InvestMacro

By William Akoto, American University School of International Service 

Hackers tried to break into at least 30 municipal water systems in Minnesota on July 26-27, 2026. Since then, Michigan and five other states have reported similar cyberattacks.

The attackers did not try to infiltrate the computers that utility offices use. Instead, they tried to seize control of small computers in equipment like pumps and valves that deliver drinking water to millions of people.

The utilities countered the attacks by shutting down the control computers and sending personnel out into the field to operate equipment manually. Utility officials have said that water remained safe to drink.

As a scholar who researches cyber conflict, I find that the methods used in these incidents are typical of international cyberattacks. Initial suspicion has fallen on hackers allegedly aligned with Iran, but the U.S. government has yet to attribute the attack to anyone.

How can someone from far away seize control of a water system and possibly shut off the flow or taint the water?

Controlling the water machinery

There are about 152,000 public drinking water systems in the United States, according to the federal government. A municipality gets its water from lakes, reservoirs, rivers or underground aquifers.

Pumps move water through pipes to a treatment plant that filters and disinfects it. More pumps push the treated water into storage tanks, then through distribution pipes to homes and businesses. The entire system can span many square miles.

The hackers accessed small computers called programmable logic controllers at the water systems that operate all sorts of industrial equipment. The programmable logic controllers read sensors that measure conditions such as water pressure, water chemistry, tank levels and equipment status, and automatically operate pumps, valves and alarms. A household thermostat is a useful comparison: It reads the temperature and tells the heating or cooling system what to do.

The programmable logic controllers also transmit operational data to a utility’s central computer system. Workers use dashboards to monitor the information and send commands back to the controllers. The two-way communications can travel through wired networks, over radio or cellular links, or through internet connections.

Many utilities operate with small staffs, so remote connections allow an employee to monitor a distant pump or tank, receive an alarm after hours or let a vendor diagnose equipment without traveling to every site.

Controllers that use the internet may access it directly, or go through protective firewalls, secure gateways or virtual private networks. Direct access is more vulnerable because there are fewer defensive barriers. A hacker can find a controller by scanning the internet and finding its Internet Protocol, or IP, address, then try a weak or stolen password or exploit a known security flaw.

To reach a controller through a secure gateway or encrypted service, a hacker would have to steal remote-access credentials, or break into the gateway or private network, or get control of an operator’s workstation. The hacker could then use that foothold to reach the controller.

Attempted access can also be part of an intruder’s longer-term strategy to collect information, test defenses or establish entry for a later date.

How an attack works

Attacks on industrial control systems often follow a familiar sequence. Infiltration often begins with a quiet search for access. Attackers scan internet addresses for controllers, dashboards and outside companies that provide remote access services, looking for targets that are linked directly to the internet.

Next, the attacker looks for a default or stolen password to log in, an unpatched vulnerability or a misconfigured remote-access service. Sophisticated malware is not always necessary: In 2023, U.S. officials reported that Iranian-linked hackers targeted internet-connected Unitronics programmable logic controllers used by water utilities. Some utilities were still using the manufacturer’s default password, according to the Cybersecurity and Infrastructure Security Agency.

Finally, the attacker exploits the access they have gained. This could mean changing a password, issuing commands or attempting to alter the controller’s software. Researchers at the National Institute of Standards and Technology note that an intruder could replace legitimate control instructions with malicious commands. An attacker could also sneak into an office computer through phishing, then access the controller network.

Industrial equipment in service for decades is extremely vulnerable because it may not support modern security features, and utilities may delay updates because they want to avoid interrupting operations.

Reports thus far indicate that hackers accessed the Minnesota water systems through controllers that communicate over the internet directly. A July 30 FBI and Environmental Protection Agency advisory stated that attackers remotely accessed Rockwell Automation MicroLogix programmable logic controllers that were connected directly to the internet, and changed their IP addresses and passwords.

Defensive moves that utilities can take

The most immediate step that utilities can take to protect themselves is to remove controllers and human dashboards from direct connection to the internet. Following the Minnesota attacks, the Cybersecurity and Infrastructure Security Agency urged water utilities to place this equipment behind properly configured firewalls and other safeguards.

When remote access is necessary, utilities should route communications through a secure gateway or VPN, require multiple levels of authentication, and limit how much access each user has. Utilities should change default passwords, disable unused remote-access services and install vendor-approved updates to connected equipment.

In their guidance on internet-exposed dashboards, the cybersecurity agency also recommends separating operational networks from email and other business systems. This measure makes it harder for attackers to move between the two systems.

Finally, utilities should back up controller programs, log remote-access activity and practice restoring systems and operating manually.

A matter of resources

Rural water utilities with limited resources are a significant vulnerability in the United States’ critical infrastructure.

A group of volunteer cybersecurity experts is providing guidance to water utilities, but their reach is limited. Smaller utilities may need government funding or shared cybersecurity services to be able to defend themselves.

About the Author: 

William Akoto, Assistant Professor of Global Security, American University School of International Service

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Patrick J. Schena, Tufts University 

Creating a government fund to own AI stock and benefit all Americans would require many hard choices.

Should the U.S. government require artificial intelligence companies to transfer half of their stock to a sovereign wealth fund – a government-run fund that invests surplus state revenues for long-term savings and economic stability?

About 7 in 10 Americans who were asked this question in a June 2026 survey answered that it should. The survey was conducted around the time that Sen. Bernie Sanders of Vermont introduced related legislation.

His measure is intended to create government oversight over disruptive AI threats, while allowing all Americans to benefit from the value the technology creates.

And, on July 13, about 200 economists and computer scientists, including 16 Nobel Prize winners, also raised their voices about the disruptions posed by AI, warning that “AI may become radically more powerful over the next 10 years.” What AI does to the economy, they continued, “could bring risks, including large-scale job displacement, as well as opportunities such as major gains in living standards.” The letter’s signatories called for AI use that “complements humans and benefits society.”

Signs that Americans are becoming increasingly wary of AI are multiplying. Concerns not only revolve around what AI might do to the job market, but also around the broader economy.

Sovereign wealth funds

I’ve spent nearly 20 years studying sovereign wealth funds. They are typically set up as government entities, staffed by professional investors and tasked with concrete investment goals. Specific laws define how they receive the cash they invest. Similar rules define when cash can be withdrawn and why.

Although sovereign wealth funds serve a wide range of purposes, their most basic function, broadly defined, is to hold and invest government savings to support the current and future needs of citizens.

If the U.S. were to create a sovereign wealth fund to harness AI and buffer the economy and workers from long-term harms that AI could cause, it would hardly be the first country to do so.

Canada, the U.K., South Korea, Saudi Arabia and several other countries have already begun to introduce AI-focused sovereign wealth funds into their long-term government planning. While the idea is certainly appealing and simple, creating a sovereign wealth fund for this purpose is definitely not.

