Homer: Kids, there's three ways to do things. The right way, the wrong way and the Max Power way!

Bart: Isn't that the wrong way?

Homer: Yeah, but faster!

– "Homer to the Max," The Simpsons, season 10, episode 13

It's easy to be cynical about artificial intelligence (AI) these days.

If you use a newer computer, you've likely been pestered with Microsoft's insertion of Copilot buttons and AI features into the latest version of Windows. To put it mildly, this has not been popular — users have been vociferously objecting to the nonconsensual addition of the often buggy AI on their systems for months. This pushback has been strong enough to compel Microsoft to be "more intentional" about how it integrates the company's increasingly unpopular AI software into future versions of the world's most popular desktop operating system.

If you follow the news, meanwhile, you might remember when Elon Musk's Department of Government Efficiency (DOGE) team used AI last year to gain access to sensitive federal systems, surveil workers and cancel any programs that Musk and his allies thought smelled too much like diversity, equity and inclusion, or DEI. Videos of a deposition of a few former members of DOGE who were involved in those efforts were recently posted online and backed up to the Internet Archive. In the videos, they explained how they used ChatGPT to "reduce wasteful spending and noncritical spend" by cutting programs that were "not for the benefit of humankind," such as a documentary about violence against women during the Holocaust.

Given these and hundreds of other examples, it's no surprise that people don't trust AI.

Consequently, when it was recently announced that the Nevada Department of Employment, Training and Rehabilitation (DETR) is rolling out an artificial intelligence tool to process appeals on unemployment benefit decisions, it also was no surprise that people were skeptical of the project. Lawmakers including state Sens. Dina Neal (D-North Las Vegas) and Skip Daly (D-Sparks) expressed concerns about the project's security and transparency. City Cast Las Vegas host Sonja Cho Swanson and her guests — The Indy Opinion Editor Andrew Kiraly and Elle Hope, poet and founder of Spotlight Poetry — shared their mixed feelings as well.

As for The Indy's Indy Voices section, the current score is one column against with no columns in favor — and no, this column will not tie things up.

To be clear, I'm not opposed to bringing additional efficiency to Nevada's unemployment claims processes, nor am I opposed to using AI to do so. 

Nevada's unemployment system was stress-tested during the pandemic and, as in many states, it did not exactly pass with flying colors. Many Nevadans who didn't receive unemployment benefits received erroneous overpayment notices instead. High-level executives within DETR were repeatedly replaced, often under protest. Vendors refused to renew contracts with the agency. The pandemic proved that the unemployment system needed to be modernized.

Having said that, modernizing the agency's software can't improve its ability to serve Nevadans unless the policies and regulations that unemployment agencies in this country operate under are modernized as well.

To understand why, I strongly recommend reading Jennifer Pahlka's Recoding America, which starts with the founder of the U.S. Digital Service's (USDS) efforts to modernize California's similarly dysfunctional unemployment system during the pandemic. In short, the USDS team realized that even though California's unemployment system had several technological deficiencies, the issues plaguing the system were more fundamental than its reliance on some old mainframe software. 

The training manual used by new employees was more than 800 pages long — and didn't include the various procedural workarounds that experienced staff routinely used to push claims through. Staff needed at least 17 years to master the policies and processes that govern unemployment insurance in the state. Those policies and processes, meanwhile, frequently changed day by day.

The complexity of California's unemployment system is not unique. During the pandemic, New Jersey's labor commissioner brought a cardboard box to a legislative hearing that contained all of the unemployment regulations issued by the federal government — it held 7,119 pages, including more than 1,500 pages that were added during the pandemic. 

That intractable body of regulations applies just as much to Nevada as it does to New Jersey and California — and each state, including Nevada, adds its own. 

Nevada, for example, adds more than 100 pages of statutes and administrative codes. Additionally, since the Legislature only updates its online law library every two years, changes to either — such as due to the passage of amending legislation or adopted regulations — must be identified and implemented as well.

As Pahlka explains, all of this administrivia inevitably produces a cascade of rigidity. Faced with thousands of pages of requirements, many of which conflict with or supersede each other in unclear ways, government agencies are forced to optimize for process fidelity — in other words, their ability to implement and follow the rules, no matter how arcane and contradictory — over outcomes.

This doesn't just apply to unemployment systems. It applies to defense contractors, local zoning codes and everything in between.

When you have 7,119 pages of federal rules to implement, plus however many rules are applied at the state level, it doesn't matter how modern any agency's software is. Replace the creaky mainframe with a distributed cloud-hosted system, replace a pile of terminal macros written in long-lost languages with artificial intelligence, spend tens of millions of dollars replacing anything you want — so long as the rules remain in place and remain every bit as rigid and inflexible as they were when programmers first tried to encode them in COBOL decades ago, constituents will experience the same frustrating results.

Could artificial intelligence make it possible to achieve those broken and contradictory results sooner? Undoubtedly so. Will that improve outcomes for anyone? Probably not. Service delivery failure will just be achieved the Max Power way — wrong but faster.

That's why, paraphrasing from a recent Niskanen Center report on using AI in government, AI's value to government can only be realized by redesigning processes and reducing structural bottlenecks, not from adopting the tool itself. That's true for all software, whether it's "artificially intelligent" or not. That, however, requires agencies to have the flexibility to redesign their processes — and it requires policymakers, especially at the federal level, to reduce the thousands of pages of structural bottlenecks they created into something intelligible for people and AI alike.

David Colborne ran for public office twice. He is now an IT manager, the father of two sons and a recurring opinion columnist for The Nevada Independent. You can follow him on Mastodon @[email protected], on Bluesky @davidcolborne.bsky.social, on Threads @davidcolbornenvor email him at [email protected]. You can also message him on Signal at dcolborne.64.