Bottom line: AI can cut the cost of making work. It is raising the cost of trusting work. That puts more pressure on the human who checks the answer, owns the choice, and keeps the team steady.
This week made the shift hard to miss. OpenAI released GPT-6 Astra, its most capable model yet. It can do longer, harder work, and OpenAI says it crossed a critical cyber capability line. One day later, Gimlet Labs announced a $300 million round to make AI inference faster and cheaper. Better thinking machines are coming. The money behind them is coming too.
But another number tells the deeper story. Plaud surveyed 800 U.S. knowledge workers who use AI. Most said AI helped their productivity. Yet the study headline said the gain did not remove mental load. People are making more, but they are not always feeling lighter.
of AI-using knowledge workers reported productivity gains in Plaud's study published September 4, 2026. Faster output did not mean less mental load.
The new hidden cost is proof
We used to read a good memo and think, “This person understands the problem.” Now a great memo may prove only that the person can open a tool. The work looks smarter. The worker may be smarter too. We just cannot tell from the output alone.
That creates a new tax. Every fast answer needs a human to test the facts, read the risk, explain the choice, and stand behind it. The more work AI makes, the more chances there are to miss a quiet error. Speed without judgment moves a bad choice faster.
OpenAI's own safety note makes this plain. Astra was safer in many tests, but it was also harder to watch in some hostile tests. More power came with a harder proof problem. Business now has the same problem at a smaller scale: who really knows, who only sounds right, and who will speak up before the error gets costly?
More output can mean less connection
Microsoft researchers found the same split in real work. Heavy AI users had 21.2% more actions in tools tied to production, but just 7.1% more communication actions over 20 weeks. That can save time. It can also pull people toward solo document work and away from the talks where trust, context, and new ideas grow.
This is the psychology under the profit line. A person can finish more and still feel tired. A team can ship more and still understand less. A company can buy speed, then spend the gain on rework, weak handoffs, fear, and bad calls.
Here is the trap most scorecards miss. AI can lift output today while cutting the skills that protect output tomorrow. If people stop asking, checking, and talking, the firm gets faster on paper and more fragile in real life.
The New York Fed found that AI use reached 61% of service firms and 51% of manufacturers in its region. Yet the median share of workers using AI inside those firms was only 17% and 7%. The tool is spreading faster than the human habits needed to share it well. The gap is not access. The gap is the operating system around access.
Return on Individual is the real score
AI made information free. So information is not the moat. Actual Intelligence is the moat that is left: character to own the result, critical thinking to test it, EQ to read the room, grit to stay with the hard part, and purpose to aim the speed at something worth doing.
This is Josh's trillion-dollar EQ point. The largest value will not come from one more answer. It will come from people who turn cheap answers into trusted action. That is why ROI now means Return on Individual.
Measure the return in plain terms. Did AI give the person back time, money, and energy? Did that person use the gain to make a better call, serve a customer, coach a teammate, or go home with peace? Did the business gain profits without draining passion, partnership, and purpose?
If not, the tool did not create a full return. It only created more output.
Start with the human system. Take the free Actual Intelligence diagnostic, then test the shared system with the free Culture Test at actualintelligenceos.com.