Alaska’s oil revenue surpluses

In the U.S., many states have had sovereign wealth-like funds for many years. These have helped states to manage surplus revenues or invest in specific projects. New Mexico and Wyoming are two examples.

Alaska, perhaps the best known, is a third. Its US$91 billion Permanent Fund has accumulated the state’s oil revenue surpluses since 1976. Today, the Permanent Fund is completely integrated into its state revenue system – meaning that it helps fund Alaska’s budget. The state government uses it to buffer state finances when oil prices drop below levels that allow Alaska to generate surplus earnings.

Since 1982, the fund has made payments – through what it calls dividends – to Alaska residents age 1 year old and older. These individual payments vary year to year, peaking at more than $3,200 per recipient in 2022. It’s set at $1,200 for 2026.

Taking stakes in companies

Although President Donald Trump first floated the concept of a U.S. sovereign wealth fund in February 2025, his administration has made little progress advancing that idea.

The administration has instead been an active investor in several strategic sectors, including defense, energy, semiconductors and critical minerals. The 30 such deals it has struck since January 2025 total $27 billion.

They include the federal government’s investments in U.S. Steel and Intel. These investments were completed independently by various federal agencies. They include buying stock in private companies, which means that the U.S. government benefits if the companies pay dividends to shareholders. The government also benefits if the shares go up in value and the stock is sold for a profit.

In June, Vice President JD Vance indicated that the White House would support extending this strategy to include the government owning stock in AI companies.

Today, the income generated from U.S. government stock holdings is returned to the government, but without any specific use defined upfront.

If income from AI investments were to be treated in the same way, I see no rules or programs currently in place that would specifically direct those funds to offset negative impacts from AI, including those related to employment and income levels.

Could a sovereign wealth fund play that role?

Challenges to investing in AI for the public good

Before establishing any kind of fund intended to offset damage to the economy or to workers that AI may cause, the government would have to answer several tough questions.

1. Is it investing in companies or getting their stock through taxes or some other arrangement?
2. Who decides which companies it will invest in?
3. Where would the money come from to finance these investments?
4. How would the government earn income from these investments?
5. Investing in any new technology is risky. AI companies are high-growth and cash-hungry, with uncertain prospects. What happens if AI companies are not competitive or profitable and their stock prices go down?

Even if the government began to own shares in AI companies, that would not necessarily mean those companies would be paying for any disruption their products may be causing to the U.S. labor market and economy. That would require policymakers to agree on how these investments should be made, how risks to these returns ought to be managed, and how any income that the government may earn from these investments could be used.

Establishing a sovereign wealth fund to capture wealth created by AI to benefit all Americans could be a start. Making it deliver on the expectations that would accompany its creation would require consensus, discipline and strong governance to do effectively. All of which, I am afraid, are in short supply in today’s Washington.

About the Author:

Patrick J. Schena, Professor of Practice and International Business, Tufts University

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Pawan Jain, University of Michigan Flint 

Finding good financial advice can be stressful – and expensive. That’s one reason why chatbots have become an increasingly popular and free alternative.

But using artificial intelligence to answer your pressing money questions also carries hidden dangers. I’m a finance professor who has been closely watching the spread of AI into personal finance, and I recently warned that AI is riskiest when it sounds most confident. I advised readers to bring in a human professional for high-stakes financial decisions.

One response came back that caught my attention: What if you can’t afford an adviser?

For many households, that’s the reality. A traditional adviser can cost hundreds of dollars an hour, which makes little sense when your savings are modest. For those people, AI isn’t a second opinion but the only financial adviser they’ll ever have. So the useful question isn’t whether they should use it; it’s how to get something actually helpful out of it without being misled.

My answer: For people who can’t afford ongoing advice, AI is genuinely useful for budgeting, paying down debt and low-cost investing.

The skill lies in using AI well. Here are some simple guidelines to get accurate and actionable answers when you engage with a chatbot:

AI is good for everyday money questions

The gap between AI and a human adviser is narrowest for commonsense guidance. Build a small emergency fund. Pay down high interest debt. Contribute enough to your workplace retirement plan to capture the full employer match, which is free money. Invest in low-cost, diversified index funds. Don’t panic-sell when markets fall.

None of this is controversial. It’s consensus advice, and it’s precisely where AI is reliable and its confidence is earned.

Here are a few examples of specific questions that a chatbot handles well:

• My daughter is 6, and I would like to have around US$60,000 saved by the time she starts college. How much do I need to set aside each month to get there, and what rate of return does that assume?
• I’m carrying credit card debt at 22% interest, a car loan at 7%, a student loan at 5%, and I’ve allotted $300 a month to put toward them. Which takes priority?
• My employer matches 50 cents on the dollar up to 6% of my paycheck for my retirement fund, and I earn $58,000 annually. How much do I need to contribute to capture the full match, and what’s that worth for a year?
• My retirement plan has a fund charging 0.85% a year and an index option charging 0.05%. If I put away $40,000 over 25 years, what does that gap cost me?
• I’m the only earner in my household, my income varies month to month, and my expenses run about $3,200 per month. How large should my emergency fund be, and where should I invest it?
• My retirement plan offers a 2055 target date fund and an S&P 500 index fund. What’s the difference, and what factors should I consider while choosing one of the two?

Notice what these questions have in common. Each is a general question with a well-established answer, and every fact specific to you is one you supply rather than one the tool has to guess. Give it the facts and ask it to do the reasoning, and its responses are genuinely strong.

For these questions that are 90% of your financial life, AI is a useful free guide, and for someone who had no professional guidance at all, that’s a real gain. The danger lives in the other 10%, and the key is telling the two apart. A typical chatbot will answer anything you type – including questions it should not – in fluent prose that can be confidently wrong.

So slow down whenever a decision is large, irreversible, tax-related or being sold to you. Examples include negotiating a windfall or inheritance; making early withdrawals from or conversions between retirement accounts; claiming Social Security; taking on major debt; and weighing any pitch for a product like an annuity or whole life insurance.

Inherited retirement accounts alone carry withdrawal and tax rules that a chatbot can state with total confidence and still get wrong for your specific case, at a cost of thousands of dollars.

Financial advisers are pricey, and chatbots can be a good free substitute if you have basic questions.
Badhan Ganesh on Unsplash, CC BY

Five habits that make AI safer

Once the list of questions is set, here are some precautions to take once you engage with a chatbot.

Make it ask you questions first. Open with, “Before you advise me, ask me the questions a good financial planner would ask.” Generic answers come from under-specified questions, and you learn which details will actually produce a more useful outcome.

Ask it to argue against itself. After any recommendation, reply: “Give me the strongest case against this, and the situations where it would be wrong for me.” If it can’t engage in response, that’s a sign the bot is entering a more dangerous mode. This one precaution does more than any other to signal for you to be careful.

