What if everything turns out alright?
Among the many disastrous consequences predicted of artificial intelligence is a jobs famine: widespread worker displacement that devastates the economy and brings about a second Great Depression.
But that hasn’t happened. In fact, the US economy is doing well and, with a 4.1% unemployment rate, operating roughly at full employment, despite other headwinds like persistent inflation. And while whether humanity is “early” in the adoption phase of artificial intelligence is debatable — I argue that, certainly, from an exposure standpoint, AI is a familiar technology at this point — the notion that AI has displaced large swaths of the workforce is not supported in the current data.
Given the buoyancy of the economy, should we reduce the odds of that happening? We can’t definitively say that it won’t, of course, but we can use the data at our disposal now — in September 2026, nearly four years after ChatGPT’s launch — to update our likelihoods. And one likelihood that I believe is underestimated, sitting somewhere between The Prosperous Society and total economic Armageddon, is that everything turns out alright.
Thanks to the sponsors of this week’s episode of the Mobile Dev Memo podcast:
- INCRMNTAL. True attribution measures incrementality, always on.
- Branch. Branch is an AI-powered MMP, connecting every paid, owned, and organic touchpoint so growth teams can see exactly where to put their dollars to bring users in the door and keep them coming back
Interested in sponsoring the Mobile Dev Memo podcast? Contact Mobile Dev Memo advertising.
The Mobile Dev Memo podcast is available on:
Transcript
The Economist published an article earlier this month titled, “The jobs apocalypse is postponed, and an AI jobs boom is here,” in which the publication estimates that AI has created roughly one million new jobs in the United States. Quoting from that piece:
But AI-related layoffs get lost in a churning jobs market where employers shed roughly 1.7 million workers in a typical month. And the evidence so far is that AI is already creating a lot of jobs to replace those it has destroyed. The vast sums pouring into data centers and power generation have set off a race for construction and infrastructure workers. AI startups are hiring like there is no tomorrow. Incumbents racing to keep up are creating new AI roles. And by making some workers more productive, AI may be increasing demand for their services. Add it all up and The Economist estimates that AI has so far created around one million new jobs in America. That easily exceeds the roughly 200,000 layoffs attributed to AI since mid-2023 and appears more than enough to offset weaker hiring in many back-office roles. America’s AI infrastructure splurge has created many of them.”
The numbers are starting to add up. Preliminary research by Gad Levanon, chief economist at the Burning Glass Institute, uses the research outfit’s career history database, which mostly relies on LinkedIn career histories, to identify jobs that would not exist without AI, whether at AI-native firms or because they are AI-specific roles at other companies. He reckons roughly 1% of professional jobs are now AI jobs, on the order of one million positions in America. In computer occupations and life sciences, including researchers using AI to discover new drugs, the share is 4 to 5%. LinkedIn’s own analysis points to roughly 640,000 new AI-specific jobs between 2023 and 2025.
Some of the occupations once believed to be most vulnerable to AI appear to be benefiting from this effect. Between 2023 and 2025, employment among paralegals rose by about 11%, and among market research analysts by 6%, compared with a national average of around 2%. Despite dire warnings of imminent layoffs, professional services are projected to keep growing rapidly thanks to demand for AI systems and consulting, according to BLS forecasts.
This certainly feels like a narrative violation. Doesn’t it?
In a speech on September 15th, Senator Bernie Sanders stated that the economic impact of that is that AI has the potential to eliminate tens of millions of jobs, wiping out entire professions and making it harder for younger people to enter into the workforce. In a press release in March announcing the Artificial Intelligence Data Center Moratorium Act, Sanders noted that this bill will stop a global race to see which country is the first to eliminate hundreds of millions of jobs or the first to build an AI that destroys the planet.
Bernie Sanders certainly isn’t alone in his pessimism. In May, Representative Greg Casar from my home state of Texas said, “we should have good AI development but we don’t need a dystopian future where millions of Americans are out of work; we face potential unemployment on par with the Great Depression.”
In December 2025, Senator Mark Warner of Virginia stated, “unemployment among recent college graduates hit 9% in November, but that will seem like a low jobless rate among young Americans in the years ahead. I fear the unemployment rate among recent college grads could reach 25% over the next three to five years if the AI issue is not addressed.”
