The Narrative Has Turned

The dominant public conversation about artificial intelligence has changed substantially in the past six weeks. Six months ago, the mainstream framing across major media was existential risk, imminent displacement, and the AGI 2027 timeline. Today, the New York Times opinion page runs pieces calling LLMs "plausibility engines" that cannot reason. Nobel laureates publish papers quantifying the modest macroeconomic effects. The largest AI investor in the world tells his employees that the agentic acceleration his spending was premised on has not materialized. And the first serious empirical study using the right dataset shows that firms adopting AI hire more workers, not fewer, with entry-level headcount growing fastest of all.

The narrative has turned. The turn is worth documenting because it happened faster than most narrative shifts on major policy questions, and because the evidence driving the shift is now sufficient to draw preliminary conclusions about what AI has actually been doing in the American economy over the past two years.

What the Evidence Shows

Ara Kharazian at Ramp Economics Lab published a working paper on June 30 using the joined Ramp corporate spending and Revelio Labs workforce dataset that Stanford recently identified as the ideal data source for the AI jobs question. The finding is direct. Firms that adopt AI intensively grow headcount 10.2 percent over the two years following adoption, and entry-level headcount grows 12 percent. The entry-level workforce share at these firms increases relative to the control group. The finding requires the standard caveats about selection effects that all such studies require, but the direction of the effect is inconsistent with the displacement hypothesis at any reasonable magnitude of selection. AI-adopting firms are hiring more entry-level workers, not fewer.

The Kharazian result is empirical confirmation of what the macro data has been showing for eighteen months. U-6 unemployment released this week at 7.9 percent for June, down from 8.1 the month before, and near the lowest level in modern American history. The Atlanta Fed nowcasts 1.2 percent real GDP growth. Real personal income remains at all-time highs. Real wages continue to beat inflation. Every specific quantitative prediction the displacement scenario made has been falsified by the data at this point.

The capability side of the story matches the labor side. Mark Zuckerberg told Meta employees on July 3 that the company's AI agent progress is slower than expected. Executives were "super optimistic" about coding tools when they planned the January reorganization that shed 8,000 jobs. The bets "haven't come to fruition yet." Meta is spending 145 billion dollars this year on AI infrastructure. The CEO is telling his employees the capability the spending was justified by has not arrived on the schedule that justified it.

Zuckerberg is not alone. The pattern of walking back internal capability projections while extending public timelines has been visible across the major labs for at least two quarters. The gap between the loud public confidence and the quieter internal recalibration is closing, and it is closing in the direction of the recalibration.

The Discourse Response

The mainstream media coverage has adjusted to the evidence. Cal Newport in the New York Times on June 17 called out "doom trolling" as a corporate communication strategy that has become morally indefensible. The Economist on June 25 published a briefing on AI model values that used charts showing frontier models clustering in specific ideological positions consistent with the demographic composition of their training data. A New York Times op-ed today frames LLMs as plausibility engines rather than reasoning machines, illustrated with concrete deployment failures at Meta, Air Canada, and McDonald's.

Six months ago, these framings would have been fringe skeptical positions confined to substacks and technical blogs. Today they are the mainstream editorial framing at the paper of record and the leading business weekly. That is a substantial shift in the Overton window on the AI question, and it happened in a period measured in weeks rather than months or years.

The academic literature has also arrived. Daron Acemoglu's peer-reviewed paper in Economic Policy, published earlier this year and cited increasingly through 2026, estimates that AI will produce a total factor productivity gain of no more than 0.71 percent over ten years. This is roughly one-tenth of the projections that Goldman Sachs and McKinsey were promoting to investors. The most recent Nobel laureate in economics, working from a fully specified general equilibrium model, has quantified the productivity effect as modest. The number is now available for anyone making serious investment or policy decisions to use.

What Has Not Changed

AI is useful. The models keep improving. Companies are integrating them productively into workflows. The Kharazian data specifically shows that intensive adopters gain productivity that translates into growth rather than headcount reduction. The technology is real and it is producing real value.

The narrative shift is not that AI has failed. It is that AI has integrated into the economy the way most successful technologies have integrated into economies. Modest productivity gains, distributed unevenly across sectors and firms, absorbed through hiring adjustments rather than through mass displacement, with the specific occupations facing pressure being identifiable in advance and much narrower than the discourse projected. This is what the technology has actually been doing while the discourse was projecting something else.

The concerns that survive the narrative shift are also worth taking seriously. Deployment reliability failures like the Meta account takeovers matter. The plausibility engine framing captures something real about the reasoning limits of current systems. The concentration of AI capability in a small number of companies raises the regulatory capture concerns that critics from David Sacks to the New York Times editorial board have identified. The distributional effects on specific worker categories are real even if the aggregate effects are modest. These are ordinary policy questions of the kind human institutions have addressed for every previous transformative technology, and they are the questions that deserve attention now that the catastrophic framing is receding.

What the Shift Predicts

Narrative shifts on major policy questions often overshoot in both directions. The catastrophic phase overshot. The correction phase may also overshoot toward dismissiveness. The clean version of what the evidence supports is neither the AGI 2027 displacement scenario nor the AI-is-nothing dismissal. It is that AI is a genuinely useful general-purpose technology being integrated into the economy at a manageable pace, producing modest aggregate productivity gains that are distributed unevenly, with specific occupational pressures that policy can identify and address without disrupting the broader employment picture.

The macro data will continue to reflect the boring integration story. The capability improvements will continue at a rate that is impressive in absolute terms and disappointing relative to the projections that justified the current investment cycle. The mark-to-market moment for the largest concentrated technology investment in American history is coming, on the timeline of capital markets rather than the timeline of press releases. And the discourse will continue to adjust to the evidence as more of it accumulates.

The plateau is here. The evidence is now coming from multiple independent directions, from academic researchers, from empirical studies, from CEOs at the largest investing companies, and from the editorial pages of the major newspapers. The eighteen-month look-back on the loudest displacement predictions is coming clean, and it is coming clean in the direction the data has been pointing since it started arriving. What happens next depends less on the technology and more on how quickly capital allocation, policy response, and worker preparation adjust to the picture that is now visible.

This is the update that the evidence supports. It is smaller than the doom scenario and larger than the dismissal. It is the shape of the economy actually integrating a useful new technology, and it is what the data has been telling us all along.

About the Author

Sean Richey, Ph.D., is a Professor of Political Science at Georgia State University specializing in AI information environments and digital political communication.

Expert Witness & Consulting Services

Dr. Richey provides expert witness testimony, case review and analysis for counsel, survey methodology evaluation, and policy consulting on AI-associated information environments. Visit my website or email consulting@seanrichey.com.