Make it show its assumptions. If the bot projects that your savings will grow to an impressive number, ask what assumption it made and what would change it. You’ll learn that it assumes steady returns every year, no missed contributions and no fees. That means the projection is just information, not a promise.

Verify the facts. Contribution limits, tax brackets and deadlines all change, and this is exactly where AI can be subtly out of date. Check the IRS or the Social Security Administration directly. If one number drives your decision, don’t take it on a chatbot’s word.

Never share identifying details. Don’t offer account information, Social Security numbers or logins. Describe your situation in general terms. Good advice doesn’t require handing over data that can be used against you.

Robo-advisers are a low-cost middle ground

If your main need is simply to invest, robo-advisers sit between a free chatbot and an expensive human. These automated platforms build and rebalance a diversified portfolio of low-cost funds based on your goals and risk tolerance, typically for an annual fee of 0.25% to 0.50% of assets under management, often with little or no minimum balance.

In a 2022 study of a large robo-advisory platform, my co-authors and I found that typical users skew young and male, and are often small investors – exactly the people that traditional advice has priced out. We also found that new account sign-ups rise during periods of high market volatility. People reach for these tools when markets turn turbulent, which is when automated discipline helps most and a panicked move costs most.

The catch: Robo-advisers run on preset models and the inputs you give them. They won’t handle a divorce, an inheritance or a complicated tax year. If your life changes and you don’t update your inputs, the strategy won’t adjust on its own.

What a human adds that AI does not

Let’s say you can afford a professional. What advantage does human advice offer over a bot?

It’s not just about the material results. In a new study, I found that the value of human advice shows up as lower financial anxiety rather than higher returns. It wasn’t about simply having an adviser. Rather, the comprehensiveness of the planning put clients at ease: detailed attention across estate, investment, retirement, risk management and tax questions, as well as advice about the psychology of money.

That benefit, I found, was concentrated almost entirely among households with lower financial literacy, where the effect was roughly six times larger than among the most financially sophisticated, for whom it was negligible.

Two further findings are practical. My research found that clients of Certified Financial Board-certified advisers reported more comprehensive service across all six planning areas, so the credential is a reasonable proxy for what actually helps. And AI use showed no sign of substituting for human advisers, and it did not predict lower financial anxiety.

Reassurance is the one function current AI does not supply.

A professional doesn’t have to mean expensive

Hiring a human for help doesn’t require splurging on an adviser on retainer. You can save here with fee-only planners, who charge a flat hourly rate to answer one focused question for a few hundred dollars, through professional and certification organizations such as the CFP Board, the Garrett Planning Network, the XY Planning Network, or the National Association of Personal Financial Advisors.

In addition, many workplace retirement plans and credit unions offer free coaching. For handling debt, nonprofit credit counseling is available through the National Foundation of Credit Counseling. And if your annual income is $69,000 or less, the IRS’ free Volunteer Income Tax Assistance program provides tax help at no cost.

So the choice isn’t between AI or a pricey adviser. It’s AI for the everyday questions, a robo-adviser for hands-off investing if you want one, and a single affordable hour of human time at the red-flag moments, ideally with someone who does comprehensive planning rather than a single transaction. That last point isn’t just a preference – it’s what the evidence says actually reduces the worry.

About the Author:

Pawan Jain, Associate Professor of Finance, University of Michigan Flint

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Jeffrey Young, Georgia Institute of Technology 

You’ve probably heard artificial intelligence models described as “open” or “closed.” These are not descriptions of the model’s personality. Large language model AIs like the one under the hood of ChatGPT don’t have actual personalities, despite appearances.

The labels refer to whether all of the information about how an AI model works is publicly available and the model can be modified, or whether the model’s developer keeps its inner workings secret and the model itself private property.

Open-source software

The concept of open-source software originated in the free software movement of the 1980s and ’90s. The movement’s founders believed that software creators and users had the right to “four freedoms” – to run the program, to study and modify it, to distribute copies of the original, and to distribute copies of subsequently modified versions. The fundamental requirement was that the source code – the basic instructions – for a program should be made available.

In the late 1990s, software developers associated with projects such as the Netscape web browser and the Linux operating system coined and promoted the term “open source” to refer to these ideals.

As part of the evolving movement, certain organizations developed open-source licenses that specified how a particular piece of source code could be used and distributed, including the Gnu General Public License, Apache License, MIT License and the Berkeley Software Distribution. Each type of license also specified any potential restrictions on how software patents applied to the source code.

Open source or open weight?

The open-source idea has risen to prominence again in the past several years as artificial intelligence large language models have surged, notably OpenAI’s ChatGPT, released in 2022. Developers first train new models on large datasets, then deploy the models for use by other people.

Meta was one of the first large companies to release an open-source large language model, called LLaMa. The company released LLaMa on Feb. 24, 2023, and made available the “inference” source code – the instructions that run the model. And it released the so-called weights, the encoded knowledge the model learned during training. However, open-source organizations such as the Open Source Initiative have stated that the LLaMa licensing guidelines prohibit commercial reuse, which the initiative maintains is not truly open source.

Other companies have released “open weight” models, such as DeepSeek from DeepSeek AI and Qwen from Alibaba. The models have less restrictive terms for reuse, and the AI community has adopted them rapidly. Still, many developers believe that a true open-source AI model must not only include the source code and weights but also the data that is used to train the model.

A lot to open up

The Open Source Initiative’s definition of a fully open-source AI model includes the training data as a key element. Some developers wonder, however, how feasible it is to distribute the enormous datasets required.

About the Author: 

Jeffrey Young, Principal Research Scientist, Partnership for an Advanced Computing Environment, Georgia Institute of Technology

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Theodore J. Kury, University of Florida 

Many major tech companies have pledged to pay their fair share of the costs associated with generating and transmitting more electricity to serve large data centers. But ratepayers across the United States are worried about the potential costs they might have to bear. That’s because it’s not immediately clear how the cost of data centers’ energy will be calculated. The effects of price increases are likely just beginning, and their full effects may not be felt for years.

For example, a recent report by the organization that monitors the PJM market, an area that encompasses all or part of 14 mid-Atlantic and Midwest states, concluded that expected power demand from data centers was a primary reason for US$23 billion in customer price increases that will last until at least the end of 2028.

I have studied the programs states have launched to address the needs of these large electricity customers. Prices are set by state utility commissions, who determine which customers’ rates will increase by how much to pay for new investments in electricity infrastructure. It’s not simple.

The complexity of setting prices

Setting a price for electricity is straightforward in principle but complicated in execution. Regulators identify the costs to provide service, allocate the costs to customers and design prices to recover those costs.