And in a video posted to Facebook in May, Senator Elizabeth Warren of Massachusetts said, “if we’re going to build an AI future that works for everyone, then we need to tax AI and invest in people, so if millions of workers get fired because of AI, those workers don’t go bankrupt just from a visit to the doctor.”
Those are certainly frightening premonitions of an AI-blighted future, but they raise an important question: what kind of a timeline is required to evaluate these predictions?
ChatGPT was first released to the public in November 2022, so in two months, the product will be four years old. Roughly a year after it was introduced, in November 2023, OpenAI announced that ChatGPT had reached 100 million weekly active users, or WAU. By August 2024, the user base numbered 200 million WAU. By December 2024, a little over two years from launch, 300 million WAU. In October 2025, just shy of three years from launch, 800 million WAU. And on August 31st, OpenAI revealed that ChatGPT had surpassed 1 billion WAU, counting roughly 12% of humanity as users of the product.
While the LLM-powered chatbot phenomenon isn’t old by conventional consumer tech standards, its adoption is fairly widespread. If you loosen the definition of a chatbot, you might count the 2.5 billion monthly users of Google’s AI Overviews among the relevant group, which amounts to 30% of humanity. Enough people have been exposed to AI to have a sense for how it will be used. While I acknowledge that there is a capabilities component to these outcomes, they cannot be entirely binary. Do all white-collar jobs disappear once AI models reach some exact threshold of intelligence? Are there no gradations of impact? The Economist paints a different employment picture.
The broader economic data is harder to dismiss because the United States continues to generate employment while households sustain a level of spending that would be difficult to reconcile with an economy in the early stages of catastrophe or even just of decline. There are qualifications, particularly for recent college graduates, but the big picture matters. The economy into which AI is being deployed right now remains capable of supporting substantial demand for human labor. Consider the August employment report released by the Bureau of Labor Statistics earlier this month. The economy added 162,000 payroll jobs, and the unemployment rate held at 4.1%, which can be considered at or very close to full employment. The gain followed subdued hiring, with the same release reporting average monthly job growth of just 31,000 over the preceding year, although unemployment remained low as AI adoption expanded. If widespread AI adoption should already be producing meaningful economic displacement, then we’re simply not seeing it, which deserves considerably more attention than another executive or politician predicting the eventual disappearance of work.
The growth data requires a little more unpacking because the headline number understates the strength of private domestic demand. Real GDP expanded 1.5% annualized in the second quarter, following 2.1% in the first, which represents continued expansion albeit at a fairly modest pace. I wouldn’t describe that headline growth rate as spectacular, and extrapolating an economic boom from it would be a stretch. GDP includes movements in international trade and government spending that can mask what households and businesses are doing domestically. The Bureau of Economic Analysis publishes a metric that it calls Real Final Sales to Private Domestic Purchasers, which combines consumer spending with private fixed investment. That measure grew at 4.2% annualized in the second quarter, after adjusting for inflation, and it provides a considerably more optimistic picture of underlying demand. Consumers and businesses were increasing their purchases at a healthy clip even as other components weighed on the aggregate growth number. For the argument at hand, that continued willingness to spend is salient, certainly in the context of an impending AI-driven economic implosion.
It is not like headwinds don’t exist. The Iran war has disrupted energy markets, and the price of oil represents a material constraint on what households can afford elsewhere. In September, the Energy Information Administration projected that Brent crude would average around $90 per barrel in the second half of the year, raising its forecast by $8 from the previous month. That price point is about 30% above the 2025 average and 12% above the 2024 average. Persistent inflation is also a friction, although it is related to energy prices. Consumer prices were 3.4% higher in August than a year earlier, and gasoline prices increased 3.9% in the month alone. Core inflation, which excludes food and energy, was lower at 2.4% over the year, but keep in mind that the headline and core inflation numbers measure year-over-year increases, and inflation is coming off of a multi-decade high in 2022.
