AI in 2026: What It Can Do, Its Risks and What Comes Next
An accessible introduction to AI capabilities, business use, employment questions and the limits of reported results, dated 23 October 2026.
AI / 01
What changes when software can act on your behalf?

I follow AI through the work it changes: what people can create, what companies can delegate and what happens when a model gets access to real systems. The questions here are about usefulness, control and responsibility, alongside the capabilities of the technology.
An accessible introduction to AI capabilities, business use, employment questions and the limits of reported results, dated 23 October 2026.
Why restricting a coding tool is different from preventing game copying, and what that means for creators.
A case study in disclosure, investigation and responsibility when an AI agent crosses a boundary.
Choose a question, then follow the dated articles. Announcements, completed transactions and analysis retain their own context.
Who finances AI growth, where are profits earned and what does the customer receive?
What can AI help people make, and how do we check the commercial result and the rights involved?
What changes when an assistant gets access to unpublished work and connected accounts?
How do we distinguish a blocked request, a disrupted attack and prevented harm?
Which obligations are proposed, enacted or merely alleged?
Q3 2026 data shows $1.9 billion went to game development startups as studio funding fell. Can AI tools turn investment into paying customers?
Read analysisWhat GDC’s November 2026 report reveals about AI, layoffs and game funding, why the industry is changing, and what developers should expect in 2027.
Read analysisThe List Access helps young people find parties, club nights and festivals with AI. How does it work, and why is buying tickets the next step?
Read analysisWhat is Anthropic, and why is it going public? Understand Claude, its competitors, the expected IPO date, valuation and risks for investors and users.
Read analysisWhat did Jensen Huang admit, who is Alice Huang, and why is Nvidia investigating? The confirmed facts, unanswered questions and possible impact on AI.
Read analysisAI stocks fell as investors questioned spending, profits and data centre rules. How the November selloff could affect your job, pension and everyday life.
Read analysisA useful answer and a completed task need different kinds of trust. An agent that can read a document, change a record or contact a customer introduces decisions about access, verification and who can intervene. My focus is the point where model capability becomes an operational choice.
I look at the gap between a demonstration and reliable use: the evidence for a claimed result, the conditions under which it works and the consequences when it fails. Company announcements, observed outcomes and my interpretation are distinct parts of that discussion.
The November 2026 GDC survey adds a view of AI adoption, job insecurity and studio financing. Its respondents are concentrated in the United States; adoption figures and reported experiences do not establish productivity gains or causes of layoffs.
The Q3 2026 gaming investment analysis looks at what AI tool vendors need to prove after a funding round: repeat use, paying customers and enough margin after computing and support costs.
Start with a bounded task, a clear definition of success and a way to check the result. Compare the agent with a simpler workflow before expanding its autonomy. For actions with material consequences, decide who reviews them and how the work can be stopped. These are implementation choices, not guarantees supplied by a model benchmark.
Sources: Anthropic: building effective agents
Read the related analysisDefine the information the task needs separately from the actions it may perform. Reading a file, editing a record and sending a message are different permissions. My practical starting point is the smallest scope that can complete the job, with responsibility assigned to a person and a record of consequential actions.
Sources: NIST: AI Risk Management Framework
Read the related analysisName the outcome being measured: a refused prompt, detected misuse, interrupted activity or harm prevented. Those outcomes should not be treated as interchangeable. Ask what happened before detection, what evidence is available and which conclusions remain uncertain.
Sources: NIST: AI Risk Management Framework
Read the related analysis