First, regulators identify the costs that a utility company incurs to provide service. Regulators look at the value of the assets the utility company invests in, such as power plants, transmission lines and substations, as well as its day-to-day operating expenses, such as salaries, fuel, replacement parts and electricity it purchases from other sources. Then these costs are allocated to categories of customers, such as residential, commercial and industrial.

Ideally, costs are allocated to the customers who cause them, but that can be complicated to determine. For example, imagine a data center is built in an area that lacks existing power lines and is located 50 yards from a nearby electric substation. It’s clear that the data center should pay to run a 50-yard power line from the substation to the data center.

But what if the power company needs to upgrade the substation to handle the increased needs of the data center? Or secure additional sources of electricity? In these cases, the investments are part of the electricity grid that everyone uses. These costs will likely be shared among all customers.

Cost analysts review each line of a utility company’s costs, often thousands of items, and determine how each cost will be allocated. Each decision incorporates one basic idea: What’s your share?

For instance, if a group of customers uses 20% of the electricity delivered by the utility, they would be allocated 20% of the costs associated with energy delivery. Other cost items may be allocated based on the number of customers or how much electricity customers use at particular points in time, but the idea is the same.

Finally, the analysts set prices that are designed to recover the costs allocated to each customer group. So, the costs that are allocated to you are directly reflected in the electricity prices that you pay.

Flexibility and a potential loophole

One common criterion for figuring out how much a customer should pay is based on what is called “coincident peak demand” – the amount a customer group uses at the moment when all customers are collectively using the largest amount of electricity. Costs associated with overall peak usage are typically split proportionally – but this opens an opportunity for data centers to exploit the system.

Data centers often are able to fine-tune their electricity consumption, using more one minute and less another, in ways that residential users can’t easily replicate. Computerized systems can automatically adjust the amount of work a data center is doing, while a homeowner would either have to race around shutting off appliances to meaningfully reduce the amount of power their home was using or invest in a device that does.

Their flexibility means data centers may be able to learn to predict when system loads will peak and consume little to no power in just the right period to avoid contributing to peak loads, as has happened with cryptocurrency-mining operations in Texas. So when regulators look at their usage to determine prices, data centers may be able to avoid paying any costs allocated through coincident peak demand, even if they use large amounts of electricity at other times.

Who speaks for you?

When utility regulators decide how costs should be allocated to each customer group, they solicit input from different groups. The utility company initially submits its own proposal for how it thinks costs should be allocated across its system.

Large industrial customer groups representing customers such as factories will also submit their own proposals for how to allocate costs and set rates. Retail customer groups representing large and small stores will submit theirs. And large data centers, with the resources to hire experts in cost allocation, will submit theirs as well. Some states have specific state-government agencies to do some of this work on behalf of particular commercial groups, such as Pennsylvania’s Office of Small Business Advocate.

Regulators don’t always get a good sense of residential customers’ voices, though. Every state except Georgia, Idaho and Louisiana has an office of the consumer advocate that represents customer interests in proceedings before the state utility regulator. But they are often charged with representing all customers in the state without bias, meaning they cannot advocate for outcomes that would impose costs on one group of customers in favor of another.

So while every state’s consumer advocate is concerned with keeping the utility’s costs as low as possible, they may be barred by law from adopting a position on how those costs should be allocated. This lack of representation in this aspect of rate-setting for average households may lead to situations where the data centers’ advocates argue for minimal costs to be allocated to them – but nobody advocates on behalf of residents to examine or refute that argument.

Citizens left holding the bag

There are other risks for residential customers, too. Utilities’ investments in electricity infrastructure last for many years. But not every proposed data center will get built, and some may use less energy than originally projected. Technology may even change, making some data centers obsolete after a year or two of operations.

If those events happen, then any costs the utility company incurred to provide enough electricity will be spread among all the other customers.

The allocation process may be even more complicated for municipal utilities regulated by city councils or independent boards, or cooperative utilities regulated by elected boards in rural communities. These groups may not have full-time staff who are utility or regulatory experts, yet they face the same decision-making challenges as trained professionals and might have to retain outside experts to aid in the process.

Consumers need to be aware of the importance of cost allocation and how it affects their electricity rates. I believe they should provide public comments to the regulators and speak during open hearings, as there may not be anyone else effectively advocating for their interests.

About the Author:

Theodore J. Kury, Director of Energy Studies, University of Florida

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Rachel Mural, Harvard Kennedy School 

As the race to build data centers across the United States accelerates, local governments worry that the tech industry mantra of “move fast and break things” means their communities are at risk of being broken.

I’m a Harvard researcher studying the relationship between data centers and energy. I’ve closely monitored how local governments respond to proposals or even just concerns about the potential for data centers in their communities. What I’ve found is a complex story of community needs, political tensions and corporate power – all interacting with local, state and national democratic processes.

Promises and potential

Technology companies stay competitive by being ready to provide data and communications services even before customer demand rises. Data centers already power online communications, shopping and banking systems. Now, expanding demand for artificial intelligence has led to over 1,000 pending data center proposals across the country.

Federal actions also drive development. The Trump administration has identified data center build-out as a strategic priority. The administration has promoted data center capacity as a measure of American strength and signaled that federal regulations on data centers may be eased.

At the community level, technology companies claim that data centers bring jobs, economic revitalization, digital connectivity and economic growth to local communities.

Not great neighbors

So far, however, data centers’ benefits are overshadowed by more visible harms.

Nearby residents experience higher air pollution and excess noise. Data processing also uses a lot of water to cool the buildings and their equipment.

Simultaneously, electricity prices continue to outpace inflation, burdening families across the country. These trends reflect, in part, the costly infrastructure investments needed to power data centers.

The local movement

My research has found that local governments across the U.S. are trying to avoid or reduce these harms.

Some counties and cities that don’t have specific zoning rules and regulations for data center development are using short-term moratoriums. These pauses in data center permitting and construction give communities time to consider how to define new laws and regulations about the facilities’ location, electricity use, water conservation and noise buffering.

Speaking about his town’s decision to impose a one-year data center moratorium, Rick Bella, the town council president in Merrillville, Indiana, about 40 miles southeast of Chicago, stressed a desire to “evaluate real-world impacts and learn from a project developing right next door before determining what may or may not be appropriate for Merrillville.”

Other places want to block data centers altogether. In April 2026, for example, the Ypsilanti Community Utilities Authority near Detroit, Michigan, passed a yearlong halt to the “delivery, commitment, reservation, extension, or approval of water and sewer services” for data centers. The move blocks data centers, including one under development by the University of Michigan and Los Alamos National Laboratory, from getting the water they need to operate.

Separately, towns across Ohio, Wisconsin, Maryland, Nevada and California have put questions related to data centers on their local ballots. Through these referendums, voters can weigh in on construction bans, tax incentives and zoning ordinances.

Power struggles

While public attitudes around data centers have remained largely nonpartisan, local and state officials don’t always see eye to eye.