This nuance also applies to employment, where a promising headline number and real distress in subsectors of the economy can coexist. The information industry lost 23,000 jobs in August, while manufacturing added 16,000. Restaurants and bars contributed 59,000 jobs, and local government education added another 42,000, so the headline gain certainly shouldn’t be presented as evidence of uniformly robust private sector hiring. Recent graduates face an acutely grim job market. The New York Fed’s measure put their unemployment rate at about 5.6% in the second quarter, with underemployment at 42%. Underemployment here means working in an occupation that typically doesn’t require a college degree, which can represent a substantial torment to someone who just completed a college degree. A graduate struggling to establish a career probably doesn’t find solace in low headline unemployment.
AI certainly could be contributing to that struggle, especially in tasks that employers can automate that were previously handled by recent graduates. But quantifying the size of the impact from AI in those situations invites a counterfactual: how would hiring have evolved without AI, given everything else affecting those employers? A company might recruit fewer people for any number of reasons: financing has become more expensive or it expanded too aggressively during COVID. Attributing every drop in headcount to AI or automation is too blunt of an approach. The same discipline should apply when counting jobs created by AI, particularly when the categories overlap.
Productivity provides another useful perspective here because producing more value from each hour of work is central to the economic case for adopting AI. Nonfarm business productivity was 2.2% higher in the second quarter than a year earlier, according to the BLS. Within the quarter, productivity increased at a 1.4% annualized rate, while hours worked also increased, albeit slightly. The economy was obtaining more output per hour without an accompanying contraction in the total hours worked across that broad business sector measure. A September working paper from David Autor, a professor of economics at MIT, and several researchers at Google, provides a more concrete view of how AI can improve productivity, at least locally within the context of a profession.
The study examined patent lawyers, whose work requires specialized judgment. That makes it a useful setting for asking whether AI enables professionals to produce better work in a very specific field that requires deep domain expertise, and whether using it actually augments or amplifies their own output. The researchers conducted a three-month randomized controlled trial involving 133 practicing lawyers at 11 American intellectual property law firms, with 91 completing the full study. Roughly two-thirds received access to Inflow, which is an AI writing assistant developed by Google, and the remainder served as a control group that didn’t use generative AI writing tools at all. Randomization accounted for firm and experience, and independent patent attorneys evaluated the submissions without knowing which lawyers had access to AI. Google funded the experiment, and the participating firms had existing patent drafting relationships with the company.
Ten days after receiving access, participants completed a drafting assignment using materials describing an invention, including notes attributed to its inventor. The participants produced patent claims with accompanying descriptions, which the evaluators assessed for legal quality. Around 90 days into the experiment, the lawyers completed a second, more demanding drafting assignment relating to a different invention, allowing the researchers to assess whether the advantage associated with AI persisted after extended use. It did. AI access improved drafting quality by 0.34 standard deviations in the first assessment and 0.38 in the second. Those figures describe improvements in evaluated quality, so they shouldn’t be taken at face value as percentage increases in productivity. Notably, the gains were larger among junior lawyers, and the overall improvement primarily reflected fewer poor submissions and more middling ones. In other words, AI made the work more reliable, even without producing a larger share of exceptional drafts.
There were also modest time savings, again concentrated among junior lawyers. Participants with AI reported spending about 10 fewer minutes on the first assignment against a control group baseline of 112 minutes. The estimated savings was again 10 minutes on the second assignment, although that result was not statistically significant. Senior lawyers generally produced better drafts without working faster, which matters because an improvement in the value of an hour’s work can take the form of higher quality even when time spent remains unchanged. The final assessment introduced an important complication because everyone was asked to review and mark up an existing patent draft without AI. This assessed whether three months of AI-assisted practice had improved the judgment lawyers could exercise independently. The previously assisted group performed better overall, but the average benefit was concentrated entirely among senior lawyers, whose scores improved by 0.45 standard deviations. Junior lawyers showed no average improvement, with their results becoming more dispersed; more poor scores were offset by more good ones.
The authors suggest that existing expertise may help professionals convert AI assistance into durable learning. That’s a worthy qualification for firms deciding on how to train their staff. For the productivity argument, the study offers causal evidence that access to AI can improve the quality of professional output, with some workers also completing assignments faster. It does not measure job creation or establish AI’s contribution in national productivity growth, and I should mention that the working paper has not yet been peer-reviewed. But it illustrates a mechanism through which AI can increase the value produced by people already doing skilled work, while showing why developing their expertise remains an imperative for their firms.
