This Week in Putting AI to Work (May 15, 2026)

Slow progress adopting fast-moving AI, sharp thinking about fuzzy technologies, real fake AI art, and other news from another busy week.

There’s Good News

All most AI alignment research so far has a pronounced negative framing: How can we keep AI from doing all these bad things? A large team recently put out a paper that reframes the issue as one of “positive alignment”

The first two sentences of the abstract set up the topic well, and throw a bit of shade at the alignment community’s work to date:

Existing alignment research is dominated by concerns about safety and preventing harm: safeguards, controllability, and compliance. This paradigm of alignment parallels early psychology's focus on mental illness: necessary but incomplete.

So what would help complete it? The paper’s proposed approach is positive alignment, or

the development of AI systems that (i) actively support human and ecological flourishing in a pluralistic, polycentric, context-sensitive, and user-authored way while (ii) remaining safe and cooperative.

In other words, instead of playing whack-a-mole with AI alignment problems, why not take an approach to alignment that’s more holistic and optimistic? The paper includes

design principles for promoting disagreement and decentralization through contextual grounding, community customization, continual adaptation, and polycentric governance; that is, many legitimate centers of oversight rather than one institutional or moral chokepoint.

I like this positivist agenda a lot, and I’m looking forward to seeing how it unfolds. I particularly like that it is “pluralistic, polycentric, [and based on] many legitimate centers of oversight.” A.I. alignment research, to me, has always had kind of a totalitarian flavor: “we must set up a giant central bureaucracy to protect humans from A.I.” We have enough experience with both totalitarianism and bureaucracy to know that that is just not going to work. It’s going to go horribly wrong, in fact. It’s great that the nascent positive alignment movement explicitly rejects this approach and starts from the point of view that what will work here is what has demonstrably been working elsewhere: decentralized and pluralistic control, oversight, and ownership (see, for example, democracies and market-oriented economies).

And Bad News

UBS published research looking at the state of AI adoption in enterprises. They had the clever idea to compare the percentage of firms reporting that they were doing AI at scale at time T to the percentage of firms that said 12 months prior to T that they would be doing AI at scale at T. The results provide yet more evidence that human overconfidence has not yet been vanquished.

A summary of the UBS report¹ makes clear how slowly enterprises are scaling up their AI — much more slowly than they themselves expected.

As of March 2026, only 19% of companies had achieved scaled deployment, advancing just 9 percentage points in two years. A year ago, as many as 84% of surveyed companies expected to reach this benchmark within 12 months, yet only 5% actually did—a gap between expectations and reality that continues to widen.

Progress with AI technologies is stupefyingly fast; organizational success with them is clearly not. But why not? According to the report, there are

Six Major Bottlenecks: It’s Not Just a Technology Problem

Why is the enterprise AI deployment journey so painfully slow? The survey lists six key obstacles repeatedly cited by respondents:

• Unclear Return on Investment (ROI) (53%): More than half of companies still cannot clearly quantify the specific returns from their AI investments, creating pervasive uncertainty in budget approvals and resource allocation.

• Compliance and Regulatory Issues (48%): As countries tighten data security and AI regulations, compliance costs and risks have become significant deterrents for enterprises.

• Integration Complexity (45%): This metric has risen notably from 37%-38% in the two prior survey rounds. It reflects that as companies attempt to embed AI into existing legacy IT architectures, the engineering, organizational, and process restructuring difficulties far exceed expectations.

• Qualified Talent Shortage (42%): Hybrid talent that understands both AI technology and vertical industry domain knowledge is extremely scarce.

• Data Availability and Quality Issues (42%): Many companies are discovering that their data infrastructure is insufficient to support large-scale model training and inference.

• Lack of Data Privacy (37%): Striking a balance between leveraging data and protecting privacy remains a persistent and difficult challenge.

I find it interesting that the number one problem preventing faster AI scale-up, and the only one cited by a majority of respondents, was unclear ROI. I find it so interesting, in fact, that I co-founded a startup to help with exactly that problem 😏. If you’d like more clarity — if you’d like to both know and grow the ROI of your AI — get in touch with us at Workhelix.

And a Whole Lot of Uncertainty

I want to highlight three excellent recent essays on the labor market impacts of AI. All of them stress different aspects of the same point: It is early days for this powerful, protean technology. There is a huge amount of uncertainty ahead, and not a lot of obvious massive technological unemployment yet. When it comes to making policy around AI jobs and wages, then, we should proceed with humility and in an evidence-driven way.

First up is political scientist Andy Hall on the politics of AI-driven “jobless prosperity” in an extended X post and on his Substack

1. The backlash to AI isn’t here yet. There is anxiety among American voters, but there is no populist backlash yet, because the job losses haven’t started yet—and we don’t even know if they ever will. AI is not in the top 20 issues Americans say they care most about, and the AI policy issue with the most energy right now, data center opposition, reflects not just AI but also NIMBYism, as @mattyglesias has pointed out.