Officials in Hood County, Texas, for example, rejected a proposal for a six-month moratorium after a state senator urged the Texas attorney general to intervene and prevent the measure.

In 2025, West Virginia passed a bill that reduces local governments’ zoning and regulatory powers in relation to data centers and microgrids. A similar bill in New Hampshire’s legislature was defeated in May 2026.

Tech companies are also flexing their legal and financial muscles. For example, data center developers sued Saline Township, Michigan, and Chatham County, North Carolina, seeking to overturn their local zoning decisions, to be able to proceed with data center construction.

Changing tides

Local pushback comes at a pivotal moment for artificial intelligence technology itself.

As seen in objections to the internet’s expanding AI “slop,” backlash over AI-generated Super Bowl ads, worries about an AI-related financial bubble and complaints about Google’s pivot to AI-directed search, Americans are reckoning with AI’s role in society.

Further, many people are questioning the role of technology broadly. Increasing numbers of teens and adults are addicted to their smartphones, emotionally and psychologically dependent on their availability. Parents and teachers are questioning the usefulness of various types of digital technologies in classrooms. Even the pope has warned that technology must serve humanity – and not the other way around.

Americans are responding to this moment through the power of their voices and votes.

Technology companies may view moratoriums and new regulations as delays in project development. But the town hall discussions, community coalitions, public petitions and even farmers’ unions reflect American democracy at work.

In Sunbury, Ohio, local officials considered a moratorium only after witnessing the scope of public protest over a proposed data center.

In April 2026, voters in Festus, Missouri, removed several City Council members after they supported a new data center despite resident pushback.

The question of whether a community wants or should have a data center does not have a universal answer. I believe it’s a question that deserves deliberate processes, transparency and consideration.

To me, these local-level actions reflect a desire to slow down. There is little question that data centers and AI will be part of our collective future. Today, communities are asking for a fair say in what their futures will be.

About the Author:

Rachel Mural, Senior Research Associate in Environment and Natural Resources and Science, Technology, and Public Policy, Harvard Kennedy School

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Alex Krasnok, Florida International University 

Every bridge has parts that drivers never see: steel buried in concrete, welds tucked under girders, and soil packed around foundations below the waterline. A bridge can look fine from the road while rust spreads around steel hidden inside concrete. A small fatigue crack can lengthen. A flood can wash soil away from a pier. By the time cracks, loose concrete or lane closures appear, the cheapest repair window may already have closed.

When it comes to these damaged bridges, this problem is national. The United States has more than 624,000 highway bridges. About 220,000 need major repair or replacement, and 41,677 are rated poor, also called structurally deficient. While “poor” does not mean unsafe, it does mean at least one key bridge element received a poor rating, indicating deterioration or cracking that will require significant repair.

As a researcher who studies photonics and quantum sensing, I work on devices that measure faint or hidden signals. My lab applies physics to develop devices, including quantum sensors. Advanced sensors of this type might one day be able to help engineers pinpoint where to look to determine whether hidden damage in infrastructure is worsening. However, they cannot replace human inspectors.

The Dames Point Bridge spans a river in Jacksonville, Fla.
Jonathan Zander/Wikimedia Commons, CC BY

Inspections keep bridges safe, but are snapshots

Federal bridge inspections – rooted in National Bridge Inspection Standards mandated by Congress in 1968 – exist because past failures showed that small defects can threaten large structures.

Under current federal rules, many bridges must be inspected in, at most, 24-month intervals. Higher-risk bridges, such as those carrying heavy interstate traffic, those with aging structures or known defects, or those built over saltwater, may require shorter intervals. Lower-risk bridges with lighter traffic and sound materials may qualify for longer intervals.

Those inspections remain essential, but they are snapshots. A bridge may change during the months between visits. Corrosion can spread below a deck that looks sound. A small crack can sit inside a weld. A river can displace soil from a foundation while the roadway above looks unchanged. Sensors extend inspections by tracking these change that form between scheduled checks.

Hidden damage can grow quietly

The three common hidden threats to bridges are corrosion, fatigue and scour. Corrosion begins when water, oxygen and salts reach steel. A concrete layer usually protects steel, but cracks, salt spray and chloride ions from seawater or deicing salts can break that protection. The rust then expands, much like ice widening a crack in a sidewalk. It pushes the concrete outward and can cause the material to come loose or the layers to separate.

Fatigue damage is the bridge version of bending a paper clip back and forth. Just as a paper clip eventually snaps after repeated bending, a bridge’s steel components weaken and break down under continuous cycles of stress. Thousands of heavy vehicles can make tiny cracks grow near welds, bolted connections or older steel details.

Scour damage is different: Moving water removes soil around the bridge’s foundations. The bridge above can look stable, while the support below loses the ground it needs.

Waiting costs more

The earlier engineers can identify damage to aging bridges, the more time and options they have to fix them. The average U.S. bridge is about 47 years old. Many bridges are near or past the 50-year life they were designed for, and about 45% have exceeded their planned design lives.

Typically, it’s less costly to preserve bridges in fair condition than those already in poor condition. Making all the identified necessary U.S. bridge repairs would cost about US$467 billion.

Past failures show why small details matter. As one example, the 2007 I-35W bridge collapse in Minneapolis was partially due to undersized gusset plates – steel plates that connect the intersecting beams in a bridge’s structural framework – along with added weight and construction loads. The collapse killed 13 people and injured 145.

Sensors alone are not a cure for such failures, but better measurements can help engineers notice when important details are changing.

Sensors help engineers look, listen and scan

Sensor systems are easiest to categorize based on what they do.

Some sensors see: Drones can photograph cracks and loose concrete, infrared cameras can show heat patterns linked to damaged deck zones, and LiDAR, short for light detection and ranging, can build three-dimensional maps.

Some sensors listen: Ultrasonic testing and impact-echo probes send sound waves into concrete or steel, acoustic emission sensors listen for active cracking, and accelerometers track how a bridge vibrates.

Some sensors scan below the surface. Specialized radio tools try to locate hidden steel, trapped moisture, empty pockets or crumbling layers inside the concrete. Meanwhile, magnetic and electrical instruments attempt to guess whether that buried steel is rusting away.

The value of sensors often comes from combining methods. One bridge deck inspection robot uses subsurface radar, electrical tools that measure moisture, and a standard camera to collect data. It then builds simple visual maps showing the exact health of the bridge deck. Fiber-optic sensing could be another route. Researchers have shown that existing telecommunication cables can record bridge vibration signatures.

Sensors are evidence, not verdicts

While instruments provide crucial clues about a structure’s condition, they do not automatically dictate the solution. Engineers still need to examine the bridge design, inspection history, traffic loads, weather, material condition and measurement uncertainty before deciding whether to repair, restrict traffic or close a bridge.