The employment consequences of those productivity gains depend partly on how much additional demand they generate, and I think that question gets lost in predictions that assume the quantity of work available remains essentially fixed. If a business can deliver its existing output with fewer employees, then reducing headcount may be the rational response, and some businesses will choose that. But a lower cost of production can also make additional output commercially possible and allow the business to serve customers it previously could not reach profitably. Whether employment grows depends partly on how demand responds to that opportunity. The productivity data does not tell us how much of this improvement is derived from AI, and assigning the entire increase to chatbots would overstate that impact. But the data does describe an economic situation that is resonant with the optimistic argument: greater efficiency can accompany continued demand for labor.
The investment described in The Economist’s reporting adds another dimension to that interpretation because building the data infrastructure required to supply AI involves spending in the physical economy, with employment effects that extend beyond frontier labs and the largest advertising platforms. Of course, there are limitations to that dimension, since constructing a data center and operating one incur different staffing requirements. An investment cycle can support employment before the ultimate returns on that investment are established, and to be fair, they may never be. I am wary of projecting today’s construction activity indefinitely into the future, just as I would be wary of treating every announced automation initiative as proof that the economy has permanently lost the capacity to create jobs.
There is also a timing problem that any honest account of technological change has to acknowledge because the new opportunity elsewhere in the economy provides limited immediate relief to someone whose existing occupation is contracting. That friction deserves a thoughtful policy response, even when the headline numbers seem rosy. But what seems defensible today is a picture of an economy with considerable underlying strength. The weakness facing some white-collar workers warrants concern, and stronger AI capabilities could make that weakness more pronounced. Predictions of imminent mass unemployment need to be weighed against the reality of the economy as it exists now, including the spending that sustains employment and the additional activity that improved productivity can make possible. So far, that economy continues to expand while accommodating increasingly widespread use of AI. Four years after the introduction of ChatGPT, there is no employment catastrophe.
The absence of an employment catastrophe leaves space for a more specific and more useful account of the damage AI may be doing. A September paper by Federal Reserve Board economists titled, “Artificial Intelligence and Labor Market Reallocation,” examines that question, and its findings deserve attention even from someone who is broadly optimistic about AI. The authors combine individual records from the current population survey with job openings data from JOLTS, then connect those observations to measures of AI exposure and adoption. Exposure describes how much of an occupation’s work could potentially be performed efficiently using AI, using input from task assessments from OpenAI researchers and others. Adoption is approximated through the share of job postings within an industry that request AI or machine learning skills, using data from Lightcast.
The adoption measure is imperfect since a business can use AI without advertising for AI expertise, but the authors check its relationship with census survey evidence. Their flexible approach lets them follow the movement of workers through the labor market, including whether an unemployed person finds work and whether an employed person loses it. Those flows can reveal deterioration that would not be perceptible in an aggregate unemployment rate. The methodology is more observational in this paper than in the randomized patent lawyer experiment. The authors’ main result is that workers facing both high AI exposure and high adoption experienced more severe challenges in finding jobs after the introduction of LLMs. Their unemployment rate rose by slightly less than one percentage point between the end of 2022 and July 2026, compared with roughly 0.3 percentage points for workers with low exposure and low adoption.
The deterioration principally materialized through reduced hiring, while differences in job loss rates were much less pronounced. The paper also finds that more highly exposed employees see their work tasks change more frequently, and that the earlier positive relationships between AI and nominal wage growth or hours worked weaken after LLMs arrive. In the aggregate, the results suggest that firms are changing how they use labor while also becoming more reluctant to increase headcount. The authors estimate that AI-related reallocation has raised the natural rate of unemployment, which is the underlying rate reflecting frictional and structural unemployment, by about 0.1 to 0.2 percentage points, although that estimate depends on their modeling assumptions and comes with what they describe as “considerable uncertainty.”