Hall stresses two points

First, there is way more uncertainty than I appreciated about how the economics of AGI might play out, and there is stronger evidence than I appreciated that job losses from AI have not meaningfully started yet.

And second, if AGI plays out the way the labs are predicting, the politics will be very hard to forecast, because it will be the politics of “jobless prosperity,” with jobs falling while the economy grows. We have very little experience with this happening at this kind of scale, and it will break our typical models of politics.

For both of these reasons, we should all be really humble in making pronouncements about the politics of AGI. I hope my piece will be read in this light, as an attempt to reason about something that is super important but also super hard to forecast accurately.

Economist Anton Shenk captures just how vast are the differences in recent forecasts of AI’s economic impact

In the three years since OpenAI launched ChatGPT, economists and AI researchers have published forecasts projecting that, over the next decade, AI will add to annual growth by amounts ranging from as little as 0.1% to as much as 30%. By 2035, the gap between these forecasts nears a quadrillion dollars: an amount that exceeds a decade’s worth of current global output.

Here’s a picture of that range

So what do we do when uncertainty is this high? I like Shenk’s answer here: instead of generating yet more long-range predictions, let’s instead monitor the situation closely as it unfolds and figure out which metrics actually tell us how fast things are changing

The quadrillion-dollar disagreement hinges on what metrics count most… The correct policy response is not to average the extremes, pick a camp, or wait for certainty. It is to build monitoring infrastructure that tells us which mechanisms are actually operating in the economy, then adapt in real time as evidence accumulates.

Economist Ben Jones, who has done a lot of excellent work about the sources of growth and prosperity has a new essay that makes some important and wildly counterintuitive points. Among them are:

Computers and software are everywhere, and they’re profoundly changing the business world. Yet they’re still a tiny piece of the economy: “[US] business investment in computer equipment and software now peaks at around 4% of GDP. Similarly, households devote tiny shares of their budgets to computers and software — we spend many multiples more on housing, transportation, health care, or even restaurant meals” (emphasis added)

It’s not that computers do the valuable stuff. It’s that value accrues to the stuff that computers can’t do: “Here’s the critical and perhaps surprising result: the economy is largely a story of what we do badly, not what we do well. This follows the same logic whereby amazing machines make their output cheap. The related implication is that the worse we are at something, the more expensive it will be. The expensive parts of the economy tend to be the things we need to do but haven’t figured out how to improve. And these so-called bottlenecks are everywhere. Combine harvesters have made harvesting corn the wide part of the bottle. Getting corn from the farm to a corn chip in your hand? Therein lie many tasks that constitute the narrower part of the bottle. And that’s where the payments go.”

Jones’ whole essay is mind-expanding. Give it a read.

And Motivated Reasoning

A lot of the discourse around AI these days reminds me of Jonathan Haidt’s deep insight that among us weird and wonderful human beings, judging and justifying are separate processes. “Intuitions come first, strategic reasoning second” as he puts it. We immediately, intuitively, and subconsciously know what we want to say about a topic, then dragoon our conscious reasoning powers into crafting an explanation for why that intuition is correct.

This two-step reasoning is especially likely when a topic activates our righteous mind - our moral beliefs. And AI touches on a lot of those beliefs. Many people believe that AI is cold and inhuman and can never create true art or other expressions of the human soul, whatever that is. Many people also believe that AI has a tendency to be racist because it’s created by people and power structures that have tendencies to be racist. We’ve seen examples of both kinds of motivated reasoning recently.

X account @SHL0MS put up this post on May 12

They got a lot of responses (more than 12k pages worth?!?!), many of which offered withering critiques of this pathetic attempt:

You’ve seen the punchline here coming, right? The posted image was actually a nearly complete crop of a real Monet painting. The people gleefully disparaging it thought they were engaged in art criticism, but they weren’t. They were unconsciously following to their moral intuitions and crafting post-hoc rationalizations for them (Once they realized this, many respondents deleted their original critiques. But SHL0MS kept a few gems like this one.).

As AI races ahead and brings many changes along with it, it’s also going to bring a huge amount of motivated reasoning. We’ve got to be vigilant about it, call it out when it appears, and strive to do better. Fellow Substacker Kelsey Piper of The Argument did just as she debunked the dog-whistle claim that “studies have shown that automated vehicles are less able to detect people of color.”

Here’s hoping we see more of this kind of work. The stakes are too high here to settle for motivated reason.

And Now a Word from Our Sponsor

This weekly roundup is brought to you by Workhelix, the startup I cofounded to help organizations know and grow the ROI of their AI. If that topic is top of mind for you, please get in touch.

1. A desultory search didn’t yield a copy of the report itself.