Field data is messy. Wet concrete can blur radar results. Traffic, wind and temperature can mask vibration changes.

The best systems answer narrow questions: Where is the concrete deck beginning to split into horizontal layers underneath the surface? Is this crack actively widening? Is a suspension cable losing its structural strength because its inner steel wires are rusting away? Is the fast-moving water washing away the critical soil supporting the bridge’s underwater foundations after a storm?

Quantum sensors are a frontier

Quantum sensors may help when the signs of structural distress are weak, buried or noisy. These devices use quantum systems, such as atoms or electron spins, as highly sensitive probes.

By measuring how these atomic properties shift in response to extremely subtle changes in gravity, motion or magnetic fields, the sensors can detect flaws that traditional instruments miss.

For bridges, the nearest-term opportunity is likely magnetic inspection. My team and I co-authored a review, which has not yet been peer-reviewed, on quantum magnetometers for infrastructure inspection. These sensors identify signals from induction responses, magnetic flux leakage, stress, corrosion and operational currents.

In plain terms, these sensors may help map weak magnetic fields near steel, cables or electrical conductors. Changes or disruptions in these local magnetic fields can reveal hidden rust, snapped wire strands inside a thick suspension cable, or abnormal stress points in the steel before a crack even forms.

The hard part is not building a record-setting sensor in a quiet lab, but rather making a device that works on a noisy bridge, near traffic, weather, steel and electrical interference. Quantum sensors will matter only where they beat cheaper classical tools in real inspection conditions.

The goal is not to make every bridge smart. The goal is to make damage harder to hide. Sensors give engineers more ways to see inside concrete, steel, soil and water, turning some surprise closures into repairs planned months earlier.

The public may never notice the best use of bridge sensors. That is the point: The safest infrastructure technology often works before a problem becomes visible from the road.

About the Author:

Alex Krasnok, Assistant Professor of Electrical & Computer Engineering, Florida International University

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Matt Motta, Boston University and Robert Ralston, University of Birmingham 

Arrests for betting on the U.S. military operation that removed Venezuelan leader Nicolás Maduro. Death threats from gamblers to a journalist reporting on an Iranian missile attack on Israel. Fears of government officials manipulating world events – including the Iran war – to make a quick buck.

These are some of many concerns that experts have raised about how prediction markets – online marketplaces that allow people to bet on world events – might be affecting national security in the U.S. and abroad.

But prediction markets may not be only influencing international affairs. They could also affect the 2026 midterm elections.

We are social scientists who study gambling, public policy and national security. Here are four things you need to know about how prediction markets may be changing American politics:

Prediction markets turn politics into a game

Prediction markets offer people the opportunity to bet on political events by purchasing “shares” – like stock in a company – of different potential outcomes. If an outcome takes place, the market pays out for each share purchased by those who guessed correctly. More betting activity in favor of an outcome raises its price and lowers its payout, and vice versa.

Prediction markets are different from casinos and online sportsbooks because there is no “house” – like a casino – that determines the size of the payout for correctly guessing who will win or lose a sporting event. In a prediction market, players “bet” against one another, not the house. The markets make money by charging transaction fees on each trade.

Betting on prediction markets allows users to turn many aspects of U.S. politics into a game. For example, betting on election outcomes is very popular on prediction markets. Kalshi – a popular prediction market platform – has a portion of its site specifically designated for election-related markets. That includes the chance to bet on the eventual winner of the 2028 presidential election, the margin of victory in the 2026 South Dakota primary elections and which of two Dan Sullivans could become Alaska’s next senator.

Kalshi also offers opportunities to bet on nonelection outcomes, like whether or not the Supreme Court will ban transgender girls and women from competing on “female sports teams,” or whether the government will confirm before September 2026 that aliens exist.

The gamification of politics through prediction market betting is not new. Predictit, a self-described “political prediction market,” has been operating in the U.S. for over a decade.

What has changed in recent years, however, is that prediction markets are no longer an obscure pastime enjoyed by political junkies. Prediction markets have become quite popular, and media organizations are even integrating betting market data in their political analysis. For example, Kalshi is CNN’s “official prediction markets partner.” In a segment called “The Odds,” CNN commentators often use Kalshi data to make predictions about candidates’ electoral performance.

Insider trading could affect US elections

Insider trading on prediction markets occurs when people with nonpublic information – like internal polling, military intelligence, etc. – place wagers on events. While some prediction markets are trying to crack down on the practice, insider trading could already be affecting the upcoming U.S. midterm elections.

In spring 2026, for example, NPR documented several cases where campaign staffers working on statewide campaigns admitted to using inside information about candidates’ performance in the polls to “buy low” on their candidate’s electoral prospects prior to the release of favorable polling data. Additionally, although prediction markets usually prohibit betting on one’s own campaign, both Democrats and Republicans running for political office have come under fire for betting on their own campaigns.

Betting on one’s own campaign could create a scenario where a candidate’s electoral performance seems more robust than it actually is to prediction market users or watchers, including media organizations who report on prediction market data.

This may in turn generate more favorable media coverage, which could affect public sentiment toward the candidate. Unlike polling, which is not typically prone to the same kind of meddling by campaigns, betting on one’s own campaign could ultimately change voters’ minds regarding the viability of a candidate.

Policymakers are paying attention

Given concerns about insider trading and its potential consequences, we asked Americans whether U.S. government officials should be forbidden from trading on prediction markets. In a nationally representative online survey of 1,000 U.S. adults conducted via the survey platform Verasight in March 2026, we found that nearly 70% supported banning government officials from trading on prediction markets, while 20% supported a more limited trading ban when government officials have “inside” information.

Lawmakers in Washington are beginning to respond to public opinion. The Senate recently banned senators and their staff from trading on prediction markets, although how this policy will be implemented remains uncertain. However, members of the House, employees of the executive branch, military officials and other government employees can still bet on prediction markets.

Some lawmakers have proposed limiting trading when government officials have insider information about an event, such as internal polling or fundraising data that members of the public do not have access to.

Others in Congress have made an effort to ban all trading on “death markets,” which include war, assassinations and related topics. Known as the “DEATH BETS Act” – its title is an acronym that stands for “Discouraging Exploitative Assassination, Tragedy, and Harm Betting in Event Trading Systems Act – the legislation has been introduced but is pending committee review.

State governments are also taking action to regulate prediction markets.

Massachusetts, for example, is suing Kalshi for allowing “backdoor betting” on sports.

Backdoor betting refers to wagering through less regulated channels like prediction markets, rather than highly regulated state casinos and sportsbooks. Backdoor betting has been estimated to cost states over US$1 billion in tax revenue since prediction markets first began allowing sports wagering in early 2025.

Minnesota became the first state to ban prediction markets altogether, while Illinois has sent cease and desist letters to prediction market operators that it claims are operating without adhering to state gambling laws.