That is an economically meaningful cost, and it confounds any claim that AI’s employment dividend is straightforward or clear-cut. But it also provides a sense of scale because the estimated aggregate effect remains modest relative to the more apocalyptic premonitions I cited earlier. The authors even identify recent improvement in job finding and job switching among highly exposed workers. Reading this alongside the patent lawyer experiment, it seems clear that the impact of AI isn’t uniformly good or bad or evenly applied. What’s important to ponder is how these effects develop as businesses adapt over time. Considering that requires a willingness to examine imperfect evidence before the current economic moment’s activity becomes historical record.
Alex Imas, a professor at the University of Chicago’s Booth School of Business who is currently on leave and who serves as the director of AGI Economics at Google DeepMind, recently made a profound argument on X in a similar vein:
“A few personal thoughts on reading empirical AI papers on the economy. Economists have gotten used to reading papers with super clean identification, arguing about the validity of an instrument, making sure parallel trend assumptions are satisfied. This is what gets you into a top journal, and it is very important research, no question here. But it also takes years and sometimes decades to get these types of papers right. People often don’t find a good instrument to answer a specific causal question decades after the natural experiment. We will eventually have this type of research for AI as well, and it is absolutely necessary. But we also need signals right now, even if they are noisier than what we are used to. We need papers where we can trust that researchers did their best methodologically, while at the same time acknowledging that the space is moving way too fast to wait for perfect identification. This will allow us to accumulate enough signals coming at the same question using different angles—for example, to say yes, X is likely happening in the economy. The AI exposure and early career hiring papers are a good example of this. There is no silver bullet paper with super clean identification, but at this point we have several independent teams researching the same general conclusion—enough where we can say there seems to be a slowdown in AI-exposed early career hiring.”
When Imas refers to identification in this context, he is talking about the identifiability of a causal effect, which concerns whether the evidence and assumptions allow us to distinguish AI’s contribution from other forces that might be contributing to the same outcome. If hiring is reduced in occupations that are exposed to AI, we need a credible and defensible account of how hiring would have evolved in its absence. An instrument attempts to isolate variation in adoption that affects the outcome through adoption itself, while a parallel trends assumption asks whether comparison groups would otherwise have followed similar trajectories.
I value excruciatingly thorough economic research, and those requirements protect us from mistaking a convenient narrative for a causal explanation. But I also think a technology advancing as quickly as AI deserves contemporary pulse checks because decisions about its deployment are being made continuously in real time. If economic Armageddon does arrive, an impeccable paper in Econometrica three years after the fact, read by candlelight in a cave, will be of little comfort. Imas’s argument also burdens optimists like myself with an obligation because the example he chooses is evidence of reduced early career hiring. We should take that signal just as seriously as we should examine evidence of new jobs and rising productivity, even while acknowledging identification problems.
That brings me back to the objection that we are simply too early to say anything meaningful, which seems increasingly difficult to reconcile with the scale and breadth of AI’s reach into our economy. AI is already used by an appreciable proportion of humanity. While there is room for much deeper adoption, especially where organizations must redesign processes before a better model becomes useful in practice, we can still observe effects. Those observations allow us to rerate the likelihood of various outcomes. We are far enough into deployment to update our priors, even though we cannot say for certain what the world will look like in 10 years. We never can.
The current evidence—what we see now—at least tells us that in September 2026, AI hasn’t caused the economy to grind to a halt, and in fact, in aggregate, it is quite healthy. It leaves us with reasons to monitor weak hiring closely and reasons to take the expansionary possibilities just as seriously. There is more than just a downside scenario to account for. We have seen professional output improve, and we have seen a growing commercial ecosystem develop while the broader economy continues to operate more or less at full employment. Those observations deserve to influence our future expectations alongside whatever harm is already visible in particular parts of the labor market.
There is a broad range of potential outcomes between Armageddon and Utopia. One potential long-term stable state is that everything ends up being alright: significant dislocation in certain segments of the economy, increased productivity in certain types of knowledge work, greater levels of automation that produce efficiencies for firms that may or may not be passed along, all in the shadow of a private infrastructure investment boom that evokes the ambition of the New Deal. Certainly a policy intervention may smooth rough adjustments and reduce the human cost of the transition. But humanity advances, and material conditions improve.
Comments: