At the International Monetary Fund-World Bank meetings in Bangkok last month, Lesetja Kganyago, governor of the South African Reserve Bank, said that r* is covered by clouds so thick that we know it is higher but not how high. The image captures how central bankers think about the natural rate of interest. There is a number up there, and better data and models will eventually reveal it.
I have argued that r* is less a fact of nature than a convention with real consequences, capable of becoming self-validating once embedded in policy decisions. In a model by Bank for International Settlements economists, central banks and markets confuse the effects of their own actions with useful information. Bundesbank President Joachim Nagel warned that central banks looking for r* in market prices might be looking in a mirror. AI now puts this view to a hard test.
On 16 November, the Federal Reserve’s productivity and jobs task force will hold a field hearing, where Chair Kevin Warsh wants to learn whether a productivity miracle is coming, or will come slowly. The hearing should also ask what the Fed measures when it estimates r* in the middle of an AI boom.
Two readings of the AI boom
Central bankers increasingly read the boom as a demand shock. Warsh wrote in November 2025 that AI would be a significant disinflationary force, yet in Bangkok he said that it is hitting the demand side before the supply side. Mārtiņš Kazāks, governor of the Bank of Latvia spoke of a global fight for money, and the Bank of Japan has cited AI-related demand among the reasons for its rate rises. Investment comes before the productivity it promises, and r* is expected to rise with it.
Markets tell a less clear story. PIMCO looked at 43 days of major AI model releases since January 2023. The 5-year, 5-year-forward real rate fell by about 100 basis points cumulatively on those days, although it rose by about the same amount over the whole period. Forward rates, however, also carry expected policy rates and risk and liquidity premia, and PIMCO attributes around half of the fall in long-term yields to term premia. This evidence does not isolate movements in r*, but it challenges the presumption that AI must lead markets to price in higher real rates. Shinichi Uchida, Bank of Japan deputy governor, concluded that there is no clear answer yet.
My own account in 2025 focused on AI as a source of supply shocks, first inflationary and later disinflationary. It did not address the investment boom that central bankers now see first, and it took the final disinflationary phase for granted.
What depends on beliefs
Two claims must be kept apart. The first is modest: beliefs affect market rates and estimates of r*, and standard theory can accommodate this. The second is stronger: under certain conditions, shared beliefs may help determine the equilibrium rate itself. Different conventional beliefs about the same technological opportunity may induce different investment decisions and financing commitments, and thus different paths of capital accumulation and potentially different equilibrium real rates. This claim still has to be demonstrated.
The mechanisms also differ. A belief is self-fulfilling when it changes behaviour so that the expected outcome occurs. AI has not got there yet. Firms build datacentres because they expect demand for AI, and lenders finance them because they expect revenues to follow. The spending raises activity and the demand for capital, and this may reinforce the optimism before any return materialises. A belief is self-confirming when agents see nothing that contradicts it along the observed path, while its predictions for paths not taken remain untested. This may fit AI better. John Maynard Keynes called the underlying practice a convention: we assume that the existing state of affairs will continue, except in so far as we have specific reasons to expect a change.
The obvious objection is that expectations must ultimately be disciplined by fundamentals. Yet fundamentals do not stand apart from the path the economy takes. Investment decisions shape capital and debt, and thus alter the conditions against which expectations are later judged. The telecom boom illustrates this, although it does not prove multiple natural rates. Capital spending by publicly traded US telecom service companies rose from $47bn in 1995 to $121bn in 2000, on ebullient expectations of demand, and dropped by 25% in 2001.
What remains to be shown
The stronger claim has not been proved. A formal account would have to cover optimism as well as pessimism. It would need a financial sector to trace AI beliefs through debt, equity and interest rates. And it would have to be tested on data.
Conventional beliefs can also be tested, and AI is being tested. Warsh reports that since the Fed’s September rate rise investors have become more discriminating and some credit spreads have widened. As for my essay for the London School of Economics, it showed that r* estimates are heavily revised, which documents uncertainty and does not by itself prove self-validation. The stronger claim therefore remains a hypothesis.
Central banks do not observe the prevailing belief from outside. A central bank that revises its estimate of r* upwards may reinforce private beliefs about prospective returns. Its policy response can work the opposite way, if it raises the real policy rate relative to the prevailing natural rate, which no one observes. The net effect depends on how information and monetary transmission interact.
None of this makes r* arbitrary. But central banks cannot fully stand outside beliefs that their own estimates, communications and policy choices may help sustain. This is a problem of identification even before any theory of multiple equilibria is established.
What central banks should do
First, central banks should complement their point estimates of r* with scenario-contingent assessments that state their assumptions on AI, including a case in which its returns disappoint. This would show how much of an estimate rests on the prevailing belief. Publishing scenarios would itself influence that belief, and central banks should weigh this too.
Second, policy should be tested against the failure of the belief. G20 officials in Bangkok warned that a slight reassessment of AI earnings prospects could cause profound market ructions. A rate path that looks right if AI delivers should also be judged by what it would do if the convention broke, with the debt already issued and the capacity already built. This test belongs as much to financial stability committees as to monetary policy committees.
The Fed’s task force should add one question to its list: what would the Fed’s estimate of r* be if the market’s view of AI changed?
Kganyago is right that the sky is cloudy. If part of what lies behind the clouds is a shared belief, estimates of r* could fall quickly when that belief breaks, and the economy’s path would change with them. A central bank that mistakes a convention for a star may find it has gone out while it is still steering by it.
Biagio Bossone is an adviser to international financial institutions and national central banks.
Interested in this topic? Subscribe to OMFIF’s newsletter for more.