Trump wants control over prediction markets

In a recent Truth Social post, President Donald Trump blasted the idea that states should be able to regulate prediction markets. Referencing their recent regulatory actions, Trump referred to Minnesota Governor Tim Walz and Illinois Governor JB Pritzker as “SCUM” in the post.

Trump also expressed enthusiasm for prediction markets in the post, saying that the U.S. is “at the top” of a “new form of Financial Market.” The president and his family have deep financial ties to the industry. For example, Donald Trump Jr. serves as a prediction market adviser to Kalshi and Polymarket and is an investor in Polymarket.

Following Trump’s post, the administration began reviewing a proposal to give the Commodity Futures Trading Commission the exclusive authority to regulate prediction markets.

While the CFTC has repeatedly asserted regulatory authority over prediction markets, some – like former CFTC Chairman Gary Gensler – believe that states, not the CFTC, should be in charge.

About the Authors: 

Matt Motta, Associate Professor of Health Law, Policy and Management, Boston University and Robert Ralston, Lecturer in Political Science and International Studies, University of Birmingham

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Neha Gour, George Mason University; Ed Maibach, George Mason University, and Luis Ortiz, George Mason University 

Every internet search, streamed video and AI-generated response depends on a data center somewhere. Driven by rapid growth in artificial intelligence, cloud computing and cryptocurrency, data centers have become the backbone of the modern digital economy. But though their key role is in enabling virtual and remote experiences, data centers are physical buildings in real communities around the nation and the globe.

The United States hosts more than 4,000 data centers – more than any other country. The U.S. Department of Energy expects that, taken together, all U.S. data centers will consume as much as 12% of all U.S. electricity by 2028. In 2023, data centers consumed about 4.4% of total U.S. electricity – roughly 176 terawatt-hours.

In the U.S., Virginia has more data centers than any other state – over 600, two-thirds of which are in the northern Virginia suburbs of Washington, D.C. In 2023, the state’s data centers consumed about 26% of Virginia’s total electricity supply – a higher share than in any other state.

We study science communication, climate science and public health, so we wanted to understand how data centers in Virginia affect the people who live near them and the broader public.

We found that the data centers that already exist affect nearby residents and the nation as a whole in five main areas: air quality, water quality, noise levels, land use and energy costs.

Air pollution

Data centers generally operate 24/7 and consume enormous amounts of electricity, which must be generated somewhere – either near the data center or farther away.

When fossil fuels are burned to generate that power, they emit a wide range of air pollutants, including those linked to lung disease, cardiovascular disease, stroke and neurological conditions. They also emit heat-trapping pollution that causes global warming and climate change, which, in turn, worsens air pollution further.

Generating power for U.S. data centers in 2023 emitted the equivalent of 2.2% of the nation’s greenhouse gas emissions. Other air pollutants emitted from fossil-fuel combustion are associated with increased risk of ADHD and autism in children and risks of Parkinson’s and Alzheimer’s diseases in older adults.

Unless the energy powering data centers comes from clean energy sources, such as solar, wind or geothermal, generating that electricity also pollutes the air. People who live near fossil-fuel burning power plants, whether in communities that also host data centers or in distant states, are exposed to air pollution. And during electrical outages, on-site diesel generators kick in, releasing large amounts of air pollution that can harm data center employees and nearby residents alike.

Water consumption and pollution

Data centers require vast quantities of water to cool their servers. Globally, they are projected to consume between 4.2 billion and 6.6 billion cubic meters of water annually by 2027. In the United States, data centers already rank among the top 10 industrial water users.

In northern Virginia, data center water use has risen sharply. In Loudoun County alone, just northwest of D.C., potable water use by data centers more than doubled between 2019 and 2023, while facilities across northern Virginia consumed nearly 2 billion gallons of water in 2023.

This demand can strain local rivers, aquifers and municipal water systems, even in regions like the mid-Atlantic that are not usually prone to drought, but especially in regions like the U.S. Southwest that face persistent droughts.

Noise pollution

Data centers’ continuous operation means that cooling systems, including air chillers and cooling fans, generate a persistent humming sound around the clock – as do any generators that are in use to provide power.

In northern Virginia, some residents have complained about an industrial-scale “drone” or “hum.” Measurements at the data centers that were the subject of complaints found noise levels were between 40 and 59 decibels on residential property.

Those noise levels are quieter than a conversation with someone 3 feet away and not loud enough to damage people’s hearing or violate local noise ordinances. But they are close to levels the EPA says reduce people’s ability to work, sleep and exercise. Some people have complained that data center noise has given them trouble sleeping and concentrating, and some have said they avoid using their homes’ outdoor spaces, where the noise is louder.

Land use and community well-being

Data center expansion often targets land near green spaces, agricultural areas or rural communities where developers can secure affordable land with access to existing electricity supplies.

Converting green space into industrial facilities can diminish health benefits associated with being in and near natural environments, including opportunities for physical activity and improved mental well-being.

In Virginia, residents living near data center construction have reported increased exposure to truck traffic and diesel exhaust, which can contribute to respiratory and cardiovascular health risks, especially in children and older adults. While these effects are typical of large construction projects, they can be amplified when several data centers are clustered together.

In places like Prince William County, Virginia, developers have proposed data centers on roughly 2,400 acres of undeveloped land in the Rural Crescent, an area designated by the county’s planners to remain relatively undeveloped. Those data centers could transform open space and rural farmland into industrial zones, disrupting communities with long-standing ties to the land.

Rising energy costs

As data centers increase electricity demand, they put upward pressure on energy prices across the grid. A 2024 Virginia legislative report found that the state’s typical residential electricity bill could rise by $14 to $37 per month by 2040 because of grid strain tied to data center growth – a 9% to 25% increase over current average bills, and a figure that does not factor in potential inflation.

These higher costs are paid by all consumers, but they place a greater burden on families that are most economically distressed, who also tend to have more health problems. Lower-income families spend a higher share of their budget on electricity, and when bills rise, the consequences can include reduced access to adequate heating and cooling, increased risks of heat-related illness and cold-related cardiovascular stress, as well as difficult choices between paying for energy and food or healthcare.

What can be done

Many of these health harms can be mitigated through better planning and design.

Increasing the share of renewable energy used to power data centers would help reduce air pollution and associated health harms.

Using recycled water in targeted systems that cool individual server rows or racks rather than whole buildings can significantly reduce cooling energy demand, with some studies estimating reductions of up to 29%.

On noise, a Leesburg, Virginia, data center reduced low-frequency tonal noise by reengineering its fan mounts.

And on energy costs, requiring large-scale data centers to cover more of the grid costs they create could help protect residential customers from higher electricity bills.

The world’s digital infrastructure runs through data centers, and that is not changing. We believe that expanding this infrastructure without protecting the health of surrounding communities is an unacceptable option.

