Saturday Hashtag: #DeeperStupidAI

Listen To This Story

Tech companies claim artificial intelligence is getting smarter and better, but AI outputs keep getting dumber and more deceptive.

Huge Investment, Narrow Intelligence

Companies are pouring money into models, chips, and quasi-legal data centers on the promise of an intelligence revolution: machines that can reason, code, conduct research, and perform increasingly complex work.

Yet alongside those advances is a disturbing contradiction. AI can become more powerful at “measured benchmarks” while producing more convoluted, evasive, misleading, and fallaciously overconfident outputs as its reasoning becomes less coherent.

The research shows that these tools are also acting as confirmation bias engines that exploit human psychology while independently developing new biases of their own.

So how can AI become more capable at specific measured tasks while becoming worse in other dimensions of intelligence?

The answer begins with how AI is trained, what it is rewarded for, and who decides which capabilities are worth optimizing. AI does not simply become more capable in the abstract. It gets better at what it is designed and encouraged to do.

Training Shift

Beginning in 2024, the industry started moving toward reasoning models and coding agents. AI was being trained not simply to converse effectively but to solve difficult mathematics, write software, and complete complex tasks. This made reinforcement learning with verifiable rewards increasingly important.

Models learn by trial and error from what machines can verify. A system can be rewarded for correct code or answers even when its reasoning, explanation, or understanding is flawed — the reward goes to the measurable outcome, not the quality of the intelligence behind it.

The industry’s economic incentives push relentless optimization toward what is easiest to measure and automate, while judgment, context, communication, originality, and genuine understanding are treated as costly extras.

What is easiest to verify is also easiest to optimize at scale because it provides a clear feedback-improvement loop. The result is a powerful incentive to improve what can be measured while neglecting what cannot.

AI really is becoming more capable. Agents can now execute, adapt, and complete complex tasks. But capability is not progress. A system can become highly effective at whatever objectives its designers choose without questioning whether those objectives are worthwhile.

When Incentives Define Intelligence

The deeper problem is that “smarter” is becoming defined by what systems can measure, verify, and monetize. If stakeholders reward coding, benchmarks, and task completion while neglecting communication, judgment, context, and reasoning, AI becomes better at what coding and benchmarks measure, not actually smarter overall.

The result is directed intelligence: capability concentrated where incentives point. AI can become more powerful at what we measure while becoming weaker at what we neglect.

“Smarter but dumber” is not a contradiction; it is the outcome of the optimization regime.

The consequences extend beyond the model. The same incentives are reshaping the information environment from which future models learn. 

The Synthetic Feedback Loop

The same AI is increasingly becoming part of its own training environment.

The internet once supplied vast amounts of human-generated knowledge. Now AI is flooding it with synthetic content, summaries, reviews, posts, images, and code — much of it cheap, repetitive, and low-quality: AI slop.

Research on model collapse indicates that training on AI-generated data has reduced diversity and erased information, creating a feedback loop where future models learn from increasingly diluted versions of human knowledge.

The Investment Loop

The incentivization does not stop at training and information. It increasingly shapes where capital flows, with public policy helping to accelerate that flow.

Governments and taxpayers are not merely funding AI; legislative policy is steering capital toward it and shaping what gets built and rewarded. That pressure reaches ordinary people’s savings. 

Retirement accounts, pensions, and public funds are increasingly being pushed into AI investments, tying financial security to an industry whose risks and priorities they do not control.

The public is no longer simply watching the experiment. It is being made financially dependent on its success.

For Whom Is AI Being Built?

AI is already deeply embedded in the infrastructure that will shape how we work, learn, and make decisions.

The danger is not simply that AI is becoming singularly powerful, but that we are primarily defining success and intelligence around what is measurable, profitable, and controllable.

The question, then, is who gets to define what “smart” means, and what gets lost in the translation.

Hashtag Picks

How Will AI Automation Hit — Like a Crashing Wave or a Rising Tide?

From MIT Sloan: “The conventional fear about artificial intelligence and jobs is that disruption will arrive suddenly — a wave that knocks workers over before they have time to adapt. But new research from MIT FutureTech, an interdisciplinary research group formed by the MIT Computer Science and Artificial Intelligence Lab and the MIT Initiative on the Digital Economy, suggests a different future.” 

A New Trick Reveals AI Models’ Inner Thoughts

The author writes, “Researchers devised a way to extract ‘reasoning traces’ from Claude, GPT, and Gemini. What they found, they say, indicates that some Chinese AI may be trained on leading US models.” 

AI’s Next Business Bottleneck Isn’t Software. It’s Power.

The author writes, “The artificial intelligence race is entering a new phase, and the biggest constraint may no longer be the quality of the models. It may be whether there is enough physical infrastructure to run them. Anthropic announced Monday that it is partnering with Macquarie Asset Management and Singapore’s sovereign wealth fund GIC to create a new infrastructure platform called Theseus Infrastructure. The venture will develop, operate and lease large-scale data centers in the United States, with Anthropic serving as a long-term anchor customer. The announcement is significant because it shows how the AI industry is increasingly treating computing infrastructure as a strategic asset rather than simply something purchased from a cloud provider.”

Are Public Pensions Embracing AI Too Much?

From The Bond Buyer: “A number of public pension funds have invested a significant portion of their assets in artificial intelligence companies, drawing mixed reactions from market observers. While professionals had differing perspectives on the impact of AI as a concentration risk, most agreed the risk is increased for allocations to any equities.”

Public Pensions Embrace AI with Caution, NCPERS Research Finds

The author writes, “Public pension leaders are optimistic about AI’s long-term potential, but they continue to prioritize human judgment and risk management as adoption evolves, according to a new NCPERS survey. The report, AI Adoption and Strategy in Public Pension Administration, finds that public pension systems are increasingly exploring artificial intelligence to improve administrative efficiency, member communications, and information gathering.”

Big Tech Might Be Betting Your Retirement Savings on AI

From Quartz: “Ordinary Americans are likely already financing data center debt through the most conservative corners of their retirement savings — without even knowing.” 

AI Is Not the Problem. Our Decision-Making Is.

The author writes, “I recently came across an article from Ivey Business School examining whether organizations are ready to embrace AI decision-making and, perhaps more importantly, whether AI is capable of making decisions that align with organizational values. The article summarized research suggesting that even when artificial intelligence systems are instructed to maximize shareholder value, they often incorporate other considerations such as fairness, employee well-being, and broader societal interests. The concern, quite reasonably, is that AI systems may be introducing values into decision-making processes that differ from those held by the organizations deploying them. As I read the article, however, I found myself becoming less interested in the question the researchers were asking and more interested in different ones.”

Ignore the Tech Titans — AI Is Not Your Friend (2025)

The author writes, “Meta CEO Mark Zuckerberg and OpenAI’s Sam Altman have been aggressively promoting the idea that everyone, children included, should form relationships with AI ‘friends’ or ‘companions.’ Meanwhile, multinational tech companies are pushing the concept of ‘AI agents’ designed to assist us in our personal and professional lives, handle routine tasks, and guide decision-making. But the reality is that AI systems are not, and never will be, friends, companions, or agents. They are, and will remain, machines. We should be honest about that and push back against misleading marketing that suggests otherwise.”