Workplace AI Spreads Wide Yet Runs Shallow

Start with the famous wager. A few years back several AI voices predicted half of white-collar roles would vanish within two years. Manyika took the other side, betting that not even a tenth of jobs would be automated away inside a year. The clock ran out and employment rose instead. He restated that calm optimism in his Platformer interview while the forecasters pushed their dates forward again.

Roles lost, roles grown, roles reshaped

His frame is simple: some roles shrink, some expand, most get reshaped. The shrinking set is small, posts whose repetitive slices move to models. The expanding set surges in demand, with electricians as a live example. The reshaped set is the crowd: same title, new content. MIT economist David Autor backs the pattern, noting most of today's occupations did not exist in the nineteen forties. Web designer was a joke in 1995 and a profession now.

Google's ATLAS research puts numbers on the picture, scoring close to fifteen million exchanges across the Gemini app, AI Mode and developer interfaces. Reach is enormous yet depth is thin. As Axios summarized, AI touches occupations covering eighty eight percent of American employment, though the median occupation uses it in just twenty one percent of duties. Everyone samples, few go deep.

Direction matters more than volume, and the direction is assistance rather than autonomy. In thinking-heavy duties, fewer than one in ten exchanges aim to hand a whole assignment to the system. Most exchanges draft, review, brainstorm or look things up. Picture an apprentice and a master: the system carries the draft, the person signs off. Fully autonomous execution stays optimal only for a small minority of duties.

Why so slow? Because the technology is uneven. Brilliant on one prompt, stuck on the next, with every output needing a check. Visual reasoning and counting remain weak spots, and team workflows never caught up; groups run new instruments through old routines. The span of work agent setups can steer keeps growing, yet real life is messier than benchmarks. Capability sprints, adoption strolls.

One slogan needs correcting: exposure to models does not equal ruin. It sounds clever and the data rejects it. Radiologists stood closest to the technology and ended with higher output and rising demand, not unemployment. Bank tellers tell the same tale: cash dispensers arrived, teller employment grew and the role changed shape. Call it demand elasticity; easier service pulls in more customers.

Pay packets and fresh occupations

The sharper question is not headcount but paychecks. When models lift you into higher-value output, your pay should follow. When models seize the valuable slice and leave you the scraps, your pay gets squeezed. Economists call this labor's share. As he told CharterWorks years ago, Manyika fears that gap widening. Productivity climbed for four decades while wages trailed; intelligent systems could stretch the gap further.

Part of the fix sits with workers. Make your contribution visible and bargain; employers cannot track this pace alone. The shape of work is shifting too: independent specialists selling outcomes to many firms keep multiplying. Models hand them speed and they price by result. My advice is blunt: own the valuable slice, delegate the rest.

Fresh occupations are coming, though we will name them late. History always works that way: the title lags, the work leads. The ATLAS team watches for exactly this emergence of new firms and duties. Founders should test rather than predict: agent supervisors, orchestration watchers, novel evaluation services. Needs of large and small firms change daily, and the opening sits there.

Three signals and the lab crunch

Manyika names three signals that could rewrite the story. First, the growing span of work that agent setups can steer unaided; the horizon research by MIT economists and the long-assignment measurements from METR track it. Second, smoothing of the unevenness; should systems reach similar steadiness across duties, balances shift. Third, robotics: sixty percent of the world's workers do physical duties, and beyond driverless cars no leap has landed. Should robots catch up, every calculation restarts.

A scientist's day splits the same way: half general systems, half specialist ones. Joint work with MIT finds researchers reaching for Gemini-style general systems for prose and search, and for specialist systems such as AlphaFold for proteins and genes. Most use both together. AlphaFold now serves four million researchers across one hundred ninety countries, and neglected illnesses become studyable for the first time. The instrument belongs to everyone now.

Across from that sits a fresh crunch: proposing ideas turned cheap while proving them stayed dear. DeepMind's GNoME program proposed two point two million crystal structures; a few hundred thousand looked plausible for production and only a few thousand reached synthesis. A Yale partnership on cancer pathways kept merely a handful of model proposals after bench checks. Picture a sieve: plenty pours in above, little passes below. Science now moves at sieve speed.

Even so, science both quickens and reaches new ground. Surveying every single-letter genome variant at once is possible now, once unthinkable. In wiring diagrams, a first complete fruit fly brain map emerged with the Howard Hughes partnership. I see a new microscope here: not only pace but fresh sight. AI is both a general-purpose technology and an invention engine for inventions; my analogy is the Enlightenment, with telescopes and method opening unseen places.

Health today, quantum tomorrow, access for all

Health effects live in the present, not the future. A breast screening program with one hundred eighty thousand women in Britain caught signs that readers had missed, earlier. Lung and tuberculosis work delivered accurate diagnoses in Zambia and India. The AMIE helper went through trials with a Harvard-linked hospital, and eye screening passed six hundred thousand cases in India on the way to six million. Nobody thanks the model; patients remember the diagnosis. Indirect credit, concrete result.

Resilience work grows quietly too. Flood forecasts now cover one hundred fifty countries where more than two billion people live, after the India pilot held. Android sensors pushed warnings to twelve million people before the Venezuela quake. Weather models served during Hurricane Melissa. These are not headlines but phone alerts, and the alerts save lives.

Moonshots run on two fronts. One is the cell: a predictor for inner cell dynamics, with contributions to the Broad cell atlas effort. The other is quantum AI: black hole structure, electromagnetism, traversable wormholes, stubborn molecules. The Nature wormhole experiment and the Quantum Echoes advance described on the Google Blog mark that front. My excitement is plain: this happens in labs, not slide decks.

On general intelligence the answer depends on definition. Read the 1956 Dartmouth text today and it sounds like a general-intelligence charter; Shane Legg popularized the term in 2001. Under a minimal reading of human-like cognitive breadth we stand close; under a fuller scientific reading we stand far. The DeepMind Institute exists for those deeper puzzles. Recursive self-improvement quickens the pace, so capability and safety must advance side by side. The Springer survey of agent control agrees: boundaries alone fall short, oversight belongs inside the design.

The underrated hazard is connection. Testing one system in isolation misses the point once models plug into grids and infrastructure; evaluation must cover the whole. Then comes the non-use hazard raised by the global south at the UN body: places without care or schools gain nothing should models never arrive. The Brazil question bites here: taxes at home, benefits everywhere, and no participation without power, links, and language. As intelligence cheapens, inventiveness appreciates; calculators never ended arithmetic, they enlarged meaning. I read it this way: systems compute, people signify.