About the Authors:

Neha Gour, Ph.D. Candidate in Science Communication, George Mason University; Ed Maibach, Distinguished University Professor Emeritus of Communication, George Mason University, and Luis Ortiz, Assistant Professor of Atmospheric, Oceanic and Earth Sciences, George Mason University

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Ji Y. Son, California State University, Los Angeles and Alice Xu, University of California, Los Angeles 

The nonprofit ARC Prize Foundation on May 1, 2026, released the results of a new benchmark: a test of an AI system’s ability to solve a game. The results were striking – humans scored 100%, while the most advanced AI systems scored under 1%.

At first glance, this may be surprising to users of AI who are impressed by its polished essays, codebases and multistep projects generated in seconds. How can these brilliant AI systems struggle with these simple Tetris-shape puzzles?

That confusion points to a risk: AI is becoming integrated into everyday life faster than people can make sense of it.

We are cognitive psychologists who study how to teach difficult concepts. To recognize the limits and risks of today’s AI agent systems, it’s important for people to grasp that the systems can both accomplish superhuman feats and make mistakes few humans would. To that end, we propose a new way to think about AIs: as button-pushing explorers.

Mental models for AI

We teach college students, a group rapidly incorporating AI tools into their daily routines. That gives us regular opportunities to ask what they think is going on with AI. The answers vary widely. One student said that someone at OpenAI or Anthropic is reading and approving every response the system generates. Another, more succinctly, said, “It’s magic.”

These responses illustrate two tempting ways of making sense of AI. At one extreme, AI is treated as an inscrutable black box – a powerful but ultimately mysterious force. At another, people explain it using the same assumptions they use to understand other humans: that its outputs reflect reasoning or judgment.

The worry is that these misinterpretations don’t go away as users gain more experience interacting with AI, and they might get reinforced. When AI performs well, its output can feel like evidence of understanding or confirmation that it really is something like magic. That apparent success makes it harder to question what the system is actually doing. Biases can seem logical or inevitable; harmful behavior can look like a deliberate choice or even fate, as if it could not have gone any other way.

Saying that AI models are shaped by patterns in data, training processes and system design is true, but that’s too abstract to tell people when to trust the systems’ outputs or when they might fail. To help people avoid misplaced trust in AI, AI literacy efforts will need to include some mechanistic understanding of what produces their behavior – explanations that are perhaps not perfectly accurate but useful. Statistician George Box once wrote, “All models are wrong, but some are useful.”

Researchers have come up with several mental models for large language models. One is “stochastic parrot,” which shows that the models use statistical methods – stochastic refers to probabilities – to mimic responses with no understanding of meaning. Another is “bag of words,” which emphasizes that the models are collections of words – for example, all English words found on the internet – with a mechanism for giving you the best set of words based on your prompt.

These ways of thinking about large language models were never meant to be complete accounts of the systems. But the metaphors serve an important cognitive purpose: They push back against the idea that fluent language is necessarily caused by humanlike understanding.

But as the AI systems people use are increasingly powerful agents capable of stringing together actions on their own, it’s important for people to have a different kind of mental model: one that explains how they act. One place to find such a model is in earlier research on AI systems that learned to play Atari 2600 games. These systems didn’t understand the games the way humans do, but they still managed to rack up a lot of points.

The simple loop: Act, observe, adjust

Imagine a neural network, a relatively simple kind of AI model, placed into a video game it has never seen before. It does not “understand” the game like a human would. It has no idea whether it’s shooting space invaders or navigating an ancient pyramid. It doesn’t know the goals or rules.

Instead, it learns to play through a simple loop: Take an action – move left, jump, shoot – observe what changes, and then adjust. If an action leads to a good outcome, such as gaining points, it adjusts to become more likely to take similar actions in similar situations. If it leads to a bad outcome, such as losing a life, it adjusts in the opposite direction.

Even this simple mechanism can produce surprisingly capable behavior. Over time, by repeating this loop, the neural networks learned to play a wide range of Atari games – but not all games.

There is one game that famously stumped these early neural networks: Montezuma’s Revenge. To make progress, a player must carry out a long sequence of actions – climbing ladders, avoiding obstacles, retrieving keys – before receiving any reward at all. Unlike simpler games, most actions offer very little immediate feedback. The game required something like goal-directed, long-term planning.

Early neural networks would try a few actions, receive no reward and fail to make further progress through Montezuma’s underground pyramid. From the system’s perspective, all actions looked equally useless. But researchers made a breakthrough by changing the feedback signal. Instead of rewarding only success, they also rewarded the system for doing something new. The rewards were for visiting parts of the game it had not seen before or trying actions it had not previously taken. This tweak encouraged exploration.

With that change, performance improved dramatically. The neural network began navigating obstacles, taking multiple steps toward goals and adapting when things went wrong. From the outside, this kind of behavior can look like planning or problem-solving. But what looks like planning was not caused by sophisticated planning abilities. The underlying mechanism is still the same simple loop: act, observe, adjust.

This kind of system isn’t a stochastic parrot or a bag of words. It’s closer to a button-pushing explorer: something that doesn’t understand the world in a human sense but moves forward by pushing buttons, seeing what happens and adjusting what it does next.

From video games to modern AI agents

Today’s AI systems can do far more than play games like Montezuma’s Revenge. They can coordinate tools, write and run code, and carry out multistep projects. The range of possible actions is much larger, and the environments in which they operate are increasingly complex.

But these agents are still fundamentally button-pushing explorers. The behavior can be sophisticated, but the process that produces it is not. Humans can often infer how a new environment works after just a few observations. Systems that rely on these feedback loops cannot. They need to try many actions and see what happens before they can make progress.

This helps explain both the strengths of these AI systems and some of their most concerning failures. What these agents learn depends on what is being rewarded. And in real-world systems, those reward signals are often imperfect.

AI systems that conduct negotiations aim to maximize their client’s interests, sometimes with deceptive tactics. Rental pricing software used by landlords ends up price fixing. Marketing tools generate persuasive but misleading reviews.

These systems aren’t trying to be evil or greedy. They are adjusting to the signals they are given. From the button-pushing explorer perspective, these failures are downright predictable.

Effective AI literacy means holding two ideas at once: These systems can do surprisingly complex things, and they are not doing them the way humans do. If AI is seen as humanlike or magical, its outputs feel authoritative. But if it is understood, even imperfectly, as a button-pushing explorer shaped by feedback, people are likely to ask better questions: Why is it doing this? What shaped this behavior? What might it be missing?

That’s the difference between being impressed by AI and being able to reason about it.

About the Author:

Ji Y. Son, Professor of Psychology, California State University, Los Angeles and Alice Xu, Ph.D. Student in Developmental Psychology, University of California, Los Angeles

This article is republished from The Conversation under a Creative Commons license. Read the original article.