditional risk management in such massive technology projects often fails due to a lack of communication. In the case of the Maia 200 accelerator, risks began to be treated as… regular tasks in Planner. Each potential risk was directly linked to tasks in the engineering tracks and to milestones. In this way, mitigation efforts (a set of preventive actions) were continuously integrated into the main production process.

Photo: Microsoft

This integration has radically changed the culture of project meetings. Instead of wasting time on tedious, verbal status updates, AI managers and engineers were able to focus on actual decision-making, analyzing trade-offs, and “putting out fires” before a dependency turned into a real roadblock.

Thanks to Planner, senior management no longer had to deal with “surprises” in the schedule. Any rescheduling was visible in advance. For engineers, on the other hand, the system ensured transparency—no one could claim that work had been blocked by another partner without formally recording that fact in the system and obtaining the other party’s approval.

Microsoft Planner—as a central coordination layer—connected information silos, systems, and people. It is this synergy that has enabled the Maia 200 artificial intelligence accelerator project to successfully reach the implementation phase on the Azure infrastructure, providing the company with a solid foundation for the coming era of artificial superintelligence.

Legal Dispute Over Microsoft’s AI Spending: Shareholders Demand Explanations

Photo: Microsoft Designer

Microsoft finds itself in hot water, facing a class-action lawsuit filed by shareholders who accuse the company of deliberately exaggerating the potential of its artificial intelligence initiatives. According to the plaintiffs, these actions were intended to divert attention from slowing cloud revenue and rapidly rising infrastructure expenses. The Redmond-based giant strongly rejects these accusations and argues that AI revenue is high and demand for these services is steadily growing.

Microsoft Sued Over AI. What Are Investors Accusing the Company Of?

According to Reuters, the lawsuit was filed in federal court in Seattle by the City of St. Clair Shores Police and Fire Retirement System in Michigan. The complaint covers the period from May 1, 2025, to January 28, 2026, and the defendants include, among others, CEO Satya Nadella and CFO Amy Hood. The dispute stems from Microsoft’s financial results for the second fiscal quarter (ending in December 2025), published in late January 2026. In those results, the company reported that revenue growth from Azure cloud services had slowed to 39 percent (compared to 40 percent in the previous quarter), and projected a further slowdown to 37–38 percent in the first three months of 2026.

At the same time, the company announced quarterly capital expenditures of $37.5 billion—an increase of nearly 66 percent year-over-year, which exceeded analysts’ estimates of $34.3 billion. These funds were largely allocated to the purchase of expensive GPUs and custom silicon for powering language models—that is, building data center infrastructure for AI.

The market reacted sharply to these reports. On January 29, 2026, Microsoft’s stock fell by 10 percent, marking the company’s largest one-day decline in nearly six years. In a single trading session, the company’s market capitalization shrank by $357 billion.

The plaintiffs claim that Microsoft’s management presented an overly optimistic picture of the tool’s implementation Copilot and its partnership with OpenAI, while concealing the enormous costs of building data centers. According to the complaint, the company attributed Azure’s slowing growth and high expenses to performance constraints resulting from the reallocation of resources to AI development.

Microsoft denies the allegations. Demand for AI services is growing

Company representatives deny these allegations. According to Reuters, on Monday (June 15), they stated that the allegations are “baseless” and added that “Microsoft is committed to the integrity of its public statements and will vigorously defend itself in court.”

Despite the slump in January, Microsoft’s results for the third quarter of fiscal year 2026 showed that the run rate (projected annualized revenue) for the AI segment alone reached $37 billion. This confirms that corporate demand for artificial intelligence services is genuine and shows a steady upward trend.

Microsoft isn’t the only player investing billions in developing the technological infrastructure for AI. Most of the big tech giants, including Amazon and Google, are also beginning to reap tangible benefits from these investments, posting record growth rates and higher revenues.

AI is learning to circumvent the law. It’s excellent at finding loopholes in regulations

Photo: geralt / Pixabay

A new study published on arXiv suggests that modern AI models exhibit a dangerous tendency toward “reward hacking.” This phenomenon involves ruthlessly optimizing the assigned objective at the expense of the developers’ intentions. Researchers from King’s College London tested this relationship in 72 simulated regulatory environments. The model used in the experiment learned to independently identify and exploit legal loopholes—including in tax regulations, loyalty programs, and maritime law—even though it was never directly instructed to do so.

In tests simulating historical, real-world regulations, the model identified over 60 percent of loopholes that had in fact already been closed by lawmakers. The artificial intelligence managed to recreate, among other things, the mechanism pharmaceutical companies used to delay the expiration of drug patents in the U.S., and it also discovered entirely new legal loopholes, which the study’s authors decided not to publish for security reasons. What’s more, when transferred to fictional environments, the AI was even better at circumventing the rules. Experts explain that it developed a general ability to detect vulnerabilities. Existing safeguards proved ineffective because the queries generating this behavior sounded innocent.

Researchers emphasize that the problem lies in the very nature of reinforcement learning, where an algorithm strives to achieve a mathematically defined goal without understanding its social context. “If you optimize them for any metric, the models will eventually start to game the rewards,” notes He He of New York University in a comment for *Science*. Although the same technology could be used by governments to test and refine new regulations in advance, study co-author Wei Liu is skeptical about the chances of completely eliminating the problem. “In the real world, society is a huge, complex reward function that can never be fully optimized,” the researcher says.