Five questions that will shape AI over the next decade

Predictions about AI date quickly. Questions last longer. These are five we think will shape the next decade, and that anyone building with AI should be thinking about now.

1. How will we know when a system is good enough?

As AI takes on more consequential work, “good enough” needs a definition that holds up under scrutiny. That means standards written for specific tasks, and evidence that a system meets them, not general impressions of capability.

2. How much autonomy should agents have?

Agents that act on our behalf are arriving in many kinds of work. The open question is how to grant autonomy gradually, based on evidence, and how to take it back when that evidence weakens.

3. Where will expert knowledge come from?

Much of the knowledge that makes work reliable was never written down publicly. It lives in practice: how a technician handles a fragile sample, how a lawyer reads an unusual clause. Capturing that knowledge carefully, and fairly, will matter more each year.

The next leap in AI may depend less on bigger models and more on better judgment about what they should do.

4. How do we keep people skilled while AI does more?

If AI handles the routine work, people may lose the practice that built their expertise in the first place. Training, deliberate practice and roles that keep people close to the hard cases will be essential.

5. Who is accountable when AI gets it wrong?

Clear records, clear standards and clear roles make accountability possible. Without them, mistakes become no one's fault and nothing improves. Building accountability into how AI is used, not only into the models, is one of the most important design choices ahead.

What to do now

• Write down, for each AI feature you run, what good enough means and how you would prove it.
• Grant agents autonomy one task at a time, and keep the evidence that justified it.
• Invest in the experts whose knowledge your systems depend on, and keep them close to the hard cases.
• Keep a record of decisions, so that when something goes wrong you can see why and fix it.

Our view

Each of these questions comes back to the same thing: trustworthy AI needs trustworthy human judgment behind it. Making that judgment measurable and scalable is the work SoReliable exists to do.